diff --git a/PROGRESS.md b/PROGRESS.md index fe66c78..bcedcb1 100644 --- a/PROGRESS.md +++ b/PROGRESS.md @@ -14,7 +14,7 @@ | 表示、位置与残差高速公路 | 完成首版 | 81% | 真实 hidden-state / norm traces、长上下文位置外推与深层稳定性消融 | | Scaling Laws | 完成首版 | 74% | 真实拟合复现、置信区间与更多模型族对照 | | 数据工程与预训练配方 | 完成首版 | 73% | FineWeb / DCLM 逐图精读、真实去重误伤与 mixture traces | -| DeepSeek 专题 | 完成二轮 | 83% | 真实专家负载 / MLA kernel / RL 训练 traces 与独立复现 | +| DeepSeek 专题 | 三轮实证进行中 | 87% | 真实 latent-cache kernel、更大样本负载、FP8/pipeline 与 R1-like RL 复现 | | 指令微调与人类偏好 | 完成首版 | 75% | 真实偏好分歧、RM 长度偏置与 PPO/DPO 小模型复现 | | 推理与测试时扩展 | 完成首版 | 76% | 真实模型采样曲线、PRM 案例与逐篇图表精读 | | 工具使用与长程 Agent | 完成首版 | 74% | 真实环境 traces、cross-harness 对照、Agent RL 曲线与安全案例 | @@ -41,7 +41,7 @@ - [x] 完成 486 篇关键论文索引,覆盖 16 个标签专题与 Kimi/DeepSeek 聚光主线。 - [x] 完成可检索、可按专题筛选的论文库页面。 - [x] 完成 K3、语言模型前史、Transformer 基础、表示/位置/残差、DeepSeek 谱系、Scaling Laws、数据工程、长上下文、MoE、指令微调与人类偏好、推理、Agent、原生多模态、训练系统、推理服务、数值优化与评测安全十七篇首版长文。 -- [x] 完成 K3 三轴架构、八联报告实验与四联开放工件实验、语言模型前史四联实验、Transformer 四联实验、表示深度四联实验、DeepSeek 四联实验、长上下文、MoE 路由、推理三页签,以及训练系统、推理服务、Scaling、数据工程、数值、Alignment、Agent、原生多模态与评测安全专题各四页签等七十一个原创交互视图。 +- [x] 完成 K3 三轴架构、八联报告实验与四联开放工件实验、语言模型前史四联实验、Transformer 四联实验、表示深度四联实验、DeepSeek 八联实验、长上下文、MoE 路由、推理三页签,以及训练系统、推理服务、Scaling、数据工程、数值、Alignment、Agent、原生多模态与评测安全专题各四页签等七十五个原创交互视图。 - [x] 完成长上下文首版:五张成本账、26 篇一手论文、10+ 机制图与 8 策略交互实验室。 - [x] 核验 FlashAttention、DeepSeek-V2/V3.2/V4、Kimi Linear/K3 等六份论文原文,并建立长上下文研究账本。 - [x] 核验 Switch、ST-MoE、DeepSeekMoE、Loss-Free、V3、LatentMoE 与 K3 原文,并建立 MoE 研究账本。 @@ -169,11 +169,15 @@ - [x] RTX 5090 K3 形状受控计时完成:fixed / varlen、BF16 / no-state / FP32 state 共 1,800 个 CUDA Event samples;fixed BF16 mean 2.6210 ms、P95 2.6437 ms,varlen BF16 mean 2.3335 ms、P95 2.3574 ms。 - [x] 可复现 Dockerfile、运行脚本与机器可读 JSON 接入开源树;网站第四视图把作者 H20/GB200 表、本机 RTX 5090 值、exact suite、构建链、synthetic router 和 `A_log` 未决冲突分开显示。 - [x] K3 FlashKDA 执行里程碑以源提交 `2ef846f`、不可变镜像 `20260729T053852Z-2ef846f` 发布;OCI digest `sha256:08ad5b55…32e6`,NAS、VPS/Tailscale、NPM、DNS、HTTPS、证书、门户与十六套生产 Chrome 回归全链路通过;保留 `20260729T044605Z-be2b291` 回滚。 +- [x] 启动 DeepSeek 三轮真实权重执行:固定官方 V2-Lite revision `604d5664…82de0`,审计 29.261 GiB / 4 shards,并以第一分片的完整层边界执行 layer 0–6。 +- [x] RTX 5090 连续 forward 闭环:4 条固定 prompt、90 个有效 token、6 个 MoE 层产生 3,240 次真实 top-6 路由;记录每层 `[4,27,576]` MLA 投影与 HF eager 展开 K/V shapes。 +- [x] 独立复跑 31/31 exact:来源、配置、tokenization、逐层 hidden hashes、MLA shapes、aggregate loads 与全部 token routes 完全一致;计时明确不比较。 +- [x] DeepSeek 新四视图本地真实 Chrome 回归通过:逐 token 路由、覆盖/均衡分账、latent/HF eager cache 计算、27 层执行断面、键盘 tabs 与 390px 移动端均无异常。 ## 正在进行 - [ ] K3 三轮下一闸门:获得真实 token hidden states、expert load 与 cache traces,解释或修订 `A_log [128]` 工件冲突,再做 Figure 3/4/5 数值重绘和独立小模型复现。 -- [ ] DeepSeek 三轮:真实专家负载、MLA kernel、FP8 / pipeline 与 R1-like RL traces,外加独立小模型复现。 +- [ ] DeepSeek 三轮下一闸门:真实 latent-cache kernel、更大样本 expert-load 置信区间、FP8 / pipeline traces 与 R1-like RL 小模型复现。 - [ ] 表示、位置与残差二轮:真实 hidden-state / norm traces、长上下文位置外推复现与 mHC / AttnRes 深层稳定性消融。 - [ ] 评测安全二轮:真实 cross-harness / pass@k 复跑、Judge 元评测、动态污染与过拒案例。 - [ ] 推理服务二轮:真实 GPU kernel / workload traces、功耗与成本、跨 vLLM / SGLang / TensorRT-LLM 复现。 @@ -289,6 +293,9 @@ | 2026-07-29 | 论文库扩充到 486 篇 | 新增 DeepSeek-Coder/Coder-V2、ESFT、Prover-V1.5/V2 与 Engram 6 个旁支节点 | | 2026-07-29 | DeepSeek 二轮用四个独立实验闭环 | 稀疏容量、MLA 缓存、V3 协同与 RL 偏差分别操作;精确公式、作者报告和 teaching model 永久分级 | | 2026-07-29 | DeepSeek 二轮用不可变镜像 `20260729T031901Z-cd96dab` 发布 | OCI digest `sha256:4bfb3faf…fa3968`;复用 `12010→8080`、NPM host 31 / cert 41、门户 order 180 与公开 Forgejo;保留 `20260729T023329Z-a2c9298` 回滚 | +| 2026-07-29 | DeepSeek 三轮以官方 V2-Lite 第一分片建立真实执行证据 | layer 0–6 是完整连续边界;layer 7 跨分片即停止,不把部分下载写成完整生成 | +| 2026-07-29 | MLA 算法状态与框架物化格式分两张账 | V2-Lite latent 576 元素与 HF eager 5,120 元素同时报告;88.75% 不冒充 V2 论文 93.3% | +| 2026-07-29 | 专家覆盖、负载均衡与语义永久分开 | 60–64/64 used 不推出 balanced;CV/Gini/effective 同报;expert ID 不跨层连线或命名 | | 2026-07-29 | K3 二轮按 32 张对象账与完整报告顺序重建 | total/active、2.5×、KDA state、深度来源、专家路由、视觉目标、轨迹、缓存与评测协议不再压成一页组件摘要 | | 2026-07-29 | K3 原生视觉事实回到 §2.4 / §3.3 核验 | 删除“先冻结语言模型再解冻”旧表述;明确 MoonViT-V2 从头训练,视觉/文本从开始共同 NTP | | 2026-07-29 | K3 Figure 1–16 / Table 1–5 全部建立课程视觉契约 | 每张图同时写支持范围与不可外推项;作者报告、论文、推导与 toy model 使用 R/P/D/T 标签 | diff --git a/README.md b/README.md index bd7b15a..69db2db 100644 --- a/README.md +++ b/README.md @@ -19,7 +19,7 @@ 当前里程碑包含 17 专题学习地图、486 篇关键论文索引、Kimi K3 完整导读, 语言模型前史、Transformer 基础、表示/位置/残差、DeepSeek 技术谱系、Scaling Laws、数据工程、长上下文、MoE、指令微调与人类偏好、推理、工具使用与长程 Agent、原生多模态、训练系统、推理服务、数值优化,以及评测与安全深度专题, -以及 71 个覆盖核心机制的原创交互视图。K3 二轮导读以 32 张问题账、16 图 / 5 表审计、 +以及 75 个覆盖核心机制的原创交互视图。K3 二轮导读以 32 张问题账、16 图 / 5 表审计、 8 个交互实验和 100 个一手/官方节点,完整覆盖架构、预训练、后训练、系统、评测、案例与附录。 第三轮已完成开放工件与首个真实 kernel 里程碑:固定官方模型与 FlashKDA revisions,审计 96 个 checkpoint shards、 497,220 个 tensor entries、真实 KDA / MLA / MoE / MoonViT shapes 与小范围参数统计,并用 4 个新视图 @@ -27,8 +27,11 @@ `sm_120a` wheel,在 RTX 5090 上完成 6/6 官方参考 exact-match 和 K3 fixed / varlen 形状计时。详见 [K3_ARTIFACT_AUDIT.md](./research/K3_ARTIFACT_AUDIT.md) 与 [checkpoint_probe.py](./experiments/k3/checkpoint_probe.py)、[FlashKDA probe](./experiments/k3/flashkda/)。 -DeepSeek 二轮专题以 24 张问题账、10 次技术转向、 -4 个交互实验和 60 个一手/官方节点,串起 Dense、MoE、MLA、V3 协同、R1 与 V4。 +DeepSeek 三轮专题以 24 张问题账、10 次技术转向、 +8 个交互实验和 60 个一手/官方节点,串起 Dense、MoE、MLA、V3 协同、R1 与 V4; +并固定官方 V2-Lite revision,在 RTX 5090 上连续执行 7/27 层,记录 3,240 次真实专家选择、 +MLA/HF eager cache shapes 与 `31/31` exact 独立复跑。详见 +[DEEPSEEK_V2_LITE_TRACE.md](./research/DEEPSEEK_V2_LITE_TRACE.md)。 其余专题按进度账本持续扩建。 ## 本地开发 diff --git a/experiments/deepseek/README.md b/experiments/deepseek/README.md new file mode 100644 index 0000000..bb89198 --- /dev/null +++ b/experiments/deepseek/README.md @@ -0,0 +1,56 @@ +# DeepSeek real-weight execution probes + +The probes in this directory use official DeepSeek artifacts and keep their +scope deliberately narrower than a full-model benchmark. + +## DeepSeek-V2-Lite truncated trace + +`v2_lite_trace.py` executes layers 0–6 from the official BF16 checkpoint. Those +seven layers are fully contained in shard 1; layer 7 is split across shards 1 +and 2 and is therefore outside the default evidence boundary. + +Pinned model: + +```text +deepseek-ai/DeepSeek-V2-Lite@604d5664dddd88a0433dbae533b7fe9472482de0 +``` + +Required Python stack: + +```text +torch==2.11.0+cu128 +transformers==4.41.2 +safetensors==0.8.0 +``` + +The 2024 remote code does not import under Transformers 5.5 because +`is_torch_fx_available` was removed. The probe imports the official files as a +read-only local package; it does not patch the model source. + +Download the metadata, tokenizer, remote code, index, and first shard with the +Hugging Face CLI, then run: + +```bash +python experiments/deepseek/v2_lite_trace.py \ + --artifact-dir /path/to/deepseek-v2-lite \ + --output src/data/deepseek-v2-lite-trace.json +``` + +The result contains: + +- real tokenizer pieces and model-derived hidden states; +- the actual `[B,T,576]` MLA compressed projection at each executed layer; +- the expanded key/value tensors stored by the Hugging Face eager cache; +- token-level top-6 routed expert IDs and weights for six MoE layers; +- per-layer and per-prompt expert-load summaries; +- explicit boundaries against global load, expert semantics, training traces, + full-model generation, and production serving claims. + +Run the probe twice and compare deterministic evidence while excluding timing: + +```bash +python experiments/deepseek/compare_v2_lite_traces.py \ + --first /path/to/trace-1.json \ + --second /path/to/trace-2.json \ + --output src/data/deepseek-v2-lite-trace-repro.json +``` diff --git a/experiments/deepseek/compare_v2_lite_traces.py b/experiments/deepseek/compare_v2_lite_traces.py new file mode 100644 index 0000000..9d14448 --- /dev/null +++ b/experiments/deepseek/compare_v2_lite_traces.py @@ -0,0 +1,110 @@ +#!/usr/bin/env python3 +"""Compare deterministic evidence from two DeepSeek-V2-Lite trace runs.""" + +from __future__ import annotations + +import argparse +import hashlib +import json +from datetime import datetime, timezone +from pathlib import Path +from typing import Any + + +def parse_args() -> argparse.Namespace: + parser = argparse.ArgumentParser() + parser.add_argument("--first", type=Path, required=True) + parser.add_argument("--second", type=Path, required=True) + parser.add_argument("--output", type=Path, required=True) + parser.add_argument("--captured-at", default=None) + return parser.parse_args() + + +def read(path: Path) -> dict[str, Any]: + return json.loads(path.read_text()) + + +def sha256(path: Path) -> str: + return hashlib.sha256(path.read_bytes()).hexdigest() + + +def main() -> None: + args = parse_args() + first = read(args.first) + second = read(args.second) + checks: list[dict[str, Any]] = [] + + def check(name: str, left: Any, right: Any) -> None: + checks.append({"name": name, "exact": left == right}) + + check("provenance", first["provenance"], second["provenance"]) + check("configuration", first["configuration"], second["configuration"]) + check("prompt_tokenization", first["prompts"], second["prompts"]) + check( + "initial_hidden_hash", + first["execution"]["initial_hidden"]["sha256_fp32"], + second["execution"]["initial_hidden"]["sha256_fp32"], + ) + check( + "final_hidden_hash", + first["execution"]["final_hidden"]["sha256_fp32"], + second["execution"]["final_hidden"]["sha256_fp32"], + ) + for left, right in zip( + first["execution"]["layers"], + second["execution"]["layers"], + strict=True, + ): + layer = left["layer"] + check( + f"layer_{layer}_hidden_hash", + left["hidden_after"]["sha256_fp32"], + right["hidden_after"]["sha256_fp32"], + ) + check(f"layer_{layer}_mla_shapes", left["mla"], right["mla"]) + if "routing" in left: + check( + f"layer_{layer}_aggregate_load", + left["routing"]["aggregate_load"], + right["routing"]["aggregate_load"], + ) + check( + f"layer_{layer}_token_routes", + [ + prompt["token_routes"] + for prompt in left["routing"]["per_prompt"] + ], + [ + prompt["token_routes"] + for prompt in right["routing"]["per_prompt"] + ], + ) + + captured_at = args.captured_at or datetime.now(timezone.utc).isoformat() + result = { + "schema_version": 1, + "captured_at": captured_at, + "first": { + "captured_at": first["captured_at"], + "sha256": sha256(args.first), + }, + "second": { + "captured_at": second["captured_at"], + "sha256": sha256(args.second), + }, + "timing_compared": False, + "checks": checks, + "exact_checks": sum(item["exact"] for item in checks), + "total_checks": len(checks), + "all_exact": all(item["exact"] for item in checks), + } + if not result["all_exact"]: + failed = [item["name"] for item in checks if not item["exact"]] + raise AssertionError(f"trace mismatch: {failed}") + args.output.parent.mkdir(parents=True, exist_ok=True) + args.output.write_text(json.dumps(result, indent=2) + "\n") + print(json.dumps(result, indent=2)) + + +if __name__ == "__main__": + main() diff --git a/experiments/deepseek/v2_lite_trace.py b/experiments/deepseek/v2_lite_trace.py new file mode 100644 index 0000000..4449c70 --- /dev/null +++ b/experiments/deepseek/v2_lite_trace.py @@ -0,0 +1,588 @@ +#!/usr/bin/env python3 +"""Trace real DeepSeek-V2-Lite MLA states and MoE routes from official weights. + +The script deliberately executes only the contiguous layers fully contained in +the first official safetensors shard (layers 0 through 6). It loads one decoder +layer at a time, so a 32 GB workstation GPU can produce model-derived hidden +states without downloading or materializing the complete 15.7B-parameter model. + +No model source is patched. The pinned remote-code files are imported as a +read-only local package, and every decoder layer is the official class. +""" + +from __future__ import annotations + +import argparse +import gc +import hashlib +import importlib.util +import json +import math +import platform +import statistics +import subprocess +import sys +import time +import types +from datetime import datetime, timezone +from itertools import combinations +from pathlib import Path +from typing import Any + +import torch +import torch.nn.functional as F +from safetensors import safe_open +from transformers import AutoTokenizer, DynamicCache + + +DEFAULT_PROMPTS = [ + { + "id": "zh_explanation", + "label": "中文解释", + "text": "用通俗的语言解释,为什么稀疏专家模型可以拥有很多参数,但每个 token 只使用其中一小部分。", + }, + { + "id": "en_architecture", + "label": "English architecture", + "text": "Explain how a compressed key-value latent changes the memory cost of autoregressive decoding.", + }, + { + "id": "code", + "label": "Python code", + "text": "Write a Python function that returns the first repeated element in a list and explain its complexity.", + }, + { + "id": "math", + "label": "数学推理", + "text": "若正数 x 满足 x 加上它的倒数等于 3,求 x 的平方加上倒数的平方。", + }, +] + + +def parse_args() -> argparse.Namespace: + parser = argparse.ArgumentParser() + parser.add_argument("--artifact-dir", type=Path, required=True) + parser.add_argument("--output", type=Path, required=True) + parser.add_argument("--layers", type=int, default=7) + parser.add_argument("--device", default="cuda") + parser.add_argument("--captured-at", default=None) + return parser.parse_args() + + +def sha256(path: Path) -> str: + digest = hashlib.sha256() + with path.open("rb") as handle: + for block in iter(lambda: handle.read(1024 * 1024), b""): + digest.update(block) + return digest.hexdigest() + + +def tensor_sha256(tensor: torch.Tensor) -> str: + value = tensor.detach().float().contiguous().cpu().numpy() + return hashlib.sha256(value.tobytes()).hexdigest() + + +def load_official_modules(root: Path) -> tuple[Any, Any]: + """Import relative official files without Transformers' remote-code scanner.""" + package_name = "deepseek_v2_lite_official" + package = types.ModuleType(package_name) + package.__path__ = [str(root)] + sys.modules[package_name] = package + + loaded = {} + for leaf in ("configuration_deepseek", "modeling_deepseek"): + name = f"{package_name}.{leaf}" + spec = importlib.util.spec_from_file_location(name, root / f"{leaf}.py") + if spec is None or spec.loader is None: + raise RuntimeError(f"cannot load official module: {leaf}") + module = importlib.util.module_from_spec(spec) + sys.modules[name] = module + spec.loader.exec_module(module) + loaded[leaf] = module + return loaded["configuration_deepseek"], loaded["modeling_deepseek"] + + +def percentile(values: torch.Tensor, q: float) -> float: + return torch.quantile(values.float(), q).item() + + +def hidden_stats(hidden: torch.Tensor, mask: torch.Tensor) -> dict[str, Any]: + valid = hidden[mask.bool()] + rms = valid.float().square().mean(dim=-1).sqrt() + norms = valid.float().norm(dim=-1) + return { + "valid_tokens": valid.shape[0], + "hidden_rms": { + "mean": rms.mean().item(), + "p10": percentile(rms, 0.10), + "p50": percentile(rms, 0.50), + "p90": percentile(rms, 0.90), + }, + "l2_norm": { + "mean": norms.mean().item(), + "p10": percentile(norms, 0.10), + "p50": percentile(norms, 0.50), + "p90": percentile(norms, 0.90), + }, + "finite": bool(torch.isfinite(valid).all()), + "sha256_fp32": tensor_sha256(valid), + } + + +def load_metrics(load: torch.Tensor) -> dict[str, Any]: + values = load.float() + mean = values.mean() + ordered = values.sort().values + count = values.numel() + indices = torch.arange(1, count + 1, dtype=torch.float32) + denominator = count * ordered.sum() + gini = ( + ((2 * indices - count - 1) * ordered).sum() / denominator + if denominator + else torch.tensor(0.0) + ) + probabilities = values / values.sum().clamp_min(1) + nonzero = probabilities[probabilities > 0] + entropy = -(nonzero * nonzero.log()).sum() + return { + "routes": int(values.sum()), + "used_experts": int((values > 0).sum()), + "zero_experts": int((values == 0).sum()), + "mean": mean.item(), + "std": values.std().item(), + "cv": (values.std() / mean).item() if mean else 0.0, + "gini": gini.item(), + "entropy_nats": entropy.item(), + "effective_experts": math.exp(entropy.item()), + "min": values.min().item(), + "p50": percentile(values, 0.50), + "p90": percentile(values, 0.90), + "max": values.max().item(), + } + + +def route_trace( + tokenizer: Any, + input_ids: torch.Tensor, + attention_mask: torch.Tensor, + topk_ids: torch.Tensor, + topk_weights: torch.Tensor, + prompts: list[dict[str, str]], + expert_count: int, +) -> dict[str, Any]: + batch, sequence, top_k = topk_ids.shape + valid_routes = attention_mask.bool().unsqueeze(-1).expand_as(topk_ids) + aggregate_load = torch.bincount( + topk_ids[valid_routes].cpu(), minlength=expert_count + ) + + per_prompt = [] + used_sets: dict[str, set[int]] = {} + for batch_index, prompt in enumerate(prompts): + length = int(attention_mask[batch_index].sum()) + ids = input_ids[batch_index, :length].tolist() + tokens = tokenizer.convert_ids_to_tokens(ids) + routes = topk_ids[batch_index, :length].cpu() + weights = topk_weights[batch_index, :length].float().cpu() + load = torch.bincount(routes.flatten(), minlength=expert_count) + used_sets[prompt["id"]] = set(torch.nonzero(load, as_tuple=False).flatten().tolist()) + token_rows = [] + for position, (token_id, token, experts, scores) in enumerate( + zip(ids, tokens, routes.tolist(), weights.tolist(), strict=True) + ): + ordered = sorted( + zip(experts, scores, strict=True), + key=lambda pair: pair[1], + reverse=True, + ) + token_rows.append( + { + "position": position, + "token_id": token_id, + "token": token, + "decoded_piece": tokenizer.decode( + [token_id], + skip_special_tokens=False, + clean_up_tokenization_spaces=False, + ), + "experts_by_weight": [ + {"expert": expert, "weight": weight} + for expert, weight in ordered + ], + "selected_weight_sum": sum(scores), + "top1_top2_margin": ordered[0][1] - ordered[1][1], + } + ) + top_experts = sorted( + enumerate(load.tolist()), key=lambda pair: pair[1], reverse=True + )[:8] + per_prompt.append( + { + "id": prompt["id"], + "label": prompt["label"], + "tokens": length, + "load": load.tolist(), + "metrics": load_metrics(load), + "top_experts": [ + {"expert": expert, "routes": routes} + for expert, routes in top_experts + if routes + ], + "token_routes": token_rows, + } + ) + + jaccard = [] + for left, right in combinations(prompts, 2): + a = used_sets[left["id"]] + b = used_sets[right["id"]] + jaccard.append( + { + "left": left["id"], + "right": right["id"], + "used_expert_jaccard": len(a & b) / len(a | b), + "shared_experts": len(a & b), + "union_experts": len(a | b), + } + ) + + valid_weights = topk_weights[valid_routes].float().cpu().view(-1, top_k) + weight_sums = valid_weights.sum(dim=-1) + ordered_weights = valid_weights.sort(dim=-1, descending=True).values + top_experts = sorted( + enumerate(aggregate_load.tolist()), key=lambda pair: pair[1], reverse=True + )[:12] + return { + "aggregate_load": aggregate_load.tolist(), + "aggregate_metrics": load_metrics(aggregate_load), + "top_experts": [ + {"expert": expert, "routes": routes} + for expert, routes in top_experts + if routes + ], + "selected_weight_sum": { + "mean": weight_sums.mean().item(), + "p10": percentile(weight_sums, 0.10), + "p50": percentile(weight_sums, 0.50), + "p90": percentile(weight_sums, 0.90), + }, + "top1_top2_margin": { + "mean": (ordered_weights[:, 0] - ordered_weights[:, 1]).mean().item(), + "p50": percentile(ordered_weights[:, 0] - ordered_weights[:, 1], 0.50), + "p90": percentile(ordered_weights[:, 0] - ordered_weights[:, 1], 0.90), + }, + "per_prompt": per_prompt, + "prompt_pair_jaccard": jaccard, + } + + +def nvidia_smi() -> dict[str, str]: + fields = [ + "name", + "driver_version", + "memory.total", + "power.limit", + "clocks.max.sm", + ] + output = subprocess.check_output( + [ + "nvidia-smi", + f"--query-gpu={','.join(fields)}", + "--format=csv,noheader,nounits", + ], + text=True, + ).strip() + return dict(zip(fields, [item.strip() for item in output.split(",")], strict=True)) + + +def main() -> None: + args = parse_args() + root = args.artifact_dir.resolve() + shard = root / "model-00001-of-000004.safetensors" + required = [ + root / "config.json", + root / "configuration_deepseek.py", + root / "modeling_deepseek.py", + root / "model.safetensors.index.json", + root / "tokenizer.json", + shard, + ] + missing = [str(path) for path in required if not path.exists()] + if missing: + raise FileNotFoundError(f"missing official artifacts: {missing}") + if args.device.startswith("cuda") and not torch.cuda.is_available(): + raise RuntimeError("CUDA requested but unavailable") + if not 1 <= args.layers <= 7: + raise ValueError("shard 1 fully contains only layers 0 through 6") + + configuration, modeling = load_official_modules(root) + config = configuration.DeepseekV2Config.from_pretrained(root) + config._attn_implementation = "eager" + tokenizer = AutoTokenizer.from_pretrained( + root, + trust_remote_code=True, + local_files_only=True, + ) + if tokenizer.pad_token_id is None: + tokenizer.pad_token = tokenizer.eos_token + tokenizer.padding_side = "right" + + encoded = tokenizer( + [prompt["text"] for prompt in DEFAULT_PROMPTS], + add_special_tokens=True, + padding=True, + return_tensors="pt", + ) + input_ids = encoded.input_ids + attention_mask_cpu = encoded.attention_mask + device = torch.device(args.device) + + start = time.perf_counter() + with safe_open(shard, framework="pt", device="cpu") as handle: + embedding = handle.get_tensor("model.embed_tokens.weight") + hidden = F.embedding(input_ids, embedding).to(device) + embedding_ms = (time.perf_counter() - start) * 1000 + + attention_mask = attention_mask_cpu.to(device) + sequence = input_ids.shape[1] + position_ids = torch.arange(sequence, device=device).unsqueeze(0) + causal_mask = modeling._prepare_4d_causal_attention_mask( + attention_mask, + input_ids.shape, + hidden, + 0, + ) + cache = DynamicCache() + layer_results = [] + initial_hidden = hidden_stats(hidden, attention_mask) + + for layer_index in range(args.layers): + prefix = f"model.layers.{layer_index}." + load_start = time.perf_counter() + with safe_open(shard, framework="pt", device="cpu") as handle: + state = { + key[len(prefix) :]: handle.get_tensor(key) + for key in handle.keys() + if key.startswith(prefix) + } + state_numel = sum(value.numel() for value in state.values()) + state_bytes = sum(value.numel() * value.element_size() for value in state.values()) + with torch.device("meta"): + layer = modeling.DeepseekV2DecoderLayer(config, layer_index) + layer.to_empty(device="cpu") + layer.load_state_dict(state, strict=True, assign=True) + del state + load_ms = (time.perf_counter() - load_start) * 1000 + + torch.cuda.reset_peak_memory_stats(device) if device.type == "cuda" else None + transfer_start = time.perf_counter() + layer = layer.to(device=device, dtype=torch.bfloat16).eval() + if device.type == "cuda": + torch.cuda.synchronize(device) + transfer_ms = (time.perf_counter() - transfer_start) * 1000 + + captures: dict[str, torch.Tensor] = {} + + def capture_kv(_module: Any, _inputs: Any, output: torch.Tensor) -> None: + captures["compressed_kv"] = output.detach() + + handles = [ + layer.self_attn.kv_a_proj_with_mqa.register_forward_hook(capture_kv) + ] + if layer_index > 0: + + def capture_gate(_module: Any, _inputs: Any, output: Any) -> None: + captures["topk_ids"] = output[0].detach() + captures["topk_weights"] = output[1].detach() + + handles.append(layer.mlp.gate.register_forward_hook(capture_gate)) + + before = hidden_stats(hidden, attention_mask) + if device.type == "cuda": + torch.cuda.synchronize(device) + forward_start = time.perf_counter() + with torch.inference_mode(): + outputs = layer( + hidden, + attention_mask=causal_mask, + position_ids=position_ids, + past_key_value=cache, + use_cache=True, + ) + hidden = outputs[0] + if device.type == "cuda": + torch.cuda.synchronize(device) + forward_ms = (time.perf_counter() - forward_start) * 1000 + after = hidden_stats(hidden, attention_mask) + for handle in handles: + handle.remove() + + compressed = captures["compressed_kv"] + latent = compressed[..., : config.kv_lora_rank] + rope_key = compressed[..., config.kv_lora_rank :] + key_cache = cache.key_cache[layer_index] + value_cache = cache.value_cache[layer_index] + result: dict[str, Any] = { + "layer": layer_index, + "ffn": "dense" if layer_index == 0 else "moe", + "state_tensors": len(layer.state_dict()), + "state_numel": state_numel, + "state_bytes": state_bytes, + "timing_ms": { + "load_cpu": load_ms, + "transfer_to_device": transfer_ms, + "forward": forward_ms, + }, + "peak_device_allocated_mib": ( + torch.cuda.max_memory_allocated(device) / 2**20 + if device.type == "cuda" + else None + ), + "hidden_before": before, + "hidden_after": after, + "mla": { + "compressed_projection_shape": list(compressed.shape), + "latent_shape": list(latent.shape), + "rope_key_shape": list(rope_key.shape), + "latent_rms": latent.float().square().mean().sqrt().item(), + "rope_key_rms": rope_key.float().square().mean().sqrt().item(), + "eager_key_cache_shape": list(key_cache.shape), + "eager_value_cache_shape": list(value_cache.shape), + "eager_cache_bytes": key_cache.numel() * key_cache.element_size() + + value_cache.numel() * value_cache.element_size(), + }, + } + if layer_index > 0: + topk_ids = captures["topk_ids"].view( + input_ids.shape[0], input_ids.shape[1], -1 + ) + topk_weights = captures["topk_weights"].view( + input_ids.shape[0], input_ids.shape[1], -1 + ) + result["routing"] = route_trace( + tokenizer, + input_ids, + attention_mask_cpu, + topk_ids.cpu(), + topk_weights.cpu(), + DEFAULT_PROMPTS, + config.n_routed_experts, + ) + layer_results.append(result) + + del layer, compressed, latent, rope_key, key_cache, value_cache + captures.clear() + gc.collect() + if device.type == "cuda": + torch.cuda.empty_cache() + + valid_token_count = int(attention_mask_cpu.sum()) + latent_elements = config.kv_lora_rank + config.qk_rope_head_dim + eager_elements = config.num_attention_heads * ( + config.qk_nope_head_dim + config.qk_rope_head_dim + config.v_head_dim + ) + captured_at = args.captured_at or datetime.now(timezone.utc).isoformat() + index = json.loads((root / "model.safetensors.index.json").read_text()) + result = { + "schema_version": 1, + "captured_at": captured_at, + "evidence_identity": "X / official BF16 weights and tokenizer, local truncated forward", + "boundary": { + "model": "DeepSeek-V2-Lite base", + "executed_layers": list(range(args.layers)), + "total_model_layers": config.num_hidden_layers, + "full_model_generation": False, + "training_or_global_expert_load": False, + "expert_semantics_inferred": False, + "prompt_sample": "four authored prompts; descriptive trace, not population estimate", + "cache_note": "HF eager materializes expanded K/V; latent-cache arithmetic is derived from official dimensions", + }, + "provenance": { + "huggingface_model": "deepseek-ai/DeepSeek-V2-Lite", + "huggingface_revision": "604d5664dddd88a0433dbae533b7fe9472482de0", + "sha256": { + "config": sha256(root / "config.json"), + "modeling_code": sha256(root / "modeling_deepseek.py"), + "tokenizer": sha256(root / "tokenizer.json"), + "index": sha256(root / "model.safetensors.index.json"), + "shard_1": sha256(shard), + }, + "checkpoint_tensor_bytes": index["metadata"]["total_size"], + "shard_1_bytes": shard.stat().st_size, + }, + "environment": { + "python": platform.python_version(), + "platform": platform.platform(), + "libc": list(platform.libc_ver()), + "torch": torch.__version__, + "torch_cuda": torch.version.cuda, + "transformers": __import__("transformers").__version__, + "safetensors": __import__("safetensors").__version__, + "device": str(device), + "nvidia_smi": nvidia_smi() if device.type == "cuda" else None, + "matmul_allow_tf32": torch.backends.cuda.matmul.allow_tf32 + if device.type == "cuda" + else None, + }, + "configuration": { + "total_parameters_reported": "15.7B", + "activated_parameters_reported": "2.4B", + "layers": config.num_hidden_layers, + "hidden": config.hidden_size, + "attention_heads": config.num_attention_heads, + "qk_nope_head_dim": config.qk_nope_head_dim, + "qk_rope_head_dim": config.qk_rope_head_dim, + "v_head_dim": config.v_head_dim, + "kv_lora_rank": config.kv_lora_rank, + "routed_experts": config.n_routed_experts, + "active_routed_experts": config.num_experts_per_tok, + "shared_experts": config.n_shared_experts, + "expert_intermediate": config.moe_intermediate_size, + "first_dense_layers": config.first_k_dense_replace, + "router_scoring": config.scoring_func, + "router_topk_method": config.topk_method, + "normalize_selected_weights": config.norm_topk_prob, + }, + "prompts": [ + { + **prompt, + "token_ids": input_ids[index, : int(attention_mask_cpu[index].sum())].tolist(), + "tokens": tokenizer.convert_ids_to_tokens( + input_ids[index, : int(attention_mask_cpu[index].sum())].tolist() + ), + } + for index, prompt in enumerate(DEFAULT_PROMPTS) + ], + "execution": { + "batch": input_ids.shape[0], + "padded_sequence": input_ids.shape[1], + "valid_tokens": valid_token_count, + "embedding_ms": embedding_ms, + "initial_hidden": initial_hidden, + "layers": layer_results, + "final_hidden": hidden_stats(hidden, attention_mask), + }, + "cache_accounting": { + "dtype": "BF16", + "latent_elements_per_token_layer": latent_elements, + "latent_bytes_per_token_layer": latent_elements * 2, + "hf_eager_elements_per_token_layer": eager_elements, + "hf_eager_bytes_per_token_layer": eager_elements * 2, + "eager_over_latent_ratio": eager_elements / latent_elements, + "latent_reduction_vs_eager": 1 - latent_elements / eager_elements, + "components": { + "latent_content": config.kv_lora_rank, + "rope_key": config.qk_rope_head_dim, + "expanded_key": config.num_attention_heads + * (config.qk_nope_head_dim + config.qk_rope_head_dim), + "expanded_value": config.num_attention_heads * config.v_head_dim, + }, + }, + } + args.output.parent.mkdir(parents=True, exist_ok=True) + args.output.write_text(json.dumps(result, ensure_ascii=False, indent=2) + "\n") + print(json.dumps(result, ensure_ascii=False, indent=2)) + + +if __name__ == "__main__": + main() diff --git a/research/DEEPSEEK_RESEARCH.md b/research/DEEPSEEK_RESEARCH.md index d00358b..f909c8a 100644 --- a/research/DEEPSEEK_RESEARCH.md +++ b/research/DEEPSEEK_RESEARCH.md @@ -1,7 +1,7 @@ -# DeepSeek 技术谱系二轮正式研究账本 +# DeepSeek 技术谱系正式研究账本(含三轮真实权重执行) > 研究截止:2026-07-29 -> 课程角色:DeepSeek 聚光专题二轮;与 MoE、长上下文、训练系统、数值、推理、Agent、评测专题互相链接,但不替代各专题完整推导。 +> 课程角色:DeepSeek 聚光专题三轮;与 MoE、长上下文、训练系统、数值、推理、Agent、评测专题互相链接,但不替代各专题完整推导。 > 证据规则:正文事实只来自一手论文、作者官方仓库和 Kimi K3 官方报告;Grok 产物仅见 `DEEPSEEK_GROK_LEADS.md`,不承担证据。 > 简化规则:所有二维图、成本滑条和训练曲线若非论文复跑,必须标“教学模型”。 @@ -557,7 +557,7 @@ total bytes = per-token-layer · L · T · B · bytes - 至少 24 张问题账; - 至少 20 个正文目录; - 60 个一手/官方阅读节点; -- 四个独立可操作实验; +- 八个独立可操作实验:四个公式/教学实验 + 四个真实工件实验; - DeepSeekMath 必须在主时间线中; - DAPO / Dr.GRPO 必须标后续公开研究; - MLA 实验必须把 RoPE cache 算进去; @@ -568,3 +568,30 @@ total bytes = per-token-layer · L · T · B · bytes - tabs 支持键盘方向键; - toy model、作者报告和公式推导使用不同标签; - 专属 Chrome 回归并纳入全站回归。 + +## 8. 第三轮真实权重执行增量 + +第三轮固定官方 `deepseek-ai/DeepSeek-V2-Lite` revision +`604d5664dddd88a0433dbae533b7fe9472482de0`,在 RTX 5090 上用官方 tokenizer、 +模型代码与 BF16 第一分片连续执行 layer 0–6。 + +已闭合: + +- 4 条固定 prompt、90 个有效 token; +- 6 个 MoE 层、每 token top-6,共 3,240 次真实 routed-expert 选择; +- 每层 `[4,27,576]` MLA 压缩投影; +- HF eager 实际 key `[4,16,27,192]`、value `[4,16,27,128]`; +- V2-Lite 逐 token/层 BF16 的 1,152-byte latent 合同与 10,240-byte eager 物化账; +- 两次独立执行在来源、tokenization、hidden hashes、MLA shapes、loads 与全部 routes 上 + `31/31 exact`,计时排除。 + +永久边界: + +- 覆盖率不等于均衡,必须同时看 CV、Gini 与 effective experts; +- expert ID 不赋予语义,跨层同号 ID 不是同一参数; +- `88.75%` 是 V2-Lite latent vs HF eager 的 shape 算术,不冒充 V2 报告 `93.3%`; +- layer 7 因跨 shard 停止,不把 7/27 层写成完整模型生成; +- 原始权重不进 Git,工件以 revision、SHA-256、脚本、trace 与复跑比较固定。 + +完整方法、数值表与复现入口见 +`research/DEEPSEEK_V2_LITE_TRACE.md`。 diff --git a/research/DEEPSEEK_V2_LITE_TRACE.md b/research/DEEPSEEK_V2_LITE_TRACE.md new file mode 100644 index 0000000..4f13e48 --- /dev/null +++ b/research/DEEPSEEK_V2_LITE_TRACE.md @@ -0,0 +1,340 @@ +# DeepSeek-V2-Lite 真实权重执行审计 + +> 状态:DeepSeek 专题第三轮工件;2026-07-29 +> 目的:把“MLA 会压缩缓存”“MoE 会选择专家”从论文概念推进到固定官方权重、真实 tokenizer、真实 hidden states 和逐 token 路由记录。 +> 边界:这不是完整模型 benchmark、训练 trace、生产服务复现、专家语义分析或总体负载估计。 + +## 0. 一页结论 + +这次实验固定官方 `deepseek-ai/DeepSeek-V2-Lite` revision +`604d5664dddd88a0433dbae533b7fe9472482de0`,使用官方 tokenizer、配置、模型代码和 +BF16 checkpoint 第一分片,在 RTX 5090 上依次执行 embedding 与 decoder layer 0–6。 +第 0 层是 dense FFN,第 1–6 层是完整 MoE 层;第 7 层横跨第一、第二分片,因此明确停止。 + +四条最重要的结果: + +1. 4 条人工固定 prompt 共 90 个有效 token,在 6 个 MoE 层产生 + `6 layers × 90 tokens × top-6 = 3,240` 次真实 routed-expert 选择。 +2. 每个执行层都实际观测到 MLA 的 `[B,T,576]` 压缩投影;其中 512 是 + KV latent,64 是 decoupled RoPE key。 +3. 官方 Hugging Face eager 实现仍把 cache 物化为 + key `[4,16,27,192]` 和 value `[4,16,27,128]`。以 BF16 计, + 每 token、每层是 10,240 bytes;按配置保存 latent + RoPE key 只需 1,152 bytes, + 相差 `8.8889×`,后者相对前者减少 `88.75%`。 +4. 两次独立执行的来源、配置、tokenization、hidden-state hashes、MLA shapes、 + aggregate loads 与全部 token routes 共 `31/31` 项完全一致;计时明确不参与一致性检查。 + +这里最反直觉的发现不是某个“神奇专家”,而是: + +- **覆盖不等于均衡。** 90 个 token 在单层可触达 60–64/64 个专家,但各层 + CV 为 0.549–0.925,Gini 为 0.303–0.420。 +- **路由 ID 不等于语义标签。** `E8` 只表示某一层内部数组的第 8 号专家;跨层的 + `E8` 是不同参数,四条 prompt 也不足以赋予“代码专家”等语义。 +- **算法压缩不等于实现已经压缩存储。** MLA 的低秩状态合同与 HF eager cache + 的实际物化格式是两张账,必须同时报告。 + +## 1. 证据身份 + +### 1.1 官方工件 + +| 工件 | 固定值 | +|---|---| +| Hugging Face model | `deepseek-ai/DeepSeek-V2-Lite` | +| revision | `604d5664dddd88a0433dbae533b7fe9472482de0` | +| checkpoint tensor bytes | `31,412,968,448`,约 29.261 GiB | +| shard 1 bytes | `8,594,887,408`,约 8.005 GiB | +| config SHA-256 | `f346286b0f1c8b044252fd54cb4fa78b9fab6472a6e8bebb9edfe03d414ea03d` | +| modeling code SHA-256 | `7d8e5221095286eea991137760893fd7ba52727c0b4ebf48ec09e8bc56b45b9c` | +| tokenizer SHA-256 | `41f3bf64213da8c012d8bd0871a58a1fdf70463e8f08f110ddbb1082f529f669` | +| index SHA-256 | `d2cdb2f325f6682cf3ad1ad2526a9f979d857390b579380c0331d975136e0acf` | +| shard 1 SHA-256 | `0d7e9f39bde40111a4c0f390b87497dce4565cf578d916395e6b2c7851f1e8da` | + +第一分片完整容纳 embedding、layer 0–6 和 layer 7 的一部分。为了不把缺参数的 +layer 7 冒充完整执行,本实验把证据边界钉在 layer 6。 + +原始 checkpoint 分片不进入 Git;仓库只提交: + +- 可复现执行脚本; +- 机器可读 trace; +- 独立复跑比较结果; +- 研究审计与网页可视化。 + +### 1.2 模型配置 + +| 对象 | 官方配置 | +|---|---:| +| reported total / active | 15.7B / 2.4B | +| decoder layers | 27 | +| hidden size | 2048 | +| attention heads | 16 | +| qk NoPE / RoPE head dim | 128 / 64 | +| value head dim | 128 | +| KV LoRA rank | 512 | +| routed experts / active | 64 / 6 | +| shared experts | 2 | +| expert intermediate | 1408 | +| dense prefix | 1 layer | +| router | softmax + greedy top-k | +| selected top-k normalization | false | + +`max_position_embeddings=163840` 与 YaRN 配置属于 checkpoint config; +官方发布表中的上下文口径不应被本实验自行改写成同一个数字。 + +### 1.3 软件与硬件 + +| 对象 | 值 | +|---|---| +| GPU | NVIDIA GeForce RTX 5090,32,607 MiB | +| driver | 595.84 | +| Python | 3.10.14 | +| PyTorch | 2.11.0+cu128 | +| CUDA runtime | 12.8 | +| Transformers | 4.41.2 | +| safetensors | 0.8.0 | +| TF32 matmul | disabled | + +官方 2024 remote code 在当前 Transformers 5.5 环境中会因已移除的 +`is_torch_fx_available` 入口而导入失败。因此实验使用隔离的 Transformers 4.41.2。 +模型源码保持只读、不打补丁;runner 只通过本地 package 方式导入官方 +`configuration_deepseek.py` 与 `modeling_deepseek.py`。 + +这条版本边界很重要:**“模型代码公开”不等于“任意未来依赖版本都可直接运行”。** + +## 2. 实验设计 + +### 2.1 为什么选择截断的精确 forward + +完整 15.7B BF16 权重与运行时状态虽然可能勉强接近单卡边界,但没有必要为了观察 +前六层真实路由而把“全模型能否高效生成”混进同一实验。本 runner: + +1. 用官方 tokenizer 一次编码 4 条 prompt; +2. 从第一分片读取官方 embedding; +3. 每次在 meta device 构造一个官方 decoder layer; +4. 为这一层绑定真实 safetensors 参数并传到 GPU; +5. 用上层输出作为下一层输入; +6. 通过 hook 捕获实际 `kv_a_proj_with_mqa` 与实际 MoE gate 输出; +7. 释放当前层,再加载下一层。 + +因此得到的是**真实权重、真实层类、真实连续 hidden states**,而不是随机初始化、 +只读 config 或合成 router;但它仍不是完整 27 层生成。 + +### 2.2 固定 prompt + +| ID | 标签 | token | +|---|---|---:| +| `zh_explanation` | 中文解释 | 26 | +| `en_architecture` | English architecture | 18 | +| `code` | Python code | 19 | +| `math` | 数学推理 | 27 | + +batch padding 后形状为 `B=4, T=27`,有效 token 共 90。prompt 是用于建立可复现 +探针的人工样本,不是自然语料分布、benchmark 数据集或统计代表样本。 + +## 3. MLA:一张算法状态账,一张实现物化账 + +### 3.1 实际观测 + +每个执行层的 `kv_a_proj_with_mqa` 输出都是: + +```text +[batch, sequence, kv_lora_rank + qk_rope_head_dim] += [4, 27, 512 + 64] += [4, 27, 576] +``` + +这 576 个元素可以分为: + +- `c_t^KV`:512 维 joint latent; +- `k_t^R`:64 维 decoupled RoPE key。 + +后者不能因为体积小就从缓存公式中消失。 + +### 3.2 HF eager 实际保存了什么 + +本次 `use_cache=True` 的官方 Hugging Face eager 路径实际返回: + +```text +key [4, 16, 27, 192] +value [4, 16, 27, 128] +``` + +所以每 token、每层的缓存元素是: + +```text +expanded eager = 16 × (192 + 128) = 5,120 elements +latent contract = 512 + 64 = 576 elements +``` + +在 BF16 下: + +```text +eager bytes = 5,120 × 2 = 10,240 +latent bytes = 576 × 2 = 1,152 +ratio = 10,240 / 1,152 = 8.8889× +reduction = 1 - 1,152 / 10,240 = 88.75% +``` + +### 3.3 为什么这不是论文 93.3% 的复跑 + +V2 报告的 `93.3%` 是完整 V2 相对 DeepSeek 67B、在论文指定配置和比较口径下的 +作者报告结果。本实验的 `88.75%` 是 V2-Lite 配置中: + +- HF eager 展开 K/V 元素; +- 与 latent + RoPE key 理论存储元素; + +之间的逐 token、逐层 BF16 算术比较。模型、维度、基线和实现口径都不同, +二者不能互相替代。 + +进一步说,低秩架构只定义“可以保存什么最小状态”;生产系统是否真的只保存 latent, +还取决于 attention kernel、权重吸收、RoPE 分叉、paged cache 和服务框架。 + +## 4. MoE:真实路由显微镜 + +### 4.1 总量守恒 + +每个 MoE 层都满足: + +```text +90 valid tokens × top-6 = 540 routed selections +``` + +6 个层合计 `3,240` 次。两条 shared expert 等价路径始终运行,但它们不属于 router +输出的 0–63 routed expert ID,因此不写进 top-6 列表。 + +### 4.2 六层聚合负载 + +| layer | selections | used / 64 | zero | CV | Gini | effective experts | top expert | +|---:|---:|---:|---:|---:|---:|---:|---| +| 1 | 540 | 63 | 1 | 0.925 | 0.420 | 46.66 | E8 · 48 | +| 2 | 540 | 64 | 0 | 0.549 | 0.303 | 55.08 | E62 · 23 | +| 3 | 540 | 62 | 2 | 0.569 | 0.322 | 53.51 | E54 · 19 | +| 4 | 540 | 62 | 2 | 0.866 | 0.417 | 47.03 | E48 · 40 | +| 5 | 540 | 60 | 4 | 0.642 | 0.332 | 52.14 | E14 · 32 | +| 6 | 540 | 61 | 3 | 0.765 | 0.378 | 49.48 | E25 · 40 | + +三个指标回答不同问题: + +- `used experts`:这批 token 是否至少触达某专家一次; +- `CV = std / mean`:计数离散程度; +- `effective experts = exp(entropy)`:若把当前熵换算成均匀分布,相当于多少个专家。 + +因此 layer 1 触达 63 个专家,却只有 46.66 个 effective experts,且 E8 承担 48/540 +次选择。只报“63/64 被使用”会掩盖负载长尾。 + +### 4.3 prompt 之间看到的差异 + +同一层内,四条 prompt 的 used-expert set Jaccard 与 CV 会变化。例如: + +- layer 1:中文 prompt 使用 54 个专家、CV 0.898;英文架构 prompt 使用 40 个、 + CV 1.360;代码 prompt 44 个、CV 1.263;数学 prompt 45 个、CV 1.184。 +- layer 4:聚合 top expert 是 E48;英文和代码各把 E48 路由 11 次, + 中文 prompt 的 top expert 却是 E29、16 次。 +- layer 6:英文 prompt 使用 42 个专家、CV 1.439;中文为 49 个、1.031; + 数学为 48 个、0.915。 + +这些是**描述性 route traces**,说明 token 与上下文确实改变 gate 输出;它们不证明 +“E48 是代码专家”,更不允许以四条 prompt 推断训练语料上的领域分工。 + +### 4.4 top-6 权重没有重新归一 + +配置 `norm_topk_prob=false`。router 先对 64 个专家做 softmax,再选 top-6; +被选中的 6 个概率不会重新缩放到和为 1。各层选中权重和的均值: + +| layer | mean selected top-6 weight sum | +|---:|---:| +| 1 | 0.3949 | +| 2 | 0.4801 | +| 3 | 0.4898 | +| 4 | 0.4749 | +| 5 | 0.4832 | +| 6 | 0.4440 | + +所以网页显示某 token 的六个权重和小于 1 是真实配置行为,不是漏算。 + +## 5. 可复现性闸门 + +正式 trace: + +```text +src/data/deepseek-v2-lite-trace.json +SHA-256 81ff4ab34d20121fe42219c566dd4780f4c7f74252a31b596f90e5280e1f1aef +captured 2026-07-29T05:55:35.599912+00:00 +``` + +第二次独立执行: + +```text +SHA-256 b5e2e7567e17765195313473499522a4529b9d37ea96963ba2b9455c3c0b86e6 +captured 2026-07-29T05:56:32.527770+00:00 +``` + +比较结果: + +```text +31 / 31 exact +timing compared: false +``` + +精确检查覆盖: + +- provenance、configuration、prompt tokenization; +- initial / final hidden hashes; +- layer 0–6 hidden hashes 与 MLA shapes; +- layer 1–6 aggregate loads; +- layer 1–6 全部 token routes。 + +计时会受首次 kernel、频率、温度、后台进程与同步位置影响,因此没有伪装成 deterministic +evidence。本轮只把计时保留在原始 trace,网站不把它包装成性能 benchmark。 + +## 6. 绝对禁止的外推 + +| 本实验能够说 | 本实验不能说 | +|---|---| +| 固定 revision 的前 7 层产生了这些 hidden states | 完整模型会生成什么回答 | +| 四条 prompt 在 6 层产生这些 top-6 routes | 训练集或线上请求的总体专家负载 | +| 某层某 expert ID 被选中多少次 | 该专家具有什么稳定语义 | +| HF eager 物化了展开 K/V | 所有 MLA serving kernel 都这样缓存 | +| V2-Lite 配置可导出 576 元素 latent state | 完整 V2 的 93.3% 已被复跑 | +| 两次相同环境执行 31/31 exact | 跨 GPU、依赖版本和 kernel 仍 bitwise identical | +| layer 0–6 在第一分片内完整 | layer 7 或 layer 8–26 已执行 | + +跨层连线也被禁止:layer 1 的 E8 与 layer 2 的 E8 是不同参数对象。可视化只能在同一层内 +比较 expert load,不能画一条 “E8 专长” 纵贯 6 层。 + +## 7. 复现入口 + +环境与命令见: + +- `experiments/deepseek/README.md` +- `experiments/deepseek/v2_lite_trace.py` +- `experiments/deepseek/compare_v2_lite_traces.py` + +执行一次: + +```bash +python experiments/deepseek/v2_lite_trace.py \ + --artifact-dir /path/to/deepseek-v2-lite \ + --output /path/to/trace.json +``` + +比较两次: + +```bash +python experiments/deepseek/compare_v2_lite_traces.py \ + --first /path/to/trace-1.json \ + --second /path/to/trace-2.json \ + --output /path/to/repro.json +``` + +## 8. 下一道研究闸门 + +本轮已经把 DeepSeek 第三轮从“计划真实 trace”推进到真实 tokenizer、MLA state、 +MoE routes 与确定性复跑,但仍有四块未闭合: + +1. 用真正只保存 latent 的 MLA kernel 对照 HF eager materialization; +2. 扩大公开语料样本,给 aggregate load 置信区间,并考察 batch / padding / sequence + 对 expert parallel 的影响; +3. 复跑 FP8、pipeline 或通信 trace,而不是继续用教学 schedule; +4. 做可审计的 R1-like 小模型训练实验,记录 rollout、reward、KL、长度与梯度统计。 + +在这些闸门完成之前,DeepSeek 专题保持“第三轮进行中”,不写成完整独立复现。 diff --git a/scripts/check-deepseek-browser.mjs b/scripts/check-deepseek-browser.mjs index e80b70b..37a02d5 100644 --- a/scripts/check-deepseek-browser.mjs +++ b/scripts/check-deepseek-browser.mjs @@ -73,6 +73,9 @@ const overview = await evaluate(`(() => ({ paperLinks: document.querySelectorAll("[data-deepseek-paper-chain] a").length, labTabs: document.querySelectorAll("[data-ds-tab]").length, labPanels: document.querySelectorAll("[data-ds-panel]").length, + artifactTabs: document.querySelectorAll("[data-artifact-tab]").length, + artifactPanels: document.querySelectorAll("[data-artifact-panel]").length, + artifactLayers: document.querySelectorAll(".layer-evidence > span").length, branches: document.querySelectorAll(".branch-grid > a").length, followups: document.querySelectorAll(".lineage-row.followup").length, navLinks: document.querySelectorAll(".top-nav a").length, @@ -195,12 +198,111 @@ const rl = await evaluate(`(() => { }; })()`); +const artifactRoute = await evaluate(`(() => { + const root = document.querySelector("[data-dsv2-lab]"); + const read = () => ({ + panel: root.querySelector("[data-artifact-panel]:not([hidden])").dataset.artifactPanel, + position: root.querySelector("[data-route-position]").textContent.trim(), + piece: root.querySelector("[data-route-piece]").textContent.trim(), + tokenId: root.querySelector("[data-route-token-id]").textContent.trim(), + weightSum: root.querySelector("[data-route-weight-sum]").textContent.trim(), + chosen: [...root.querySelectorAll("[data-route-experts] article")].map((node) => ({ + expert: node.querySelector("b").textContent.trim(), + weight: node.querySelector("small").textContent.trim(), + })), + heatCells: root.querySelectorAll("[data-route-heatmap] > span").length, + selectedCells: root.querySelectorAll("[data-route-heatmap] > span.selected").length, + routes: root.querySelector("[data-route-count]").textContent.trim(), + used: root.querySelector("[data-route-used]").textContent.trim(), + cv: root.querySelector("[data-route-cv]").textContent.trim(), + effective: root.querySelector("[data-route-effective]").textContent.trim(), + }); + const initial = read(); + const layer = root.querySelector("[data-route-layer]"); + const prompt = root.querySelector("[data-route-prompt]"); + const token = root.querySelector("[data-route-token]"); + layer.value = "4"; + layer.dispatchEvent(new Event("change", { bubbles: true })); + prompt.value = "en_architecture"; + prompt.dispatchEvent(new Event("change", { bubbles: true })); + token.value = String(token.options.length - 1); + token.dispatchEvent(new Event("change", { bubbles: true })); + const switched = read(); + return { initial, switched, tokenOptions: token.options.length }; +})()`); + +const artifactLoad = await evaluate(`(() => { + const root = document.querySelector("[data-dsv2-lab]"); + root.querySelector('[data-artifact-tab="load"]').click(); + const read = () => ({ + panel: root.querySelector("[data-artifact-panel]:not([hidden])").dataset.artifactPanel, + used: root.querySelector("[data-load-used]").textContent.trim(), + zero: root.querySelector("[data-load-zero]").textContent.trim(), + cv: root.querySelector("[data-load-cv]").textContent.trim(), + gini: root.querySelector("[data-load-gini]").textContent.trim(), + effective: root.querySelector("[data-load-effective]").textContent.trim(), + rows: root.querySelectorAll("[data-load-rows] > div").length, + jaccards: root.querySelectorAll("[data-load-jaccard] > article").length, + }); + const layer1 = read(); + root.querySelector('[data-load-layer="2"]').click(); + const layer2 = read(); + root.querySelector('[data-load-layer="4"]').click(); + const layer4 = read(); + return { layer1, layer2, layer4 }; +})()`); + +const artifactCache = await evaluate(`(() => { + const root = document.querySelector("[data-dsv2-lab]"); + root.querySelector('[data-artifact-tab="cache"]').click(); + const read = () => ({ + panel: root.querySelector("[data-artifact-panel]:not([hidden])").dataset.artifactPanel, + latent: root.querySelector("[data-cache-latent-bytes]").textContent.trim(), + eager: root.querySelector("[data-cache-eager-bytes]").textContent.trim(), + ratio: root.querySelector("[data-cache-ratio]").textContent.trim(), + reduction: root.querySelector("[data-cache-reduction]").textContent.trim(), + }); + const trace = read(); + const context = root.querySelector("[data-cache-context]"); + const batch = root.querySelector("[data-cache-batch]"); + const layers = root.querySelector("[data-cache-layers]"); + context.value = "1048576"; + context.dispatchEvent(new Event("change", { bubbles: true })); + batch.value = "8"; + batch.dispatchEvent(new Event("input", { bubbles: true })); + layers.value = "27"; + layers.dispatchEvent(new Event("input", { bubbles: true })); + const million = read(); + return { trace, million }; +})()`); + +const artifactEvidence = await evaluate(`(() => { + const root = document.querySelector("[data-dsv2-lab]"); + root.querySelector('[data-artifact-tab="evidence"]').click(); + const result = { + panel: root.querySelector("[data-artifact-panel]:not([hidden])").dataset.artifactPanel, + layers: root.querySelectorAll(".layer-evidence > span").length, + executed: root.querySelectorAll(".layer-evidence > span.executed").length, + split: root.querySelectorAll(".layer-evidence > span.split").length, + unloaded: root.querySelectorAll(".layer-evidence > span.unloaded").length, + exact: root.querySelector(".repro-gate strong").textContent.trim(), + dependency: root.querySelector(".dependency-split").textContent.replaceAll(/\\s+/g, " ").trim(), + boundary: root.querySelector('[data-artifact-panel="evidence"] .artifact-boundary').textContent.replaceAll(/\\s+/g, " ").trim(), + }; + const first = root.querySelector('[data-artifact-tab="route"]'); + first.focus(); + first.dispatchEvent(new KeyboardEvent("keydown", { key: "ArrowRight", bubbles: true })); + result.keyboardSelected = root.querySelector('[data-artifact-tab][aria-selected="true"]').dataset.artifactTab; + result.keyboardVisible = root.querySelector("[data-artifact-panel]:not([hidden])").dataset.artifactPanel; + return result; +})()`); + await evaluate(`(() => { - document.querySelector("[data-deepseek-lab]").scrollIntoView({ block: "start", behavior: "instant" }); + document.querySelector("[data-dsv2-lab]").scrollIntoView({ block: "start", behavior: "instant" }); window.scrollBy(0, -82); })()`); await pause(180); -await screenshot("/tmp/llm-atlas-deepseek-lab-desktop.png"); +await screenshot("/tmp/llm-atlas-deepseek-artifact-desktop.png"); await navigate("/"); const home = await evaluate(`(() => ({ @@ -233,6 +335,7 @@ await command("Emulation.setDeviceMetricsOverride", { await navigate("/deepseek/"); const mobile = await evaluate(`(() => { const root = document.querySelector("[data-deepseek-lab]"); + const artifact = document.querySelector("[data-dsv2-lab]"); root.scrollIntoView({ block: "start", behavior: "instant" }); const toggle = document.querySelector("#menu-toggle"); toggle?.click(); @@ -242,8 +345,10 @@ const mobile = await evaluate(`(() => { menuOpen: toggle.getAttribute("aria-expanded"), mobileLinks: document.querySelectorAll("#mobile-nav a").length, tabs: root.querySelectorAll("[data-ds-tab]").length, + artifactTabs: artifact.querySelectorAll("[data-artifact-tab]").length, + artifactHeatCells: artifact.querySelectorAll("[data-route-heatmap] > span").length, offenders: [...document.querySelectorAll("body *")] - .filter((node) => !node.closest(".paper-chain, .advantage-table, .precision-table, .mapping-table, [data-deepseek-lab]")) + .filter((node) => !node.closest(".paper-chain, .advantage-table, .precision-table, .mapping-table, [data-deepseek-lab], [data-dsv2-lab]")) .filter((node) => node.getBoundingClientRect().right > document.documentElement.clientWidth + 1) .slice(0, 12) .map((node) => ({ @@ -256,21 +361,23 @@ const mobile = await evaluate(`(() => { })()`); await evaluate(`(() => { document.querySelector("#menu-toggle")?.click(); + document.querySelector("[data-dsv2-lab]").scrollIntoView({ block: "start", behavior: "instant" }); window.scrollBy(0, -82); })()`); await pause(180); -await screenshot("/tmp/llm-atlas-deepseek-mobile.png"); +await screenshot("/tmp/llm-atlas-deepseek-artifact-mobile.png"); -const report = { overview, capacity, cache, codesign, rl, home, papers, mobile, exceptions }; +const report = { overview, capacity, cache, codesign, rl, artifactRoute, artifactLoad, artifactCache, artifactEvidence, home, papers, mobile, exceptions }; console.log(JSON.stringify(report, null, 2)); const numeric = (text) => Number.parseFloat(text.replaceAll(",", "")); const failures = []; if (!overview.title.includes("为什么转向")) failures.push("专题标题异常"); -if (overview.sections !== 25 || overview.tocLinks !== 25) failures.push("二十四个编号专题加阅读链的目录结构异常"); +if (overview.sections !== 26 || overview.tocLinks !== 26) failures.push("二十五个编号专题加阅读链的目录结构异常"); if (overview.ledgers !== 24 || overview.waves !== 10) failures.push("二十四张问题账或十次转向结构异常"); if (overview.paperLinks !== 60 || overview.branches !== 5 || overview.followups !== 1) failures.push("论文链、旁支或公开后续标记异常"); if (overview.labTabs !== 4 || overview.labPanels !== 4) failures.push("四联实验结构异常"); +if (overview.artifactTabs !== 4 || overview.artifactPanels !== 4 || overview.artifactLayers !== 27) failures.push("真实权重四联实验结构异常"); if (overview.navLinks !== 20 || home.navLinks !== 20 || mobile.mobileLinks !== 20 || overview.activeNav !== "DeepSeek") failures.push("全站导航未同步 DeepSeek"); if (overview.documentOverflow > 1 || mobile.documentOverflow > 1) failures.push("桌面或移动端存在文档级横向溢出"); if (capacity.initial.panel !== "capacity" || capacity.initial.total !== "32.1× FFN" || capacity.initial.active !== "1.13× FFN") failures.push("V3 稀疏容量初始账异常"); @@ -287,9 +394,19 @@ if (rl.initial.signal !== "GROUP-RELATIVE SIGNAL" || rl.same.signal !== "ZERO GR if (!rl.dapo.provenance.includes("2503.14476") || !rl.dapo.algorithm.includes("FOLLOW-UP") || !rl.dr.provenance.includes("2503.20783")) failures.push("DAPO / Dr.GRPO 来源边界异常"); if (!rl.r1.includes("cold start") || !rl.distill.includes("没有重演")) failures.push("R1 / distill 身份切换异常"); if (rl.keyboardSelected !== "cache" || rl.keyboardVisible !== "cache") failures.push("实验键盘 tab 导航异常"); +if (artifactRoute.initial.panel !== "route" || artifactRoute.initial.chosen.length !== 6 || artifactRoute.initial.heatCells !== 64 || artifactRoute.initial.selectedCells !== 6) failures.push("真实 top-6 路由结构异常"); +if (artifactRoute.initial.routes !== "156" || artifactRoute.initial.used !== "54 / 64" || artifactRoute.initial.weightSum !== "0.4540") failures.push("Layer 1 中文 token 路由初值异常"); +if (artifactRoute.switched.chosen.length !== 6 || artifactRoute.switched.heatCells !== 64 || artifactRoute.tokenOptions !== 18 || numeric(artifactRoute.switched.weightSum) >= 1) failures.push("路由层 / prompt / token 切换异常"); +if (artifactLoad.layer1.panel !== "load" || artifactLoad.layer1.used !== "63" || artifactLoad.layer1.cv !== "0.925" || artifactLoad.layer1.rows !== 5 || artifactLoad.layer1.jaccards !== 6) failures.push("Layer 1 聚合负载账异常"); +if (artifactLoad.layer2.used !== "64" || artifactLoad.layer2.cv !== "0.549" || artifactLoad.layer4.used !== "62" || artifactLoad.layer4.gini !== "0.417") failures.push("跨层负载统计切换异常"); +if (artifactCache.trace.panel !== "cache" || artifactCache.trace.latent !== "850.50 KiB" || artifactCache.trace.eager !== "7.38 MiB" || artifactCache.trace.ratio !== "8.89×" || artifactCache.trace.reduction !== "88.75%") failures.push("V2-Lite trace 缓存实现账异常"); +if (!artifactCache.million.latent.includes("GiB") || !artifactCache.million.eager.includes("TiB")) failures.push("V2-Lite 百万 Token 缓存外推异常"); +if (artifactEvidence.panel !== "evidence" || artifactEvidence.layers !== 27 || artifactEvidence.executed !== 7 || artifactEvidence.split !== 1 || artifactEvidence.unloaded !== 19 || artifactEvidence.exact !== "31 / 31") failures.push("真实工件执行边界或复跑闸门异常"); +if (!artifactEvidence.dependency.includes("Transformers 5.5") || !artifactEvidence.dependency.includes("4.41.2") || !artifactEvidence.boundary.includes("完整 27 层生成")) failures.push("依赖版本或未覆盖边界异常"); +if (artifactEvidence.keyboardSelected !== "load" || artifactEvidence.keyboardVisible !== "load") failures.push("真实工件实验键盘 tab 导航异常"); if (home.releaseCards !== 17 || !home.firstRelease.includes("47 页不再压成摘要") || home.firstHref !== "/k3/" || home.paperCount !== "486") failures.push("首页 DeepSeek 首发入口或论文数异常"); if (papers.total !== 486 || !papers.hasFilter || papers.visible < 20 || !papers.hasCoder || !papers.hasEngram) failures.push("论文库 DeepSeek 聚光异常"); -if (!mobile.menuVisible || mobile.menuOpen !== "true" || mobile.tabs !== 4) failures.push("移动端导航或实验异常"); +if (!mobile.menuVisible || mobile.menuOpen !== "true" || mobile.tabs !== 4 || mobile.artifactTabs !== 4 || mobile.artifactHeatCells !== 64) failures.push("移动端导航或实验异常"); if (mobile.offenders.length) failures.push(`移动端越界元素:${JSON.stringify(mobile.offenders)}`); if (exceptions.length) failures.push(`浏览器异常:${exceptions.join(" | ")}`); diff --git a/src/components/DeepSeekArtifactLab.astro b/src/components/DeepSeekArtifactLab.astro new file mode 100644 index 0000000..23e018c --- /dev/null +++ b/src/components/DeepSeekArtifactLab.astro @@ -0,0 +1,1084 @@ +--- +import rawTrace from "@/data/deepseek-v2-lite-trace.json"; +import rawRepro from "@/data/deepseek-v2-lite-trace-repro.json"; + +const trace = rawTrace as any; +const repro = rawRepro as any; +const compact = { + prompts: trace.prompts.map((prompt: any) => ({ + id: prompt.id, + label: prompt.label, + text: prompt.text, + })), + layers: trace.execution.layers.slice(1).map((layer: any) => ({ + layer: layer.layer, + load: layer.routing.aggregate_load, + metrics: layer.routing.aggregate_metrics, + selectedWeight: layer.routing.selected_weight_sum, + topExperts: layer.routing.top_experts, + prompts: layer.routing.per_prompt.map((prompt: any) => ({ + id: prompt.id, + label: prompt.label, + tokens: prompt.tokens, + load: prompt.load, + metrics: prompt.metrics, + topExperts: prompt.top_experts, + tokenRoutes: prompt.token_routes, + })), + jaccard: layer.routing.prompt_pair_jaccard, + })), + cache: trace.cache_accounting, + execution: { + batch: trace.execution.batch, + paddedSequence: trace.execution.padded_sequence, + validTokens: trace.execution.valid_tokens, + }, +}; +const compactJson = JSON.stringify(compact).replaceAll("<", "\\u003c"); +const shardFraction = trace.provenance.shard_1_bytes / trace.provenance.checkpoint_tensor_bytes; +--- + +
+
+
+

REAL-WEIGHT LAB / DEEPSEEK-V2-LITE

+

这一次不是调公式:让官方权重真的走过七层

+
+

+ 固定官方 revision、tokenizer、模型代码和 BF16 第一分片;RTX 5090 连续执行 layer 0–6, + 捕获真实 MLA 状态与 3,240 次 routed-expert 选择。所有结论都带证据身份与停止线。 +

+
+ +
+ X本机执行真实权重与 hidden states + O官方工件配置、代码与 checkpoint + D确定推导从 shape 计算 bytes + U未覆盖全模型、训练与生产 kernel +
+ +
+ + + + +
+ +
+
+
X / TOKEN-LEVEL ROUTES

先选层,再选 prompt 和 token

+

同一个 expert ID 只在当前层内有意义;跨层同号专家是不同参数,图中不会把它们连成“专长轨迹”。

+
+ +
+ + + +
+ +
+ PROMPT +

+
+ +
+
+ TOKEN #0 + + ID +
+ +
+ ACTUAL TOP-6 / WEIGHT DESCENDING +
+
+ +
+ SELECTED WEIGHT SUM + + 配置不把 top-6 重新归一到 1 +
+
+ +
+
+ 64 ROUTED EXPERTS / CURRENT PROMPT + +
+

颜色表示这条 prompt 在当前层的选择次数;空白表示本小样本未触达,不表示专家失效。

+
+
+ +
+
ROUTES

tokens × top-6

+
USED / 64

至少被选一次

+
CV

std ÷ mean

+
EFFECTIVE

exp(route entropy)

+
+ +
+ X / DESCRIPTIVE TRACE +

这是 4 条固定 prompt、90 个有效 token 的前六个 MoE 层;不能据此命名专家、估计线上总体负载或判断训练均衡。

+
+
+ + + + + + + +
+ 可复现入口 + experiments/deepseek/v2_lite_trace.py · + src/data/deepseek-v2-lite-trace.json · + research/DEEPSEEK_V2_LITE_TRACE.md +
+ + +
+ + + + diff --git a/src/data/deepseek-v2-lite-trace-repro.json b/src/data/deepseek-v2-lite-trace-repro.json new file mode 100644 index 0000000..5e94b26 --- /dev/null +++ b/src/data/deepseek-v2-lite-trace-repro.json @@ -0,0 +1,142 @@ +{ + "schema_version": 1, + "captured_at": "2026-07-29T05:57:27.848243+00:00", + "first": { + "captured_at": "2026-07-29T05:55:35.599912+00:00", + "sha256": "81ff4ab34d20121fe42219c566dd4780f4c7f74252a31b596f90e5280e1f1aef" + }, + "second": { + "captured_at": "2026-07-29T05:56:32.527770+00:00", + "sha256": "b5e2e7567e17765195313473499522a4529b9d37ea96963ba2b9455c3c0b86e6" + }, + "timing_compared": false, + "checks": [ + { + "name": "provenance", + "exact": true + }, + { + "name": "configuration", + "exact": true + }, + { + "name": "prompt_tokenization", + "exact": true + }, + { + "name": "initial_hidden_hash", + "exact": true + }, + { + "name": "final_hidden_hash", + "exact": true + }, + { + "name": "layer_0_hidden_hash", + "exact": true + }, + { + "name": "layer_0_mla_shapes", + "exact": true + }, + { + "name": "layer_1_hidden_hash", + "exact": true + }, + { + "name": "layer_1_mla_shapes", + "exact": true + }, + { + "name": "layer_1_aggregate_load", + "exact": true + }, + { + "name": "layer_1_token_routes", + "exact": true + }, + { + "name": "layer_2_hidden_hash", + "exact": true + }, + { + "name": "layer_2_mla_shapes", + "exact": true + }, + { + "name": "layer_2_aggregate_load", + "exact": true + }, + { + "name": "layer_2_token_routes", + "exact": true + }, + { + "name": "layer_3_hidden_hash", + "exact": true + }, + { + "name": "layer_3_mla_shapes", + "exact": true + }, + { + "name": "layer_3_aggregate_load", + "exact": true + }, + { + "name": "layer_3_token_routes", + "exact": true + }, + { + "name": "layer_4_hidden_hash", + "exact": true + }, + { + "name": "layer_4_mla_shapes", + "exact": true + }, + { + "name": "layer_4_aggregate_load", + "exact": true + }, + { + "name": "layer_4_token_routes", + "exact": true + }, + { + "name": "layer_5_hidden_hash", + "exact": true + }, + { + "name": "layer_5_mla_shapes", + "exact": true + }, + { + "name": "layer_5_aggregate_load", + "exact": true + }, + { + "name": "layer_5_token_routes", + "exact": true + }, + { + "name": "layer_6_hidden_hash", + "exact": true + }, + { + "name": "layer_6_mla_shapes", + "exact": true + }, + { + "name": "layer_6_aggregate_load", + "exact": true + }, + { + "name": "layer_6_token_routes", + "exact": true + } + ], + "exact_checks": 31, + "total_checks": 31, + "all_exact": true +} diff --git a/src/data/deepseek-v2-lite-trace.json b/src/data/deepseek-v2-lite-trace.json new file mode 100644 index 0000000..64eceb7 --- /dev/null +++ b/src/data/deepseek-v2-lite-trace.json @@ -0,0 +1,23332 @@ +{ + "schema_version": 1, + "captured_at": "2026-07-29T05:55:35.599912+00:00", + "evidence_identity": "X / official BF16 weights and tokenizer, local truncated forward", + "boundary": { + "model": "DeepSeek-V2-Lite base", + "executed_layers": [ + 0, + 1, + 2, + 3, + 4, + 5, + 6 + ], + "total_model_layers": 27, + "full_model_generation": false, + "training_or_global_expert_load": false, + "expert_semantics_inferred": false, + "prompt_sample": "four authored prompts; descriptive trace, not population estimate", + "cache_note": "HF eager materializes expanded K/V; latent-cache arithmetic is derived from official dimensions" + }, + "provenance": { + "huggingface_model": "deepseek-ai/DeepSeek-V2-Lite", + "huggingface_revision": "604d5664dddd88a0433dbae533b7fe9472482de0", + "sha256": { + "config": "f346286b0f1c8b044252fd54cb4fa78b9fab6472a6e8bebb9edfe03d414ea03d", + "modeling_code": "7d8e5221095286eea991137760893fd7ba52727c0b4ebf48ec09e8bc56b45b9c", + "tokenizer": "41f3bf64213da8c012d8bd0871a58a1fdf70463e8f08f110ddbb1082f529f669", + "index": "d2cdb2f325f6682cf3ad1ad2526a9f979d857390b579380c0331d975136e0acf", + "shard_1": "0d7e9f39bde40111a4c0f390b87497dce4565cf578d916395e6b2c7851f1e8da" + }, + "checkpoint_tensor_bytes": 31412968448, + "shard_1_bytes": 8594887408 + }, + "environment": { + "python": "3.10.14", + "platform": "Linux-7.0.0-28-generic-x86_64-with-glibc2.43", + "libc": [ + "glibc", + "2.43" + ], + "torch": "2.11.0+cu128", + "torch_cuda": "12.8", + "transformers": "4.41.2", + "safetensors": "0.8.0", + "device": "cuda", + "nvidia_smi": { + "name": "NVIDIA GeForce RTX 5090", + "driver_version": "595.84", + "memory.total": "32607", + "power.limit": "600.00", + "clocks.max.sm": "3105" + }, + "matmul_allow_tf32": false + }, + "configuration": { + "total_parameters_reported": "15.7B", + "activated_parameters_reported": "2.4B", + "layers": 27, + "hidden": 2048, + "attention_heads": 16, + "qk_nope_head_dim": 128, + "qk_rope_head_dim": 64, + "v_head_dim": 128, + "kv_lora_rank": 512, + "routed_experts": 64, + "active_routed_experts": 6, + "shared_experts": 2, + "expert_intermediate": 1408, + "first_dense_layers": 1, + "router_scoring": "softmax", + "router_topk_method": "greedy", + "normalize_selected_weights": false + }, + "prompts": [ + { + "id": "zh_explanation", + "label": "中文解释", + "text": "用通俗的语言解释,为什么稀疏专家模型可以拥有很多参数,但每个 token 只使用其中一小部分。", + "token_ids": [ + 100000, + 1082, + 83983, + 91476, + 17882, + 19304, + 10285, + 21844, + 20923, + 10900, + 39322, + 1876, + 9476, + 3599, + 21273, + 19304, + 1705, + 13712, + 10728, + 207, + 2141, + 5118, + 6802, + 51905, + 5781, + 398 + ], + "tokens": [ + "<|begin▁of▁sentence|>", + "ç͍", + "éĢļä¿Ĺ", + "çļĦè¯Ńè¨Ģ", + "è§£éĩĬ", + "ï¼Į", + "为ä»Ģä¹Ī", + "ç¨Ģ", + "çĸı", + "ä¸ĵå®¶", + "模åŀĭ", + "åı¯ä»¥", + "æĭ¥æľī", + "å¾Īå¤ļ", + "åıĤæķ°", + "ï¼Į", + "ä½Ĩ", + "æ¯ı个", + "Ġtoken", + "Ġ", + "åıª", + "使ç͍", + "åħ¶ä¸Ń", + "ä¸Ģå°ı", + "éĥ¨åĪĨ", + "ãĢĤ" + ] + }, + { + "id": "en_architecture", + "label": "English architecture", + "text": "Explain how a compressed key-value latent changes the memory cost of autoregressive decoding.", + "token_ids": [ + 100000, + 99297, + 946, + 245, + 29640, + 2131, + 12, + 2834, + 28937, + 4183, + 254, + 4887, + 2295, + 280, + 77939, + 79184, + 39671, + 13 + ], + "tokens": [ + "<|begin▁of▁sentence|>", + "Explain", + "Ġhow", + "Ġa", + "Ġcompressed", + "Ġkey", + "-", + "value", + "Ġlatent", + "Ġchanges", + "Ġthe", + "Ġmemory", + "Ġcost", + "Ġof", + "Ġautore", + "gressive", + "Ġdecoding", + "." + ] + }, + { + "id": "code", + "label": "Python code", + "text": "Write a Python function that returns the first repeated element in a list and explain its complexity.", + "token_ids": [ + 100000, + 9081, + 245, + 12974, + 1157, + 344, + 7578, + 254, + 1022, + 9842, + 3699, + 279, + 245, + 1525, + 285, + 6716, + 895, + 13913, + 13 + ], + "tokens": [ + "<|begin▁of▁sentence|>", + "Write", + "Ġa", + "ĠPython", + "Ġfunction", + "Ġthat", + "Ġreturns", + "Ġthe", + "Ġfirst", + "Ġrepeated", + "Ġelement", + "Ġin", + "Ġa", + "Ġlist", + "Ġand", + "Ġexplain", + "Ġits", + "Ġcomplexity", + "." + ] + }, + { + "id": "math", + "label": "数学推理", + "text": "若正数 x 满足 x 加上它的倒数等于 3,求 x 的平方加上倒数的平方。", + "token_ids": [ + 100000, + 8246, + 2206, + 2266, + 1376, + 207, + 13211, + 1376, + 207, + 14855, + 12985, + 8270, + 2266, + 37360, + 207, + 18, + 19304, + 2989, + 1376, + 207, + 337, + 18718, + 14855, + 8270, + 51971, + 18718, + 398 + ], + "tokens": [ + "<|begin▁of▁sentence|>", + "èĭ¥", + "æŃ£", + "æķ°", + "Ġx", + "Ġ", + "满足", + "Ġx", + "Ġ", + "åĬłä¸Ĭ", + "å®ĥçļĦ", + "åĢĴ", + "æķ°", + "çŃīäºİ", + "Ġ", + "3", + "ï¼Į", + "æ±Ĥ", + "Ġx", + "Ġ", + "çļĦ", + "å¹³æĸ¹", + "åĬłä¸Ĭ", + "åĢĴ", + "æķ°çļĦ", + "å¹³æĸ¹", + "ãĢĤ" + ] + } + ], + "execution": { + "batch": 4, + "padded_sequence": 27, + "valid_tokens": 90, + "embedding_ms": 88.24645899585448, + "initial_hidden": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 0.09717819839715958, + "p10": 0.06574996560811996, + "p50": 0.09712168574333191, + "p90": 0.13157525658607483 + }, + "l2_norm": { + "mean": 4.397783279418945, + "p10": 2.97550368309021, + "p50": 4.395225524902344, + "p90": 5.954416275024414 + }, + "finite": true, + "sha256_fp32": "837950812a7e04233f5a10439c8e5b2017241797652a1f06cac94c122e41b8e5" + }, + "layers": [ + { + "layer": 0, + "ffn": "dense", + "state_tensors": 10, + "state_numel": 81007104, + "state_bytes": 162014208, + "timing_ms": { + "load_cpu": 506.36647900682874, + "transfer_to_device": 24.61088399286382, + "forward": 125.80606798292138 + }, + "peak_device_allocated_mib": 205.18603515625, + "hidden_before": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 0.09717819839715958, + "p10": 0.06574996560811996, + "p50": 0.09712168574333191, + "p90": 0.13157525658607483 + }, + "l2_norm": { + "mean": 4.397783279418945, + "p10": 2.97550368309021, + "p50": 4.395225524902344, + "p90": 5.954416275024414 + }, + "finite": true, + "sha256_fp32": "837950812a7e04233f5a10439c8e5b2017241797652a1f06cac94c122e41b8e5" + }, + "hidden_after": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 0.10381482541561127, + "p10": 0.07184165716171265, + "p50": 0.10495559871196747, + "p90": 0.13329632580280304 + }, + "l2_norm": { + "mean": 4.698122978210449, + "p10": 3.2511825561523438, + "p50": 4.749748229980469, + "p90": 6.032303333282471 + }, + "finite": true, + "sha256_fp32": "7c907ba1a1d9c0e9c47523d9c1267f2bef2c8fde8a7cc317d6be36ff41394e7e" + }, + "mla": { + "compressed_projection_shape": [ + 4, + 27, + 576 + ], + "latent_shape": [ + 4, + 27, + 512 + ], + "rope_key_shape": [ + 4, + 27, + 64 + ], + "latent_rms": 1.287002682685852, + "rope_key_rms": 2.812988758087158, + "eager_key_cache_shape": [ + 4, + 16, + 27, + 192 + ], + "eager_value_cache_shape": [ + 4, + 16, + 27, + 128 + ], + "eager_cache_bytes": 1105920 + } + }, + { + "layer": 1, + "ffn": "moe", + "state_tensors": 203, + "state_numel": 584847872, + "state_bytes": 1169695744, + "timing_ms": { + "load_cpu": 24.967125995317474, + "transfer_to_device": 128.2450080034323, + "forward": 89.96552301687188 + }, + "peak_device_allocated_mib": 1171.78759765625, + "hidden_before": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 0.10381482541561127, + "p10": 0.07184165716171265, + "p50": 0.10495559871196747, + "p90": 0.13329632580280304 + }, + "l2_norm": { + "mean": 4.698122978210449, + "p10": 3.2511825561523438, + "p50": 4.749748229980469, + "p90": 6.032303333282471 + }, + "finite": true, + "sha256_fp32": "7c907ba1a1d9c0e9c47523d9c1267f2bef2c8fde8a7cc317d6be36ff41394e7e" + }, + "hidden_after": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 0.12936507165431976, + "p10": 0.08545936644077301, + "p50": 0.12583057582378387, + "p90": 0.1601526141166687 + }, + "l2_norm": { + "mean": 5.854394435882568, + "p10": 3.8674492835998535, + "p50": 5.694441795349121, + "p90": 7.247680187225342 + }, + "finite": true, + "sha256_fp32": "bc8c6a2a556f09b333c171b50788905902a2aa18e0bbac784837a7604d931fad" + }, + "mla": { + "compressed_projection_shape": [ + 4, + 27, + 576 + ], + "latent_shape": [ + 4, + 27, + 512 + ], + "rope_key_shape": [ + 4, + 27, + 64 + ], + "latent_rms": 2.390777587890625, + "rope_key_rms": 3.4520363807678223, + "eager_key_cache_shape": [ + 4, + 16, + 27, + 192 + ], + "eager_value_cache_shape": [ + 4, + 16, + 27, + 128 + ], + "eager_cache_bytes": 1105920 + }, + "routing": { + "aggregate_load": [ + 7, + 6, + 9, + 6, + 7, + 8, + 22, + 5, + 48, + 2, + 5, + 17, + 0, + 6, + 5, + 2, + 1, + 29, + 4, + 1, + 3, + 8, + 6, + 4, + 25, + 4, + 13, + 7, + 4, + 9, + 5, + 3, + 23, + 20, + 9, + 5, + 4, + 8, + 7, + 10, + 3, + 10, + 11, + 15, + 7, + 11, + 1, + 6, + 8, + 7, + 5, + 8, + 5, + 10, + 3, + 3, + 12, + 16, + 3, + 12, + 5, + 2, + 4, + 6 + ], + "aggregate_metrics": { + "routes": 540, + "used_experts": 63, + "zero_experts": 1, + "mean": 8.4375, + "std": 7.800844669342041, + "cv": 0.9245445728302002, + "gini": 0.42042824625968933, + "entropy_nats": 3.842954635620117, + "effective_experts": 46.66314354467963, + "min": 0.0, + "p50": 6.0, + "p90": 16.699996948242188, + "max": 48.0 + }, + "top_experts": [ + { + "expert": 8, + "routes": 48 + }, + { + "expert": 17, + "routes": 29 + }, + { + "expert": 24, + "routes": 25 + }, + { + "expert": 32, + "routes": 23 + }, + { + "expert": 6, + "routes": 22 + }, + { + "expert": 33, + "routes": 20 + }, + { + "expert": 11, + "routes": 17 + }, + { + "expert": 57, + "routes": 16 + }, + { + "expert": 43, + "routes": 15 + }, + { + "expert": 26, + "routes": 13 + }, + { + "expert": 56, + "routes": 12 + }, + { + "expert": 59, + "routes": 12 + } + ], + "selected_weight_sum": { + "mean": 0.3948917090892792, + "p10": 0.31607890129089355, + "p50": 0.3835059404373169, + "p90": 0.48514068126678467 + }, + "top1_top2_margin": { + "mean": 0.04313327372074127, + "p50": 0.024309832602739334, + "p90": 0.1049143448472023 + }, + "per_prompt": [ + { + "id": "zh_explanation", + "label": "中文解释", + "tokens": 26, + "load": [ + 2, + 2, + 3, + 0, + 3, + 2, + 3, + 2, + 13, + 1, + 3, + 6, + 0, + 1, + 1, + 0, + 1, + 3, + 1, + 0, + 1, + 1, + 2, + 1, + 6, + 3, + 3, + 3, + 2, + 2, + 2, + 1, + 6, + 3, + 4, + 1, + 0, + 4, + 1, + 2, + 2, + 6, + 4, + 5, + 4, + 4, + 1, + 3, + 4, + 0, + 1, + 3, + 0, + 3, + 0, + 0, + 5, + 5, + 0, + 3, + 1, + 1, + 1, + 5 + ], + "metrics": { + "routes": 156, + "used_experts": 54, + "zero_experts": 10, + "mean": 2.4375, + "std": 2.1885342597961426, + "cv": 0.8978602290153503, + "gini": 0.4509214758872986, + "entropy_nats": 3.77732253074646, + "effective_experts": 43.69888254607027, + "min": 0.0, + "p50": 2.0, + "p90": 5.0, + "max": 13.0 + }, + "top_experts": [ + { + "expert": 8, + "routes": 13 + }, + { + "expert": 11, + "routes": 6 + }, + { + "expert": 24, + "routes": 6 + }, + { + "expert": 32, + "routes": 6 + }, + { + "expert": 41, + "routes": 6 + }, + { + "expert": 43, + "routes": 5 + }, + { + "expert": 56, + "routes": 5 + }, + { + "expert": 57, + "routes": 5 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 57, + "weight": 0.19667254388332367 + }, + { + "expert": 60, + "weight": 0.09833430498838425 + }, + { + "expert": 26, + "weight": 0.04889551177620888 + }, + { + "expert": 8, + "weight": 0.04098263755440712 + }, + { + "expert": 11, + "weight": 0.03667396679520607 + }, + { + "expert": 56, + "weight": 0.03248872235417366 + } + ], + "selected_weight_sum": 0.45404768735170364, + "top1_top2_margin": 0.09833823889493942 + }, + { + "position": 1, + "token_id": 1082, + "token": "ç͍", + "decoded_piece": "用", + "experts_by_weight": [ + { + "expert": 41, + "weight": 0.06996385753154755 + }, + { + "expert": 24, + "weight": 0.06454664468765259 + }, + { + "expert": 28, + "weight": 0.052981361746788025 + }, + { + "expert": 38, + "weight": 0.04313444346189499 + }, + { + "expert": 47, + "weight": 0.03537776321172714 + }, + { + "expert": 8, + "weight": 0.03254995867609978 + } + ], + "selected_weight_sum": 0.29855402931571007, + "top1_top2_margin": 0.0054172128438949585 + }, + { + "position": 2, + "token_id": 83983, + "token": "éĢļä¿Ĺ", + "decoded_piece": "通俗", + "experts_by_weight": [ + { + "expert": 63, + "weight": 0.10692810267210007 + }, + { + "expert": 21, + "weight": 0.10059474408626556 + }, + { + "expert": 14, + "weight": 0.05983731523156166 + }, + { + "expert": 32, + "weight": 0.04131235554814339 + }, + { + "expert": 31, + "weight": 0.03579572215676308 + }, + { + "expert": 8, + "weight": 0.03260975703597069 + } + ], + "selected_weight_sum": 0.37707799673080444, + "top1_top2_margin": 0.006333358585834503 + }, + { + "position": 3, + "token_id": 91476, + "token": "çļĦè¯Ńè¨Ģ", + "decoded_piece": "的语言", + "experts_by_weight": [ + { + "expert": 22, + "weight": 0.1252635419368744 + }, + { + "expert": 59, + "weight": 0.10231700539588928 + }, + { + "expert": 48, + "weight": 0.08574634790420532 + }, + { + "expert": 1, + "weight": 0.035869430750608444 + }, + { + "expert": 34, + "weight": 0.03414113447070122 + }, + { + "expert": 8, + "weight": 0.03363116458058357 + } + ], + "selected_weight_sum": 0.41696862503886223, + "top1_top2_margin": 0.022946536540985107 + }, + { + "position": 4, + "token_id": 17882, + "token": "è§£éĩĬ", + "decoded_piece": "解释", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.11453470587730408 + }, + { + "expert": 22, + "weight": 0.10259787738323212 + }, + { + "expert": 8, + "weight": 0.0590984970331192 + }, + { + "expert": 10, + "weight": 0.051855940371751785 + }, + { + "expert": 32, + "weight": 0.038944680243730545 + }, + { + "expert": 41, + "weight": 0.036984872072935104 + } + ], + "selected_weight_sum": 0.40401657298207283, + "top1_top2_margin": 0.01193682849407196 + }, + { + "position": 5, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.17547623813152313 + }, + { + "expert": 51, + "weight": 0.1141868308186531 + }, + { + "expert": 4, + "weight": 0.03966223821043968 + }, + { + "expert": 43, + "weight": 0.0280931256711483 + }, + { + "expert": 57, + "weight": 0.026378853246569633 + }, + { + "expert": 56, + "weight": 0.026011040434241295 + } + ], + "selected_weight_sum": 0.40980832651257515, + "top1_top2_margin": 0.061289407312870026 + }, + { + "position": 6, + "token_id": 10285, + "token": "为ä»Ģä¹Ī", + "decoded_piece": "为什么", + "experts_by_weight": [ + { + "expert": 5, + "weight": 0.08941874653100967 + }, + { + "expert": 43, + "weight": 0.08715063333511353 + }, + { + "expert": 63, + "weight": 0.08338002115488052 + }, + { + "expert": 42, + "weight": 0.05639657378196716 + }, + { + "expert": 8, + "weight": 0.044801801443099976 + }, + { + "expert": 25, + "weight": 0.03526010364294052 + } + ], + "selected_weight_sum": 0.3964078798890114, + "top1_top2_margin": 0.0022681131958961487 + }, + { + "position": 7, + "token_id": 21844, + "token": "ç¨Ģ", + "decoded_piece": "稀", + "experts_by_weight": [ + { + "expert": 16, + "weight": 0.1019749641418457 + }, + { + "expert": 53, + "weight": 0.09891179949045181 + }, + { + "expert": 44, + "weight": 0.06054722145199776 + }, + { + "expert": 35, + "weight": 0.0462147556245327 + }, + { + "expert": 6, + "weight": 0.03941645100712776 + }, + { + "expert": 33, + "weight": 0.03171882405877113 + } + ], + "selected_weight_sum": 0.37878401577472687, + "top1_top2_margin": 0.0030631646513938904 + }, + { + "position": 8, + "token_id": 20923, + "token": "çĸı", + "decoded_piece": "疏", + "experts_by_weight": [ + { + "expert": 34, + "weight": 0.19530633091926575 + }, + { + "expert": 11, + "weight": 0.08023042231798172 + }, + { + "expert": 53, + "weight": 0.06899507343769073 + }, + { + "expert": 18, + "weight": 0.06873473525047302 + }, + { + "expert": 29, + "weight": 0.058919522911310196 + }, + { + "expert": 33, + "weight": 0.048458561301231384 + } + ], + "selected_weight_sum": 0.5206446461379528, + "top1_top2_margin": 0.11507590860128403 + }, + { + "position": 9, + "token_id": 10900, + "token": "ä¸ĵå®¶", + "decoded_piece": "专家", + "experts_by_weight": [ + { + "expert": 63, + "weight": 0.09651511907577515 + }, + { + "expert": 10, + "weight": 0.09186359494924545 + }, + { + "expert": 39, + "weight": 0.0887473002076149 + }, + { + "expert": 61, + "weight": 0.0633588582277298 + }, + { + "expert": 40, + "weight": 0.05882406234741211 + }, + { + "expert": 8, + "weight": 0.04215756803750992 + } + ], + "selected_weight_sum": 0.4414665028452873, + "top1_top2_margin": 0.004651524126529694 + }, + { + "position": 10, + "token_id": 39322, + "token": "模åŀĭ", + "decoded_piece": "模型", + "experts_by_weight": [ + { + "expert": 2, + "weight": 0.1100841760635376 + }, + { + "expert": 48, + "weight": 0.10468249022960663 + }, + { + "expert": 1, + "weight": 0.06450875848531723 + }, + { + "expert": 59, + "weight": 0.054316893219947815 + }, + { + "expert": 41, + "weight": 0.049031782895326614 + }, + { + "expert": 39, + "weight": 0.0398552268743515 + } + ], + "selected_weight_sum": 0.4224793277680874, + "top1_top2_margin": 0.005401685833930969 + }, + { + "position": 11, + "token_id": 1876, + "token": "åı¯ä»¥", + "decoded_piece": "可以", + "experts_by_weight": [ + { + "expert": 63, + "weight": 0.08753079921007156 + }, + { + "expert": 43, + "weight": 0.07256003469228745 + }, + { + "expert": 25, + "weight": 0.06117047369480133 + }, + { + "expert": 8, + "weight": 0.055068738758563995 + }, + { + "expert": 7, + "weight": 0.02229606918990612 + }, + { + "expert": 45, + "weight": 0.022259222343564034 + } + ], + "selected_weight_sum": 0.3208853378891945, + "top1_top2_margin": 0.014970764517784119 + }, + { + "position": 12, + "token_id": 9476, + "token": "æĭ¥æľī", + "decoded_piece": "拥有", + "experts_by_weight": [ + { + "expert": 32, + "weight": 0.06885721534490585 + }, + { + "expert": 41, + "weight": 0.058269087225198746 + }, + { + "expert": 23, + "weight": 0.05805779621005058 + }, + { + "expert": 17, + "weight": 0.05642109364271164 + }, + { + "expert": 37, + "weight": 0.045869726687669754 + }, + { + "expert": 5, + "weight": 0.03334243223071098 + } + ], + "selected_weight_sum": 0.32081735134124756, + "top1_top2_margin": 0.010588128119707108 + }, + { + "position": 13, + "token_id": 3599, + "token": "å¾Īå¤ļ", + "decoded_piece": "很多", + "experts_by_weight": [ + { + "expert": 0, + "weight": 0.10499579459428787 + }, + { + "expert": 27, + "weight": 0.09095507115125656 + }, + { + "expert": 63, + "weight": 0.06519477069377899 + }, + { + "expert": 30, + "weight": 0.05123397707939148 + }, + { + "expert": 17, + "weight": 0.027353990823030472 + }, + { + "expert": 44, + "weight": 0.026716141030192375 + } + ], + "selected_weight_sum": 0.36644974537193775, + "top1_top2_margin": 0.014040723443031311 + }, + { + "position": 14, + "token_id": 21273, + "token": "åıĤæķ°", + "decoded_piece": "参数", + "experts_by_weight": [ + { + "expert": 2, + "weight": 0.09879104048013687 + }, + { + "expert": 48, + "weight": 0.09676965326070786 + }, + { + "expert": 45, + "weight": 0.06424219161272049 + }, + { + "expert": 24, + "weight": 0.04395608976483345 + }, + { + "expert": 41, + "weight": 0.04187783598899841 + }, + { + "expert": 17, + "weight": 0.04052548110485077 + } + ], + "selected_weight_sum": 0.38616229221224785, + "top1_top2_margin": 0.002021387219429016 + }, + { + "position": 15, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 51, + "weight": 0.116199791431427 + }, + { + "expert": 8, + "weight": 0.1023869439959526 + }, + { + "expert": 4, + "weight": 0.04513482749462128 + }, + { + "expert": 57, + "weight": 0.028082912787795067 + }, + { + "expert": 11, + "weight": 0.026392823085188866 + }, + { + "expert": 47, + "weight": 0.022527199238538742 + } + ], + "selected_weight_sum": 0.34072449803352356, + "top1_top2_margin": 0.013812847435474396 + }, + { + "position": 16, + "token_id": 1705, + "token": "ä½Ĩ", + "decoded_piece": "但", + "experts_by_weight": [ + { + "expert": 25, + "weight": 0.07215771824121475 + }, + { + "expert": 42, + "weight": 0.06972325593233109 + }, + { + "expert": 9, + "weight": 0.05969814583659172 + }, + { + "expert": 43, + "weight": 0.0547492578625679 + }, + { + "expert": 11, + "weight": 0.025664115324616432 + }, + { + "expert": 28, + "weight": 0.02441459707915783 + } + ], + "selected_weight_sum": 0.3064070902764797, + "top1_top2_margin": 0.002434462308883667 + }, + { + "position": 17, + "token_id": 13712, + "token": "æ¯ı个", + "decoded_piece": "每个", + "experts_by_weight": [ + { + "expert": 27, + "weight": 0.09334097057580948 + }, + { + "expert": 30, + "weight": 0.08088334649801254 + }, + { + "expert": 45, + "weight": 0.06753778457641602 + }, + { + "expert": 44, + "weight": 0.033187661319971085 + }, + { + "expert": 24, + "weight": 0.028868872672319412 + }, + { + "expert": 37, + "weight": 0.028245503082871437 + } + ], + "selected_weight_sum": 0.33206413872539997, + "top1_top2_margin": 0.012457624077796936 + }, + { + "position": 18, + "token_id": 10728, + "token": "Ġtoken", + "decoded_piece": " token", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.21899603307247162 + }, + { + "expert": 2, + "weight": 0.10515932738780975 + }, + { + "expert": 33, + "weight": 0.08899100124835968 + }, + { + "expert": 7, + "weight": 0.05321361497044563 + }, + { + "expert": 32, + "weight": 0.04883015528321266 + }, + { + "expert": 26, + "weight": 0.04874962940812111 + } + ], + "selected_weight_sum": 0.5639397613704205, + "top1_top2_margin": 0.11383670568466187 + }, + { + "position": 19, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 57, + "weight": 0.09224173426628113 + }, + { + "expert": 13, + "weight": 0.08393412828445435 + }, + { + "expert": 8, + "weight": 0.06542550027370453 + }, + { + "expert": 34, + "weight": 0.042790114879608154 + }, + { + "expert": 26, + "weight": 0.03504375368356705 + }, + { + "expert": 50, + "weight": 0.02880442701280117 + } + ], + "selected_weight_sum": 0.3482396584004164, + "top1_top2_margin": 0.008307605981826782 + }, + { + "position": 20, + "token_id": 2141, + "token": "åıª", + "decoded_piece": "只", + "experts_by_weight": [ + { + "expert": 42, + "weight": 0.08134366571903229 + }, + { + "expert": 6, + "weight": 0.06718657910823822 + }, + { + "expert": 24, + "weight": 0.06551655381917953 + }, + { + "expert": 56, + "weight": 0.0429602786898613 + }, + { + "expert": 32, + "weight": 0.035062018781900406 + }, + { + "expert": 53, + "weight": 0.032046057283878326 + } + ], + "selected_weight_sum": 0.3241151534020901, + "top1_top2_margin": 0.014157086610794067 + }, + { + "position": 21, + "token_id": 5118, + "token": "使ç͍", + "decoded_piece": "使用", + "experts_by_weight": [ + { + "expert": 56, + "weight": 0.16159342229366302 + }, + { + "expert": 10, + "weight": 0.055032674223184586 + }, + { + "expert": 24, + "weight": 0.05452797934412956 + }, + { + "expert": 32, + "weight": 0.03965990990400314 + }, + { + "expert": 11, + "weight": 0.037650369107723236 + }, + { + "expert": 41, + "weight": 0.03701045364141464 + } + ], + "selected_weight_sum": 0.3854748085141182, + "top1_top2_margin": 0.10656074807047844 + }, + { + "position": 22, + "token_id": 6802, + "token": "åħ¶ä¸Ń", + "decoded_piece": "其中", + "experts_by_weight": [ + { + "expert": 62, + "weight": 0.07383304089307785 + }, + { + "expert": 24, + "weight": 0.0634578987956047 + }, + { + "expert": 42, + "weight": 0.06159329041838646 + }, + { + "expert": 45, + "weight": 0.05860455334186554 + }, + { + "expert": 37, + "weight": 0.0438292920589447 + }, + { + "expert": 43, + "weight": 0.02614068053662777 + } + ], + "selected_weight_sum": 0.327458756044507, + "top1_top2_margin": 0.010375142097473145 + }, + { + "position": 23, + "token_id": 51905, + "token": "ä¸Ģå°ı", + "decoded_piece": "一小", + "experts_by_weight": [ + { + "expert": 27, + "weight": 0.10564204305410385 + }, + { + "expert": 20, + "weight": 0.10040134191513062 + }, + { + "expert": 44, + "weight": 0.07950114458799362 + }, + { + "expert": 6, + "weight": 0.046730685979127884 + }, + { + "expert": 0, + "weight": 0.04516390711069107 + }, + { + "expert": 46, + "weight": 0.04194054380059242 + } + ], + "selected_weight_sum": 0.41937966644763947, + "top1_top2_margin": 0.005240701138973236 + }, + { + "position": 24, + "token_id": 5781, + "token": "éĥ¨åĪĨ", + "decoded_piece": "部分", + "experts_by_weight": [ + { + "expert": 47, + "weight": 0.25300896167755127 + }, + { + "expert": 59, + "weight": 0.05960262939333916 + }, + { + "expert": 37, + "weight": 0.057047683745622635 + }, + { + "expert": 34, + "weight": 0.04516291618347168 + }, + { + "expert": 29, + "weight": 0.03590259701013565 + }, + { + "expert": 40, + "weight": 0.03268502280116081 + } + ], + "selected_weight_sum": 0.4834098108112812, + "top1_top2_margin": 0.1934063322842121 + }, + { + "position": 25, + "token_id": 398, + "token": "ãĢĤ", + "decoded_piece": "。", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.08622618764638901 + }, + { + "expert": 57, + "weight": 0.08423644304275513 + }, + { + "expert": 51, + "weight": 0.05857071280479431 + }, + { + "expert": 8, + "weight": 0.0584087148308754 + }, + { + "expert": 11, + "weight": 0.023649701848626137 + }, + { + "expert": 56, + "weight": 0.023337118327617645 + } + ], + "selected_weight_sum": 0.3344288785010576, + "top1_top2_margin": 0.0019897446036338806 + } + ] + }, + { + "id": "en_architecture", + "label": "English architecture", + "tokens": 18, + "load": [ + 0, + 0, + 3, + 2, + 1, + 1, + 7, + 0, + 11, + 0, + 0, + 4, + 0, + 0, + 2, + 0, + 0, + 7, + 3, + 1, + 0, + 3, + 0, + 0, + 1, + 0, + 2, + 0, + 1, + 1, + 1, + 2, + 9, + 8, + 1, + 2, + 0, + 0, + 1, + 3, + 0, + 1, + 0, + 3, + 2, + 0, + 0, + 0, + 3, + 2, + 2, + 1, + 3, + 0, + 2, + 1, + 2, + 3, + 0, + 2, + 2, + 1, + 0, + 1 + ], + "metrics": { + "routes": 108, + "used_experts": 40, + "zero_experts": 24, + "mean": 1.6875, + "std": 2.2947490215301514, + "cv": 1.3598512411117554, + "gini": 0.6241319179534912, + "entropy_nats": 3.399005651473999, + "effective_experts": 29.93432009696054, + "min": 0.0, + "p50": 1.0, + "p90": 3.0, + "max": 11.0 + }, + "top_experts": [ + { + "expert": 8, + "routes": 11 + }, + { + "expert": 32, + "routes": 9 + }, + { + "expert": 33, + "routes": 8 + }, + { + "expert": 6, + "routes": 7 + }, + { + "expert": 17, + "routes": 7 + }, + { + "expert": 11, + "routes": 4 + }, + { + "expert": 2, + "routes": 3 + }, + { + "expert": 18, + "routes": 3 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 57, + "weight": 0.19667254388332367 + }, + { + "expert": 60, + "weight": 0.09833430498838425 + }, + { + "expert": 26, + "weight": 0.04889551177620888 + }, + { + "expert": 8, + "weight": 0.04098263755440712 + }, + { + "expert": 11, + "weight": 0.03667396679520607 + }, + { + "expert": 56, + "weight": 0.03248872235417366 + } + ], + "selected_weight_sum": 0.45404768735170364, + "top1_top2_margin": 0.09833823889493942 + }, + { + "position": 1, + "token_id": 99297, + "token": "Explain", + "decoded_piece": "Explain", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.12304119020700455 + }, + { + "expert": 32, + "weight": 0.07146582752466202 + }, + { + "expert": 6, + "weight": 0.04816020652651787 + }, + { + "expert": 56, + "weight": 0.046724919229745865 + }, + { + "expert": 48, + "weight": 0.04111320152878761 + }, + { + "expert": 17, + "weight": 0.04073270410299301 + } + ], + "selected_weight_sum": 0.3712380491197109, + "top1_top2_margin": 0.05157536268234253 + }, + { + "position": 2, + "token_id": 946, + "token": "Ġhow", + "decoded_piece": " how", + "experts_by_weight": [ + { + "expert": 43, + "weight": 0.11726353317499161 + }, + { + "expert": 8, + "weight": 0.10638604313135147 + }, + { + "expert": 5, + "weight": 0.10123571753501892 + }, + { + "expert": 49, + "weight": 0.09169016033411026 + }, + { + "expert": 63, + "weight": 0.05507644638419151 + }, + { + "expert": 6, + "weight": 0.026508130133152008 + } + ], + "selected_weight_sum": 0.4981600306928158, + "top1_top2_margin": 0.010877490043640137 + }, + { + "position": 3, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.12597927451133728 + }, + { + "expert": 52, + "weight": 0.08409375697374344 + }, + { + "expert": 21, + "weight": 0.07757996022701263 + }, + { + "expert": 26, + "weight": 0.07270722091197968 + }, + { + "expert": 6, + "weight": 0.052257198840379715 + }, + { + "expert": 11, + "weight": 0.02372206375002861 + } + ], + "selected_weight_sum": 0.43633947521448135, + "top1_top2_margin": 0.04188551753759384 + }, + { + "position": 4, + "token_id": 29640, + "token": "Ġcompressed", + "decoded_piece": " compressed", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.10477275401353836 + }, + { + "expert": 32, + "weight": 0.07370883226394653 + }, + { + "expert": 33, + "weight": 0.06914308667182922 + }, + { + "expert": 8, + "weight": 0.048680469393730164 + }, + { + "expert": 6, + "weight": 0.04684734344482422 + }, + { + "expert": 54, + "weight": 0.04026190936565399 + } + ], + "selected_weight_sum": 0.3834143951535225, + "top1_top2_margin": 0.031063921749591827 + }, + { + "position": 5, + "token_id": 2131, + "token": "Ġkey", + "decoded_piece": " key", + "experts_by_weight": [ + { + "expert": 2, + "weight": 0.09092872589826584 + }, + { + "expert": 17, + "weight": 0.06525559723377228 + }, + { + "expert": 33, + "weight": 0.05327434092760086 + }, + { + "expert": 6, + "weight": 0.0475538969039917 + }, + { + "expert": 39, + "weight": 0.043563321232795715 + }, + { + "expert": 18, + "weight": 0.03889021649956703 + } + ], + "selected_weight_sum": 0.3394660986959934, + "top1_top2_margin": 0.02567312866449356 + }, + { + "position": 6, + "token_id": 12, + "token": "-", + "decoded_piece": "-", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.09165064245462418 + }, + { + "expert": 60, + "weight": 0.08469503372907639 + }, + { + "expert": 57, + "weight": 0.07083211094141006 + }, + { + "expert": 55, + "weight": 0.050766441971063614 + }, + { + "expert": 8, + "weight": 0.038136690855026245 + }, + { + "expert": 43, + "weight": 0.02105005271732807 + } + ], + "selected_weight_sum": 0.35713097266852856, + "top1_top2_margin": 0.0069556087255477905 + }, + { + "position": 7, + "token_id": 2834, + "token": "value", + "decoded_piece": "value", + "experts_by_weight": [ + { + "expert": 39, + "weight": 0.11641892790794373 + }, + { + "expert": 2, + "weight": 0.09496383368968964 + }, + { + "expert": 3, + "weight": 0.04481128975749016 + }, + { + "expert": 14, + "weight": 0.042960286140441895 + }, + { + "expert": 8, + "weight": 0.04264334961771965 + }, + { + "expert": 35, + "weight": 0.035825781524181366 + } + ], + "selected_weight_sum": 0.37762346863746643, + "top1_top2_margin": 0.02145509421825409 + }, + { + "position": 8, + "token_id": 28937, + "token": "Ġlatent", + "decoded_piece": " latent", + "experts_by_weight": [ + { + "expert": 6, + "weight": 0.06952936947345734 + }, + { + "expert": 32, + "weight": 0.06664305180311203 + }, + { + "expert": 33, + "weight": 0.06500348448753357 + }, + { + "expert": 8, + "weight": 0.04819929972290993 + }, + { + "expert": 21, + "weight": 0.04369500279426575 + }, + { + "expert": 17, + "weight": 0.04272771626710892 + } + ], + "selected_weight_sum": 0.3357979245483875, + "top1_top2_margin": 0.0028863176703453064 + }, + { + "position": 9, + "token_id": 4183, + "token": "Ġchanges", + "decoded_piece": " changes", + "experts_by_weight": [ + { + "expert": 34, + "weight": 0.0739561915397644 + }, + { + "expert": 2, + "weight": 0.05987396091222763 + }, + { + "expert": 61, + "weight": 0.05065435543656349 + }, + { + "expert": 32, + "weight": 0.04441293329000473 + }, + { + "expert": 29, + "weight": 0.04319792985916138 + }, + { + "expert": 54, + "weight": 0.041305456310510635 + } + ], + "selected_weight_sum": 0.31340082734823227, + "top1_top2_margin": 0.014082230627536774 + }, + { + "position": 10, + "token_id": 254, + "token": "Ġthe", + "decoded_piece": " the", + "experts_by_weight": [ + { + "expert": 50, + "weight": 0.1116655096411705 + }, + { + "expert": 8, + "weight": 0.09895200282335281 + }, + { + "expert": 21, + "weight": 0.0893455520272255 + }, + { + "expert": 44, + "weight": 0.02281578816473484 + }, + { + "expert": 38, + "weight": 0.022700324654579163 + }, + { + "expert": 48, + "weight": 0.021918442100286484 + } + ], + "selected_weight_sum": 0.3673976194113493, + "top1_top2_margin": 0.012713506817817688 + }, + { + "position": 11, + "token_id": 4887, + "token": "Ġmemory", + "decoded_piece": " memory", + "experts_by_weight": [ + { + "expert": 33, + "weight": 0.09330408275127411 + }, + { + "expert": 59, + "weight": 0.08162695169448853 + }, + { + "expert": 17, + "weight": 0.06619293242692947 + }, + { + "expert": 3, + "weight": 0.060682181268930435 + }, + { + "expert": 48, + "weight": 0.04300126060843468 + }, + { + "expert": 32, + "weight": 0.03936149552464485 + } + ], + "selected_weight_sum": 0.38416890427470207, + "top1_top2_margin": 0.011677131056785583 + }, + { + "position": 12, + "token_id": 2295, + "token": "Ġcost", + "decoded_piece": " cost", + "experts_by_weight": [ + { + "expert": 33, + "weight": 0.215052992105484 + }, + { + "expert": 32, + "weight": 0.061788905411958694 + }, + { + "expert": 59, + "weight": 0.054121699184179306 + }, + { + "expert": 19, + "weight": 0.04691912233829498 + }, + { + "expert": 50, + "weight": 0.04069184884428978 + }, + { + "expert": 6, + "weight": 0.03785664215683937 + } + ], + "selected_weight_sum": 0.45643121004104614, + "top1_top2_margin": 0.15326408669352531 + }, + { + "position": 13, + "token_id": 280, + "token": "Ġof", + "decoded_piece": " of", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.084696464240551 + }, + { + "expert": 8, + "weight": 0.05794258788228035 + }, + { + "expert": 24, + "weight": 0.045855727046728134 + }, + { + "expert": 28, + "weight": 0.044749513268470764 + }, + { + "expert": 49, + "weight": 0.04419646039605141 + }, + { + "expert": 35, + "weight": 0.0432145893573761 + } + ], + "selected_weight_sum": 0.32065534219145775, + "top1_top2_margin": 0.026753876358270645 + }, + { + "position": 14, + "token_id": 77939, + "token": "Ġautore", + "decoded_piece": " autore", + "experts_by_weight": [ + { + "expert": 30, + "weight": 0.09902932494878769 + }, + { + "expert": 33, + "weight": 0.07717865705490112 + }, + { + "expert": 44, + "weight": 0.0736161321401596 + }, + { + "expert": 31, + "weight": 0.05943404510617256 + }, + { + "expert": 32, + "weight": 0.050912320613861084 + }, + { + "expert": 39, + "weight": 0.04211832955479622 + } + ], + "selected_weight_sum": 0.4022888094186783, + "top1_top2_margin": 0.021850667893886566 + }, + { + "position": 15, + "token_id": 79184, + "token": "gressive", + "decoded_piece": "gressive", + "experts_by_weight": [ + { + "expert": 33, + "weight": 0.1954519897699356 + }, + { + "expert": 11, + "weight": 0.10430377721786499 + }, + { + "expert": 18, + "weight": 0.05457714945077896 + }, + { + "expert": 14, + "weight": 0.052528709173202515 + }, + { + "expert": 32, + "weight": 0.051252055913209915 + }, + { + "expert": 17, + "weight": 0.03450470045208931 + } + ], + "selected_weight_sum": 0.4926183819770813, + "top1_top2_margin": 0.09114821255207062 + }, + { + "position": 16, + "token_id": 39671, + "token": "Ġdecoding", + "decoded_piece": " decoding", + "experts_by_weight": [ + { + "expert": 33, + "weight": 0.10569825768470764 + }, + { + "expert": 17, + "weight": 0.049603477120399475 + }, + { + "expert": 31, + "weight": 0.044465046375989914 + }, + { + "expert": 41, + "weight": 0.040463220328092575 + }, + { + "expert": 32, + "weight": 0.039218440651893616 + }, + { + "expert": 18, + "weight": 0.03692803531885147 + } + ], + "selected_weight_sum": 0.3163764774799347, + "top1_top2_margin": 0.056094780564308167 + }, + { + "position": 17, + "token_id": 13, + "token": ".", + "decoded_piece": ".", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.10390158742666245 + }, + { + "expert": 8, + "weight": 0.07251034677028656 + }, + { + "expert": 57, + "weight": 0.06104576960206032 + }, + { + "expert": 51, + "weight": 0.04427094757556915 + }, + { + "expert": 11, + "weight": 0.02940911054611206 + }, + { + "expert": 43, + "weight": 0.029390856623649597 + } + ], + "selected_weight_sum": 0.34052861854434013, + "top1_top2_margin": 0.031391240656375885 + } + ] + }, + { + "id": "code", + "label": "Python code", + "tokens": 19, + "load": [ + 1, + 2, + 3, + 2, + 1, + 1, + 7, + 1, + 12, + 1, + 1, + 3, + 0, + 0, + 0, + 0, + 0, + 4, + 0, + 0, + 1, + 4, + 3, + 1, + 7, + 0, + 3, + 0, + 1, + 3, + 1, + 0, + 8, + 5, + 0, + 1, + 4, + 1, + 1, + 2, + 0, + 0, + 1, + 3, + 0, + 0, + 0, + 3, + 1, + 5, + 2, + 2, + 0, + 2, + 1, + 1, + 2, + 2, + 0, + 1, + 1, + 0, + 2, + 0 + ], + "metrics": { + "routes": 114, + "used_experts": 44, + "zero_experts": 20, + "mean": 1.78125, + "std": 2.25, + "cv": 1.263157844543457, + "gini": 0.5923793911933899, + "entropy_nats": 3.49080753326416, + "effective_experts": 32.812434142501516, + "min": 0.0, + "p50": 1.0, + "p90": 4.0, + "max": 12.0 + }, + "top_experts": [ + { + "expert": 8, + "routes": 12 + }, + { + "expert": 32, + "routes": 8 + }, + { + "expert": 6, + "routes": 7 + }, + { + "expert": 24, + "routes": 7 + }, + { + "expert": 33, + "routes": 5 + }, + { + "expert": 49, + "routes": 5 + }, + { + "expert": 17, + "routes": 4 + }, + { + "expert": 21, + "routes": 4 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 57, + "weight": 0.19667254388332367 + }, + { + "expert": 60, + "weight": 0.09833430498838425 + }, + { + "expert": 26, + "weight": 0.04889551177620888 + }, + { + "expert": 8, + "weight": 0.04098263755440712 + }, + { + "expert": 11, + "weight": 0.03667396679520607 + }, + { + "expert": 56, + "weight": 0.03248872235417366 + } + ], + "selected_weight_sum": 0.45404768735170364, + "top1_top2_margin": 0.09833823889493942 + }, + { + "position": 1, + "token_id": 9081, + "token": "Write", + "decoded_piece": "Write", + "experts_by_weight": [ + { + "expert": 24, + "weight": 0.09549997746944427 + }, + { + "expert": 32, + "weight": 0.08619722723960876 + }, + { + "expert": 8, + "weight": 0.08242639154195786 + }, + { + "expert": 56, + "weight": 0.07498472929000854 + }, + { + "expert": 22, + "weight": 0.0488131120800972 + }, + { + "expert": 17, + "weight": 0.032170750200748444 + } + ], + "selected_weight_sum": 0.4200921878218651, + "top1_top2_margin": 0.00930275022983551 + }, + { + "position": 2, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 49, + "weight": 0.14923442900180817 + }, + { + "expert": 8, + "weight": 0.13033752143383026 + }, + { + "expert": 21, + "weight": 0.08116056770086288 + }, + { + "expert": 6, + "weight": 0.061656370759010315 + }, + { + "expert": 26, + "weight": 0.057049237191677094 + }, + { + "expert": 29, + "weight": 0.0237296000123024 + } + ], + "selected_weight_sum": 0.5031677260994911, + "top1_top2_margin": 0.018896907567977905 + }, + { + "position": 3, + "token_id": 12974, + "token": "ĠPython", + "decoded_piece": " Python", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.10350564122200012 + }, + { + "expert": 36, + "weight": 0.059758465737104416 + }, + { + "expert": 30, + "weight": 0.05859985575079918 + }, + { + "expert": 1, + "weight": 0.05378400906920433 + }, + { + "expert": 8, + "weight": 0.0489177405834198 + }, + { + "expert": 47, + "weight": 0.04772283136844635 + } + ], + "selected_weight_sum": 0.3722885437309742, + "top1_top2_margin": 0.043747175484895706 + }, + { + "position": 4, + "token_id": 1157, + "token": "Ġfunction", + "decoded_piece": " function", + "experts_by_weight": [ + { + "expert": 24, + "weight": 0.1556571125984192 + }, + { + "expert": 32, + "weight": 0.11224940419197083 + }, + { + "expert": 8, + "weight": 0.08262474089860916 + }, + { + "expert": 2, + "weight": 0.061063818633556366 + }, + { + "expert": 39, + "weight": 0.0245226863771677 + }, + { + "expert": 62, + "weight": 0.023945074528455734 + } + ], + "selected_weight_sum": 0.460062837228179, + "top1_top2_margin": 0.043407708406448364 + }, + { + "position": 5, + "token_id": 344, + "token": "Ġthat", + "decoded_piece": " that", + "experts_by_weight": [ + { + "expert": 43, + "weight": 0.1358845978975296 + }, + { + "expert": 8, + "weight": 0.12254359573125839 + }, + { + "expert": 5, + "weight": 0.06546720862388611 + }, + { + "expert": 50, + "weight": 0.0648581013083458 + }, + { + "expert": 49, + "weight": 0.02445332705974579 + }, + { + "expert": 36, + "weight": 0.022628234699368477 + } + ], + "selected_weight_sum": 0.43583506532013416, + "top1_top2_margin": 0.01334100216627121 + }, + { + "position": 6, + "token_id": 7578, + "token": "Ġreturns", + "decoded_piece": " returns", + "experts_by_weight": [ + { + "expert": 24, + "weight": 0.13557755947113037 + }, + { + "expert": 32, + "weight": 0.09798909723758698 + }, + { + "expert": 8, + "weight": 0.04169568419456482 + }, + { + "expert": 23, + "weight": 0.03891124948859215 + }, + { + "expert": 37, + "weight": 0.0314752459526062 + }, + { + "expert": 38, + "weight": 0.02870868146419525 + } + ], + "selected_weight_sum": 0.37435751780867577, + "top1_top2_margin": 0.037588462233543396 + }, + { + "position": 7, + "token_id": 254, + "token": "Ġthe", + "decoded_piece": " the", + "experts_by_weight": [ + { + "expert": 21, + "weight": 0.14209777116775513 + }, + { + "expert": 8, + "weight": 0.06827035546302795 + }, + { + "expert": 50, + "weight": 0.04493793100118637 + }, + { + "expert": 6, + "weight": 0.03718099743127823 + }, + { + "expert": 53, + "weight": 0.031235428526997566 + }, + { + "expert": 36, + "weight": 0.02947692945599556 + } + ], + "selected_weight_sum": 0.3531994130462408, + "top1_top2_margin": 0.07382741570472717 + }, + { + "position": 8, + "token_id": 1022, + "token": "Ġfirst", + "decoded_piece": " first", + "experts_by_weight": [ + { + "expert": 3, + "weight": 0.11077141761779785 + }, + { + "expert": 9, + "weight": 0.0720604807138443 + }, + { + "expert": 47, + "weight": 0.06530284881591797 + }, + { + "expert": 6, + "weight": 0.05896109715104103 + }, + { + "expert": 42, + "weight": 0.041370950639247894 + }, + { + "expert": 33, + "weight": 0.03513072803616524 + } + ], + "selected_weight_sum": 0.3835975229740143, + "top1_top2_margin": 0.03871093690395355 + }, + { + "position": 9, + "token_id": 9842, + "token": "Ġrepeated", + "decoded_piece": " repeated", + "experts_by_weight": [ + { + "expert": 32, + "weight": 0.18357861042022705 + }, + { + "expert": 24, + "weight": 0.11116449534893036 + }, + { + "expert": 33, + "weight": 0.05987098067998886 + }, + { + "expert": 6, + "weight": 0.05746008828282356 + }, + { + "expert": 17, + "weight": 0.043144918978214264 + }, + { + "expert": 29, + "weight": 0.029252201318740845 + } + ], + "selected_weight_sum": 0.48447129502892494, + "top1_top2_margin": 0.07241411507129669 + }, + { + "position": 10, + "token_id": 3699, + "token": "Ġelement", + "decoded_piece": " element", + "experts_by_weight": [ + { + "expert": 39, + "weight": 0.11392679065465927 + }, + { + "expert": 2, + "weight": 0.055260807275772095 + }, + { + "expert": 32, + "weight": 0.05243296921253204 + }, + { + "expert": 33, + "weight": 0.04332876577973366 + }, + { + "expert": 6, + "weight": 0.035943593829870224 + }, + { + "expert": 29, + "weight": 0.03192679211497307 + } + ], + "selected_weight_sum": 0.33281971886754036, + "top1_top2_margin": 0.058665983378887177 + }, + { + "position": 11, + "token_id": 279, + "token": "Ġin", + "decoded_piece": " in", + "experts_by_weight": [ + { + "expert": 24, + "weight": 0.08486376702785492 + }, + { + "expert": 53, + "weight": 0.06781504303216934 + }, + { + "expert": 8, + "weight": 0.06028468906879425 + }, + { + "expert": 49, + "weight": 0.03633364662528038 + }, + { + "expert": 35, + "weight": 0.03177472949028015 + }, + { + "expert": 11, + "weight": 0.03020312637090683 + } + ], + "selected_weight_sum": 0.3112750016152859, + "top1_top2_margin": 0.017048723995685577 + }, + { + "position": 12, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 26, + "weight": 0.10673930495977402 + }, + { + "expert": 21, + "weight": 0.10209112614393234 + }, + { + "expert": 8, + "weight": 0.06482338160276413 + }, + { + "expert": 20, + "weight": 0.03735273331403732 + }, + { + "expert": 28, + "weight": 0.03692882880568504 + }, + { + "expert": 1, + "weight": 0.029585134238004684 + } + ], + "selected_weight_sum": 0.37752050906419754, + "top1_top2_margin": 0.004648178815841675 + }, + { + "position": 13, + "token_id": 1525, + "token": "Ġlist", + "decoded_piece": " list", + "experts_by_weight": [ + { + "expert": 24, + "weight": 0.10538846254348755 + }, + { + "expert": 32, + "weight": 0.06087666004896164 + }, + { + "expert": 36, + "weight": 0.059287186712026596 + }, + { + "expert": 2, + "weight": 0.04374944046139717 + }, + { + "expert": 47, + "weight": 0.04220329970121384 + }, + { + "expert": 17, + "weight": 0.028590604662895203 + } + ], + "selected_weight_sum": 0.340095654129982, + "top1_top2_margin": 0.04451180249452591 + }, + { + "position": 14, + "token_id": 285, + "token": "Ġand", + "decoded_piece": " and", + "experts_by_weight": [ + { + "expert": 51, + "weight": 0.1155504509806633 + }, + { + "expert": 8, + "weight": 0.05517224967479706 + }, + { + "expert": 24, + "weight": 0.03504233807325363 + }, + { + "expert": 43, + "weight": 0.026940295472741127 + }, + { + "expert": 22, + "weight": 0.025131845846772194 + }, + { + "expert": 49, + "weight": 0.02402118220925331 + } + ], + "selected_weight_sum": 0.2818583622574806, + "top1_top2_margin": 0.06037820130586624 + }, + { + "position": 15, + "token_id": 6716, + "token": "Ġexplain", + "decoded_piece": " explain", + "experts_by_weight": [ + { + "expert": 33, + "weight": 0.14533041417598724 + }, + { + "expert": 32, + "weight": 0.11240343749523163 + }, + { + "expert": 48, + "weight": 0.07190889120101929 + }, + { + "expert": 6, + "weight": 0.0551474392414093 + }, + { + "expert": 17, + "weight": 0.03574245423078537 + }, + { + "expert": 54, + "weight": 0.034908734261989594 + } + ], + "selected_weight_sum": 0.4554413706064224, + "top1_top2_margin": 0.032926976680755615 + }, + { + "position": 16, + "token_id": 895, + "token": "Ġits", + "decoded_piece": " its", + "experts_by_weight": [ + { + "expert": 62, + "weight": 0.08592283725738525 + }, + { + "expert": 0, + "weight": 0.08066461980342865 + }, + { + "expert": 49, + "weight": 0.06557497382164001 + }, + { + "expert": 55, + "weight": 0.053587011992931366 + }, + { + "expert": 21, + "weight": 0.045721616595983505 + }, + { + "expert": 22, + "weight": 0.024450529366731644 + } + ], + "selected_weight_sum": 0.35592158883810043, + "top1_top2_margin": 0.005258217453956604 + }, + { + "position": 17, + "token_id": 13913, + "token": "Ġcomplexity", + "decoded_piece": " complexity", + "experts_by_weight": [ + { + "expert": 32, + "weight": 0.11456479132175446 + }, + { + "expert": 33, + "weight": 0.07901880145072937 + }, + { + "expert": 6, + "weight": 0.07786156237125397 + }, + { + "expert": 3, + "weight": 0.06110609322786331 + }, + { + "expert": 59, + "weight": 0.044746607542037964 + }, + { + "expert": 10, + "weight": 0.041917551308870316 + } + ], + "selected_weight_sum": 0.4192154072225094, + "top1_top2_margin": 0.035545989871025085 + }, + { + "position": 18, + "token_id": 13, + "token": ".", + "decoded_piece": ".", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.10835051536560059 + }, + { + "expert": 8, + "weight": 0.07074465602636337 + }, + { + "expert": 57, + "weight": 0.056041520088911057 + }, + { + "expert": 51, + "weight": 0.04474358260631561 + }, + { + "expert": 11, + "weight": 0.03019094280898571 + }, + { + "expert": 43, + "weight": 0.027098262682557106 + } + ], + "selected_weight_sum": 0.33716947957873344, + "top1_top2_margin": 0.03760585933923721 + } + ] + }, + { + "id": "math", + "label": "数学推理", + "tokens": 27, + "load": [ + 4, + 2, + 0, + 2, + 2, + 4, + 5, + 2, + 12, + 0, + 1, + 4, + 0, + 5, + 2, + 2, + 0, + 15, + 0, + 0, + 1, + 0, + 1, + 2, + 11, + 1, + 5, + 4, + 0, + 3, + 1, + 0, + 0, + 4, + 4, + 1, + 0, + 3, + 4, + 3, + 1, + 3, + 6, + 4, + 1, + 7, + 0, + 0, + 0, + 0, + 0, + 2, + 2, + 5, + 0, + 1, + 3, + 6, + 3, + 6, + 1, + 0, + 1, + 0 + ], + "metrics": { + "routes": 162, + "used_experts": 45, + "zero_experts": 19, + "mean": 2.53125, + "std": 2.997187852859497, + "cv": 1.1840741634368896, + "gini": 0.5760030746459961, + "entropy_nats": 3.53518009185791, + "effective_experts": 34.301191523325606, + "min": 0.0, + "p50": 2.0, + "p90": 5.6999969482421875, + "max": 15.0 + }, + "top_experts": [ + { + "expert": 17, + "routes": 15 + }, + { + "expert": 8, + "routes": 12 + }, + { + "expert": 24, + "routes": 11 + }, + { + "expert": 45, + "routes": 7 + }, + { + "expert": 42, + "routes": 6 + }, + { + "expert": 57, + "routes": 6 + }, + { + "expert": 59, + "routes": 6 + }, + { + "expert": 6, + "routes": 5 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 57, + "weight": 0.19667254388332367 + }, + { + "expert": 60, + "weight": 0.09833430498838425 + }, + { + "expert": 26, + "weight": 0.04889551177620888 + }, + { + "expert": 8, + "weight": 0.04098263755440712 + }, + { + "expert": 11, + "weight": 0.03667396679520607 + }, + { + "expert": 56, + "weight": 0.03248872235417366 + } + ], + "selected_weight_sum": 0.45404768735170364, + "top1_top2_margin": 0.09833823889493942 + }, + { + "position": 1, + "token_id": 8246, + "token": "èĭ¥", + "decoded_piece": "若", + "experts_by_weight": [ + { + "expert": 43, + "weight": 0.1394088715314865 + }, + { + "expert": 25, + "weight": 0.06345965713262558 + }, + { + "expert": 24, + "weight": 0.042812079191207886 + }, + { + "expert": 56, + "weight": 0.04217054322361946 + }, + { + "expert": 42, + "weight": 0.04117071256041527 + }, + { + "expert": 53, + "weight": 0.029628943651914597 + } + ], + "selected_weight_sum": 0.3586508072912693, + "top1_top2_margin": 0.07594921439886093 + }, + { + "position": 2, + "token_id": 2206, + "token": "æŃ£", + "decoded_piece": "正", + "experts_by_weight": [ + { + "expert": 58, + "weight": 0.10335297882556915 + }, + { + "expert": 3, + "weight": 0.0671335905790329 + }, + { + "expert": 44, + "weight": 0.04563439264893532 + }, + { + "expert": 6, + "weight": 0.037449806928634644 + }, + { + "expert": 53, + "weight": 0.03589992597699165 + }, + { + "expert": 17, + "weight": 0.034529875963926315 + } + ], + "selected_weight_sum": 0.32400057092309, + "top1_top2_margin": 0.036219388246536255 + }, + { + "position": 3, + "token_id": 2266, + "token": "æķ°", + "decoded_piece": "数", + "experts_by_weight": [ + { + "expert": 59, + "weight": 0.08278294652700424 + }, + { + "expert": 24, + "weight": 0.06642577052116394 + }, + { + "expert": 38, + "weight": 0.053992412984371185 + }, + { + "expert": 1, + "weight": 0.0396278016269207 + }, + { + "expert": 14, + "weight": 0.03648533672094345 + }, + { + "expert": 29, + "weight": 0.03367812559008598 + } + ], + "selected_weight_sum": 0.3129923939704895, + "top1_top2_margin": 0.0163571760058403 + }, + { + "position": 4, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 26, + "weight": 0.21853145956993103 + }, + { + "expert": 8, + "weight": 0.15359973907470703 + }, + { + "expert": 15, + "weight": 0.07538707554340363 + }, + { + "expert": 45, + "weight": 0.06006627157330513 + }, + { + "expert": 59, + "weight": 0.04238240048289299 + }, + { + "expert": 41, + "weight": 0.022744085639715195 + } + ], + "selected_weight_sum": 0.572711031883955, + "top1_top2_margin": 0.064931720495224 + }, + { + "position": 5, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.12307935208082199 + }, + { + "expert": 57, + "weight": 0.11029956489801407 + }, + { + "expert": 27, + "weight": 0.04811596870422363 + }, + { + "expert": 0, + "weight": 0.04188108816742897 + }, + { + "expert": 43, + "weight": 0.041408829391002655 + }, + { + "expert": 13, + "weight": 0.035461392253637314 + } + ], + "selected_weight_sum": 0.40024619549512863, + "top1_top2_margin": 0.012779787182807922 + }, + { + "position": 6, + "token_id": 13211, + "token": "满足", + "decoded_piece": "满足", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.2006656676530838 + }, + { + "expert": 24, + "weight": 0.09593425691127777 + }, + { + "expert": 38, + "weight": 0.051401205360889435 + }, + { + "expert": 5, + "weight": 0.043837904930114746 + }, + { + "expert": 37, + "weight": 0.03912341967225075 + }, + { + "expert": 10, + "weight": 0.036207932978868484 + } + ], + "selected_weight_sum": 0.467170387506485, + "top1_top2_margin": 0.10473141074180603 + }, + { + "position": 7, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 26, + "weight": 0.16054129600524902 + }, + { + "expert": 8, + "weight": 0.14097601175308228 + }, + { + "expert": 59, + "weight": 0.04995854198932648 + }, + { + "expert": 45, + "weight": 0.047472260892391205 + }, + { + "expert": 41, + "weight": 0.031187430024147034 + }, + { + "expert": 39, + "weight": 0.0243584793061018 + } + ], + "selected_weight_sum": 0.4544940199702978, + "top1_top2_margin": 0.019565284252166748 + }, + { + "position": 8, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.10879295319318771 + }, + { + "expert": 57, + "weight": 0.10352718830108643 + }, + { + "expert": 27, + "weight": 0.04323926195502281 + }, + { + "expert": 13, + "weight": 0.037018291652202606 + }, + { + "expert": 34, + "weight": 0.034377746284008026 + }, + { + "expert": 0, + "weight": 0.028740879148244858 + } + ], + "selected_weight_sum": 0.35569632053375244, + "top1_top2_margin": 0.005265764892101288 + }, + { + "position": 9, + "token_id": 14855, + "token": "åĬłä¸Ĭ", + "decoded_piece": "加上", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.1886313408613205 + }, + { + "expert": 24, + "weight": 0.10988873988389969 + }, + { + "expert": 5, + "weight": 0.07265309244394302 + }, + { + "expert": 42, + "weight": 0.05920105054974556 + }, + { + "expert": 37, + "weight": 0.03127415105700493 + }, + { + "expert": 45, + "weight": 0.02951677516102791 + } + ], + "selected_weight_sum": 0.4911651499569416, + "top1_top2_margin": 0.0787426009774208 + }, + { + "position": 10, + "token_id": 12985, + "token": "å®ĥçļĦ", + "decoded_piece": "它的", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.17231044173240662 + }, + { + "expert": 62, + "weight": 0.11317699402570724 + }, + { + "expert": 30, + "weight": 0.06843745708465576 + }, + { + "expert": 55, + "weight": 0.06510318070650101 + }, + { + "expert": 5, + "weight": 0.03418071195483208 + }, + { + "expert": 8, + "weight": 0.027257278561592102 + } + ], + "selected_weight_sum": 0.4804660640656948, + "top1_top2_margin": 0.05913344770669937 + }, + { + "position": 11, + "token_id": 8270, + "token": "åĢĴ", + "decoded_piece": "倒", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.32550475001335144 + }, + { + "expert": 33, + "weight": 0.0801854208111763 + }, + { + "expert": 6, + "weight": 0.07861076295375824 + }, + { + "expert": 53, + "weight": 0.05385090410709381 + }, + { + "expert": 58, + "weight": 0.046686459332704544 + }, + { + "expert": 42, + "weight": 0.04537269100546837 + } + ], + "selected_weight_sum": 0.6302109882235527, + "top1_top2_margin": 0.24531932920217514 + }, + { + "position": 12, + "token_id": 2266, + "token": "æķ°", + "decoded_piece": "数", + "experts_by_weight": [ + { + "expert": 29, + "weight": 0.203230619430542 + }, + { + "expert": 24, + "weight": 0.061504725366830826 + }, + { + "expert": 59, + "weight": 0.05149833485484123 + }, + { + "expert": 17, + "weight": 0.04692880064249039 + }, + { + "expert": 14, + "weight": 0.042546626180410385 + }, + { + "expert": 23, + "weight": 0.038524165749549866 + } + ], + "selected_weight_sum": 0.4442332722246647, + "top1_top2_margin": 0.14172589406371117 + }, + { + "position": 13, + "token_id": 37360, + "token": "çŃīäºİ", + "decoded_piece": "等于", + "experts_by_weight": [ + { + "expert": 43, + "weight": 0.08651761710643768 + }, + { + "expert": 17, + "weight": 0.08569129556417465 + }, + { + "expert": 24, + "weight": 0.08219180256128311 + }, + { + "expert": 27, + "weight": 0.07018200308084488 + }, + { + "expert": 45, + "weight": 0.03885370492935181 + }, + { + "expert": 42, + "weight": 0.036339085549116135 + } + ], + "selected_weight_sum": 0.39977550879120827, + "top1_top2_margin": 0.000826321542263031 + }, + { + "position": 14, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.14133651554584503 + }, + { + "expert": 57, + "weight": 0.08895187824964523 + }, + { + "expert": 13, + "weight": 0.06798499077558517 + }, + { + "expert": 0, + "weight": 0.031565990298986435 + }, + { + "expert": 26, + "weight": 0.028622372075915337 + }, + { + "expert": 34, + "weight": 0.024421988055109978 + } + ], + "selected_weight_sum": 0.3828837350010872, + "top1_top2_margin": 0.0523846372961998 + }, + { + "position": 15, + "token_id": 18, + "token": "3", + "decoded_piece": "3", + "experts_by_weight": [ + { + "expert": 13, + "weight": 0.062128130346536636 + }, + { + "expert": 8, + "weight": 0.06102198734879494 + }, + { + "expert": 37, + "weight": 0.036797475069761276 + }, + { + "expert": 40, + "weight": 0.032095663249492645 + }, + { + "expert": 1, + "weight": 0.030832836404442787 + }, + { + "expert": 52, + "weight": 0.029817776754498482 + } + ], + "selected_weight_sum": 0.25269386917352676, + "top1_top2_margin": 0.0011061429977416992 + }, + { + "position": 16, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 51, + "weight": 0.1388871967792511 + }, + { + "expert": 8, + "weight": 0.08805988729000092 + }, + { + "expert": 4, + "weight": 0.05442273989319801 + }, + { + "expert": 43, + "weight": 0.02597079984843731 + }, + { + "expert": 11, + "weight": 0.02435963973402977 + }, + { + "expert": 56, + "weight": 0.0232080165296793 + } + ], + "selected_weight_sum": 0.3549082800745964, + "top1_top2_margin": 0.05082730948925018 + }, + { + "position": 17, + "token_id": 2989, + "token": "æ±Ĥ", + "decoded_piece": "求", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.13024798035621643 + }, + { + "expert": 24, + "weight": 0.12015853077173233 + }, + { + "expert": 53, + "weight": 0.04987603798508644 + }, + { + "expert": 35, + "weight": 0.049373287707567215 + }, + { + "expert": 22, + "weight": 0.04563625529408455 + }, + { + "expert": 38, + "weight": 0.040090806782245636 + } + ], + "selected_weight_sum": 0.4353828988969326, + "top1_top2_margin": 0.0100894495844841 + }, + { + "position": 18, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 26, + "weight": 0.14243966341018677 + }, + { + "expert": 8, + "weight": 0.06603427976369858 + }, + { + "expert": 59, + "weight": 0.06250512599945068 + }, + { + "expert": 45, + "weight": 0.04692105948925018 + }, + { + "expert": 41, + "weight": 0.033770523965358734 + }, + { + "expert": 7, + "weight": 0.026090526953339577 + } + ], + "selected_weight_sum": 0.3777611795812845, + "top1_top2_margin": 0.07640538364648819 + }, + { + "position": 19, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 57, + "weight": 0.08014163374900818 + }, + { + "expert": 34, + "weight": 0.061893001198768616 + }, + { + "expert": 13, + "weight": 0.0616314522922039 + }, + { + "expert": 8, + "weight": 0.04130232706665993 + }, + { + "expert": 27, + "weight": 0.03659188747406006 + }, + { + "expert": 0, + "weight": 0.024267001077532768 + } + ], + "selected_weight_sum": 0.30582730285823345, + "top1_top2_margin": 0.018248632550239563 + }, + { + "position": 20, + "token_id": 337, + "token": "çļĦ", + "decoded_piece": "的", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.054228734225034714 + }, + { + "expert": 38, + "weight": 0.052332378923892975 + }, + { + "expert": 20, + "weight": 0.04236849769949913 + }, + { + "expert": 15, + "weight": 0.03419002890586853 + }, + { + "expert": 17, + "weight": 0.032592903822660446 + }, + { + "expert": 7, + "weight": 0.0297713503241539 + } + ], + "selected_weight_sum": 0.2454838939011097, + "top1_top2_margin": 0.001896355301141739 + }, + { + "position": 21, + "token_id": 18718, + "token": "å¹³æĸ¹", + "decoded_piece": "平方", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.13682566583156586 + }, + { + "expert": 24, + "weight": 0.12559771537780762 + }, + { + "expert": 33, + "weight": 0.05148334428668022 + }, + { + "expert": 45, + "weight": 0.043595705181360245 + }, + { + "expert": 6, + "weight": 0.042157530784606934 + }, + { + "expert": 39, + "weight": 0.028853029012680054 + } + ], + "selected_weight_sum": 0.4285129904747009, + "top1_top2_margin": 0.01122795045375824 + }, + { + "position": 22, + "token_id": 14855, + "token": "åĬłä¸Ĭ", + "decoded_piece": "加上", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.1608467400074005 + }, + { + "expert": 24, + "weight": 0.07378645241260529 + }, + { + "expert": 42, + "weight": 0.06305347383022308 + }, + { + "expert": 3, + "weight": 0.03693674877285957 + }, + { + "expert": 11, + "weight": 0.03622017800807953 + }, + { + "expert": 5, + "weight": 0.033631253987550735 + } + ], + "selected_weight_sum": 0.4044748470187187, + "top1_top2_margin": 0.08706028759479523 + }, + { + "position": 23, + "token_id": 8270, + "token": "åĢĴ", + "decoded_piece": "倒", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.22711610794067383 + }, + { + "expert": 33, + "weight": 0.08942194283008575 + }, + { + "expert": 6, + "weight": 0.06970909237861633 + }, + { + "expert": 58, + "weight": 0.05923660099506378 + }, + { + "expert": 53, + "weight": 0.05873405933380127 + }, + { + "expert": 42, + "weight": 0.04915972426533699 + } + ], + "selected_weight_sum": 0.553377527743578, + "top1_top2_margin": 0.13769416511058807 + }, + { + "position": 24, + "token_id": 51971, + "token": "æķ°çļĦ", + "decoded_piece": "数的", + "experts_by_weight": [ + { + "expert": 29, + "weight": 0.17918023467063904 + }, + { + "expert": 59, + "weight": 0.05802371725440025 + }, + { + "expert": 24, + "weight": 0.05564861372113228 + }, + { + "expert": 34, + "weight": 0.04281933978199959 + }, + { + "expert": 23, + "weight": 0.042292799800634384 + }, + { + "expert": 17, + "weight": 0.04076436161994934 + } + ], + "selected_weight_sum": 0.4187290668487549, + "top1_top2_margin": 0.12115651741623878 + }, + { + "position": 25, + "token_id": 18718, + "token": "å¹³æĸ¹", + "decoded_piece": "平方", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.12711961567401886 + }, + { + "expert": 24, + "weight": 0.12048730254173279 + }, + { + "expert": 33, + "weight": 0.05697174370288849 + }, + { + "expert": 6, + "weight": 0.045253265649080276 + }, + { + "expert": 45, + "weight": 0.03940247744321823 + }, + { + "expert": 39, + "weight": 0.030314432457089424 + } + ], + "selected_weight_sum": 0.41954883746802807, + "top1_top2_margin": 0.006632313132286072 + }, + { + "position": 26, + "token_id": 398, + "token": "ãĢĤ", + "decoded_piece": "。", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.09412042051553726 + }, + { + "expert": 51, + "weight": 0.06351388990879059 + }, + { + "expert": 8, + "weight": 0.06225387006998062 + }, + { + "expert": 57, + "weight": 0.06197313964366913 + }, + { + "expert": 17, + "weight": 0.027688290923833847 + }, + { + "expert": 11, + "weight": 0.025522006675601006 + } + ], + "selected_weight_sum": 0.33507161773741245, + "top1_top2_margin": 0.030606530606746674 + } + ] + } + ], + "prompt_pair_jaccard": [ + { + "left": "zh_explanation", + "right": "en_architecture", + "used_expert_jaccard": 0.5666666666666667, + "shared_experts": 34, + "union_experts": 60 + }, + { + "left": "zh_explanation", + "right": "code", + "used_expert_jaccard": 0.6610169491525424, + "shared_experts": 39, + "union_experts": 59 + }, + { + "left": "zh_explanation", + "right": "math", + "used_expert_jaccard": 0.6779661016949152, + "shared_experts": 40, + "union_experts": 59 + }, + { + "left": "en_architecture", + "right": "code", + "used_expert_jaccard": 0.5555555555555556, + "shared_experts": 30, + "union_experts": 54 + }, + { + "left": "en_architecture", + "right": "math", + "used_expert_jaccard": 0.46551724137931033, + "shared_experts": 27, + "union_experts": 58 + }, + { + "left": "code", + "right": "math", + "used_expert_jaccard": 0.5892857142857143, + "shared_experts": 33, + "union_experts": 56 + } + ] + } + }, + { + "layer": 2, + "ffn": "moe", + "state_tensors": 203, + "state_numel": 584847872, + "state_bytes": 1169695744, + "timing_ms": { + "load_cpu": 24.617862014565617, + "transfer_to_device": 123.9008959964849, + "forward": 4.285834002075717 + }, + "peak_device_allocated_mib": 1173.2724609375, + "hidden_before": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 0.12936507165431976, + "p10": 0.08545936644077301, + "p50": 0.12583057582378387, + "p90": 0.1601526141166687 + }, + "l2_norm": { + "mean": 5.854394435882568, + "p10": 3.8674492835998535, + "p50": 5.694441795349121, + "p90": 7.247680187225342 + }, + "finite": true, + "sha256_fp32": "bc8c6a2a556f09b333c171b50788905902a2aa18e0bbac784837a7604d931fad" + }, + "hidden_after": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 0.22556711733341217, + "p10": 0.10850249230861664, + "p50": 0.1559184491634369, + "p90": 0.19672976434230804 + }, + "l2_norm": { + "mean": 10.208003044128418, + "p10": 4.910262107849121, + "p50": 7.056063175201416, + "p90": 8.902973175048828 + }, + "finite": true, + "sha256_fp32": "76c1ba59ea5ef19f16162513c7c36e6e8d17a9af9fa3fdedbf657e80882fa0a3" + }, + "mla": { + "compressed_projection_shape": [ + 4, + 27, + 576 + ], + "latent_shape": [ + 4, + 27, + 512 + ], + "rope_key_shape": [ + 4, + 27, + 64 + ], + "latent_rms": 2.27437686920166, + "rope_key_rms": 3.476457118988037, + "eager_key_cache_shape": [ + 4, + 16, + 27, + 192 + ], + "eager_value_cache_shape": [ + 4, + 16, + 27, + 128 + ], + "eager_cache_bytes": 1105920 + }, + "routing": { + "aggregate_load": [ + 13, + 3, + 1, + 13, + 5, + 8, + 6, + 9, + 9, + 10, + 1, + 6, + 7, + 6, + 4, + 12, + 10, + 13, + 11, + 4, + 10, + 2, + 11, + 5, + 8, + 3, + 13, + 12, + 5, + 4, + 14, + 4, + 14, + 5, + 14, + 18, + 8, + 3, + 7, + 7, + 10, + 4, + 7, + 9, + 5, + 13, + 9, + 19, + 8, + 8, + 4, + 5, + 13, + 2, + 3, + 13, + 5, + 11, + 10, + 4, + 17, + 9, + 23, + 11 + ], + "aggregate_metrics": { + "routes": 540, + "used_experts": 64, + "zero_experts": 0, + "mean": 8.4375, + "std": 4.6321001052856445, + "cv": 0.5489896535873413, + "gini": 0.30329862236976624, + "entropy_nats": 4.008847713470459, + "effective_experts": 55.08336216433062, + "min": 1.0, + "p50": 8.0, + "p90": 13.699996948242188, + "max": 23.0 + }, + "top_experts": [ + { + "expert": 62, + "routes": 23 + }, + { + "expert": 47, + "routes": 19 + }, + { + "expert": 35, + "routes": 18 + }, + { + "expert": 60, + "routes": 17 + }, + { + "expert": 30, + "routes": 14 + }, + { + "expert": 32, + "routes": 14 + }, + { + "expert": 34, + "routes": 14 + }, + { + "expert": 0, + "routes": 13 + }, + { + "expert": 3, + "routes": 13 + }, + { + "expert": 17, + "routes": 13 + }, + { + "expert": 26, + "routes": 13 + }, + { + "expert": 45, + "routes": 13 + } + ], + "selected_weight_sum": { + "mean": 0.48013466596603394, + "p10": 0.35627180337905884, + "p50": 0.4594435691833496, + "p90": 0.587585985660553 + }, + "top1_top2_margin": { + "mean": 0.10077717900276184, + "p50": 0.0558306984603405, + "p90": 0.1981295347213745 + }, + "per_prompt": [ + { + "id": "zh_explanation", + "label": "中文解释", + "tokens": 26, + "load": [ + 3, + 1, + 0, + 4, + 0, + 2, + 1, + 4, + 3, + 3, + 0, + 3, + 2, + 1, + 2, + 4, + 2, + 5, + 3, + 1, + 4, + 2, + 2, + 2, + 4, + 1, + 3, + 4, + 2, + 3, + 4, + 2, + 4, + 0, + 3, + 4, + 2, + 1, + 1, + 3, + 0, + 1, + 2, + 1, + 3, + 5, + 3, + 6, + 3, + 3, + 0, + 2, + 2, + 0, + 2, + 2, + 2, + 2, + 4, + 2, + 3, + 4, + 6, + 3 + ], + "metrics": { + "routes": 156, + "used_experts": 57, + "zero_experts": 7, + "mean": 2.4375, + "std": 1.457057237625122, + "cv": 0.5977670550346375, + "gini": 0.32912659645080566, + "entropy_nats": 3.9409890174865723, + "effective_experts": 51.469480352336895, + "min": 0.0, + "p50": 2.0, + "p90": 4.0, + "max": 6.0 + }, + "top_experts": [ + { + "expert": 47, + "routes": 6 + }, + { + "expert": 62, + "routes": 6 + }, + { + "expert": 17, + "routes": 5 + }, + { + "expert": 45, + "routes": 5 + }, + { + "expert": 3, + "routes": 4 + }, + { + "expert": 7, + "routes": 4 + }, + { + "expert": 15, + "routes": 4 + }, + { + "expert": 20, + "routes": 4 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 3, + "weight": 0.7993167042732239 + }, + { + "expert": 62, + "weight": 0.018388060852885246 + }, + { + "expert": 38, + "weight": 0.010282253846526146 + }, + { + "expert": 63, + "weight": 0.009434774518013 + }, + { + "expert": 35, + "weight": 0.009029458276927471 + }, + { + "expert": 60, + "weight": 0.008924801833927631 + } + ], + "selected_weight_sum": 0.8553760536015034, + "top1_top2_margin": 0.7809286434203386 + }, + { + "position": 1, + "token_id": 1082, + "token": "ç͍", + "decoded_piece": "用", + "experts_by_weight": [ + { + "expert": 14, + "weight": 0.20668703317642212 + }, + { + "expert": 42, + "weight": 0.17426803708076477 + }, + { + "expert": 61, + "weight": 0.07382696866989136 + }, + { + "expert": 63, + "weight": 0.053661659359931946 + }, + { + "expert": 57, + "weight": 0.03640357777476311 + }, + { + "expert": 35, + "weight": 0.027232389897108078 + } + ], + "selected_weight_sum": 0.5720796659588814, + "top1_top2_margin": 0.03241899609565735 + }, + { + "position": 2, + "token_id": 83983, + "token": "éĢļä¿Ĺ", + "decoded_piece": "通俗", + "experts_by_weight": [ + { + "expert": 18, + "weight": 0.2451930046081543 + }, + { + "expert": 60, + "weight": 0.08248794078826904 + }, + { + "expert": 17, + "weight": 0.060274556279182434 + }, + { + "expert": 27, + "weight": 0.042924247682094574 + }, + { + "expert": 37, + "weight": 0.03560378775000572 + }, + { + "expert": 3, + "weight": 0.035133421421051025 + } + ], + "selected_weight_sum": 0.5016169585287571, + "top1_top2_margin": 0.16270506381988525 + }, + { + "position": 3, + "token_id": 91476, + "token": "çļĦè¯Ńè¨Ģ", + "decoded_piece": "的语言", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.23428620398044586 + }, + { + "expert": 26, + "weight": 0.1792253851890564 + }, + { + "expert": 45, + "weight": 0.08116163313388824 + }, + { + "expert": 30, + "weight": 0.06720003485679626 + }, + { + "expert": 36, + "weight": 0.03637100011110306 + }, + { + "expert": 43, + "weight": 0.023322662338614464 + } + ], + "selected_weight_sum": 0.6215669196099043, + "top1_top2_margin": 0.055060818791389465 + }, + { + "position": 4, + "token_id": 17882, + "token": "è§£éĩĬ", + "decoded_piece": "解释", + "experts_by_weight": [ + { + "expert": 9, + "weight": 0.133388951420784 + }, + { + "expert": 55, + "weight": 0.10470598191022873 + }, + { + "expert": 39, + "weight": 0.09774415194988251 + }, + { + "expert": 11, + "weight": 0.0421929694712162 + }, + { + "expert": 12, + "weight": 0.03626151382923126 + }, + { + "expert": 32, + "weight": 0.02716749534010887 + } + ], + "selected_weight_sum": 0.44146106392145157, + "top1_top2_margin": 0.028682969510555267 + }, + { + "position": 5, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.13226112723350525 + }, + { + "expert": 46, + "weight": 0.12259065359830856 + }, + { + "expert": 0, + "weight": 0.07745791226625443 + }, + { + "expert": 47, + "weight": 0.04579097405076027 + }, + { + "expert": 62, + "weight": 0.03493405878543854 + }, + { + "expert": 14, + "weight": 0.03442758321762085 + } + ], + "selected_weight_sum": 0.4474623091518879, + "top1_top2_margin": 0.009670473635196686 + }, + { + "position": 6, + "token_id": 10285, + "token": "为ä»Ģä¹Ī", + "decoded_piece": "为什么", + "experts_by_weight": [ + { + "expert": 59, + "weight": 0.12769539654254913 + }, + { + "expert": 61, + "weight": 0.08412055671215057 + }, + { + "expert": 7, + "weight": 0.046987637877464294 + }, + { + "expert": 44, + "weight": 0.04405679926276207 + }, + { + "expert": 42, + "weight": 0.03699975833296776 + }, + { + "expert": 57, + "weight": 0.026510879397392273 + } + ], + "selected_weight_sum": 0.3663710281252861, + "top1_top2_margin": 0.04357483983039856 + }, + { + "position": 7, + "token_id": 21844, + "token": "ç¨Ģ", + "decoded_piece": "稀", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.16432742774486542 + }, + { + "expert": 63, + "weight": 0.0821835920214653 + }, + { + "expert": 24, + "weight": 0.054694924503564835 + }, + { + "expert": 31, + "weight": 0.037129003554582596 + }, + { + "expert": 35, + "weight": 0.036085568368434906 + }, + { + "expert": 22, + "weight": 0.03494443744421005 + } + ], + "selected_weight_sum": 0.4093649536371231, + "top1_top2_margin": 0.08214383572340012 + }, + { + "position": 8, + "token_id": 20923, + "token": "çĸı", + "decoded_piece": "疏", + "experts_by_weight": [ + { + "expert": 49, + "weight": 0.2578503489494324 + }, + { + "expert": 56, + "weight": 0.12817910313606262 + }, + { + "expert": 24, + "weight": 0.07802081108093262 + }, + { + "expert": 11, + "weight": 0.03407106548547745 + }, + { + "expert": 6, + "weight": 0.026755359023809433 + }, + { + "expert": 41, + "weight": 0.022954754531383514 + } + ], + "selected_weight_sum": 0.547831442207098, + "top1_top2_margin": 0.12967124581336975 + }, + { + "position": 9, + "token_id": 10900, + "token": "ä¸ĵå®¶", + "decoded_piece": "专家", + "experts_by_weight": [ + { + "expert": 20, + "weight": 0.17112381756305695 + }, + { + "expert": 55, + "weight": 0.0664329081773758 + }, + { + "expert": 26, + "weight": 0.05617460608482361 + }, + { + "expert": 45, + "weight": 0.05288673937320709 + }, + { + "expert": 54, + "weight": 0.051459167152643204 + }, + { + "expert": 27, + "weight": 0.038296010345220566 + } + ], + "selected_weight_sum": 0.4363732486963272, + "top1_top2_margin": 0.10469090938568115 + }, + { + "position": 10, + "token_id": 39322, + "token": "模åŀĭ", + "decoded_piece": "模型", + "experts_by_weight": [ + { + "expert": 54, + "weight": 0.17979325354099274 + }, + { + "expert": 36, + "weight": 0.10962751507759094 + }, + { + "expert": 28, + "weight": 0.0830262154340744 + }, + { + "expert": 12, + "weight": 0.0828174576163292 + }, + { + "expert": 51, + "weight": 0.04263872653245926 + }, + { + "expert": 30, + "weight": 0.03515750914812088 + } + ], + "selected_weight_sum": 0.5330606773495674, + "top1_top2_margin": 0.0701657384634018 + }, + { + "position": 11, + "token_id": 1876, + "token": "åı¯ä»¥", + "decoded_piece": "可以", + "experts_by_weight": [ + { + "expert": 15, + "weight": 0.1157049909234047 + }, + { + "expert": 5, + "weight": 0.11091289669275284 + }, + { + "expert": 47, + "weight": 0.050560325384140015 + }, + { + "expert": 7, + "weight": 0.0373777411878109 + }, + { + "expert": 51, + "weight": 0.032718632370233536 + }, + { + "expert": 32, + "weight": 0.025750335305929184 + } + ], + "selected_weight_sum": 0.37302492186427116, + "top1_top2_margin": 0.0047920942306518555 + }, + { + "position": 12, + "token_id": 9476, + "token": "æĭ¥æľī", + "decoded_piece": "拥有", + "experts_by_weight": [ + { + "expert": 24, + "weight": 0.17254002392292023 + }, + { + "expert": 39, + "weight": 0.08969754725694656 + }, + { + "expert": 9, + "weight": 0.07656068354845047 + }, + { + "expert": 34, + "weight": 0.038461942225694656 + }, + { + "expert": 56, + "weight": 0.03243434801697731 + }, + { + "expert": 32, + "weight": 0.02500019408762455 + } + ], + "selected_weight_sum": 0.4346947390586138, + "top1_top2_margin": 0.08284247666597366 + }, + { + "position": 13, + "token_id": 3599, + "token": "å¾Īå¤ļ", + "decoded_piece": "很多", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.11956122517585754 + }, + { + "expert": 3, + "weight": 0.11295492947101593 + }, + { + "expert": 52, + "weight": 0.06735346466302872 + }, + { + "expert": 58, + "weight": 0.061125051230192184 + }, + { + "expert": 11, + "weight": 0.042656976729631424 + }, + { + "expert": 20, + "weight": 0.027563519775867462 + } + ], + "selected_weight_sum": 0.43121516704559326, + "top1_top2_margin": 0.006606295704841614 + }, + { + "position": 14, + "token_id": 21273, + "token": "åıĤæķ°", + "decoded_piece": "参数", + "experts_by_weight": [ + { + "expert": 45, + "weight": 0.09754917025566101 + }, + { + "expert": 30, + "weight": 0.07696481049060822 + }, + { + "expert": 27, + "weight": 0.04299183189868927 + }, + { + "expert": 26, + "weight": 0.04288526996970177 + }, + { + "expert": 29, + "weight": 0.036022163927555084 + }, + { + "expert": 18, + "weight": 0.029702916741371155 + } + ], + "selected_weight_sum": 0.3261161632835865, + "top1_top2_margin": 0.020584359765052795 + }, + { + "position": 15, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.10675214231014252 + }, + { + "expert": 0, + "weight": 0.10003915429115295 + }, + { + "expert": 46, + "weight": 0.07917536050081253 + }, + { + "expert": 62, + "weight": 0.03613792732357979 + }, + { + "expert": 15, + "weight": 0.029818829149007797 + }, + { + "expert": 47, + "weight": 0.0273029413074255 + } + ], + "selected_weight_sum": 0.3792263548821211, + "top1_top2_margin": 0.006712988018989563 + }, + { + "position": 16, + "token_id": 1705, + "token": "ä½Ĩ", + "decoded_piece": "但", + "experts_by_weight": [ + { + "expert": 44, + "weight": 0.0980173721909523 + }, + { + "expert": 61, + "weight": 0.08442847430706024 + }, + { + "expert": 59, + "weight": 0.08428655564785004 + }, + { + "expert": 21, + "weight": 0.036362823098897934 + }, + { + "expert": 23, + "weight": 0.031637802720069885 + }, + { + "expert": 48, + "weight": 0.03130587935447693 + } + ], + "selected_weight_sum": 0.3660389073193073, + "top1_top2_margin": 0.01358889788389206 + }, + { + "position": 17, + "token_id": 13712, + "token": "æ¯ı个", + "decoded_piece": "每个", + "experts_by_weight": [ + { + "expert": 61, + "weight": 0.11883475631475449 + }, + { + "expert": 7, + "weight": 0.11768975853919983 + }, + { + "expert": 44, + "weight": 0.04394533485174179 + }, + { + "expert": 16, + "weight": 0.03521232306957245 + }, + { + "expert": 29, + "weight": 0.02718379534780979 + }, + { + "expert": 20, + "weight": 0.025340577587485313 + } + ], + "selected_weight_sum": 0.36820654571056366, + "top1_top2_margin": 0.001144997775554657 + }, + { + "position": 18, + "token_id": 10728, + "token": "Ġtoken", + "decoded_piece": " token", + "experts_by_weight": [ + { + "expert": 45, + "weight": 0.14425362646579742 + }, + { + "expert": 35, + "weight": 0.12026972323656082 + }, + { + "expert": 60, + "weight": 0.07732635736465454 + }, + { + "expert": 27, + "weight": 0.06987600773572922 + }, + { + "expert": 30, + "weight": 0.030445804819464684 + }, + { + "expert": 62, + "weight": 0.03043285384774208 + } + ], + "selected_weight_sum": 0.47260437346994877, + "top1_top2_margin": 0.023983903229236603 + }, + { + "position": 19, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 62, + "weight": 0.16929486393928528 + }, + { + "expert": 47, + "weight": 0.09825362265110016 + }, + { + "expert": 15, + "weight": 0.04029633849859238 + }, + { + "expert": 48, + "weight": 0.032645758241415024 + }, + { + "expert": 19, + "weight": 0.026754116639494896 + }, + { + "expert": 58, + "weight": 0.022297902032732964 + } + ], + "selected_weight_sum": 0.3895426020026207, + "top1_top2_margin": 0.07104124128818512 + }, + { + "position": 20, + "token_id": 2141, + "token": "åıª", + "decoded_piece": "只", + "experts_by_weight": [ + { + "expert": 5, + "weight": 0.1217007115483284 + }, + { + "expert": 15, + "weight": 0.06380197405815125 + }, + { + "expert": 34, + "weight": 0.039825357496738434 + }, + { + "expert": 25, + "weight": 0.03771115466952324 + }, + { + "expert": 47, + "weight": 0.036976952105760574 + }, + { + "expert": 1, + "weight": 0.033029451966285706 + } + ], + "selected_weight_sum": 0.3330456018447876, + "top1_top2_margin": 0.057898737490177155 + }, + { + "position": 21, + "token_id": 5118, + "token": "使ç͍", + "decoded_piece": "使用", + "experts_by_weight": [ + { + "expert": 9, + "weight": 0.1564585566520691 + }, + { + "expert": 24, + "weight": 0.15290194749832153 + }, + { + "expert": 49, + "weight": 0.09143153578042984 + }, + { + "expert": 39, + "weight": 0.045543499290943146 + }, + { + "expert": 32, + "weight": 0.03213364630937576 + }, + { + "expert": 16, + "weight": 0.030746884644031525 + } + ], + "selected_weight_sum": 0.5092160701751709, + "top1_top2_margin": 0.0035566091537475586 + }, + { + "position": 22, + "token_id": 6802, + "token": "åħ¶ä¸Ń", + "decoded_piece": "其中", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.18916168808937073 + }, + { + "expert": 3, + "weight": 0.14847588539123535 + }, + { + "expert": 48, + "weight": 0.06800840049982071 + }, + { + "expert": 18, + "weight": 0.03237553685903549 + }, + { + "expert": 20, + "weight": 0.029806064441800117 + }, + { + "expert": 34, + "weight": 0.029624512419104576 + } + ], + "selected_weight_sum": 0.497452087700367, + "top1_top2_margin": 0.040685802698135376 + }, + { + "position": 23, + "token_id": 51905, + "token": "ä¸Ģå°ı", + "decoded_piece": "一小", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.2096928209066391 + }, + { + "expert": 58, + "weight": 0.18414442241191864 + }, + { + "expert": 22, + "weight": 0.06322769820690155 + }, + { + "expert": 21, + "weight": 0.03188531473278999 + }, + { + "expert": 31, + "weight": 0.026658499613404274 + }, + { + "expert": 28, + "weight": 0.023066453635692596 + } + ], + "selected_weight_sum": 0.5386752095073462, + "top1_top2_margin": 0.02554839849472046 + }, + { + "position": 24, + "token_id": 5781, + "token": "éĥ¨åĪĨ", + "decoded_piece": "部分", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.19247254729270935 + }, + { + "expert": 29, + "weight": 0.09449118375778198 + }, + { + "expert": 45, + "weight": 0.08716513961553574 + }, + { + "expert": 49, + "weight": 0.050964564085006714 + }, + { + "expert": 58, + "weight": 0.04497525840997696 + }, + { + "expert": 23, + "weight": 0.02731938287615776 + } + ], + "selected_weight_sum": 0.4973880760371685, + "top1_top2_margin": 0.09798136353492737 + }, + { + "position": 25, + "token_id": 398, + "token": "ãĢĤ", + "decoded_piece": "。", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.13403064012527466 + }, + { + "expert": 0, + "weight": 0.1253703534603119 + }, + { + "expert": 46, + "weight": 0.07578480988740921 + }, + { + "expert": 62, + "weight": 0.05519406497478485 + }, + { + "expert": 47, + "weight": 0.027481039986014366 + }, + { + "expert": 13, + "weight": 0.02377881109714508 + } + ], + "selected_weight_sum": 0.44163971953094006, + "top1_top2_margin": 0.008660286664962769 + } + ] + }, + { + "id": "en_architecture", + "label": "English architecture", + "tokens": 18, + "load": [ + 2, + 1, + 1, + 2, + 3, + 2, + 3, + 1, + 1, + 0, + 0, + 1, + 1, + 2, + 0, + 2, + 1, + 2, + 2, + 1, + 0, + 0, + 3, + 0, + 1, + 0, + 6, + 4, + 2, + 0, + 3, + 1, + 3, + 0, + 2, + 5, + 2, + 0, + 1, + 0, + 6, + 1, + 1, + 4, + 0, + 3, + 1, + 1, + 1, + 1, + 2, + 1, + 1, + 1, + 1, + 5, + 0, + 3, + 0, + 0, + 4, + 2, + 4, + 4 + ], + "metrics": { + "routes": 108, + "used_experts": 49, + "zero_experts": 15, + "mean": 1.6875, + "std": 1.5520083904266357, + "cv": 0.9197086691856384, + "gini": 0.48755785822868347, + "entropy_nats": 3.7096519470214844, + "effective_experts": 40.83958971219074, + "min": 0.0, + "p50": 1.0, + "p90": 4.0, + "max": 6.0 + }, + "top_experts": [ + { + "expert": 26, + "routes": 6 + }, + { + "expert": 40, + "routes": 6 + }, + { + "expert": 35, + "routes": 5 + }, + { + "expert": 55, + "routes": 5 + }, + { + "expert": 27, + "routes": 4 + }, + { + "expert": 43, + "routes": 4 + }, + { + "expert": 60, + "routes": 4 + }, + { + "expert": 62, + "routes": 4 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 3, + "weight": 0.7993167042732239 + }, + { + "expert": 62, + "weight": 0.018388060852885246 + }, + { + "expert": 38, + "weight": 0.010282253846526146 + }, + { + "expert": 63, + "weight": 0.009434774518013 + }, + { + "expert": 35, + "weight": 0.009029458276927471 + }, + { + "expert": 60, + "weight": 0.008924801833927631 + } + ], + "selected_weight_sum": 0.8553760536015034, + "top1_top2_margin": 0.7809286434203386 + }, + { + "position": 1, + "token_id": 99297, + "token": "Explain", + "decoded_piece": "Explain", + "experts_by_weight": [ + { + "expert": 63, + "weight": 0.1933971792459488 + }, + { + "expert": 42, + "weight": 0.1331486701965332 + }, + { + "expert": 57, + "weight": 0.07104431837797165 + }, + { + "expert": 35, + "weight": 0.06700316816568375 + }, + { + "expert": 61, + "weight": 0.061555638909339905 + }, + { + "expert": 60, + "weight": 0.047345709055662155 + } + ], + "selected_weight_sum": 0.5734946839511395, + "top1_top2_margin": 0.06024850904941559 + }, + { + "position": 2, + "token_id": 946, + "token": "Ġhow", + "decoded_piece": " how", + "experts_by_weight": [ + { + "expert": 5, + "weight": 0.11682333797216415 + }, + { + "expert": 53, + "weight": 0.09969403594732285 + }, + { + "expert": 16, + "weight": 0.07592637836933136 + }, + { + "expert": 51, + "weight": 0.05120022967457771 + }, + { + "expert": 34, + "weight": 0.03427540510892868 + }, + { + "expert": 3, + "weight": 0.029783086851239204 + } + ], + "selected_weight_sum": 0.40770247392356396, + "top1_top2_margin": 0.01712930202484131 + }, + { + "position": 3, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 35, + "weight": 0.24449819326400757 + }, + { + "expert": 62, + "weight": 0.07239027321338654 + }, + { + "expert": 7, + "weight": 0.05962592363357544 + }, + { + "expert": 18, + "weight": 0.05310329422354698 + }, + { + "expert": 24, + "weight": 0.04907490685582161 + }, + { + "expert": 17, + "weight": 0.047838661819696426 + } + ], + "selected_weight_sum": 0.5265312530100346, + "top1_top2_margin": 0.17210792005062103 + }, + { + "position": 4, + "token_id": 29640, + "token": "Ġcompressed", + "decoded_piece": " compressed", + "experts_by_weight": [ + { + "expert": 18, + "weight": 0.11223237961530685 + }, + { + "expert": 27, + "weight": 0.09656558930873871 + }, + { + "expert": 17, + "weight": 0.04231010004878044 + }, + { + "expert": 30, + "weight": 0.03794880956411362 + }, + { + "expert": 43, + "weight": 0.034158121794462204 + }, + { + "expert": 63, + "weight": 0.03112885169684887 + } + ], + "selected_weight_sum": 0.3543438520282507, + "top1_top2_margin": 0.015666790306568146 + }, + { + "position": 5, + "token_id": 2131, + "token": "Ġkey", + "decoded_piece": " key", + "experts_by_weight": [ + { + "expert": 30, + "weight": 0.17843440175056458 + }, + { + "expert": 26, + "weight": 0.11469165980815887 + }, + { + "expert": 50, + "weight": 0.06070755422115326 + }, + { + "expert": 40, + "weight": 0.05988939851522446 + }, + { + "expert": 36, + "weight": 0.05858447402715683 + }, + { + "expert": 6, + "weight": 0.05206485837697983 + } + ], + "selected_weight_sum": 0.5243723466992378, + "top1_top2_margin": 0.0637427419424057 + }, + { + "position": 6, + "token_id": 12, + "token": "-", + "decoded_piece": "-", + "experts_by_weight": [ + { + "expert": 0, + "weight": 0.15919552743434906 + }, + { + "expert": 19, + "weight": 0.10320810228586197 + }, + { + "expert": 35, + "weight": 0.09070276468992233 + }, + { + "expert": 62, + "weight": 0.042428430169820786 + }, + { + "expert": 4, + "weight": 0.04033466428518295 + }, + { + "expert": 60, + "weight": 0.03656792640686035 + } + ], + "selected_weight_sum": 0.47243741527199745, + "top1_top2_margin": 0.05598742514848709 + }, + { + "position": 7, + "token_id": 2834, + "token": "value", + "decoded_piece": "value", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.14338308572769165 + }, + { + "expert": 40, + "weight": 0.1320078819990158 + }, + { + "expert": 43, + "weight": 0.13028068840503693 + }, + { + "expert": 45, + "weight": 0.04761816933751106 + }, + { + "expert": 26, + "weight": 0.04696355387568474 + }, + { + "expert": 1, + "weight": 0.043221261352300644 + } + ], + "selected_weight_sum": 0.5434746406972408, + "top1_top2_margin": 0.011375203728675842 + }, + { + "position": 8, + "token_id": 28937, + "token": "Ġlatent", + "decoded_piece": " latent", + "experts_by_weight": [ + { + "expert": 55, + "weight": 0.11452669650316238 + }, + { + "expert": 6, + "weight": 0.10020633041858673 + }, + { + "expert": 40, + "weight": 0.058247704058885574 + }, + { + "expert": 22, + "weight": 0.05524905398488045 + }, + { + "expert": 13, + "weight": 0.038477592170238495 + }, + { + "expert": 26, + "weight": 0.0318424291908741 + } + ], + "selected_weight_sum": 0.39854980632662773, + "top1_top2_margin": 0.014320366084575653 + }, + { + "position": 9, + "token_id": 4183, + "token": "Ġchanges", + "decoded_piece": " changes", + "experts_by_weight": [ + { + "expert": 26, + "weight": 0.10610464960336685 + }, + { + "expert": 55, + "weight": 0.08787750452756882 + }, + { + "expert": 32, + "weight": 0.08229339122772217 + }, + { + "expert": 40, + "weight": 0.06280581653118134 + }, + { + "expert": 45, + "weight": 0.047021713107824326 + }, + { + "expert": 43, + "weight": 0.04179764166474342 + } + ], + "selected_weight_sum": 0.4279007166624069, + "top1_top2_margin": 0.018227145075798035 + }, + { + "position": 10, + "token_id": 254, + "token": "Ġthe", + "decoded_piece": " the", + "experts_by_weight": [ + { + "expert": 22, + "weight": 0.10047860443592072 + }, + { + "expert": 55, + "weight": 0.09288082271814346 + }, + { + "expert": 52, + "weight": 0.08753325045108795 + }, + { + "expert": 49, + "weight": 0.03979819640517235 + }, + { + "expert": 28, + "weight": 0.033622682094573975 + }, + { + "expert": 11, + "weight": 0.03130070120096207 + } + ], + "selected_weight_sum": 0.3856142573058605, + "top1_top2_margin": 0.007597781717777252 + }, + { + "position": 11, + "token_id": 4887, + "token": "Ġmemory", + "decoded_piece": " memory", + "experts_by_weight": [ + { + "expert": 27, + "weight": 0.10118205845355988 + }, + { + "expert": 45, + "weight": 0.08818523585796356 + }, + { + "expert": 26, + "weight": 0.06513181328773499 + }, + { + "expert": 36, + "weight": 0.03309033811092377 + }, + { + "expert": 57, + "weight": 0.03305121511220932 + }, + { + "expert": 61, + "weight": 0.02862332947552204 + } + ], + "selected_weight_sum": 0.34926399029791355, + "top1_top2_margin": 0.012996822595596313 + }, + { + "position": 12, + "token_id": 2295, + "token": "Ġcost", + "decoded_piece": " cost", + "experts_by_weight": [ + { + "expert": 40, + "weight": 0.2086726278066635 + }, + { + "expert": 55, + "weight": 0.1317712515592575 + }, + { + "expert": 32, + "weight": 0.08160752803087234 + }, + { + "expert": 12, + "weight": 0.06133318692445755 + }, + { + "expert": 5, + "weight": 0.04480573907494545 + }, + { + "expert": 2, + "weight": 0.040977973490953445 + } + ], + "selected_weight_sum": 0.5691683068871498, + "top1_top2_margin": 0.076901376247406 + }, + { + "position": 13, + "token_id": 280, + "token": "Ġof", + "decoded_piece": " of", + "experts_by_weight": [ + { + "expert": 15, + "weight": 0.14979112148284912 + }, + { + "expert": 48, + "weight": 0.11800755560398102 + }, + { + "expert": 22, + "weight": 0.1067730262875557 + }, + { + "expert": 30, + "weight": 0.0327712818980217 + }, + { + "expert": 34, + "weight": 0.02830866351723671 + }, + { + "expert": 13, + "weight": 0.02398701384663582 + } + ], + "selected_weight_sum": 0.45963866263628006, + "top1_top2_margin": 0.0317835658788681 + }, + { + "position": 14, + "token_id": 77939, + "token": "Ġautore", + "decoded_piece": " autore", + "experts_by_weight": [ + { + "expert": 27, + "weight": 0.10697006434202194 + }, + { + "expert": 63, + "weight": 0.05533495172858238 + }, + { + "expert": 57, + "weight": 0.05411042273044586 + }, + { + "expert": 60, + "weight": 0.03916148468852043 + }, + { + "expert": 31, + "weight": 0.03792927414178848 + }, + { + "expert": 35, + "weight": 0.03450075909495354 + } + ], + "selected_weight_sum": 0.32800695672631264, + "top1_top2_margin": 0.05163511261343956 + }, + { + "position": 15, + "token_id": 79184, + "token": "gressive", + "decoded_piece": "gressive", + "experts_by_weight": [ + { + "expert": 28, + "weight": 0.2267676740884781 + }, + { + "expert": 4, + "weight": 0.13144974410533905 + }, + { + "expert": 32, + "weight": 0.06287088990211487 + }, + { + "expert": 50, + "weight": 0.05779385194182396 + }, + { + "expert": 41, + "weight": 0.05445825681090355 + }, + { + "expert": 43, + "weight": 0.033917136490345 + } + ], + "selected_weight_sum": 0.5672575533390045, + "top1_top2_margin": 0.09531792998313904 + }, + { + "position": 16, + "token_id": 39671, + "token": "Ġdecoding", + "decoded_piece": " decoding", + "experts_by_weight": [ + { + "expert": 55, + "weight": 0.11397071182727814 + }, + { + "expert": 40, + "weight": 0.08910854160785675 + }, + { + "expert": 26, + "weight": 0.07447721809148788 + }, + { + "expert": 6, + "weight": 0.05076982080936432 + }, + { + "expert": 54, + "weight": 0.046305228024721146 + }, + { + "expert": 27, + "weight": 0.04167233407497406 + } + ], + "selected_weight_sum": 0.4163038544356823, + "top1_top2_margin": 0.024862170219421387 + }, + { + "position": 17, + "token_id": 13, + "token": ".", + "decoded_piece": ".", + "experts_by_weight": [ + { + "expert": 0, + "weight": 0.17542369663715363 + }, + { + "expert": 8, + "weight": 0.10671209543943405 + }, + { + "expert": 46, + "weight": 0.07208603620529175 + }, + { + "expert": 62, + "weight": 0.0618680976331234 + }, + { + "expert": 47, + "weight": 0.05220293253660202 + }, + { + "expert": 15, + "weight": 0.040070801973342896 + } + ], + "selected_weight_sum": 0.5083636604249477, + "top1_top2_margin": 0.06871160119771957 + } + ] + }, + { + "id": "code", + "label": "Python code", + "tokens": 19, + "load": [ + 3, + 0, + 0, + 5, + 2, + 1, + 0, + 2, + 1, + 3, + 0, + 2, + 2, + 1, + 0, + 3, + 2, + 4, + 3, + 1, + 3, + 0, + 2, + 1, + 0, + 1, + 4, + 0, + 1, + 1, + 5, + 0, + 2, + 2, + 2, + 3, + 2, + 2, + 1, + 3, + 2, + 2, + 1, + 4, + 1, + 4, + 2, + 3, + 2, + 1, + 0, + 1, + 4, + 0, + 0, + 3, + 1, + 2, + 1, + 1, + 3, + 1, + 3, + 2 + ], + "metrics": { + "routes": 114, + "used_experts": 52, + "zero_experts": 12, + "mean": 1.78125, + "std": 1.3269938230514526, + "cv": 0.7449789643287659, + "gini": 0.4098135828971863, + "entropy_nats": 3.8270628452301025, + "effective_experts": 45.9274439238954, + "min": 0.0, + "p50": 2.0, + "p90": 3.6999969482421875, + "max": 5.0 + }, + "top_experts": [ + { + "expert": 3, + "routes": 5 + }, + { + "expert": 30, + "routes": 5 + }, + { + "expert": 17, + "routes": 4 + }, + { + "expert": 26, + "routes": 4 + }, + { + "expert": 43, + "routes": 4 + }, + { + "expert": 45, + "routes": 4 + }, + { + "expert": 52, + "routes": 4 + }, + { + "expert": 0, + "routes": 3 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 3, + "weight": 0.7993167042732239 + }, + { + "expert": 62, + "weight": 0.018388060852885246 + }, + { + "expert": 38, + "weight": 0.010282253846526146 + }, + { + "expert": 63, + "weight": 0.009434774518013 + }, + { + "expert": 35, + "weight": 0.009029458276927471 + }, + { + "expert": 60, + "weight": 0.008924801833927631 + } + ], + "selected_weight_sum": 0.8553760536015034, + "top1_top2_margin": 0.7809286434203386 + }, + { + "position": 1, + "token_id": 9081, + "token": "Write", + "decoded_piece": "Write", + "experts_by_weight": [ + { + "expert": 63, + "weight": 0.22308281064033508 + }, + { + "expert": 42, + "weight": 0.10125719755887985 + }, + { + "expert": 39, + "weight": 0.07881859689950943 + }, + { + "expert": 57, + "weight": 0.07489147037267685 + }, + { + "expert": 35, + "weight": 0.06411762535572052 + }, + { + "expert": 60, + "weight": 0.03780627250671387 + } + ], + "selected_weight_sum": 0.5799739733338356, + "top1_top2_margin": 0.12182561308145523 + }, + { + "position": 2, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.11058355867862701 + }, + { + "expert": 62, + "weight": 0.09710631519556046 + }, + { + "expert": 17, + "weight": 0.09567485749721527 + }, + { + "expert": 55, + "weight": 0.07556460052728653 + }, + { + "expert": 51, + "weight": 0.040634334087371826 + }, + { + "expert": 35, + "weight": 0.03968478739261627 + } + ], + "selected_weight_sum": 0.45924845337867737, + "top1_top2_margin": 0.013477243483066559 + }, + { + "position": 3, + "token_id": 12974, + "token": "ĠPython", + "decoded_piece": " Python", + "experts_by_weight": [ + { + "expert": 43, + "weight": 0.3106133043766022 + }, + { + "expert": 60, + "weight": 0.10412774980068207 + }, + { + "expert": 18, + "weight": 0.07865095138549805 + }, + { + "expert": 20, + "weight": 0.03615188226103783 + }, + { + "expert": 4, + "weight": 0.03441538289189339 + }, + { + "expert": 30, + "weight": 0.03303976729512215 + } + ], + "selected_weight_sum": 0.5969990380108356, + "top1_top2_margin": 0.2064855545759201 + }, + { + "position": 4, + "token_id": 1157, + "token": "Ġfunction", + "decoded_piece": " function", + "experts_by_weight": [ + { + "expert": 12, + "weight": 0.2627643644809723 + }, + { + "expert": 36, + "weight": 0.20828509330749512 + }, + { + "expert": 30, + "weight": 0.07966912537813187 + }, + { + "expert": 43, + "weight": 0.04565059021115303 + }, + { + "expert": 11, + "weight": 0.04058252274990082 + }, + { + "expert": 25, + "weight": 0.032550934702157974 + } + ], + "selected_weight_sum": 0.6695026308298111, + "top1_top2_margin": 0.05447927117347717 + }, + { + "position": 5, + "token_id": 344, + "token": "Ġthat", + "decoded_piece": " that", + "experts_by_weight": [ + { + "expert": 34, + "weight": 0.18105126917362213 + }, + { + "expert": 15, + "weight": 0.12537729740142822 + }, + { + "expert": 9, + "weight": 0.06263353675603867 + }, + { + "expert": 22, + "weight": 0.04046915844082832 + }, + { + "expert": 57, + "weight": 0.03560767322778702 + }, + { + "expert": 47, + "weight": 0.035325679928064346 + } + ], + "selected_weight_sum": 0.4804646149277687, + "top1_top2_margin": 0.05567397177219391 + }, + { + "position": 6, + "token_id": 7578, + "token": "Ġreturns", + "decoded_piece": " returns", + "experts_by_weight": [ + { + "expert": 39, + "weight": 0.07301884144544601 + }, + { + "expert": 32, + "weight": 0.06769263744354248 + }, + { + "expert": 9, + "weight": 0.057351455092430115 + }, + { + "expert": 7, + "weight": 0.057337403297424316 + }, + { + "expert": 5, + "weight": 0.05577944219112396 + }, + { + "expert": 16, + "weight": 0.0453062430024147 + } + ], + "selected_weight_sum": 0.3564860224723816, + "top1_top2_margin": 0.005326204001903534 + }, + { + "position": 7, + "token_id": 254, + "token": "Ġthe", + "decoded_piece": " the", + "experts_by_weight": [ + { + "expert": 22, + "weight": 0.1945454478263855 + }, + { + "expert": 55, + "weight": 0.0892791748046875 + }, + { + "expert": 52, + "weight": 0.08212577551603317 + }, + { + "expert": 49, + "weight": 0.03960113599896431 + }, + { + "expert": 3, + "weight": 0.038635849952697754 + }, + { + "expert": 28, + "weight": 0.03487828001379967 + } + ], + "selected_weight_sum": 0.4790656641125679, + "top1_top2_margin": 0.105266273021698 + }, + { + "position": 8, + "token_id": 1022, + "token": "Ġfirst", + "decoded_piece": " first", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.14187313616275787 + }, + { + "expert": 18, + "weight": 0.05546113848686218 + }, + { + "expert": 52, + "weight": 0.038507647812366486 + }, + { + "expert": 3, + "weight": 0.03536592423915863 + }, + { + "expert": 40, + "weight": 0.030138704925775528 + }, + { + "expert": 58, + "weight": 0.029435299336910248 + } + ], + "selected_weight_sum": 0.33078185096383095, + "top1_top2_margin": 0.08641199767589569 + }, + { + "position": 9, + "token_id": 9842, + "token": "Ġrepeated", + "decoded_piece": " repeated", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.1300642341375351 + }, + { + "expert": 45, + "weight": 0.09043307602405548 + }, + { + "expert": 56, + "weight": 0.07752960920333862 + }, + { + "expert": 26, + "weight": 0.06140988692641258 + }, + { + "expert": 29, + "weight": 0.03912041708827019 + }, + { + "expert": 30, + "weight": 0.03831544145941734 + } + ], + "selected_weight_sum": 0.4368726648390293, + "top1_top2_margin": 0.039631158113479614 + }, + { + "position": 10, + "token_id": 3699, + "token": "Ġelement", + "decoded_piece": " element", + "experts_by_weight": [ + { + "expert": 26, + "weight": 0.13783960044384003 + }, + { + "expert": 45, + "weight": 0.12804515659809113 + }, + { + "expert": 40, + "weight": 0.09556166082620621 + }, + { + "expert": 30, + "weight": 0.09431081265211105 + }, + { + "expert": 43, + "weight": 0.06392165273427963 + }, + { + "expert": 20, + "weight": 0.04370862618088722 + } + ], + "selected_weight_sum": 0.5633875094354153, + "top1_top2_margin": 0.009794443845748901 + }, + { + "position": 11, + "token_id": 279, + "token": "Ġin", + "decoded_piece": " in", + "experts_by_weight": [ + { + "expert": 9, + "weight": 0.12035012990236282 + }, + { + "expert": 15, + "weight": 0.11592075228691101 + }, + { + "expert": 34, + "weight": 0.05035826563835144 + }, + { + "expert": 11, + "weight": 0.04238616302609444 + }, + { + "expert": 48, + "weight": 0.030062539502978325 + }, + { + "expert": 41, + "weight": 0.029798714444041252 + } + ], + "selected_weight_sum": 0.3888765648007393, + "top1_top2_margin": 0.004429377615451813 + }, + { + "position": 12, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 0, + "weight": 0.1413310319185257 + }, + { + "expert": 17, + "weight": 0.06695874035358429 + }, + { + "expert": 37, + "weight": 0.06325921416282654 + }, + { + "expert": 3, + "weight": 0.035412270575761795 + }, + { + "expert": 13, + "weight": 0.02976227179169655 + }, + { + "expert": 41, + "weight": 0.029165809974074364 + } + ], + "selected_weight_sum": 0.36588933877646923, + "top1_top2_margin": 0.0743722915649414 + }, + { + "position": 13, + "token_id": 1525, + "token": "Ġlist", + "decoded_piece": " list", + "experts_by_weight": [ + { + "expert": 30, + "weight": 0.1110549345612526 + }, + { + "expert": 37, + "weight": 0.09031160920858383 + }, + { + "expert": 18, + "weight": 0.07663130760192871 + }, + { + "expert": 26, + "weight": 0.06359056383371353 + }, + { + "expert": 43, + "weight": 0.049891188740730286 + }, + { + "expert": 45, + "weight": 0.04514267295598984 + } + ], + "selected_weight_sum": 0.4366222769021988, + "top1_top2_margin": 0.020743325352668762 + }, + { + "position": 14, + "token_id": 285, + "token": "Ġand", + "decoded_piece": " and", + "experts_by_weight": [ + { + "expert": 0, + "weight": 0.11731312423944473 + }, + { + "expert": 46, + "weight": 0.07047951221466064 + }, + { + "expert": 15, + "weight": 0.05180724337697029 + }, + { + "expert": 48, + "weight": 0.044003911316394806 + }, + { + "expert": 47, + "weight": 0.03136441856622696 + }, + { + "expert": 23, + "weight": 0.03092145547270775 + } + ], + "selected_weight_sum": 0.3458896651864052, + "top1_top2_margin": 0.04683361202478409 + }, + { + "position": 15, + "token_id": 6716, + "token": "Ġexplain", + "decoded_piece": " explain", + "experts_by_weight": [ + { + "expert": 44, + "weight": 0.15685030817985535 + }, + { + "expert": 39, + "weight": 0.11473170667886734 + }, + { + "expert": 59, + "weight": 0.07455103099346161 + }, + { + "expert": 61, + "weight": 0.0551246777176857 + }, + { + "expert": 16, + "weight": 0.034474410116672516 + }, + { + "expert": 32, + "weight": 0.028114324435591698 + } + ], + "selected_weight_sum": 0.4638464581221342, + "top1_top2_margin": 0.04211860150098801 + }, + { + "position": 16, + "token_id": 895, + "token": "Ġits", + "decoded_piece": " its", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.1131797656416893 + }, + { + "expert": 20, + "weight": 0.10879005491733551 + }, + { + "expert": 33, + "weight": 0.061025477945804596 + }, + { + "expert": 55, + "weight": 0.057621393352746964 + }, + { + "expert": 7, + "weight": 0.03110535442829132 + }, + { + "expert": 3, + "weight": 0.02715742588043213 + } + ], + "selected_weight_sum": 0.3988794721662998, + "top1_top2_margin": 0.00438971072435379 + }, + { + "position": 17, + "token_id": 13913, + "token": "Ġcomplexity", + "decoded_piece": " complexity", + "experts_by_weight": [ + { + "expert": 33, + "weight": 0.08646200597286224 + }, + { + "expert": 26, + "weight": 0.08283549547195435 + }, + { + "expert": 45, + "weight": 0.06919881701469421 + }, + { + "expert": 52, + "weight": 0.060465697199106216 + }, + { + "expert": 36, + "weight": 0.05625320225954056 + }, + { + "expert": 12, + "weight": 0.04574577510356903 + } + ], + "selected_weight_sum": 0.4009609930217266, + "top1_top2_margin": 0.003626510500907898 + }, + { + "position": 18, + "token_id": 13, + "token": ".", + "decoded_piece": ".", + "experts_by_weight": [ + { + "expert": 0, + "weight": 0.19963113963603973 + }, + { + "expert": 8, + "weight": 0.1043989434838295 + }, + { + "expert": 46, + "weight": 0.07004683464765549 + }, + { + "expert": 62, + "weight": 0.04874078184366226 + }, + { + "expert": 47, + "weight": 0.03005041368305683 + }, + { + "expert": 19, + "weight": 0.02090730145573616 + } + ], + "selected_weight_sum": 0.47377541474998, + "top1_top2_margin": 0.09523219615221024 + } + ] + }, + { + "id": "math", + "label": "数学推理", + "tokens": 27, + "load": [ + 5, + 1, + 0, + 2, + 0, + 3, + 2, + 2, + 4, + 4, + 1, + 0, + 2, + 2, + 2, + 3, + 5, + 2, + 3, + 1, + 3, + 0, + 4, + 2, + 3, + 1, + 0, + 4, + 0, + 0, + 2, + 1, + 5, + 3, + 7, + 6, + 2, + 0, + 4, + 1, + 2, + 0, + 3, + 0, + 1, + 1, + 3, + 9, + 2, + 3, + 2, + 1, + 6, + 1, + 0, + 3, + 2, + 4, + 5, + 1, + 7, + 2, + 10, + 2 + ], + "metrics": { + "routes": 162, + "used_experts": 53, + "zero_experts": 11, + "mean": 2.53125, + "std": 2.1892142295837402, + "cv": 0.8648747801780701, + "gini": 0.4535107910633087, + "entropy_nats": 3.7780678272247314, + "effective_experts": 43.73146330899206, + "min": 0.0, + "p50": 2.0, + "p90": 5.0, + "max": 10.0 + }, + "top_experts": [ + { + "expert": 62, + "routes": 10 + }, + { + "expert": 47, + "routes": 9 + }, + { + "expert": 34, + "routes": 7 + }, + { + "expert": 60, + "routes": 7 + }, + { + "expert": 35, + "routes": 6 + }, + { + "expert": 52, + "routes": 6 + }, + { + "expert": 0, + "routes": 5 + }, + { + "expert": 16, + "routes": 5 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 3, + "weight": 0.7993167042732239 + }, + { + "expert": 62, + "weight": 0.018388060852885246 + }, + { + "expert": 38, + "weight": 0.010282253846526146 + }, + { + "expert": 63, + "weight": 0.009434774518013 + }, + { + "expert": 35, + "weight": 0.009029458276927471 + }, + { + "expert": 60, + "weight": 0.008924801833927631 + } + ], + "selected_weight_sum": 0.8553760536015034, + "top1_top2_margin": 0.7809286434203386 + }, + { + "position": 1, + "token_id": 8246, + "token": "èĭ¥", + "decoded_piece": "若", + "experts_by_weight": [ + { + "expert": 63, + "weight": 0.13331177830696106 + }, + { + "expert": 14, + "weight": 0.12700548768043518 + }, + { + "expert": 42, + "weight": 0.10285352915525436 + }, + { + "expert": 61, + "weight": 0.07681424915790558 + }, + { + "expert": 35, + "weight": 0.06962653994560242 + }, + { + "expert": 34, + "weight": 0.06532615423202515 + } + ], + "selected_weight_sum": 0.5749377384781837, + "top1_top2_margin": 0.006306290626525879 + }, + { + "position": 2, + "token_id": 2206, + "token": "æŃ£", + "decoded_piece": "正", + "experts_by_weight": [ + { + "expert": 16, + "weight": 0.2814550995826721 + }, + { + "expert": 51, + "weight": 0.08446836471557617 + }, + { + "expert": 24, + "weight": 0.07371900975704193 + }, + { + "expert": 7, + "weight": 0.06035252660512924 + }, + { + "expert": 18, + "weight": 0.05827755853533745 + }, + { + "expert": 31, + "weight": 0.02826753258705139 + } + ], + "selected_weight_sum": 0.5865400917828083, + "top1_top2_margin": 0.19698673486709595 + }, + { + "position": 3, + "token_id": 2266, + "token": "æķ°", + "decoded_piece": "数", + "experts_by_weight": [ + { + "expert": 49, + "weight": 0.25039201974868774 + }, + { + "expert": 24, + "weight": 0.22088326513767242 + }, + { + "expert": 58, + "weight": 0.20361556112766266 + }, + { + "expert": 32, + "weight": 0.02990611642599106 + }, + { + "expert": 23, + "weight": 0.019317051395773888 + }, + { + "expert": 56, + "weight": 0.018254980444908142 + } + ], + "selected_weight_sum": 0.7423689942806959, + "top1_top2_margin": 0.02950875461101532 + }, + { + "position": 4, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 35, + "weight": 0.3537174165248871 + }, + { + "expert": 62, + "weight": 0.08283980935811996 + }, + { + "expert": 34, + "weight": 0.0720146968960762 + }, + { + "expert": 60, + "weight": 0.0697723850607872 + }, + { + "expert": 47, + "weight": 0.052806876599788666 + }, + { + "expert": 57, + "weight": 0.04913831129670143 + } + ], + "selected_weight_sum": 0.6802894957363605, + "top1_top2_margin": 0.2708776071667671 + }, + { + "position": 5, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 62, + "weight": 0.1735713630914688 + }, + { + "expert": 47, + "weight": 0.07046853005886078 + }, + { + "expert": 34, + "weight": 0.037892214953899384 + }, + { + "expert": 19, + "weight": 0.028465814888477325 + }, + { + "expert": 38, + "weight": 0.02768941968679428 + }, + { + "expert": 60, + "weight": 0.0260683111846447 + } + ], + "selected_weight_sum": 0.3641556538641453, + "top1_top2_margin": 0.10310283303260803 + }, + { + "position": 6, + "token_id": 13211, + "token": "满足", + "decoded_piece": "满足", + "experts_by_weight": [ + { + "expert": 34, + "weight": 0.2298489511013031 + }, + { + "expert": 47, + "weight": 0.0791824534535408 + }, + { + "expert": 9, + "weight": 0.05738123506307602 + }, + { + "expert": 57, + "weight": 0.055221591144800186 + }, + { + "expert": 5, + "weight": 0.04613054171204567 + }, + { + "expert": 42, + "weight": 0.037003934383392334 + } + ], + "selected_weight_sum": 0.5047687068581581, + "top1_top2_margin": 0.1506664976477623 + }, + { + "position": 7, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 35, + "weight": 0.296394020318985 + }, + { + "expert": 60, + "weight": 0.08370600640773773 + }, + { + "expert": 27, + "weight": 0.037053510546684265 + }, + { + "expert": 57, + "weight": 0.03541581705212593 + }, + { + "expert": 62, + "weight": 0.02703707478940487 + }, + { + "expert": 58, + "weight": 0.02383062243461609 + } + ], + "selected_weight_sum": 0.5034370515495539, + "top1_top2_margin": 0.21268801391124725 + }, + { + "position": 8, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 62, + "weight": 0.16342000663280487 + }, + { + "expert": 47, + "weight": 0.08331643044948578 + }, + { + "expert": 0, + "weight": 0.04279105365276337 + }, + { + "expert": 8, + "weight": 0.037890754640102386 + }, + { + "expert": 38, + "weight": 0.03473474085330963 + }, + { + "expert": 27, + "weight": 0.02922108583152294 + } + ], + "selected_weight_sum": 0.391374072059989, + "top1_top2_margin": 0.08010357618331909 + }, + { + "position": 9, + "token_id": 14855, + "token": "åĬłä¸Ĭ", + "decoded_piece": "加上", + "experts_by_weight": [ + { + "expert": 34, + "weight": 0.318281352519989 + }, + { + "expert": 9, + "weight": 0.04051169380545616 + }, + { + "expert": 15, + "weight": 0.040131304413080215 + }, + { + "expert": 5, + "weight": 0.03766006976366043 + }, + { + "expert": 57, + "weight": 0.03394027799367905 + }, + { + "expert": 47, + "weight": 0.02833012118935585 + } + ], + "selected_weight_sum": 0.4988548196852207, + "top1_top2_margin": 0.27776965871453285 + }, + { + "position": 10, + "token_id": 12985, + "token": "å®ĥçļĦ", + "decoded_piece": "它的", + "experts_by_weight": [ + { + "expert": 20, + "weight": 0.10639920830726624 + }, + { + "expert": 17, + "weight": 0.08816708624362946 + }, + { + "expert": 52, + "weight": 0.08003807067871094 + }, + { + "expert": 3, + "weight": 0.05453759431838989 + }, + { + "expert": 23, + "weight": 0.046933963894844055 + }, + { + "expert": 32, + "weight": 0.04441061615943909 + } + ], + "selected_weight_sum": 0.42048653960227966, + "top1_top2_margin": 0.01823212206363678 + }, + { + "position": 11, + "token_id": 8270, + "token": "åĢĴ", + "decoded_piece": "倒", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.16316896677017212 + }, + { + "expert": 22, + "weight": 0.060158684849739075 + }, + { + "expert": 33, + "weight": 0.05992881581187248 + }, + { + "expert": 56, + "weight": 0.037876423448324203 + }, + { + "expert": 32, + "weight": 0.03393488749861717 + }, + { + "expert": 36, + "weight": 0.03238135948777199 + } + ], + "selected_weight_sum": 0.38744913786649704, + "top1_top2_margin": 0.10301028192043304 + }, + { + "position": 12, + "token_id": 2266, + "token": "æķ°", + "decoded_piece": "数", + "experts_by_weight": [ + { + "expert": 49, + "weight": 0.18513305485248566 + }, + { + "expert": 58, + "weight": 0.1456661969423294 + }, + { + "expert": 50, + "weight": 0.11328830569982529 + }, + { + "expert": 16, + "weight": 0.05659753084182739 + }, + { + "expert": 30, + "weight": 0.04169565439224243 + }, + { + "expert": 32, + "weight": 0.03690284490585327 + } + ], + "selected_weight_sum": 0.5792835876345634, + "top1_top2_margin": 0.03946685791015625 + }, + { + "position": 13, + "token_id": 37360, + "token": "çŃīäºİ", + "decoded_piece": "等于", + "experts_by_weight": [ + { + "expert": 34, + "weight": 0.2411312758922577 + }, + { + "expert": 5, + "weight": 0.10274491459131241 + }, + { + "expert": 55, + "weight": 0.05373065173625946 + }, + { + "expert": 22, + "weight": 0.03419595956802368 + }, + { + "expert": 25, + "weight": 0.02804616279900074 + }, + { + "expert": 6, + "weight": 0.0268552228808403 + } + ], + "selected_weight_sum": 0.4867041874676943, + "top1_top2_margin": 0.13838636130094528 + }, + { + "position": 14, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 62, + "weight": 0.16179583966732025 + }, + { + "expert": 46, + "weight": 0.07096794247627258 + }, + { + "expert": 60, + "weight": 0.06854509562253952 + }, + { + "expert": 8, + "weight": 0.04157784581184387 + }, + { + "expert": 0, + "weight": 0.03384833782911301 + }, + { + "expert": 13, + "weight": 0.0296489167958498 + } + ], + "selected_weight_sum": 0.40638397820293903, + "top1_top2_margin": 0.09082789719104767 + }, + { + "position": 15, + "token_id": 18, + "token": "3", + "decoded_piece": "3", + "experts_by_weight": [ + { + "expert": 27, + "weight": 0.16428224742412567 + }, + { + "expert": 58, + "weight": 0.10361717641353607 + }, + { + "expert": 35, + "weight": 0.08202363550662994 + }, + { + "expert": 18, + "weight": 0.04421326518058777 + }, + { + "expert": 60, + "weight": 0.02995552122592926 + }, + { + "expert": 24, + "weight": 0.02580142579972744 + } + ], + "selected_weight_sum": 0.44989327155053616, + "top1_top2_margin": 0.0606650710105896 + }, + { + "position": 16, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.13362829387187958 + }, + { + "expert": 0, + "weight": 0.093042753636837 + }, + { + "expert": 38, + "weight": 0.07785298675298691 + }, + { + "expert": 47, + "weight": 0.071451835334301 + }, + { + "expert": 46, + "weight": 0.06596096605062485 + }, + { + "expert": 62, + "weight": 0.029706455767154694 + } + ], + "selected_weight_sum": 0.471643291413784, + "top1_top2_margin": 0.04058554023504257 + }, + { + "position": 17, + "token_id": 2989, + "token": "æ±Ĥ", + "decoded_piece": "求", + "experts_by_weight": [ + { + "expert": 61, + "weight": 0.23183642327785492 + }, + { + "expert": 59, + "weight": 0.16105559468269348 + }, + { + "expert": 42, + "weight": 0.05859864875674248 + }, + { + "expert": 9, + "weight": 0.04714159667491913 + }, + { + "expert": 7, + "weight": 0.03799108788371086 + }, + { + "expert": 16, + "weight": 0.03278881683945656 + } + ], + "selected_weight_sum": 0.5694121681153774, + "top1_top2_margin": 0.07078082859516144 + }, + { + "position": 18, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 35, + "weight": 0.2791516184806824 + }, + { + "expert": 60, + "weight": 0.08195051550865173 + }, + { + "expert": 27, + "weight": 0.03374714404344559 + }, + { + "expert": 53, + "weight": 0.03366595879197121 + }, + { + "expert": 62, + "weight": 0.026014486327767372 + }, + { + "expert": 20, + "weight": 0.02444468066096306 + } + ], + "selected_weight_sum": 0.47897440381348133, + "top1_top2_margin": 0.19720110297203064 + }, + { + "position": 19, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 62, + "weight": 0.1385689079761505 + }, + { + "expert": 47, + "weight": 0.04262688755989075 + }, + { + "expert": 0, + "weight": 0.041786808520555496 + }, + { + "expert": 48, + "weight": 0.03194677457213402 + }, + { + "expert": 10, + "weight": 0.028856562450528145 + }, + { + "expert": 13, + "weight": 0.02796977385878563 + } + ], + "selected_weight_sum": 0.31175571493804455, + "top1_top2_margin": 0.09594202041625977 + }, + { + "position": 20, + "token_id": 337, + "token": "çļĦ", + "decoded_piece": "的", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.13563965260982513 + }, + { + "expert": 47, + "weight": 0.057844046503305435 + }, + { + "expert": 22, + "weight": 0.05458466708660126 + }, + { + "expert": 48, + "weight": 0.04892478883266449 + }, + { + "expert": 14, + "weight": 0.03986985981464386 + }, + { + "expert": 20, + "weight": 0.031419653445482254 + } + ], + "selected_weight_sum": 0.36828266829252243, + "top1_top2_margin": 0.0777956061065197 + }, + { + "position": 21, + "token_id": 18718, + "token": "å¹³æĸ¹", + "decoded_piece": "平方", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.3164118230342865 + }, + { + "expert": 33, + "weight": 0.04724707081913948 + }, + { + "expert": 12, + "weight": 0.044920627027750015 + }, + { + "expert": 36, + "weight": 0.029304083436727524 + }, + { + "expert": 45, + "weight": 0.025028377771377563 + }, + { + "expert": 39, + "weight": 0.024010613560676575 + } + ], + "selected_weight_sum": 0.48692259564995766, + "top1_top2_margin": 0.269164752215147 + }, + { + "position": 22, + "token_id": 14855, + "token": "åĬłä¸Ĭ", + "decoded_piece": "加上", + "experts_by_weight": [ + { + "expert": 34, + "weight": 0.20498308539390564 + }, + { + "expert": 16, + "weight": 0.06528150290250778 + }, + { + "expert": 9, + "weight": 0.046608418226242065 + }, + { + "expert": 40, + "weight": 0.04563548043370247 + }, + { + "expert": 15, + "weight": 0.04310647398233414 + }, + { + "expert": 55, + "weight": 0.03512641042470932 + } + ], + "selected_weight_sum": 0.4407413713634014, + "top1_top2_margin": 0.13970158249139786 + }, + { + "position": 23, + "token_id": 8270, + "token": "åĢĴ", + "decoded_piece": "倒", + "experts_by_weight": [ + { + "expert": 1, + "weight": 0.07203083485364914 + }, + { + "expert": 32, + "weight": 0.07109834998846054 + }, + { + "expert": 18, + "weight": 0.0562785267829895 + }, + { + "expert": 22, + "weight": 0.05611051246523857 + }, + { + "expert": 44, + "weight": 0.042118486016988754 + }, + { + "expert": 17, + "weight": 0.04034203663468361 + } + ], + "selected_weight_sum": 0.3379787467420101, + "top1_top2_margin": 0.0009324848651885986 + }, + { + "position": 24, + "token_id": 51971, + "token": "æķ°çļĦ", + "decoded_piece": "数的", + "experts_by_weight": [ + { + "expert": 49, + "weight": 0.19896872341632843 + }, + { + "expert": 58, + "weight": 0.11013466864824295 + }, + { + "expert": 50, + "weight": 0.09039244800806046 + }, + { + "expert": 16, + "weight": 0.05881365388631821 + }, + { + "expert": 30, + "weight": 0.05270054563879967 + }, + { + "expert": 52, + "weight": 0.05187920853495598 + } + ], + "selected_weight_sum": 0.5628892481327057, + "top1_top2_margin": 0.08883405476808548 + }, + { + "position": 25, + "token_id": 18718, + "token": "å¹³æĸ¹", + "decoded_piece": "平方", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.2541465163230896 + }, + { + "expert": 55, + "weight": 0.06222054362297058 + }, + { + "expert": 6, + "weight": 0.048966165632009506 + }, + { + "expert": 40, + "weight": 0.03323732689023018 + }, + { + "expert": 12, + "weight": 0.031393036246299744 + }, + { + "expert": 33, + "weight": 0.02776632457971573 + } + ], + "selected_weight_sum": 0.45772991329431534, + "top1_top2_margin": 0.19192597270011902 + }, + { + "position": 26, + "token_id": 398, + "token": "ãĢĤ", + "decoded_piece": "。", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.14508213102817535 + }, + { + "expert": 0, + "weight": 0.12345562130212784 + }, + { + "expert": 46, + "weight": 0.06189163029193878 + }, + { + "expert": 62, + "weight": 0.056270845234394073 + }, + { + "expert": 47, + "weight": 0.03983943909406662 + }, + { + "expert": 15, + "weight": 0.02559591457247734 + } + ], + "selected_weight_sum": 0.45213558152318, + "top1_top2_margin": 0.021626509726047516 + } + ] + } + ], + "prompt_pair_jaccard": [ + { + "left": "zh_explanation", + "right": "en_architecture", + "used_expert_jaccard": 0.7096774193548387, + "shared_experts": 44, + "union_experts": 62 + }, + { + "left": "zh_explanation", + "right": "code", + "used_expert_jaccard": 0.8166666666666667, + "shared_experts": 49, + "union_experts": 60 + }, + { + "left": "zh_explanation", + "right": "math", + "used_expert_jaccard": 0.7741935483870968, + "shared_experts": 48, + "union_experts": 62 + }, + { + "left": "en_architecture", + "right": "code", + "used_expert_jaccard": 0.6557377049180327, + "shared_experts": 40, + "union_experts": 61 + }, + { + "left": "en_architecture", + "right": "math", + "used_expert_jaccard": 0.6721311475409836, + "shared_experts": 41, + "union_experts": 61 + }, + { + "left": "code", + "right": "math", + "used_expert_jaccard": 0.7213114754098361, + "shared_experts": 44, + "union_experts": 61 + } + ] + } + }, + { + "layer": 3, + "ffn": "moe", + "state_tensors": 203, + "state_numel": 584847872, + "state_bytes": 1169695744, + "timing_ms": { + "load_cpu": 25.33966102055274, + "transfer_to_device": 121.24854500871152, + "forward": 3.839854005491361 + }, + "peak_device_allocated_mib": 1174.7490234375, + "hidden_before": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 0.22556711733341217, + "p10": 0.10850249230861664, + "p50": 0.1559184491634369, + "p90": 0.19672976434230804 + }, + "l2_norm": { + "mean": 10.208003044128418, + "p10": 4.910262107849121, + "p50": 7.056063175201416, + "p90": 8.902973175048828 + }, + "finite": true, + "sha256_fp32": "76c1ba59ea5ef19f16162513c7c36e6e8d17a9af9fa3fdedbf657e80882fa0a3" + }, + "hidden_after": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 2.651595115661621, + "p10": 0.14781513810157776, + "p50": 0.1792253851890564, + "p90": 0.22313819825649261 + }, + "l2_norm": { + "mean": 119.99750518798828, + "p10": 6.68934965133667, + "p50": 8.110815048217773, + "p90": 10.098081588745117 + }, + "finite": true, + "sha256_fp32": "aa223453a9b51ecdb244412002e27c70edbcaeeb29a6e1e064f340622ff4044c" + }, + "mla": { + "compressed_projection_shape": [ + 4, + 27, + 576 + ], + "latent_shape": [ + 4, + 27, + 512 + ], + "rope_key_shape": [ + 4, + 27, + 64 + ], + "latent_rms": 4.113536357879639, + "rope_key_rms": 2.1909468173980713, + "eager_key_cache_shape": [ + 4, + 16, + 27, + 192 + ], + "eager_value_cache_shape": [ + 4, + 16, + 27, + 128 + ], + "eager_cache_bytes": 1105920 + }, + "routing": { + "aggregate_load": [ + 14, + 5, + 8, + 9, + 17, + 15, + 16, + 6, + 2, + 10, + 2, + 9, + 16, + 3, + 9, + 3, + 14, + 1, + 7, + 9, + 16, + 5, + 8, + 0, + 8, + 9, + 16, + 8, + 8, + 12, + 6, + 4, + 3, + 16, + 0, + 8, + 12, + 15, + 1, + 16, + 7, + 5, + 4, + 12, + 7, + 5, + 11, + 7, + 11, + 12, + 2, + 6, + 3, + 6, + 19, + 9, + 8, + 3, + 10, + 6, + 11, + 12, + 13, + 5 + ], + "aggregate_metrics": { + "routes": 540, + "used_experts": 62, + "zero_experts": 2, + "mean": 8.4375, + "std": 4.797072410583496, + "cv": 0.568541944026947, + "gini": 0.3218750059604645, + "entropy_nats": 3.979837417602539, + "effective_experts": 53.50833400703197, + "min": 0.0, + "p50": 8.0, + "p90": 16.0, + "max": 19.0 + }, + "top_experts": [ + { + "expert": 54, + "routes": 19 + }, + { + "expert": 4, + "routes": 17 + }, + { + "expert": 6, + "routes": 16 + }, + { + "expert": 12, + "routes": 16 + }, + { + "expert": 20, + "routes": 16 + }, + { + "expert": 26, + "routes": 16 + }, + { + "expert": 33, + "routes": 16 + }, + { + "expert": 39, + "routes": 16 + }, + { + "expert": 5, + "routes": 15 + }, + { + "expert": 37, + "routes": 15 + }, + { + "expert": 0, + "routes": 14 + }, + { + "expert": 16, + "routes": 14 + } + ], + "selected_weight_sum": { + "mean": 0.4898033142089844, + "p10": 0.37064844369888306, + "p50": 0.46657082438468933, + "p90": 0.5860897302627563 + }, + "top1_top2_margin": { + "mean": 0.10427366942167282, + "p50": 0.036837637424468994, + "p90": 0.19226056337356567 + }, + "per_prompt": [ + { + "id": "zh_explanation", + "label": "中文解释", + "tokens": 26, + "load": [ + 6, + 0, + 4, + 1, + 5, + 6, + 9, + 3, + 1, + 3, + 2, + 2, + 5, + 1, + 2, + 0, + 5, + 0, + 2, + 5, + 5, + 1, + 3, + 0, + 2, + 1, + 5, + 1, + 3, + 3, + 3, + 2, + 2, + 5, + 0, + 2, + 2, + 6, + 0, + 1, + 1, + 3, + 2, + 2, + 4, + 1, + 2, + 0, + 1, + 2, + 0, + 1, + 3, + 2, + 3, + 3, + 4, + 1, + 2, + 3, + 2, + 2, + 2, + 1 + ], + "metrics": { + "routes": 156, + "used_experts": 56, + "zero_experts": 8, + "mean": 2.4375, + "std": 1.8505682945251465, + "cv": 0.7592074871063232, + "gini": 0.40444710850715637, + "entropy_nats": 3.8562657833099365, + "effective_experts": 47.28843598733856, + "min": 0.0, + "p50": 2.0, + "p90": 5.0, + "max": 9.0 + }, + "top_experts": [ + { + "expert": 6, + "routes": 9 + }, + { + "expert": 0, + "routes": 6 + }, + { + "expert": 5, + "routes": 6 + }, + { + "expert": 37, + "routes": 6 + }, + { + "expert": 4, + "routes": 5 + }, + { + "expert": 12, + "routes": 5 + }, + { + "expert": 16, + "routes": 5 + }, + { + "expert": 19, + "routes": 5 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 54, + "weight": 0.9711988568305969 + }, + { + "expert": 11, + "weight": 0.0014021513052284718 + }, + { + "expert": 9, + "weight": 0.001076161745004356 + }, + { + "expert": 4, + "weight": 0.0008088500471785665 + }, + { + "expert": 5, + "weight": 0.0007386832148768008 + }, + { + "expert": 12, + "weight": 0.000706107763107866 + } + ], + "selected_weight_sum": 0.975930810905993, + "top1_top2_margin": 0.9697967055253685 + }, + { + "position": 1, + "token_id": 1082, + "token": "ç͍", + "decoded_piece": "用", + "experts_by_weight": [ + { + "expert": 6, + "weight": 0.16575224697589874 + }, + { + "expert": 13, + "weight": 0.05748995393514633 + }, + { + "expert": 58, + "weight": 0.04968932643532753 + }, + { + "expert": 14, + "weight": 0.03696934133768082 + }, + { + "expert": 4, + "weight": 0.03696615993976593 + }, + { + "expert": 5, + "weight": 0.03662595525383949 + } + ], + "selected_weight_sum": 0.38349298387765884, + "top1_top2_margin": 0.10826229304075241 + }, + { + "position": 2, + "token_id": 83983, + "token": "éĢļä¿Ĺ", + "decoded_piece": "通俗", + "experts_by_weight": [ + { + "expert": 0, + "weight": 0.09529686719179153 + }, + { + "expert": 33, + "weight": 0.07287721335887909 + }, + { + "expert": 28, + "weight": 0.05606567859649658 + }, + { + "expert": 37, + "weight": 0.05045276880264282 + }, + { + "expert": 35, + "weight": 0.03255721554160118 + }, + { + "expert": 54, + "weight": 0.030352246016263962 + } + ], + "selected_weight_sum": 0.33760198950767517, + "top1_top2_margin": 0.022419653832912445 + }, + { + "position": 3, + "token_id": 91476, + "token": "çļĦè¯Ńè¨Ģ", + "decoded_piece": "的语言", + "experts_by_weight": [ + { + "expert": 37, + "weight": 0.1348549872636795 + }, + { + "expert": 28, + "weight": 0.1058119535446167 + }, + { + "expert": 0, + "weight": 0.06803137063980103 + }, + { + "expert": 62, + "weight": 0.056730255484580994 + }, + { + "expert": 20, + "weight": 0.056706808507442474 + }, + { + "expert": 33, + "weight": 0.036494504660367966 + } + ], + "selected_weight_sum": 0.45862988010048866, + "top1_top2_margin": 0.029043033719062805 + }, + { + "position": 4, + "token_id": 17882, + "token": "è§£éĩĬ", + "decoded_piece": "解释", + "experts_by_weight": [ + { + "expert": 37, + "weight": 0.23425081372261047 + }, + { + "expert": 63, + "weight": 0.11718924343585968 + }, + { + "expert": 39, + "weight": 0.053890570998191833 + }, + { + "expert": 0, + "weight": 0.04609973728656769 + }, + { + "expert": 61, + "weight": 0.03850342333316803 + }, + { + "expert": 56, + "weight": 0.033879783004522324 + } + ], + "selected_weight_sum": 0.52381357178092, + "top1_top2_margin": 0.1170615702867508 + }, + { + "position": 5, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.19864459335803986 + }, + { + "expert": 22, + "weight": 0.15572796761989594 + }, + { + "expert": 24, + "weight": 0.0822681188583374 + }, + { + "expert": 9, + "weight": 0.0573173351585865 + }, + { + "expert": 55, + "weight": 0.056325022131204605 + }, + { + "expert": 5, + "weight": 0.03425804525613785 + } + ], + "selected_weight_sum": 0.5845410823822021, + "top1_top2_margin": 0.04291662573814392 + }, + { + "position": 6, + "token_id": 10285, + "token": "为ä»Ģä¹Ī", + "decoded_piece": "为什么", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.2489544302225113 + }, + { + "expert": 44, + "weight": 0.06371285766363144 + }, + { + "expert": 58, + "weight": 0.04518381506204605 + }, + { + "expert": 16, + "weight": 0.03755857050418854 + }, + { + "expert": 19, + "weight": 0.03596162050962448 + }, + { + "expert": 0, + "weight": 0.0341949537396431 + } + ], + "selected_weight_sum": 0.4655662477016449, + "top1_top2_margin": 0.18524157255887985 + }, + { + "position": 7, + "token_id": 21844, + "token": "ç¨Ģ", + "decoded_piece": "稀", + "experts_by_weight": [ + { + "expert": 14, + "weight": 0.13185125589370728 + }, + { + "expert": 2, + "weight": 0.0922761932015419 + }, + { + "expert": 21, + "weight": 0.07850739359855652 + }, + { + "expert": 52, + "weight": 0.05424924194812775 + }, + { + "expert": 49, + "weight": 0.041256386786699295 + }, + { + "expert": 54, + "weight": 0.03371671214699745 + } + ], + "selected_weight_sum": 0.4318571835756302, + "top1_top2_margin": 0.039575062692165375 + }, + { + "position": 8, + "token_id": 20923, + "token": "çĸı", + "decoded_piece": "疏", + "experts_by_weight": [ + { + "expert": 35, + "weight": 0.10015559196472168 + }, + { + "expert": 25, + "weight": 0.08910446614027023 + }, + { + "expert": 57, + "weight": 0.08510752767324448 + }, + { + "expert": 53, + "weight": 0.06459127366542816 + }, + { + "expert": 41, + "weight": 0.031923577189445496 + }, + { + "expert": 40, + "weight": 0.03015722520649433 + } + ], + "selected_weight_sum": 0.4010396618396044, + "top1_top2_margin": 0.011051125824451447 + }, + { + "position": 9, + "token_id": 10900, + "token": "ä¸ĵå®¶", + "decoded_piece": "专家", + "experts_by_weight": [ + { + "expert": 60, + "weight": 0.08528561145067215 + }, + { + "expert": 32, + "weight": 0.0805189460515976 + }, + { + "expert": 0, + "weight": 0.07657847553491592 + }, + { + "expert": 37, + "weight": 0.06688879430294037 + }, + { + "expert": 20, + "weight": 0.04128864407539368 + }, + { + "expert": 29, + "weight": 0.03574351966381073 + } + ], + "selected_weight_sum": 0.38630399107933044, + "top1_top2_margin": 0.0047666653990745544 + }, + { + "position": 10, + "token_id": 39322, + "token": "模åŀĭ", + "decoded_piece": "模型", + "experts_by_weight": [ + { + "expert": 20, + "weight": 0.20124658942222595 + }, + { + "expert": 28, + "weight": 0.14611957967281342 + }, + { + "expert": 59, + "weight": 0.060534000396728516 + }, + { + "expert": 32, + "weight": 0.05685422196984291 + }, + { + "expert": 43, + "weight": 0.04330537095665932 + }, + { + "expert": 37, + "weight": 0.0407714769244194 + } + ], + "selected_weight_sum": 0.5488312393426895, + "top1_top2_margin": 0.05512700974941254 + }, + { + "position": 11, + "token_id": 1876, + "token": "åı¯ä»¥", + "decoded_piece": "可以", + "experts_by_weight": [ + { + "expert": 30, + "weight": 0.10151458531618118 + }, + { + "expert": 2, + "weight": 0.0651351809501648 + }, + { + "expert": 16, + "weight": 0.05840770900249481 + }, + { + "expert": 19, + "weight": 0.05432477220892906 + }, + { + "expert": 4, + "weight": 0.05122113600373268 + }, + { + "expert": 52, + "weight": 0.043339185416698456 + } + ], + "selected_weight_sum": 0.373942568898201, + "top1_top2_margin": 0.03637940436601639 + }, + { + "position": 12, + "token_id": 9476, + "token": "æĭ¥æľī", + "decoded_piece": "拥有", + "experts_by_weight": [ + { + "expert": 6, + "weight": 0.12864404916763306 + }, + { + "expert": 52, + "weight": 0.07182051241397858 + }, + { + "expert": 30, + "weight": 0.0615667849779129 + }, + { + "expert": 19, + "weight": 0.05789991840720177 + }, + { + "expert": 2, + "weight": 0.03908246010541916 + }, + { + "expert": 56, + "weight": 0.034220051020383835 + } + ], + "selected_weight_sum": 0.3932337760925293, + "top1_top2_margin": 0.05682353675365448 + }, + { + "position": 13, + "token_id": 3599, + "token": "å¾Īå¤ļ", + "decoded_piece": "很多", + "experts_by_weight": [ + { + "expert": 6, + "weight": 0.12923511862754822 + }, + { + "expert": 12, + "weight": 0.08376506716012955 + }, + { + "expert": 53, + "weight": 0.0825040265917778 + }, + { + "expert": 26, + "weight": 0.05150343105196953 + }, + { + "expert": 56, + "weight": 0.03739800676703453 + }, + { + "expert": 10, + "weight": 0.027514372020959854 + } + ], + "selected_weight_sum": 0.4119200222194195, + "top1_top2_margin": 0.04547005146741867 + }, + { + "position": 14, + "token_id": 21273, + "token": "åıĤæķ°", + "decoded_piece": "参数", + "experts_by_weight": [ + { + "expert": 20, + "weight": 0.15276853740215302 + }, + { + "expert": 59, + "weight": 0.11104021966457367 + }, + { + "expert": 37, + "weight": 0.06659099459648132 + }, + { + "expert": 29, + "weight": 0.04421878978610039 + }, + { + "expert": 36, + "weight": 0.04282462224364281 + }, + { + "expert": 43, + "weight": 0.042411528527736664 + } + ], + "selected_weight_sum": 0.45985469222068787, + "top1_top2_margin": 0.041728317737579346 + }, + { + "position": 15, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.10195695608854294 + }, + { + "expert": 55, + "weight": 0.09532091021537781 + }, + { + "expert": 22, + "weight": 0.09049556404352188 + }, + { + "expert": 24, + "weight": 0.08353115618228912 + }, + { + "expert": 5, + "weight": 0.06399564445018768 + }, + { + "expert": 9, + "weight": 0.04496682062745094 + } + ], + "selected_weight_sum": 0.4802670516073704, + "top1_top2_margin": 0.006636045873165131 + }, + { + "position": 16, + "token_id": 1705, + "token": "ä½Ĩ", + "decoded_piece": "但", + "experts_by_weight": [ + { + "expert": 44, + "weight": 0.11888333410024643 + }, + { + "expert": 16, + "weight": 0.11220570653676987 + }, + { + "expert": 51, + "weight": 0.05584236606955528 + }, + { + "expert": 10, + "weight": 0.03397982940077782 + }, + { + "expert": 31, + "weight": 0.02580212987959385 + }, + { + "expert": 41, + "weight": 0.025049369782209396 + } + ], + "selected_weight_sum": 0.37176273576915264, + "top1_top2_margin": 0.0066776275634765625 + }, + { + "position": 17, + "token_id": 13712, + "token": "æ¯ı个", + "decoded_piece": "每个", + "experts_by_weight": [ + { + "expert": 6, + "weight": 0.10931649059057236 + }, + { + "expert": 44, + "weight": 0.07729648053646088 + }, + { + "expert": 33, + "weight": 0.037208620458841324 + }, + { + "expert": 4, + "weight": 0.03507556393742561 + }, + { + "expert": 31, + "weight": 0.03295507654547691 + }, + { + "expert": 41, + "weight": 0.030445924028754234 + } + ], + "selected_weight_sum": 0.3222981560975313, + "top1_top2_margin": 0.03202001005411148 + }, + { + "position": 18, + "token_id": 10728, + "token": "Ġtoken", + "decoded_piece": " token", + "experts_by_weight": [ + { + "expert": 46, + "weight": 0.2605063021183014 + }, + { + "expert": 20, + "weight": 0.10610145330429077 + }, + { + "expert": 3, + "weight": 0.092587411403656 + }, + { + "expert": 36, + "weight": 0.055407967418432236 + }, + { + "expert": 29, + "weight": 0.05411171168088913 + }, + { + "expert": 49, + "weight": 0.03317439556121826 + } + ], + "selected_weight_sum": 0.6018892414867878, + "top1_top2_margin": 0.15440484881401062 + }, + { + "position": 19, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 46, + "weight": 0.5464491248130798 + }, + { + "expert": 48, + "weight": 0.10585428774356842 + }, + { + "expert": 18, + "weight": 0.04159308969974518 + }, + { + "expert": 62, + "weight": 0.037740062922239304 + }, + { + "expert": 5, + "weight": 0.030981898307800293 + }, + { + "expert": 60, + "weight": 0.028572754934430122 + } + ], + "selected_weight_sum": 0.7911912184208632, + "top1_top2_margin": 0.4405948370695114 + }, + { + "position": 20, + "token_id": 2141, + "token": "åıª", + "decoded_piece": "只", + "experts_by_weight": [ + { + "expert": 19, + "weight": 0.10716968029737473 + }, + { + "expert": 30, + "weight": 0.09130243957042694 + }, + { + "expert": 16, + "weight": 0.06434521079063416 + }, + { + "expert": 26, + "weight": 0.05338294059038162 + }, + { + "expert": 6, + "weight": 0.0471247173845768 + }, + { + "expert": 44, + "weight": 0.04004571959376335 + } + ], + "selected_weight_sum": 0.4033707082271576, + "top1_top2_margin": 0.015867240726947784 + }, + { + "position": 21, + "token_id": 5118, + "token": "使ç͍", + "decoded_piece": "使用", + "experts_by_weight": [ + { + "expert": 6, + "weight": 0.10279721766710281 + }, + { + "expert": 19, + "weight": 0.07755965739488602 + }, + { + "expert": 61, + "weight": 0.06039987877011299 + }, + { + "expert": 59, + "weight": 0.04671044647693634 + }, + { + "expert": 2, + "weight": 0.04322323203086853 + }, + { + "expert": 16, + "weight": 0.03838641941547394 + } + ], + "selected_weight_sum": 0.36907685175538063, + "top1_top2_margin": 0.025237560272216797 + }, + { + "position": 22, + "token_id": 6802, + "token": "åħ¶ä¸Ń", + "decoded_piece": "其中", + "experts_by_weight": [ + { + "expert": 11, + "weight": 0.11038965731859207 + }, + { + "expert": 12, + "weight": 0.09876789152622223 + }, + { + "expert": 6, + "weight": 0.08621250838041306 + }, + { + "expert": 18, + "weight": 0.07344488799571991 + }, + { + "expert": 8, + "weight": 0.06631779670715332 + }, + { + "expert": 26, + "weight": 0.06579618155956268 + } + ], + "selected_weight_sum": 0.5009289234876633, + "top1_top2_margin": 0.011621765792369843 + }, + { + "position": 23, + "token_id": 51905, + "token": "ä¸Ģå°ı", + "decoded_piece": "一小", + "experts_by_weight": [ + { + "expert": 26, + "weight": 0.09538868069648743 + }, + { + "expert": 33, + "weight": 0.0859629362821579 + }, + { + "expert": 42, + "weight": 0.08134215325117111 + }, + { + "expert": 12, + "weight": 0.07922398298978806 + }, + { + "expert": 6, + "weight": 0.07351940125226974 + }, + { + "expert": 27, + "weight": 0.0516715943813324 + } + ], + "selected_weight_sum": 0.46710874885320663, + "top1_top2_margin": 0.009425744414329529 + }, + { + "position": 24, + "token_id": 5781, + "token": "éĥ¨åĪĨ", + "decoded_piece": "部分", + "experts_by_weight": [ + { + "expert": 26, + "weight": 0.17455773055553436 + }, + { + "expert": 6, + "weight": 0.07393549382686615 + }, + { + "expert": 12, + "weight": 0.07050646096467972 + }, + { + "expert": 33, + "weight": 0.06601278483867645 + }, + { + "expert": 56, + "weight": 0.05656343698501587 + }, + { + "expert": 42, + "weight": 0.024452252313494682 + } + ], + "selected_weight_sum": 0.46602815948426723, + "top1_top2_margin": 0.10062223672866821 + }, + { + "position": 25, + "token_id": 398, + "token": "ãĢĤ", + "decoded_piece": "。", + "experts_by_weight": [ + { + "expert": 22, + "weight": 0.11224387586116791 + }, + { + "expert": 55, + "weight": 0.11114215850830078 + }, + { + "expert": 5, + "weight": 0.09781858325004578 + }, + { + "expert": 7, + "weight": 0.08452824503183365 + }, + { + "expert": 45, + "weight": 0.07711736857891083 + }, + { + "expert": 0, + "weight": 0.041843514889478683 + } + ], + "selected_weight_sum": 0.5246937461197376, + "top1_top2_margin": 0.0011017173528671265 + } + ] + }, + { + "id": "en_architecture", + "label": "English architecture", + "tokens": 18, + "load": [ + 3, + 0, + 2, + 2, + 4, + 3, + 1, + 1, + 0, + 3, + 0, + 3, + 3, + 1, + 2, + 0, + 0, + 0, + 1, + 0, + 6, + 4, + 1, + 0, + 3, + 1, + 0, + 0, + 1, + 2, + 0, + 2, + 0, + 1, + 0, + 4, + 6, + 1, + 0, + 5, + 5, + 0, + 1, + 6, + 0, + 1, + 1, + 1, + 2, + 2, + 0, + 0, + 0, + 1, + 5, + 1, + 3, + 2, + 2, + 1, + 1, + 4, + 0, + 3 + ], + "metrics": { + "routes": 108, + "used_experts": 44, + "zero_experts": 20, + "mean": 1.6875, + "std": 1.735484004020691, + "cv": 1.0284349918365479, + "gini": 0.5465856194496155, + "entropy_nats": 3.5949339866638184, + "effective_experts": 36.41329614537717, + "min": 0.0, + "p50": 1.0, + "p90": 4.0, + "max": 6.0 + }, + "top_experts": [ + { + "expert": 20, + "routes": 6 + }, + { + "expert": 36, + "routes": 6 + }, + { + "expert": 43, + "routes": 6 + }, + { + "expert": 39, + "routes": 5 + }, + { + "expert": 40, + "routes": 5 + }, + { + "expert": 54, + "routes": 5 + }, + { + "expert": 4, + "routes": 4 + }, + { + "expert": 21, + "routes": 4 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 54, + "weight": 0.9711988568305969 + }, + { + "expert": 11, + "weight": 0.0014021513052284718 + }, + { + "expert": 9, + "weight": 0.001076161745004356 + }, + { + "expert": 4, + "weight": 0.0008088500471785665 + }, + { + "expert": 5, + "weight": 0.0007386832148768008 + }, + { + "expert": 12, + "weight": 0.000706107763107866 + } + ], + "selected_weight_sum": 0.975930810905993, + "top1_top2_margin": 0.9697967055253685 + }, + { + "position": 1, + "token_id": 99297, + "token": "Explain", + "decoded_piece": "Explain", + "experts_by_weight": [ + { + "expert": 0, + "weight": 0.23815517127513885 + }, + { + "expert": 58, + "weight": 0.09801939874887466 + }, + { + "expert": 48, + "weight": 0.067540742456913 + }, + { + "expert": 7, + "weight": 0.05258971080183983 + }, + { + "expert": 14, + "weight": 0.049972981214523315 + }, + { + "expert": 42, + "weight": 0.03837980329990387 + } + ], + "selected_weight_sum": 0.5446578077971935, + "top1_top2_margin": 0.1401357725262642 + }, + { + "position": 2, + "token_id": 946, + "token": "Ġhow", + "decoded_piece": " how", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.16547216475009918 + }, + { + "expert": 54, + "weight": 0.10329186171293259 + }, + { + "expert": 37, + "weight": 0.10013425350189209 + }, + { + "expert": 58, + "weight": 0.07191096246242523 + }, + { + "expert": 56, + "weight": 0.04293400049209595 + }, + { + "expert": 0, + "weight": 0.030487103387713432 + } + ], + "selected_weight_sum": 0.5142303463071585, + "top1_top2_margin": 0.062180303037166595 + }, + { + "position": 3, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.1549440622329712 + }, + { + "expert": 6, + "weight": 0.1339392364025116 + }, + { + "expert": 2, + "weight": 0.08629074692726135 + }, + { + "expert": 56, + "weight": 0.03606937825679779 + }, + { + "expert": 54, + "weight": 0.030499035492539406 + }, + { + "expert": 0, + "weight": 0.026507094502449036 + } + ], + "selected_weight_sum": 0.4682495538145304, + "top1_top2_margin": 0.021004825830459595 + }, + { + "position": 4, + "token_id": 29640, + "token": "Ġcompressed", + "decoded_piece": " compressed", + "experts_by_weight": [ + { + "expert": 39, + "weight": 0.09938042610883713 + }, + { + "expert": 40, + "weight": 0.07479962706565857 + }, + { + "expert": 61, + "weight": 0.07033173739910126 + }, + { + "expert": 21, + "weight": 0.05891317501664162 + }, + { + "expert": 24, + "weight": 0.04771549254655838 + }, + { + "expert": 28, + "weight": 0.024970492348074913 + } + ], + "selected_weight_sum": 0.37611095048487186, + "top1_top2_margin": 0.02458079904317856 + }, + { + "position": 5, + "token_id": 2131, + "token": "Ġkey", + "decoded_piece": " key", + "experts_by_weight": [ + { + "expert": 35, + "weight": 0.2045895755290985 + }, + { + "expert": 43, + "weight": 0.1725901961326599 + }, + { + "expert": 20, + "weight": 0.09196025133132935 + }, + { + "expert": 36, + "weight": 0.08283847570419312 + }, + { + "expert": 24, + "weight": 0.052055127918720245 + }, + { + "expert": 29, + "weight": 0.023450862616300583 + } + ], + "selected_weight_sum": 0.6274844892323017, + "top1_top2_margin": 0.0319993793964386 + }, + { + "position": 6, + "token_id": 12, + "token": "-", + "decoded_piece": "-", + "experts_by_weight": [ + { + "expert": 18, + "weight": 0.21182844042778015 + }, + { + "expert": 9, + "weight": 0.08291414380073547 + }, + { + "expert": 12, + "weight": 0.05877301096916199 + }, + { + "expert": 13, + "weight": 0.058278877288103104 + }, + { + "expert": 5, + "weight": 0.05514194443821907 + }, + { + "expert": 40, + "weight": 0.03826766088604927 + } + ], + "selected_weight_sum": 0.5052040778100491, + "top1_top2_margin": 0.12891429662704468 + }, + { + "position": 7, + "token_id": 2834, + "token": "value", + "decoded_piece": "value", + "experts_by_weight": [ + { + "expert": 20, + "weight": 0.15702477097511292 + }, + { + "expert": 36, + "weight": 0.12268862873315811 + }, + { + "expert": 57, + "weight": 0.0899953618645668 + }, + { + "expert": 43, + "weight": 0.0687999576330185 + }, + { + "expert": 35, + "weight": 0.057560667395591736 + }, + { + "expert": 49, + "weight": 0.03539532423019409 + } + ], + "selected_weight_sum": 0.5314647108316422, + "top1_top2_margin": 0.0343361422419548 + }, + { + "position": 8, + "token_id": 28937, + "token": "Ġlatent", + "decoded_piece": " latent", + "experts_by_weight": [ + { + "expert": 36, + "weight": 0.0926990807056427 + }, + { + "expert": 39, + "weight": 0.06688005477190018 + }, + { + "expert": 49, + "weight": 0.06084948405623436 + }, + { + "expert": 21, + "weight": 0.05863860249519348 + }, + { + "expert": 40, + "weight": 0.056106846779584885 + }, + { + "expert": 24, + "weight": 0.050941843539476395 + } + ], + "selected_weight_sum": 0.386115912348032, + "top1_top2_margin": 0.025819025933742523 + }, + { + "position": 9, + "token_id": 4183, + "token": "Ġchanges", + "decoded_piece": " changes", + "experts_by_weight": [ + { + "expert": 39, + "weight": 0.1537322998046875 + }, + { + "expert": 61, + "weight": 0.1268528699874878 + }, + { + "expert": 63, + "weight": 0.08116747438907623 + }, + { + "expert": 20, + "weight": 0.0690811425447464 + }, + { + "expert": 35, + "weight": 0.049206629395484924 + }, + { + "expert": 43, + "weight": 0.04574333503842354 + } + ], + "selected_weight_sum": 0.5257837511599064, + "top1_top2_margin": 0.026879429817199707 + }, + { + "position": 10, + "token_id": 254, + "token": "Ġthe", + "decoded_piece": " the", + "experts_by_weight": [ + { + "expert": 54, + "weight": 0.12472306936979294 + }, + { + "expert": 12, + "weight": 0.10463881492614746 + }, + { + "expert": 2, + "weight": 0.06110163778066635 + }, + { + "expert": 33, + "weight": 0.04021468013525009 + }, + { + "expert": 53, + "weight": 0.039409756660461426 + }, + { + "expert": 25, + "weight": 0.03180879354476929 + } + ], + "selected_weight_sum": 0.40189675241708755, + "top1_top2_margin": 0.020084254443645477 + }, + { + "position": 11, + "token_id": 4887, + "token": "Ġmemory", + "decoded_piece": " memory", + "experts_by_weight": [ + { + "expert": 43, + "weight": 0.10257595032453537 + }, + { + "expert": 20, + "weight": 0.09988170117139816 + }, + { + "expert": 63, + "weight": 0.04727212339639664 + }, + { + "expert": 3, + "weight": 0.0421602837741375 + }, + { + "expert": 21, + "weight": 0.04022814705967903 + }, + { + "expert": 36, + "weight": 0.03870483115315437 + } + ], + "selected_weight_sum": 0.37082303687930107, + "top1_top2_margin": 0.002694249153137207 + }, + { + "position": 12, + "token_id": 2295, + "token": "Ġcost", + "decoded_piece": " cost", + "experts_by_weight": [ + { + "expert": 59, + "weight": 0.19642455875873566 + }, + { + "expert": 20, + "weight": 0.16629761457443237 + }, + { + "expert": 40, + "weight": 0.04809309169650078 + }, + { + "expert": 43, + "weight": 0.04783359542489052 + }, + { + "expert": 57, + "weight": 0.0408785454928875 + }, + { + "expert": 63, + "weight": 0.034761298447847366 + } + ], + "selected_weight_sum": 0.5342887043952942, + "top1_top2_margin": 0.030126944184303284 + }, + { + "position": 13, + "token_id": 280, + "token": "Ġof", + "decoded_piece": " of", + "experts_by_weight": [ + { + "expert": 56, + "weight": 0.1839616447687149 + }, + { + "expert": 54, + "weight": 0.04788229241967201 + }, + { + "expert": 4, + "weight": 0.033062588423490524 + }, + { + "expert": 46, + "weight": 0.03048531897366047 + }, + { + "expert": 47, + "weight": 0.027693433687090874 + }, + { + "expert": 31, + "weight": 0.026338256895542145 + } + ], + "selected_weight_sum": 0.34942353516817093, + "top1_top2_margin": 0.1360793523490429 + }, + { + "position": 14, + "token_id": 77939, + "token": "Ġautore", + "decoded_piece": " autore", + "experts_by_weight": [ + { + "expert": 29, + "weight": 0.05693405494093895 + }, + { + "expert": 36, + "weight": 0.05210784822702408 + }, + { + "expert": 21, + "weight": 0.04772772267460823 + }, + { + "expert": 3, + "weight": 0.036565009504556656 + }, + { + "expert": 14, + "weight": 0.03430648520588875 + }, + { + "expert": 48, + "weight": 0.03330330550670624 + } + ], + "selected_weight_sum": 0.2609444260597229, + "top1_top2_margin": 0.004826206713914871 + }, + { + "position": 15, + "token_id": 79184, + "token": "gressive", + "decoded_piece": "gressive", + "experts_by_weight": [ + { + "expert": 11, + "weight": 0.23832271993160248 + }, + { + "expert": 40, + "weight": 0.09413538128137589 + }, + { + "expert": 39, + "weight": 0.03894481435418129 + }, + { + "expert": 35, + "weight": 0.03155319020152092 + }, + { + "expert": 61, + "weight": 0.03087766282260418 + }, + { + "expert": 31, + "weight": 0.02754705585539341 + } + ], + "selected_weight_sum": 0.46138082444667816, + "top1_top2_margin": 0.1441873386502266 + }, + { + "position": 16, + "token_id": 39671, + "token": "Ġdecoding", + "decoded_piece": " decoding", + "experts_by_weight": [ + { + "expert": 61, + "weight": 0.10946545749902725 + }, + { + "expert": 43, + "weight": 0.07026691734790802 + }, + { + "expert": 39, + "weight": 0.06993477791547775 + }, + { + "expert": 20, + "weight": 0.0477374792098999 + }, + { + "expert": 36, + "weight": 0.04432488977909088 + }, + { + "expert": 11, + "weight": 0.04071218520402908 + } + ], + "selected_weight_sum": 0.3824417069554329, + "top1_top2_margin": 0.03919854015111923 + }, + { + "position": 17, + "token_id": 13, + "token": ".", + "decoded_piece": ".", + "experts_by_weight": [ + { + "expert": 22, + "weight": 0.12724851071834564 + }, + { + "expert": 55, + "weight": 0.11848980188369751 + }, + { + "expert": 5, + "weight": 0.09815538674592972 + }, + { + "expert": 45, + "weight": 0.06956201046705246 + }, + { + "expert": 60, + "weight": 0.05623827129602432 + }, + { + "expert": 9, + "weight": 0.04838328808546066 + } + ], + "selected_weight_sum": 0.5180772691965103, + "top1_top2_margin": 0.008758708834648132 + } + ] + }, + { + "id": "code", + "label": "Python code", + "tokens": 19, + "load": [ + 4, + 1, + 1, + 1, + 4, + 3, + 3, + 0, + 1, + 1, + 0, + 2, + 6, + 0, + 1, + 0, + 3, + 1, + 0, + 2, + 4, + 0, + 1, + 0, + 1, + 3, + 3, + 1, + 3, + 5, + 1, + 0, + 1, + 5, + 0, + 2, + 3, + 7, + 0, + 3, + 0, + 0, + 0, + 4, + 0, + 1, + 0, + 2, + 2, + 1, + 2, + 1, + 0, + 2, + 8, + 2, + 0, + 0, + 3, + 2, + 2, + 4, + 0, + 1 + ], + "metrics": { + "routes": 114, + "used_experts": 45, + "zero_experts": 19, + "mean": 1.78125, + "std": 1.8384668827056885, + "cv": 1.0321217775344849, + "gini": 0.5386512875556946, + "entropy_nats": 3.610292911529541, + "effective_experts": 36.976882183839685, + "min": 0.0, + "p50": 1.0, + "p90": 4.0, + "max": 8.0 + }, + "top_experts": [ + { + "expert": 54, + "routes": 8 + }, + { + "expert": 37, + "routes": 7 + }, + { + "expert": 12, + "routes": 6 + }, + { + "expert": 29, + "routes": 5 + }, + { + "expert": 33, + "routes": 5 + }, + { + "expert": 0, + "routes": 4 + }, + { + "expert": 4, + "routes": 4 + }, + { + "expert": 20, + "routes": 4 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 54, + "weight": 0.9711988568305969 + }, + { + "expert": 11, + "weight": 0.0014021513052284718 + }, + { + "expert": 9, + "weight": 0.001076161745004356 + }, + { + "expert": 4, + "weight": 0.0008088500471785665 + }, + { + "expert": 5, + "weight": 0.0007386832148768008 + }, + { + "expert": 12, + "weight": 0.000706107763107866 + } + ], + "selected_weight_sum": 0.975930810905993, + "top1_top2_margin": 0.9697967055253685 + }, + { + "position": 1, + "token_id": 9081, + "token": "Write", + "decoded_piece": "Write", + "experts_by_weight": [ + { + "expert": 0, + "weight": 0.15111351013183594 + }, + { + "expert": 14, + "weight": 0.11959145218133926 + }, + { + "expert": 58, + "weight": 0.08922455459833145 + }, + { + "expert": 48, + "weight": 0.05587117373943329 + }, + { + "expert": 61, + "weight": 0.038532257080078125 + }, + { + "expert": 60, + "weight": 0.034965526312589645 + } + ], + "selected_weight_sum": 0.4892984740436077, + "top1_top2_margin": 0.031522057950496674 + }, + { + "position": 2, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 58, + "weight": 0.2080550640821457 + }, + { + "expert": 12, + "weight": 0.14192862808704376 + }, + { + "expert": 6, + "weight": 0.08557718247175217 + }, + { + "expert": 37, + "weight": 0.07180862873792648 + }, + { + "expert": 53, + "weight": 0.032067738473415375 + }, + { + "expert": 17, + "weight": 0.030239181593060493 + } + ], + "selected_weight_sum": 0.569676423445344, + "top1_top2_margin": 0.06612643599510193 + }, + { + "position": 3, + "token_id": 12974, + "token": "ĠPython", + "decoded_piece": " Python", + "experts_by_weight": [ + { + "expert": 29, + "weight": 0.1371878981590271 + }, + { + "expert": 3, + "weight": 0.12900573015213013 + }, + { + "expert": 0, + "weight": 0.09749352186918259 + }, + { + "expert": 54, + "weight": 0.049715057015419006 + }, + { + "expert": 33, + "weight": 0.046508997678756714 + }, + { + "expert": 24, + "weight": 0.041462209075689316 + } + ], + "selected_weight_sum": 0.5013734139502048, + "top1_top2_margin": 0.008182168006896973 + }, + { + "position": 4, + "token_id": 1157, + "token": "Ġfunction", + "decoded_piece": " function", + "experts_by_weight": [ + { + "expert": 54, + "weight": 0.20584429800510406 + }, + { + "expert": 29, + "weight": 0.14856287837028503 + }, + { + "expert": 20, + "weight": 0.0871359184384346 + }, + { + "expert": 0, + "weight": 0.08152839541435242 + }, + { + "expert": 36, + "weight": 0.051778245717287064 + }, + { + "expert": 43, + "weight": 0.02517823502421379 + } + ], + "selected_weight_sum": 0.600027970969677, + "top1_top2_margin": 0.05728141963481903 + }, + { + "position": 5, + "token_id": 344, + "token": "Ġthat", + "decoded_piece": " that", + "experts_by_weight": [ + { + "expert": 54, + "weight": 0.18090112507343292 + }, + { + "expert": 19, + "weight": 0.11123941093683243 + }, + { + "expert": 16, + "weight": 0.07971292734146118 + }, + { + "expert": 4, + "weight": 0.02637108787894249 + }, + { + "expert": 6, + "weight": 0.025111805647611618 + }, + { + "expert": 37, + "weight": 0.02362849749624729 + } + ], + "selected_weight_sum": 0.44696485437452793, + "top1_top2_margin": 0.0696617141366005 + }, + { + "position": 6, + "token_id": 7578, + "token": "Ġreturns", + "decoded_piece": " returns", + "experts_by_weight": [ + { + "expert": 19, + "weight": 0.1616554856300354 + }, + { + "expert": 61, + "weight": 0.11549834907054901 + }, + { + "expert": 54, + "weight": 0.07323762774467468 + }, + { + "expert": 30, + "weight": 0.05683310702443123 + }, + { + "expert": 39, + "weight": 0.043762173503637314 + }, + { + "expert": 37, + "weight": 0.03263093903660774 + } + ], + "selected_weight_sum": 0.4836176820099354, + "top1_top2_margin": 0.04615713655948639 + }, + { + "position": 7, + "token_id": 254, + "token": "Ġthe", + "decoded_piece": " the", + "experts_by_weight": [ + { + "expert": 54, + "weight": 0.22255076467990875 + }, + { + "expert": 12, + "weight": 0.07750241458415985 + }, + { + "expert": 33, + "weight": 0.05901200696825981 + }, + { + "expert": 37, + "weight": 0.04646346718072891 + }, + { + "expert": 25, + "weight": 0.03500976786017418 + }, + { + "expert": 11, + "weight": 0.025494523346424103 + } + ], + "selected_weight_sum": 0.4660329446196556, + "top1_top2_margin": 0.1450483500957489 + }, + { + "position": 8, + "token_id": 1022, + "token": "Ġfirst", + "decoded_piece": " first", + "experts_by_weight": [ + { + "expert": 33, + "weight": 0.23225565254688263 + }, + { + "expert": 49, + "weight": 0.10645926743745804 + }, + { + "expert": 27, + "weight": 0.05182158946990967 + }, + { + "expert": 8, + "weight": 0.04605036601424217 + }, + { + "expert": 54, + "weight": 0.03481481224298477 + }, + { + "expert": 12, + "weight": 0.027712024748325348 + } + ], + "selected_weight_sum": 0.4991137124598026, + "top1_top2_margin": 0.1257963851094246 + }, + { + "position": 9, + "token_id": 9842, + "token": "Ġrepeated", + "decoded_piece": " repeated", + "experts_by_weight": [ + { + "expert": 33, + "weight": 0.18555673956871033 + }, + { + "expert": 39, + "weight": 0.0932159349322319 + }, + { + "expert": 61, + "weight": 0.0786258727312088 + }, + { + "expert": 53, + "weight": 0.049028702080249786 + }, + { + "expert": 50, + "weight": 0.030850498005747795 + }, + { + "expert": 2, + "weight": 0.02684546820819378 + } + ], + "selected_weight_sum": 0.4641232155263424, + "top1_top2_margin": 0.09234080463647842 + }, + { + "position": 10, + "token_id": 3699, + "token": "Ġelement", + "decoded_piece": " element", + "experts_by_weight": [ + { + "expert": 20, + "weight": 0.13938002288341522 + }, + { + "expert": 43, + "weight": 0.08577956259250641 + }, + { + "expert": 36, + "weight": 0.07479023188352585 + }, + { + "expert": 33, + "weight": 0.059141431003808975 + }, + { + "expert": 29, + "weight": 0.05091153085231781 + }, + { + "expert": 35, + "weight": 0.04373912513256073 + } + ], + "selected_weight_sum": 0.453741904348135, + "top1_top2_margin": 0.053600460290908813 + }, + { + "position": 11, + "token_id": 279, + "token": "Ġin", + "decoded_piece": " in", + "experts_by_weight": [ + { + "expert": 28, + "weight": 0.14614824950695038 + }, + { + "expert": 47, + "weight": 0.11818400770425797 + }, + { + "expert": 59, + "weight": 0.09291267395019531 + }, + { + "expert": 25, + "weight": 0.05692411959171295 + }, + { + "expert": 26, + "weight": 0.05534731596708298 + }, + { + "expert": 50, + "weight": 0.03634467348456383 + } + ], + "selected_weight_sum": 0.5058610402047634, + "top1_top2_margin": 0.027964241802692413 + }, + { + "position": 12, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 28, + "weight": 0.09758827090263367 + }, + { + "expert": 47, + "weight": 0.091375432908535 + }, + { + "expert": 6, + "weight": 0.07514315843582153 + }, + { + "expert": 12, + "weight": 0.053120844066143036 + }, + { + "expert": 26, + "weight": 0.04297527298331261 + }, + { + "expert": 25, + "weight": 0.03407928720116615 + } + ], + "selected_weight_sum": 0.394282266497612, + "top1_top2_margin": 0.006212837994098663 + }, + { + "position": 13, + "token_id": 1525, + "token": "Ġlist", + "decoded_piece": " list", + "experts_by_weight": [ + { + "expert": 36, + "weight": 0.15618914365768433 + }, + { + "expert": 43, + "weight": 0.058529019355773926 + }, + { + "expert": 29, + "weight": 0.05490716174244881 + }, + { + "expert": 32, + "weight": 0.049499817192554474 + }, + { + "expert": 20, + "weight": 0.039970483630895615 + }, + { + "expert": 37, + "weight": 0.03016483783721924 + } + ], + "selected_weight_sum": 0.3892604634165764, + "top1_top2_margin": 0.0976601243019104 + }, + { + "position": 14, + "token_id": 285, + "token": "Ġand", + "decoded_piece": " and", + "experts_by_weight": [ + { + "expert": 16, + "weight": 0.140435129404068 + }, + { + "expert": 55, + "weight": 0.11578470468521118 + }, + { + "expert": 51, + "weight": 0.08345910906791687 + }, + { + "expert": 1, + "weight": 0.038944266736507416 + }, + { + "expert": 5, + "weight": 0.037622421979904175 + }, + { + "expert": 60, + "weight": 0.030940167605876923 + } + ], + "selected_weight_sum": 0.44718579947948456, + "top1_top2_margin": 0.02465042471885681 + }, + { + "position": 15, + "token_id": 6716, + "token": "Ġexplain", + "decoded_piece": " explain", + "experts_by_weight": [ + { + "expert": 37, + "weight": 0.12016983330249786 + }, + { + "expert": 58, + "weight": 0.11260108649730682 + }, + { + "expert": 61, + "weight": 0.07255109399557114 + }, + { + "expert": 16, + "weight": 0.0641496405005455 + }, + { + "expert": 39, + "weight": 0.0602947473526001 + }, + { + "expert": 4, + "weight": 0.02975073829293251 + } + ], + "selected_weight_sum": 0.45951713994145393, + "top1_top2_margin": 0.00756874680519104 + }, + { + "position": 16, + "token_id": 895, + "token": "Ġits", + "decoded_piece": " its", + "experts_by_weight": [ + { + "expert": 54, + "weight": 0.18681631982326508 + }, + { + "expert": 37, + "weight": 0.07873330265283585 + }, + { + "expert": 12, + "weight": 0.06885220855474472 + }, + { + "expert": 29, + "weight": 0.0670696347951889 + }, + { + "expert": 28, + "weight": 0.04396243393421173 + }, + { + "expert": 4, + "weight": 0.029901383444666862 + } + ], + "selected_weight_sum": 0.47533528320491314, + "top1_top2_margin": 0.10808301717042923 + }, + { + "position": 17, + "token_id": 13913, + "token": "Ġcomplexity", + "decoded_piece": " complexity", + "experts_by_weight": [ + { + "expert": 20, + "weight": 0.07373698800802231 + }, + { + "expert": 59, + "weight": 0.06951963156461716 + }, + { + "expert": 63, + "weight": 0.05116838589310646 + }, + { + "expert": 35, + "weight": 0.041290223598480225 + }, + { + "expert": 43, + "weight": 0.03023359179496765 + }, + { + "expert": 26, + "weight": 0.027252253144979477 + } + ], + "selected_weight_sum": 0.2932010740041733, + "top1_top2_margin": 0.004217356443405151 + }, + { + "position": 18, + "token_id": 13, + "token": ".", + "decoded_piece": ".", + "experts_by_weight": [ + { + "expert": 22, + "weight": 0.11643100529909134 + }, + { + "expert": 55, + "weight": 0.11100540310144424 + }, + { + "expert": 5, + "weight": 0.10169129818677902 + }, + { + "expert": 45, + "weight": 0.058091938495635986 + }, + { + "expert": 0, + "weight": 0.03980956971645355 + }, + { + "expert": 48, + "weight": 0.03621777892112732 + } + ], + "selected_weight_sum": 0.46324699372053146, + "top1_top2_margin": 0.005425602197647095 + } + ] + }, + { + "id": "math", + "label": "数学推理", + "tokens": 27, + "load": [ + 1, + 4, + 1, + 5, + 4, + 3, + 3, + 2, + 0, + 3, + 0, + 2, + 2, + 1, + 4, + 3, + 6, + 0, + 4, + 2, + 1, + 0, + 3, + 0, + 2, + 4, + 8, + 6, + 1, + 2, + 2, + 0, + 0, + 5, + 0, + 0, + 1, + 1, + 1, + 7, + 1, + 2, + 1, + 0, + 3, + 2, + 8, + 4, + 6, + 7, + 0, + 4, + 0, + 1, + 3, + 3, + 1, + 0, + 3, + 0, + 6, + 2, + 11, + 0 + ], + "metrics": { + "routes": 162, + "used_experts": 49, + "zero_experts": 15, + "mean": 2.53125, + "std": 2.442797899246216, + "cv": 0.9650559425354004, + "gini": 0.5117669701576233, + "entropy_nats": 3.6781580448150635, + "effective_experts": 39.573434406292506, + "min": 0.0, + "p50": 2.0, + "p90": 6.0, + "max": 11.0 + }, + "top_experts": [ + { + "expert": 62, + "routes": 11 + }, + { + "expert": 26, + "routes": 8 + }, + { + "expert": 46, + "routes": 8 + }, + { + "expert": 39, + "routes": 7 + }, + { + "expert": 49, + "routes": 7 + }, + { + "expert": 16, + "routes": 6 + }, + { + "expert": 27, + "routes": 6 + }, + { + "expert": 48, + "routes": 6 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 54, + "weight": 0.9711988568305969 + }, + { + "expert": 11, + "weight": 0.0014021513052284718 + }, + { + "expert": 9, + "weight": 0.001076161745004356 + }, + { + "expert": 4, + "weight": 0.0008088500471785665 + }, + { + "expert": 5, + "weight": 0.0007386832148768008 + }, + { + "expert": 12, + "weight": 0.000706107763107866 + } + ], + "selected_weight_sum": 0.975930810905993, + "top1_top2_margin": 0.9697967055253685 + }, + { + "position": 1, + "token_id": 8246, + "token": "èĭ¥", + "decoded_piece": "若", + "experts_by_weight": [ + { + "expert": 16, + "weight": 0.14951738715171814 + }, + { + "expert": 14, + "weight": 0.0980856642127037 + }, + { + "expert": 44, + "weight": 0.08160083740949631 + }, + { + "expert": 4, + "weight": 0.07467349618673325 + }, + { + "expert": 0, + "weight": 0.06317417323589325 + }, + { + "expert": 58, + "weight": 0.052538976073265076 + } + ], + "selected_weight_sum": 0.5195905342698097, + "top1_top2_margin": 0.051431722939014435 + }, + { + "position": 2, + "token_id": 2206, + "token": "æŃ£", + "decoded_piece": "正", + "experts_by_weight": [ + { + "expert": 2, + "weight": 0.10717478394508362 + }, + { + "expert": 30, + "weight": 0.10145746916532516 + }, + { + "expert": 44, + "weight": 0.0593704953789711 + }, + { + "expert": 14, + "weight": 0.055084649473428726 + }, + { + "expert": 25, + "weight": 0.05256998538970947 + }, + { + "expert": 58, + "weight": 0.05031483620405197 + } + ], + "selected_weight_sum": 0.42597221955657005, + "top1_top2_margin": 0.005717314779758453 + }, + { + "position": 3, + "token_id": 2266, + "token": "æķ°", + "decoded_piece": "数", + "experts_by_weight": [ + { + "expert": 27, + "weight": 0.16861963272094727 + }, + { + "expert": 25, + "weight": 0.11267384886741638 + }, + { + "expert": 33, + "weight": 0.09004956483840942 + }, + { + "expert": 20, + "weight": 0.07321920245885849 + }, + { + "expert": 28, + "weight": 0.06758833676576614 + }, + { + "expert": 39, + "weight": 0.06270019710063934 + } + ], + "selected_weight_sum": 0.574850782752037, + "top1_top2_margin": 0.055945783853530884 + }, + { + "position": 4, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 46, + "weight": 0.2162427455186844 + }, + { + "expert": 62, + "weight": 0.13654249906539917 + }, + { + "expert": 3, + "weight": 0.06265389919281006 + }, + { + "expert": 49, + "weight": 0.054540734738111496 + }, + { + "expert": 60, + "weight": 0.05205049738287926 + }, + { + "expert": 9, + "weight": 0.045574069023132324 + } + ], + "selected_weight_sum": 0.5676044449210167, + "top1_top2_margin": 0.07970024645328522 + }, + { + "position": 5, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 46, + "weight": 0.486960232257843 + }, + { + "expert": 48, + "weight": 0.12458961457014084 + }, + { + "expert": 60, + "weight": 0.06354541331529617 + }, + { + "expert": 62, + "weight": 0.04251669719815254 + }, + { + "expert": 18, + "weight": 0.037103284150362015 + }, + { + "expert": 5, + "weight": 0.03097430430352688 + } + ], + "selected_weight_sum": 0.7856895457953215, + "top1_top2_margin": 0.3623706176877022 + }, + { + "position": 6, + "token_id": 13211, + "token": "满足", + "decoded_piece": "满足", + "experts_by_weight": [ + { + "expert": 6, + "weight": 0.16329914331436157 + }, + { + "expert": 19, + "weight": 0.1260579377412796 + }, + { + "expert": 16, + "weight": 0.06854663044214249 + }, + { + "expert": 47, + "weight": 0.06128653883934021 + }, + { + "expert": 60, + "weight": 0.03816733509302139 + }, + { + "expert": 4, + "weight": 0.03762584179639816 + } + ], + "selected_weight_sum": 0.4949834272265434, + "top1_top2_margin": 0.03724120557308197 + }, + { + "position": 7, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 62, + "weight": 0.1666949838399887 + }, + { + "expert": 46, + "weight": 0.13348163664340973 + }, + { + "expert": 49, + "weight": 0.08109060674905777 + }, + { + "expert": 3, + "weight": 0.06654893606901169 + }, + { + "expert": 48, + "weight": 0.03289414197206497 + }, + { + "expert": 36, + "weight": 0.030425915494561195 + } + ], + "selected_weight_sum": 0.5111362207680941, + "top1_top2_margin": 0.03321334719657898 + }, + { + "position": 8, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 46, + "weight": 0.30760839581489563 + }, + { + "expert": 18, + "weight": 0.05372292920947075 + }, + { + "expert": 62, + "weight": 0.05369928479194641 + }, + { + "expert": 48, + "weight": 0.04193925857543945 + }, + { + "expert": 1, + "weight": 0.041528962552547455 + }, + { + "expert": 60, + "weight": 0.04065709561109543 + } + ], + "selected_weight_sum": 0.5391559265553951, + "top1_top2_margin": 0.2538854666054249 + }, + { + "position": 9, + "token_id": 14855, + "token": "åĬłä¸Ĭ", + "decoded_piece": "加上", + "experts_by_weight": [ + { + "expert": 26, + "weight": 0.15855351090431213 + }, + { + "expert": 51, + "weight": 0.1316342055797577 + }, + { + "expert": 6, + "weight": 0.05368504300713539 + }, + { + "expert": 16, + "weight": 0.04912524297833443 + }, + { + "expert": 1, + "weight": 0.04827090725302696 + }, + { + "expert": 47, + "weight": 0.036717962473630905 + } + ], + "selected_weight_sum": 0.4779868721961975, + "top1_top2_margin": 0.026919305324554443 + }, + { + "position": 10, + "token_id": 12985, + "token": "å®ĥçļĦ", + "decoded_piece": "它的", + "experts_by_weight": [ + { + "expert": 54, + "weight": 0.116221122443676 + }, + { + "expert": 29, + "weight": 0.10054290294647217 + }, + { + "expert": 6, + "weight": 0.07659035921096802 + }, + { + "expert": 51, + "weight": 0.06001657620072365 + }, + { + "expert": 12, + "weight": 0.05728625878691673 + }, + { + "expert": 44, + "weight": 0.049379147589206696 + } + ], + "selected_weight_sum": 0.46003636717796326, + "top1_top2_margin": 0.015678219497203827 + }, + { + "position": 11, + "token_id": 8270, + "token": "åĢĴ", + "decoded_piece": "倒", + "experts_by_weight": [ + { + "expert": 39, + "weight": 0.12973536550998688 + }, + { + "expert": 33, + "weight": 0.07649380713701248 + }, + { + "expert": 61, + "weight": 0.04296094924211502 + }, + { + "expert": 14, + "weight": 0.041707731783390045 + }, + { + "expert": 62, + "weight": 0.03551975265145302 + }, + { + "expert": 49, + "weight": 0.03439413383603096 + } + ], + "selected_weight_sum": 0.3608117401599884, + "top1_top2_margin": 0.053241558372974396 + }, + { + "position": 12, + "token_id": 2266, + "token": "æķ°", + "decoded_piece": "数", + "experts_by_weight": [ + { + "expert": 39, + "weight": 0.1611926555633545 + }, + { + "expert": 27, + "weight": 0.1397518813610077 + }, + { + "expert": 25, + "weight": 0.09240048378705978 + }, + { + "expert": 26, + "weight": 0.044278018176555634 + }, + { + "expert": 33, + "weight": 0.04391821101307869 + }, + { + "expert": 40, + "weight": 0.03572217747569084 + } + ], + "selected_weight_sum": 0.5172634273767471, + "top1_top2_margin": 0.021440774202346802 + }, + { + "position": 13, + "token_id": 37360, + "token": "çŃīäºİ", + "decoded_piece": "等于", + "experts_by_weight": [ + { + "expert": 19, + "weight": 0.1503891944885254 + }, + { + "expert": 26, + "weight": 0.10245323926210403 + }, + { + "expert": 30, + "weight": 0.060541197657585144 + }, + { + "expert": 53, + "weight": 0.05172496289014816 + }, + { + "expert": 47, + "weight": 0.035423070192337036 + }, + { + "expert": 16, + "weight": 0.02933596819639206 + } + ], + "selected_weight_sum": 0.4298676326870918, + "top1_top2_margin": 0.047935955226421356 + }, + { + "position": 14, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 46, + "weight": 0.24005234241485596 + }, + { + "expert": 62, + "weight": 0.08549993485212326 + }, + { + "expert": 48, + "weight": 0.07248266786336899 + }, + { + "expert": 18, + "weight": 0.05309956893324852 + }, + { + "expert": 22, + "weight": 0.04152917489409447 + }, + { + "expert": 55, + "weight": 0.03998137637972832 + } + ], + "selected_weight_sum": 0.5326450653374195, + "top1_top2_margin": 0.1545524075627327 + }, + { + "position": 15, + "token_id": 18, + "token": "3", + "decoded_piece": "3", + "experts_by_weight": [ + { + "expert": 27, + "weight": 0.18014958500862122 + }, + { + "expert": 62, + "weight": 0.1050652265548706 + }, + { + "expert": 49, + "weight": 0.062478359788656235 + }, + { + "expert": 46, + "weight": 0.04287721961736679 + }, + { + "expert": 38, + "weight": 0.03764505684375763 + }, + { + "expert": 15, + "weight": 0.024584602564573288 + } + ], + "selected_weight_sum": 0.45280005037784576, + "top1_top2_margin": 0.07508435845375061 + }, + { + "position": 16, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 1, + "weight": 0.10770390182733536 + }, + { + "expert": 55, + "weight": 0.10210539400577545 + }, + { + "expert": 22, + "weight": 0.08225701749324799 + }, + { + "expert": 24, + "weight": 0.07177606225013733 + }, + { + "expert": 7, + "weight": 0.068714439868927 + }, + { + "expert": 45, + "weight": 0.060516949743032455 + } + ], + "selected_weight_sum": 0.4930737651884556, + "top1_top2_margin": 0.005598507821559906 + }, + { + "position": 17, + "token_id": 2989, + "token": "æ±Ĥ", + "decoded_piece": "求", + "experts_by_weight": [ + { + "expert": 16, + "weight": 0.12940554320812225 + }, + { + "expert": 58, + "weight": 0.1288883239030838 + }, + { + "expert": 4, + "weight": 0.06063942238688469 + }, + { + "expert": 37, + "weight": 0.05899067223072052 + }, + { + "expert": 13, + "weight": 0.03491080552339554 + }, + { + "expert": 42, + "weight": 0.033715102821588516 + } + ], + "selected_weight_sum": 0.4465498700737953, + "top1_top2_margin": 0.0005172193050384521 + }, + { + "position": 18, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 62, + "weight": 0.13900531828403473 + }, + { + "expert": 46, + "weight": 0.12066206336021423 + }, + { + "expert": 3, + "weight": 0.07627727091312408 + }, + { + "expert": 49, + "weight": 0.07498259097337723 + }, + { + "expert": 29, + "weight": 0.037425458431243896 + }, + { + "expert": 48, + "weight": 0.03221821039915085 + } + ], + "selected_weight_sum": 0.480570912361145, + "top1_top2_margin": 0.018343254923820496 + }, + { + "position": 19, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 46, + "weight": 0.29581424593925476 + }, + { + "expert": 18, + "weight": 0.06302740424871445 + }, + { + "expert": 62, + "weight": 0.044536203145980835 + }, + { + "expert": 48, + "weight": 0.03947015479207039 + }, + { + "expert": 15, + "weight": 0.038034409284591675 + }, + { + "expert": 9, + "weight": 0.031326256692409515 + } + ], + "selected_weight_sum": 0.5122086741030216, + "top1_top2_margin": 0.2327868416905403 + }, + { + "position": 20, + "token_id": 337, + "token": "çļĦ", + "decoded_piece": "的", + "experts_by_weight": [ + { + "expert": 54, + "weight": 0.10390287637710571 + }, + { + "expert": 62, + "weight": 0.07279452681541443 + }, + { + "expert": 15, + "weight": 0.06581421941518784 + }, + { + "expert": 1, + "weight": 0.05750293657183647 + }, + { + "expert": 56, + "weight": 0.035950709134340286 + }, + { + "expert": 41, + "weight": 0.031022725626826286 + } + ], + "selected_weight_sum": 0.366987993940711, + "top1_top2_margin": 0.031108349561691284 + }, + { + "position": 21, + "token_id": 18718, + "token": "å¹³æĸ¹", + "decoded_piece": "平方", + "experts_by_weight": [ + { + "expert": 27, + "weight": 0.25225111842155457 + }, + { + "expert": 26, + "weight": 0.05924615636467934 + }, + { + "expert": 39, + "weight": 0.046080317348241806 + }, + { + "expert": 49, + "weight": 0.03298409655690193 + }, + { + "expert": 3, + "weight": 0.029271787032485008 + }, + { + "expert": 24, + "weight": 0.028183098882436752 + } + ], + "selected_weight_sum": 0.4480165746062994, + "top1_top2_margin": 0.19300496205687523 + }, + { + "position": 22, + "token_id": 14855, + "token": "åĬłä¸Ĭ", + "decoded_piece": "加上", + "experts_by_weight": [ + { + "expert": 26, + "weight": 0.14583228528499603 + }, + { + "expert": 51, + "weight": 0.10939821600914001 + }, + { + "expert": 16, + "weight": 0.05952773243188858 + }, + { + "expert": 47, + "weight": 0.04561968520283699 + }, + { + "expert": 60, + "weight": 0.034100957214832306 + }, + { + "expert": 41, + "weight": 0.030454030260443687 + } + ], + "selected_weight_sum": 0.4249329064041376, + "top1_top2_margin": 0.03643406927585602 + }, + { + "position": 23, + "token_id": 8270, + "token": "åĢĴ", + "decoded_piece": "倒", + "experts_by_weight": [ + { + "expert": 39, + "weight": 0.09405837953090668 + }, + { + "expert": 14, + "weight": 0.060965847223997116 + }, + { + "expert": 61, + "weight": 0.047729793936014175 + }, + { + "expert": 33, + "weight": 0.04451792314648628 + }, + { + "expert": 26, + "weight": 0.04123474285006523 + }, + { + "expert": 51, + "weight": 0.03994479775428772 + } + ], + "selected_weight_sum": 0.3284514844417572, + "top1_top2_margin": 0.03309253230690956 + }, + { + "position": 24, + "token_id": 51971, + "token": "æķ°çļĦ", + "decoded_piece": "数的", + "experts_by_weight": [ + { + "expert": 27, + "weight": 0.12274246662855148 + }, + { + "expert": 39, + "weight": 0.11193068325519562 + }, + { + "expert": 25, + "weight": 0.11071715503931046 + }, + { + "expert": 33, + "weight": 0.05617574602365494 + }, + { + "expert": 26, + "weight": 0.04331094026565552 + }, + { + "expert": 11, + "weight": 0.037420324981212616 + } + ], + "selected_weight_sum": 0.4822973161935806, + "top1_top2_margin": 0.010811783373355865 + }, + { + "position": 25, + "token_id": 18718, + "token": "å¹³æĸ¹", + "decoded_piece": "平方", + "experts_by_weight": [ + { + "expert": 27, + "weight": 0.24886038899421692 + }, + { + "expert": 26, + "weight": 0.056682538241147995 + }, + { + "expert": 39, + "weight": 0.04547945782542229 + }, + { + "expert": 62, + "weight": 0.03942480683326721 + }, + { + "expert": 49, + "weight": 0.027925729751586914 + }, + { + "expert": 3, + "weight": 0.025453777983784676 + } + ], + "selected_weight_sum": 0.443826699629426, + "top1_top2_margin": 0.19217785075306892 + }, + { + "position": 26, + "token_id": 398, + "token": "ãĢĤ", + "decoded_piece": "。", + "experts_by_weight": [ + { + "expert": 45, + "weight": 0.1179732158780098 + }, + { + "expert": 22, + "weight": 0.1001434326171875 + }, + { + "expert": 55, + "weight": 0.08874468505382538 + }, + { + "expert": 7, + "weight": 0.06503507494926453 + }, + { + "expert": 5, + "weight": 0.056857381016016006 + }, + { + "expert": 60, + "weight": 0.052832894027233124 + } + ], + "selected_weight_sum": 0.48158668354153633, + "top1_top2_margin": 0.017829783260822296 + } + ] + } + ], + "prompt_pair_jaccard": [ + { + "left": "zh_explanation", + "right": "en_architecture", + "used_expert_jaccard": 0.7543859649122807, + "shared_experts": 43, + "union_experts": 57 + }, + { + "left": "zh_explanation", + "right": "code", + "used_expert_jaccard": 0.6833333333333333, + "shared_experts": 41, + "union_experts": 60 + }, + { + "left": "zh_explanation", + "right": "math", + "used_expert_jaccard": 0.75, + "shared_experts": 45, + "union_experts": 60 + }, + { + "left": "en_architecture", + "right": "code", + "used_expert_jaccard": 0.6181818181818182, + "shared_experts": 34, + "union_experts": 55 + }, + { + "left": "en_architecture", + "right": "math", + "used_expert_jaccard": 0.6607142857142857, + "shared_experts": 37, + "union_experts": 56 + }, + { + "left": "code", + "right": "math", + "used_expert_jaccard": 0.6491228070175439, + "shared_experts": 37, + "union_experts": 57 + } + ] + } + }, + { + "layer": 4, + "ffn": "moe", + "state_tensors": 203, + "state_numel": 584847872, + "state_bytes": 1169695744, + "timing_ms": { + "load_cpu": 24.96491398778744, + "transfer_to_device": 123.27372501022182, + "forward": 3.8379190082196146 + }, + "peak_device_allocated_mib": 1176.2255859375, + "hidden_before": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 2.651595115661621, + "p10": 0.14781513810157776, + "p50": 0.1792253851890564, + "p90": 0.22313819825649261 + }, + "l2_norm": { + "mean": 119.99750518798828, + "p10": 6.68934965133667, + "p50": 8.110815048217773, + "p90": 10.098081588745117 + }, + "finite": true, + "sha256_fp32": "aa223453a9b51ecdb244412002e27c70edbcaeeb29a6e1e064f340622ff4044c" + }, + "hidden_after": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 3.085932731628418, + "p10": 0.19619415700435638, + "p50": 0.23260724544525146, + "p90": 0.2969282567501068 + }, + "l2_norm": { + "mean": 139.65338134765625, + "p10": 8.878734588623047, + "p50": 10.526601791381836, + "p90": 13.43743896484375 + }, + "finite": true, + "sha256_fp32": "5d5d80c17f47a78f419f6ec9bf16668392a8246393ebfe2b0233841be4875b6e" + }, + "mla": { + "compressed_projection_shape": [ + 4, + 27, + 576 + ], + "latent_shape": [ + 4, + 27, + 512 + ], + "rope_key_shape": [ + 4, + 27, + 64 + ], + "latent_rms": 2.3880224227905273, + "rope_key_rms": 2.2479918003082275, + "eager_key_cache_shape": [ + 4, + 16, + 27, + 192 + ], + "eager_value_cache_shape": [ + 4, + 16, + 27, + 128 + ], + "eager_cache_bytes": 1105920 + }, + "routing": { + "aggregate_load": [ + 1, + 7, + 3, + 8, + 9, + 3, + 4, + 20, + 9, + 18, + 5, + 11, + 4, + 0, + 5, + 2, + 5, + 12, + 3, + 4, + 11, + 7, + 10, + 12, + 16, + 5, + 6, + 6, + 4, + 37, + 10, + 17, + 3, + 5, + 9, + 11, + 10, + 6, + 7, + 9, + 7, + 13, + 7, + 1, + 1, + 3, + 4, + 18, + 40, + 8, + 10, + 1, + 9, + 1, + 17, + 15, + 6, + 11, + 3, + 3, + 0, + 13, + 11, + 4 + ], + "aggregate_metrics": { + "routes": 540, + "used_experts": 62, + "zero_experts": 2, + "mean": 8.4375, + "std": 7.309213638305664, + "cv": 0.8662771582603455, + "gini": 0.4173611104488373, + "entropy_nats": 3.8508799076080322, + "effective_experts": 47.03443098267191, + "min": 0.0, + "p50": 7.0, + "p90": 16.699996948242188, + "max": 40.0 + }, + "top_experts": [ + { + "expert": 48, + "routes": 40 + }, + { + "expert": 29, + "routes": 37 + }, + { + "expert": 7, + "routes": 20 + }, + { + "expert": 9, + "routes": 18 + }, + { + "expert": 47, + "routes": 18 + }, + { + "expert": 31, + "routes": 17 + }, + { + "expert": 54, + "routes": 17 + }, + { + "expert": 24, + "routes": 16 + }, + { + "expert": 55, + "routes": 15 + }, + { + "expert": 41, + "routes": 13 + }, + { + "expert": 61, + "routes": 13 + }, + { + "expert": 17, + "routes": 12 + } + ], + "selected_weight_sum": { + "mean": 0.4748785197734833, + "p10": 0.36896076798439026, + "p50": 0.45237767696380615, + "p90": 0.5860763788223267 + }, + "top1_top2_margin": { + "mean": 0.08469092845916748, + "p50": 0.03858078271150589, + "p90": 0.18687288463115692 + }, + "per_prompt": [ + { + "id": "zh_explanation", + "label": "中文解释", + "tokens": 26, + "load": [ + 0, + 3, + 1, + 1, + 5, + 2, + 1, + 3, + 5, + 6, + 0, + 5, + 0, + 0, + 1, + 1, + 2, + 3, + 0, + 0, + 3, + 2, + 3, + 3, + 4, + 2, + 1, + 4, + 2, + 16, + 3, + 2, + 0, + 1, + 3, + 2, + 2, + 1, + 3, + 4, + 0, + 5, + 1, + 0, + 1, + 1, + 0, + 6, + 11, + 2, + 1, + 0, + 4, + 0, + 6, + 5, + 2, + 1, + 0, + 0, + 0, + 5, + 4, + 1 + ], + "metrics": { + "routes": 156, + "used_experts": 49, + "zero_experts": 15, + "mean": 2.4375, + "std": 2.7306315898895264, + "cv": 1.120259165763855, + "gini": 0.5314503312110901, + "entropy_nats": 3.623417377471924, + "effective_experts": 37.46538265656308, + "min": 0.0, + "p50": 2.0, + "p90": 5.0, + "max": 16.0 + }, + "top_experts": [ + { + "expert": 29, + "routes": 16 + }, + { + "expert": 48, + "routes": 11 + }, + { + "expert": 9, + "routes": 6 + }, + { + "expert": 47, + "routes": 6 + }, + { + "expert": 54, + "routes": 6 + }, + { + "expert": 4, + "routes": 5 + }, + { + "expert": 8, + "routes": 5 + }, + { + "expert": 11, + "routes": 5 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 21, + "weight": 0.2854343056678772 + }, + { + "expert": 49, + "weight": 0.028558792546391487 + }, + { + "expert": 33, + "weight": 0.026019850745797157 + }, + { + "expert": 31, + "weight": 0.018602484837174416 + }, + { + "expert": 14, + "weight": 0.01662612520158291 + }, + { + "expert": 37, + "weight": 0.016511129215359688 + } + ], + "selected_weight_sum": 0.39175268821418285, + "top1_top2_margin": 0.2568755131214857 + }, + { + "position": 1, + "token_id": 1082, + "token": "ç͍", + "decoded_piece": "用", + "experts_by_weight": [ + { + "expert": 29, + "weight": 0.3073623776435852 + }, + { + "expert": 31, + "weight": 0.058887239545583725 + }, + { + "expert": 38, + "weight": 0.04573711007833481 + }, + { + "expert": 42, + "weight": 0.036742717027664185 + }, + { + "expert": 34, + "weight": 0.028751496225595474 + }, + { + "expert": 9, + "weight": 0.024523386731743813 + } + ], + "selected_weight_sum": 0.5020043272525072, + "top1_top2_margin": 0.24847513809800148 + }, + { + "position": 2, + "token_id": 83983, + "token": "éĢļä¿Ĺ", + "decoded_piece": "通俗", + "experts_by_weight": [ + { + "expert": 29, + "weight": 0.139014333486557 + }, + { + "expert": 47, + "weight": 0.1303897500038147 + }, + { + "expert": 54, + "weight": 0.07262816280126572 + }, + { + "expert": 25, + "weight": 0.06052057072520256 + }, + { + "expert": 9, + "weight": 0.04703924432396889 + }, + { + "expert": 27, + "weight": 0.03525257483124733 + } + ], + "selected_weight_sum": 0.4848446361720562, + "top1_top2_margin": 0.00862458348274231 + }, + { + "position": 3, + "token_id": 91476, + "token": "çļĦè¯Ńè¨Ģ", + "decoded_piece": "的语言", + "experts_by_weight": [ + { + "expert": 41, + "weight": 0.19071631133556366 + }, + { + "expert": 9, + "weight": 0.09350986033678055 + }, + { + "expert": 24, + "weight": 0.082207590341568 + }, + { + "expert": 29, + "weight": 0.06859856098890305 + }, + { + "expert": 25, + "weight": 0.06743662059307098 + }, + { + "expert": 55, + "weight": 0.059965796768665314 + } + ], + "selected_weight_sum": 0.5624347403645515, + "top1_top2_margin": 0.09720645099878311 + }, + { + "position": 4, + "token_id": 17882, + "token": "è§£éĩĬ", + "decoded_piece": "解释", + "experts_by_weight": [ + { + "expert": 41, + "weight": 0.16050276160240173 + }, + { + "expert": 9, + "weight": 0.1405077874660492 + }, + { + "expert": 29, + "weight": 0.10335756838321686 + }, + { + "expert": 20, + "weight": 0.06643447279930115 + }, + { + "expert": 61, + "weight": 0.057550542056560516 + }, + { + "expert": 50, + "weight": 0.040715765208005905 + } + ], + "selected_weight_sum": 0.5690688975155354, + "top1_top2_margin": 0.01999497413635254 + }, + { + "position": 5, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 29, + "weight": 0.17840619385242462 + }, + { + "expert": 41, + "weight": 0.13315673172473907 + }, + { + "expert": 39, + "weight": 0.13139554858207703 + }, + { + "expert": 52, + "weight": 0.1226491779088974 + }, + { + "expert": 22, + "weight": 0.06122437119483948 + }, + { + "expert": 5, + "weight": 0.035768091678619385 + } + ], + "selected_weight_sum": 0.662600114941597, + "top1_top2_margin": 0.04524946212768555 + }, + { + "position": 6, + "token_id": 10285, + "token": "为ä»Ģä¹Ī", + "decoded_piece": "为什么", + "experts_by_weight": [ + { + "expert": 29, + "weight": 0.20878379046916962 + }, + { + "expert": 9, + "weight": 0.11653995513916016 + }, + { + "expert": 30, + "weight": 0.06885909289121628 + }, + { + "expert": 41, + "weight": 0.06782285124063492 + }, + { + "expert": 8, + "weight": 0.04838939756155014 + }, + { + "expert": 17, + "weight": 0.026638401672244072 + } + ], + "selected_weight_sum": 0.5370334889739752, + "top1_top2_margin": 0.09224383533000946 + }, + { + "position": 7, + "token_id": 21844, + "token": "ç¨Ģ", + "decoded_piece": "稀", + "experts_by_weight": [ + { + "expert": 29, + "weight": 0.07411029189825058 + }, + { + "expert": 36, + "weight": 0.061866600066423416 + }, + { + "expert": 54, + "weight": 0.060064975172281265 + }, + { + "expert": 56, + "weight": 0.047518208622932434 + }, + { + "expert": 11, + "weight": 0.04366019740700722 + }, + { + "expert": 49, + "weight": 0.0434911884367466 + } + ], + "selected_weight_sum": 0.3307114616036415, + "top1_top2_margin": 0.012243691831827164 + }, + { + "position": 8, + "token_id": 20923, + "token": "çĸı", + "decoded_piece": "疏", + "experts_by_weight": [ + { + "expert": 27, + "weight": 0.09887184202671051 + }, + { + "expert": 54, + "weight": 0.0959555059671402 + }, + { + "expert": 56, + "weight": 0.08672784268856049 + }, + { + "expert": 48, + "weight": 0.07467734813690186 + }, + { + "expert": 36, + "weight": 0.05776987597346306 + }, + { + "expert": 1, + "weight": 0.04089703410863876 + } + ], + "selected_weight_sum": 0.45489944890141487, + "top1_top2_margin": 0.0029163360595703125 + }, + { + "position": 9, + "token_id": 10900, + "token": "ä¸ĵå®¶", + "decoded_piece": "专家", + "experts_by_weight": [ + { + "expert": 62, + "weight": 0.13037250936031342 + }, + { + "expert": 15, + "weight": 0.10683903843164444 + }, + { + "expert": 1, + "weight": 0.10298178344964981 + }, + { + "expert": 6, + "weight": 0.06250814348459244 + }, + { + "expert": 28, + "weight": 0.05680561065673828 + }, + { + "expert": 29, + "weight": 0.03991430625319481 + } + ], + "selected_weight_sum": 0.4994213916361332, + "top1_top2_margin": 0.023533470928668976 + }, + { + "position": 10, + "token_id": 39322, + "token": "模åŀĭ", + "decoded_piece": "模型", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.13188892602920532 + }, + { + "expert": 62, + "weight": 0.104781374335289 + }, + { + "expert": 28, + "weight": 0.09326699376106262 + }, + { + "expert": 1, + "weight": 0.08980964124202728 + }, + { + "expert": 24, + "weight": 0.06314408034086227 + }, + { + "expert": 7, + "weight": 0.04349623620510101 + } + ], + "selected_weight_sum": 0.5263872519135475, + "top1_top2_margin": 0.02710755169391632 + }, + { + "position": 11, + "token_id": 1876, + "token": "åı¯ä»¥", + "decoded_piece": "可以", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.09696580469608307 + }, + { + "expert": 29, + "weight": 0.09488947689533234 + }, + { + "expert": 23, + "weight": 0.07881587743759155 + }, + { + "expert": 57, + "weight": 0.049439214169979095 + }, + { + "expert": 48, + "weight": 0.04129276052117348 + }, + { + "expert": 9, + "weight": 0.04105379432439804 + } + ], + "selected_weight_sum": 0.40245692804455757, + "top1_top2_margin": 0.0020763278007507324 + }, + { + "position": 12, + "token_id": 9476, + "token": "æĭ¥æľī", + "decoded_piece": "拥有", + "experts_by_weight": [ + { + "expert": 23, + "weight": 0.10077859461307526 + }, + { + "expert": 61, + "weight": 0.09732674062252045 + }, + { + "expert": 8, + "weight": 0.07639233767986298 + }, + { + "expert": 48, + "weight": 0.06467738747596741 + }, + { + "expert": 20, + "weight": 0.058186598122119904 + }, + { + "expert": 29, + "weight": 0.04759038984775543 + } + ], + "selected_weight_sum": 0.4449520483613014, + "top1_top2_margin": 0.0034518539905548096 + }, + { + "position": 13, + "token_id": 3599, + "token": "å¾Īå¤ļ", + "decoded_piece": "很多", + "experts_by_weight": [ + { + "expert": 54, + "weight": 0.08523216098546982 + }, + { + "expert": 47, + "weight": 0.07970038056373596 + }, + { + "expert": 27, + "weight": 0.06723999232053757 + }, + { + "expert": 4, + "weight": 0.05728509649634361 + }, + { + "expert": 61, + "weight": 0.05710846558213234 + }, + { + "expert": 29, + "weight": 0.05230521410703659 + } + ], + "selected_weight_sum": 0.3988713100552559, + "top1_top2_margin": 0.005531780421733856 + }, + { + "position": 14, + "token_id": 21273, + "token": "åıĤæķ°", + "decoded_piece": "参数", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.13637880980968475 + }, + { + "expert": 48, + "weight": 0.13288584351539612 + }, + { + "expert": 55, + "weight": 0.09305819869041443 + }, + { + "expert": 24, + "weight": 0.08986103534698486 + }, + { + "expert": 62, + "weight": 0.06073031201958656 + }, + { + "expert": 35, + "weight": 0.046965353190898895 + } + ], + "selected_weight_sum": 0.5598795525729656, + "top1_top2_margin": 0.0034929662942886353 + }, + { + "position": 15, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.1541302353143692 + }, + { + "expert": 29, + "weight": 0.13401253521442413 + }, + { + "expert": 22, + "weight": 0.12823665142059326 + }, + { + "expert": 39, + "weight": 0.10574101656675339 + }, + { + "expert": 16, + "weight": 0.06683357059955597 + }, + { + "expert": 34, + "weight": 0.06087338924407959 + } + ], + "selected_weight_sum": 0.6498273983597755, + "top1_top2_margin": 0.02011770009994507 + }, + { + "position": 16, + "token_id": 1705, + "token": "ä½Ĩ", + "decoded_piece": "但", + "experts_by_weight": [ + { + "expert": 16, + "weight": 0.10471916198730469 + }, + { + "expert": 39, + "weight": 0.08865265548229218 + }, + { + "expert": 30, + "weight": 0.08270280063152313 + }, + { + "expert": 29, + "weight": 0.06964780390262604 + }, + { + "expert": 52, + "weight": 0.053717441856861115 + }, + { + "expert": 48, + "weight": 0.031225744634866714 + } + ], + "selected_weight_sum": 0.43066560849547386, + "top1_top2_margin": 0.016066506505012512 + }, + { + "position": 17, + "token_id": 13712, + "token": "æ¯ı个", + "decoded_piece": "每个", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.11605170369148254 + }, + { + "expert": 30, + "weight": 0.09560476988554001 + }, + { + "expert": 29, + "weight": 0.0483170710504055 + }, + { + "expert": 11, + "weight": 0.045952338725328445 + }, + { + "expert": 48, + "weight": 0.04325094819068909 + }, + { + "expert": 47, + "weight": 0.03830089792609215 + } + ], + "selected_weight_sum": 0.38747772946953773, + "top1_top2_margin": 0.020446933805942535 + }, + { + "position": 18, + "token_id": 10728, + "token": "Ġtoken", + "decoded_piece": " token", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.15618282556533813 + }, + { + "expert": 7, + "weight": 0.05938476696610451 + }, + { + "expert": 35, + "weight": 0.057606328278779984 + }, + { + "expert": 55, + "weight": 0.056335169821977615 + }, + { + "expert": 24, + "weight": 0.050878867506980896 + }, + { + "expert": 62, + "weight": 0.04674487188458443 + } + ], + "selected_weight_sum": 0.42713283002376556, + "top1_top2_margin": 0.09679805859923363 + }, + { + "position": 19, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 38, + "weight": 0.9936936497688293 + }, + { + "expert": 41, + "weight": 0.0005532832001335919 + }, + { + "expert": 29, + "weight": 0.00046768385800532997 + }, + { + "expert": 45, + "weight": 0.0003529697423800826 + }, + { + "expert": 63, + "weight": 0.0003311999316792935 + }, + { + "expert": 26, + "weight": 0.0002939346304628998 + } + ], + "selected_weight_sum": 0.9956927211314905, + "top1_top2_margin": 0.9931403665686958 + }, + { + "position": 20, + "token_id": 2141, + "token": "åıª", + "decoded_piece": "只", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.12353485822677612 + }, + { + "expert": 38, + "weight": 0.09975308179855347 + }, + { + "expert": 44, + "weight": 0.03714262694120407 + }, + { + "expert": 11, + "weight": 0.03439706191420555 + }, + { + "expert": 17, + "weight": 0.03421425074338913 + }, + { + "expert": 21, + "weight": 0.031277887523174286 + } + ], + "selected_weight_sum": 0.3603197671473026, + "top1_top2_margin": 0.023781776428222656 + }, + { + "position": 21, + "token_id": 5118, + "token": "使ç͍", + "decoded_piece": "使用", + "experts_by_weight": [ + { + "expert": 61, + "weight": 0.1071867048740387 + }, + { + "expert": 20, + "weight": 0.1065455824136734 + }, + { + "expert": 8, + "weight": 0.05777136981487274 + }, + { + "expert": 48, + "weight": 0.04117731377482414 + }, + { + "expert": 23, + "weight": 0.0401187427341938 + }, + { + "expert": 17, + "weight": 0.03573920205235481 + } + ], + "selected_weight_sum": 0.3885389156639576, + "top1_top2_margin": 0.0006411224603652954 + }, + { + "position": 22, + "token_id": 6802, + "token": "åħ¶ä¸Ń", + "decoded_piece": "其中", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.16480860114097595 + }, + { + "expert": 47, + "weight": 0.07635561376810074 + }, + { + "expert": 5, + "weight": 0.04757159948348999 + }, + { + "expert": 11, + "weight": 0.044671352952718735 + }, + { + "expert": 55, + "weight": 0.041422437876462936 + }, + { + "expert": 61, + "weight": 0.041200462728738785 + } + ], + "selected_weight_sum": 0.41603006795048714, + "top1_top2_margin": 0.08845298737287521 + }, + { + "position": 23, + "token_id": 51905, + "token": "ä¸Ģå°ı", + "decoded_piece": "一小", + "experts_by_weight": [ + { + "expert": 47, + "weight": 0.14340093731880188 + }, + { + "expert": 4, + "weight": 0.0659688413143158 + }, + { + "expert": 11, + "weight": 0.06596291810274124 + }, + { + "expert": 54, + "weight": 0.05866997316479683 + }, + { + "expert": 2, + "weight": 0.046418167650699615 + }, + { + "expert": 48, + "weight": 0.04551021382212639 + } + ], + "selected_weight_sum": 0.42593105137348175, + "top1_top2_margin": 0.07743209600448608 + }, + { + "position": 24, + "token_id": 5781, + "token": "éĥ¨åĪĨ", + "decoded_piece": "部分", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.128562793135643 + }, + { + "expert": 47, + "weight": 0.09527810662984848 + }, + { + "expert": 54, + "weight": 0.07127159088850021 + }, + { + "expert": 27, + "weight": 0.060349851846694946 + }, + { + "expert": 55, + "weight": 0.04831140115857124 + }, + { + "expert": 48, + "weight": 0.044088590890169144 + } + ], + "selected_weight_sum": 0.44786233454942703, + "top1_top2_margin": 0.033284686505794525 + }, + { + "position": 25, + "token_id": 398, + "token": "ãĢĤ", + "decoded_piece": "。", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.15205636620521545 + }, + { + "expert": 22, + "weight": 0.136318176984787 + }, + { + "expert": 34, + "weight": 0.095779187977314 + }, + { + "expert": 39, + "weight": 0.09394431859254837 + }, + { + "expert": 29, + "weight": 0.08958881348371506 + }, + { + "expert": 3, + "weight": 0.07937369495630264 + } + ], + "selected_weight_sum": 0.6470605581998825, + "top1_top2_margin": 0.015738189220428467 + } + ] + }, + { + "id": "en_architecture", + "label": "English architecture", + "tokens": 18, + "load": [ + 0, + 0, + 2, + 1, + 1, + 0, + 0, + 7, + 1, + 3, + 0, + 0, + 4, + 0, + 1, + 1, + 0, + 3, + 1, + 0, + 1, + 1, + 1, + 1, + 4, + 1, + 1, + 0, + 1, + 5, + 2, + 6, + 2, + 1, + 2, + 2, + 6, + 1, + 1, + 1, + 0, + 2, + 1, + 1, + 0, + 0, + 0, + 4, + 11, + 1, + 6, + 0, + 1, + 1, + 3, + 2, + 2, + 1, + 0, + 0, + 0, + 1, + 6, + 0 + ], + "metrics": { + "routes": 108, + "used_experts": 45, + "zero_experts": 19, + "mean": 1.6875, + "std": 2.1371617317199707, + "cv": 1.2664662599563599, + "gini": 0.5908564925193787, + "entropy_nats": 3.490825891494751, + "effective_experts": 32.81303652626308, + "min": 0.0, + "p50": 1.0, + "p90": 4.6999969482421875, + "max": 11.0 + }, + "top_experts": [ + { + "expert": 48, + "routes": 11 + }, + { + "expert": 7, + "routes": 7 + }, + { + "expert": 31, + "routes": 6 + }, + { + "expert": 36, + "routes": 6 + }, + { + "expert": 50, + "routes": 6 + }, + { + "expert": 62, + "routes": 6 + }, + { + "expert": 29, + "routes": 5 + }, + { + "expert": 12, + "routes": 4 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 21, + "weight": 0.2854343056678772 + }, + { + "expert": 49, + "weight": 0.028558792546391487 + }, + { + "expert": 33, + "weight": 0.026019850745797157 + }, + { + "expert": 31, + "weight": 0.018602484837174416 + }, + { + "expert": 14, + "weight": 0.01662612520158291 + }, + { + "expert": 37, + "weight": 0.016511129215359688 + } + ], + "selected_weight_sum": 0.39175268821418285, + "top1_top2_margin": 0.2568755131214857 + }, + { + "position": 1, + "token_id": 99297, + "token": "Explain", + "decoded_piece": "Explain", + "experts_by_weight": [ + { + "expert": 29, + "weight": 0.16727285087108612 + }, + { + "expert": 9, + "weight": 0.1338152289390564 + }, + { + "expert": 50, + "weight": 0.07673134654760361 + }, + { + "expert": 38, + "weight": 0.05953230708837509 + }, + { + "expert": 34, + "weight": 0.05908035859465599 + }, + { + "expert": 41, + "weight": 0.04080839455127716 + } + ], + "selected_weight_sum": 0.5372404865920544, + "top1_top2_margin": 0.033457621932029724 + }, + { + "position": 2, + "token_id": 946, + "token": "Ġhow", + "decoded_piece": " how", + "experts_by_weight": [ + { + "expert": 9, + "weight": 0.21537113189697266 + }, + { + "expert": 29, + "weight": 0.13318757712841034 + }, + { + "expert": 41, + "weight": 0.08081577718257904 + }, + { + "expert": 30, + "weight": 0.05205032229423523 + }, + { + "expert": 17, + "weight": 0.03154856711626053 + }, + { + "expert": 25, + "weight": 0.030927589163184166 + } + ], + "selected_weight_sum": 0.543900964781642, + "top1_top2_margin": 0.08218355476856232 + }, + { + "position": 3, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 47, + "weight": 0.1722368448972702 + }, + { + "expert": 29, + "weight": 0.13884633779525757 + }, + { + "expert": 9, + "weight": 0.11268851161003113 + }, + { + "expert": 30, + "weight": 0.043849557638168335 + }, + { + "expert": 17, + "weight": 0.023122970014810562 + }, + { + "expert": 4, + "weight": 0.020444709807634354 + } + ], + "selected_weight_sum": 0.5111889317631721, + "top1_top2_margin": 0.033390507102012634 + }, + { + "position": 4, + "token_id": 29640, + "token": "Ġcompressed", + "decoded_piece": " compressed", + "experts_by_weight": [ + { + "expert": 47, + "weight": 0.08341158926486969 + }, + { + "expert": 36, + "weight": 0.07845339179039001 + }, + { + "expert": 50, + "weight": 0.0757032185792923 + }, + { + "expert": 54, + "weight": 0.07314742356538773 + }, + { + "expert": 31, + "weight": 0.06734482198953629 + }, + { + "expert": 48, + "weight": 0.044416025280952454 + } + ], + "selected_weight_sum": 0.42247647047042847, + "top1_top2_margin": 0.004958197474479675 + }, + { + "position": 5, + "token_id": 2131, + "token": "Ġkey", + "decoded_piece": " key", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.10751821100711823 + }, + { + "expert": 7, + "weight": 0.10609526187181473 + }, + { + "expert": 32, + "weight": 0.07796686142683029 + }, + { + "expert": 24, + "weight": 0.05935220420360565 + }, + { + "expert": 31, + "weight": 0.051578737795352936 + }, + { + "expert": 62, + "weight": 0.04501781985163689 + } + ], + "selected_weight_sum": 0.4475290961563587, + "top1_top2_margin": 0.0014229491353034973 + }, + { + "position": 6, + "token_id": 12, + "token": "-", + "decoded_piece": "-", + "experts_by_weight": [ + { + "expert": 42, + "weight": 0.126819908618927 + }, + { + "expert": 26, + "weight": 0.10091755539178848 + }, + { + "expert": 2, + "weight": 0.0688774585723877 + }, + { + "expert": 32, + "weight": 0.058631282299757004 + }, + { + "expert": 36, + "weight": 0.04395686089992523 + }, + { + "expert": 31, + "weight": 0.04179288074374199 + } + ], + "selected_weight_sum": 0.4409959465265274, + "top1_top2_margin": 0.02590235322713852 + }, + { + "position": 7, + "token_id": 2834, + "token": "value", + "decoded_piece": "value", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.2559155821800232 + }, + { + "expert": 7, + "weight": 0.11273658275604248 + }, + { + "expert": 35, + "weight": 0.09117651730775833 + }, + { + "expert": 24, + "weight": 0.08211402595043182 + }, + { + "expert": 62, + "weight": 0.036896999925374985 + }, + { + "expert": 55, + "weight": 0.033982954919338226 + } + ], + "selected_weight_sum": 0.612822663038969, + "top1_top2_margin": 0.1431789994239807 + }, + { + "position": 8, + "token_id": 28937, + "token": "Ġlatent", + "decoded_piece": " latent", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.17185746133327484 + }, + { + "expert": 54, + "weight": 0.08515983074903488 + }, + { + "expert": 47, + "weight": 0.06257638335227966 + }, + { + "expert": 7, + "weight": 0.04313451051712036 + }, + { + "expert": 24, + "weight": 0.040846388787031174 + }, + { + "expert": 36, + "weight": 0.040019016712903976 + } + ], + "selected_weight_sum": 0.4435935914516449, + "top1_top2_margin": 0.08669763058423996 + }, + { + "position": 9, + "token_id": 4183, + "token": "Ġchanges", + "decoded_piece": " changes", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.09715737402439117 + }, + { + "expert": 56, + "weight": 0.09069112688302994 + }, + { + "expert": 8, + "weight": 0.056275662034749985 + }, + { + "expert": 7, + "weight": 0.0479937382042408 + }, + { + "expert": 12, + "weight": 0.04470612481236458 + }, + { + "expert": 50, + "weight": 0.04298648610711098 + } + ], + "selected_weight_sum": 0.37981051206588745, + "top1_top2_margin": 0.006466247141361237 + }, + { + "position": 10, + "token_id": 254, + "token": "Ġthe", + "decoded_piece": " the", + "experts_by_weight": [ + { + "expert": 61, + "weight": 0.08081479370594025 + }, + { + "expert": 56, + "weight": 0.06830991804599762 + }, + { + "expert": 20, + "weight": 0.06264370679855347 + }, + { + "expert": 47, + "weight": 0.053321268409490585 + }, + { + "expert": 50, + "weight": 0.043203629553318024 + }, + { + "expert": 29, + "weight": 0.04058090224862099 + } + ], + "selected_weight_sum": 0.34887421876192093, + "top1_top2_margin": 0.012504875659942627 + }, + { + "position": 11, + "token_id": 4887, + "token": "Ġmemory", + "decoded_piece": " memory", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.12293228507041931 + }, + { + "expert": 62, + "weight": 0.0906774029135704 + }, + { + "expert": 7, + "weight": 0.08118465542793274 + }, + { + "expert": 12, + "weight": 0.07018449157476425 + }, + { + "expert": 24, + "weight": 0.06667269021272659 + }, + { + "expert": 36, + "weight": 0.04963458329439163 + } + ], + "selected_weight_sum": 0.48128610849380493, + "top1_top2_margin": 0.03225488215684891 + }, + { + "position": 12, + "token_id": 2295, + "token": "Ġcost", + "decoded_piece": " cost", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.16012635827064514 + }, + { + "expert": 7, + "weight": 0.1397150307893753 + }, + { + "expert": 55, + "weight": 0.08156415820121765 + }, + { + "expert": 35, + "weight": 0.04813092201948166 + }, + { + "expert": 31, + "weight": 0.0403243824839592 + }, + { + "expert": 53, + "weight": 0.03788750618696213 + } + ], + "selected_weight_sum": 0.5077483579516411, + "top1_top2_margin": 0.020411327481269836 + }, + { + "position": 13, + "token_id": 280, + "token": "Ġof", + "decoded_piece": " of", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.11296866834163666 + }, + { + "expert": 7, + "weight": 0.10245229303836823 + }, + { + "expert": 57, + "weight": 0.09608335793018341 + }, + { + "expert": 23, + "weight": 0.05060618370771408 + }, + { + "expert": 29, + "weight": 0.03964516520500183 + }, + { + "expert": 48, + "weight": 0.03191471844911575 + } + ], + "selected_weight_sum": 0.43367038667201996, + "top1_top2_margin": 0.010516375303268433 + }, + { + "position": 14, + "token_id": 77939, + "token": "Ġautore", + "decoded_piece": " autore", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.11619260907173157 + }, + { + "expert": 2, + "weight": 0.06598556041717529 + }, + { + "expert": 36, + "weight": 0.05050069838762283 + }, + { + "expert": 62, + "weight": 0.04955204576253891 + }, + { + "expert": 12, + "weight": 0.043280262500047684 + }, + { + "expert": 15, + "weight": 0.03925475850701332 + } + ], + "selected_weight_sum": 0.3647659346461296, + "top1_top2_margin": 0.050207048654556274 + }, + { + "position": 15, + "token_id": 79184, + "token": "gressive", + "decoded_piece": "gressive", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.15186402201652527 + }, + { + "expert": 54, + "weight": 0.06799350678920746 + }, + { + "expert": 12, + "weight": 0.043805334717035294 + }, + { + "expert": 62, + "weight": 0.04156063124537468 + }, + { + "expert": 50, + "weight": 0.03523245453834534 + }, + { + "expert": 36, + "weight": 0.03416448459029198 + } + ], + "selected_weight_sum": 0.37462043389678, + "top1_top2_margin": 0.08387051522731781 + }, + { + "position": 16, + "token_id": 39671, + "token": "Ġdecoding", + "decoded_piece": " decoding", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.14867179095745087 + }, + { + "expert": 50, + "weight": 0.1063278466463089 + }, + { + "expert": 62, + "weight": 0.1001177579164505 + }, + { + "expert": 28, + "weight": 0.05450600013136864 + }, + { + "expert": 43, + "weight": 0.05367346480488777 + }, + { + "expert": 31, + "weight": 0.0394548736512661 + } + ], + "selected_weight_sum": 0.5027517341077328, + "top1_top2_margin": 0.04234394431114197 + }, + { + "position": 17, + "token_id": 13, + "token": ".", + "decoded_piece": ".", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.14864285290241241 + }, + { + "expert": 22, + "weight": 0.12844759225845337 + }, + { + "expert": 39, + "weight": 0.082750603556633 + }, + { + "expert": 34, + "weight": 0.08111844211816788 + }, + { + "expert": 3, + "weight": 0.07854798436164856 + }, + { + "expert": 18, + "weight": 0.0698799341917038 + } + ], + "selected_weight_sum": 0.589387409389019, + "top1_top2_margin": 0.020195260643959045 + } + ] + }, + { + "id": "code", + "label": "Python code", + "tokens": 19, + "load": [ + 0, + 2, + 0, + 1, + 1, + 0, + 2, + 2, + 1, + 8, + 1, + 0, + 0, + 0, + 1, + 0, + 1, + 3, + 0, + 2, + 3, + 1, + 2, + 4, + 5, + 1, + 0, + 1, + 1, + 7, + 2, + 2, + 1, + 1, + 2, + 4, + 1, + 1, + 1, + 2, + 1, + 2, + 0, + 0, + 0, + 0, + 1, + 6, + 11, + 1, + 3, + 0, + 2, + 0, + 3, + 4, + 2, + 5, + 2, + 0, + 0, + 3, + 1, + 0 + ], + "metrics": { + "routes": 114, + "used_experts": 46, + "zero_experts": 18, + "mean": 1.78125, + "std": 2.1117115020751953, + "cv": 1.185522198677063, + "gini": 0.562774121761322, + "entropy_nats": 3.5553629398345947, + "effective_experts": 35.00052074595525, + "min": 0.0, + "p50": 1.0, + "p90": 4.0, + "max": 11.0 + }, + "top_experts": [ + { + "expert": 48, + "routes": 11 + }, + { + "expert": 9, + "routes": 8 + }, + { + "expert": 29, + "routes": 7 + }, + { + "expert": 47, + "routes": 6 + }, + { + "expert": 24, + "routes": 5 + }, + { + "expert": 57, + "routes": 5 + }, + { + "expert": 23, + "routes": 4 + }, + { + "expert": 35, + "routes": 4 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 21, + "weight": 0.2854343056678772 + }, + { + "expert": 49, + "weight": 0.028558792546391487 + }, + { + "expert": 33, + "weight": 0.026019850745797157 + }, + { + "expert": 31, + "weight": 0.018602484837174416 + }, + { + "expert": 14, + "weight": 0.01662612520158291 + }, + { + "expert": 37, + "weight": 0.016511129215359688 + } + ], + "selected_weight_sum": 0.39175268821418285, + "top1_top2_margin": 0.2568755131214857 + }, + { + "position": 1, + "token_id": 9081, + "token": "Write", + "decoded_piece": "Write", + "experts_by_weight": [ + { + "expert": 29, + "weight": 0.1973877102136612 + }, + { + "expert": 31, + "weight": 0.08133529126644135 + }, + { + "expert": 9, + "weight": 0.07558465003967285 + }, + { + "expert": 50, + "weight": 0.07121656835079193 + }, + { + "expert": 38, + "weight": 0.0485970601439476 + }, + { + "expert": 34, + "weight": 0.04800603911280632 + } + ], + "selected_weight_sum": 0.5221273191273212, + "top1_top2_margin": 0.11605241894721985 + }, + { + "position": 2, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 47, + "weight": 0.2108405977487564 + }, + { + "expert": 29, + "weight": 0.12334326654672623 + }, + { + "expert": 9, + "weight": 0.1129695475101471 + }, + { + "expert": 25, + "weight": 0.06092284992337227 + }, + { + "expert": 61, + "weight": 0.03568430244922638 + }, + { + "expert": 1, + "weight": 0.02485974319279194 + } + ], + "selected_weight_sum": 0.5686203073710203, + "top1_top2_margin": 0.08749733120203018 + }, + { + "position": 3, + "token_id": 12974, + "token": "ĠPython", + "decoded_piece": " Python", + "experts_by_weight": [ + { + "expert": 24, + "weight": 0.09828471392393112 + }, + { + "expert": 1, + "weight": 0.08841812610626221 + }, + { + "expert": 29, + "weight": 0.06621435284614563 + }, + { + "expert": 48, + "weight": 0.05853234604001045 + }, + { + "expert": 46, + "weight": 0.0506991446018219 + }, + { + "expert": 40, + "weight": 0.049311794340610504 + } + ], + "selected_weight_sum": 0.4114604778587818, + "top1_top2_margin": 0.009866587817668915 + }, + { + "position": 4, + "token_id": 1157, + "token": "Ġfunction", + "decoded_piece": " function", + "experts_by_weight": [ + { + "expert": 24, + "weight": 0.18939025700092316 + }, + { + "expert": 48, + "weight": 0.06420006603002548 + }, + { + "expert": 55, + "weight": 0.05571218580007553 + }, + { + "expert": 9, + "weight": 0.045223336666822433 + }, + { + "expert": 6, + "weight": 0.039364468306303024 + }, + { + "expert": 62, + "weight": 0.03576372563838959 + } + ], + "selected_weight_sum": 0.4296540394425392, + "top1_top2_margin": 0.12519019097089767 + }, + { + "position": 5, + "token_id": 344, + "token": "Ġthat", + "decoded_piece": " that", + "experts_by_weight": [ + { + "expert": 57, + "weight": 0.11677073687314987 + }, + { + "expert": 17, + "weight": 0.10798279941082001 + }, + { + "expert": 30, + "weight": 0.09661772847175598 + }, + { + "expert": 9, + "weight": 0.0699259340763092 + }, + { + "expert": 23, + "weight": 0.06776062399148941 + }, + { + "expert": 29, + "weight": 0.06706580519676208 + } + ], + "selected_weight_sum": 0.5261236280202866, + "top1_top2_margin": 0.008787937462329865 + }, + { + "position": 6, + "token_id": 7578, + "token": "Ġreturns", + "decoded_piece": " returns", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.08541617542505264 + }, + { + "expert": 23, + "weight": 0.08021987229585648 + }, + { + "expert": 48, + "weight": 0.06539089232683182 + }, + { + "expert": 9, + "weight": 0.05553075298666954 + }, + { + "expert": 57, + "weight": 0.03810269385576248 + }, + { + "expert": 50, + "weight": 0.03640105947852135 + } + ], + "selected_weight_sum": 0.3610614463686943, + "top1_top2_margin": 0.005196303129196167 + }, + { + "position": 7, + "token_id": 254, + "token": "Ġthe", + "decoded_piece": " the", + "experts_by_weight": [ + { + "expert": 47, + "weight": 0.11266642808914185 + }, + { + "expert": 61, + "weight": 0.07434440404176712 + }, + { + "expert": 9, + "weight": 0.06673765182495117 + }, + { + "expert": 27, + "weight": 0.06281394511461258 + }, + { + "expert": 23, + "weight": 0.0450727753341198 + }, + { + "expert": 20, + "weight": 0.039807505905628204 + } + ], + "selected_weight_sum": 0.4014427103102207, + "top1_top2_margin": 0.038322024047374725 + }, + { + "position": 8, + "token_id": 1022, + "token": "Ġfirst", + "decoded_piece": " first", + "experts_by_weight": [ + { + "expert": 47, + "weight": 0.1453605592250824 + }, + { + "expert": 48, + "weight": 0.12199264764785767 + }, + { + "expert": 19, + "weight": 0.054375119507312775 + }, + { + "expert": 10, + "weight": 0.052063606679439545 + }, + { + "expert": 54, + "weight": 0.045551251620054245 + }, + { + "expert": 35, + "weight": 0.04497215151786804 + } + ], + "selected_weight_sum": 0.46431533619761467, + "top1_top2_margin": 0.02336791157722473 + }, + { + "position": 9, + "token_id": 9842, + "token": "Ġrepeated", + "decoded_piece": " repeated", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.14767202734947205 + }, + { + "expert": 19, + "weight": 0.092751644551754 + }, + { + "expert": 54, + "weight": 0.08925823867321014 + }, + { + "expert": 47, + "weight": 0.058891624212265015 + }, + { + "expert": 35, + "weight": 0.04904596880078316 + }, + { + "expert": 56, + "weight": 0.040675900876522064 + } + ], + "selected_weight_sum": 0.4782954044640064, + "top1_top2_margin": 0.05492038279771805 + }, + { + "position": 10, + "token_id": 3699, + "token": "Ġelement", + "decoded_piece": " element", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.16412603855133057 + }, + { + "expert": 55, + "weight": 0.09820627421140671 + }, + { + "expert": 7, + "weight": 0.04715190455317497 + }, + { + "expert": 24, + "weight": 0.04653613641858101 + }, + { + "expert": 36, + "weight": 0.044597916305065155 + }, + { + "expert": 35, + "weight": 0.04324425011873245 + } + ], + "selected_weight_sum": 0.44386252015829086, + "top1_top2_margin": 0.06591976433992386 + }, + { + "position": 11, + "token_id": 279, + "token": "Ġin", + "decoded_piece": " in", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.1437619924545288 + }, + { + "expert": 57, + "weight": 0.12850943207740784 + }, + { + "expert": 20, + "weight": 0.082257479429245 + }, + { + "expert": 48, + "weight": 0.06908990442752838 + }, + { + "expert": 58, + "weight": 0.06730543076992035 + }, + { + "expert": 23, + "weight": 0.03207061067223549 + } + ], + "selected_weight_sum": 0.5229948498308659, + "top1_top2_margin": 0.015252560377120972 + }, + { + "position": 12, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 47, + "weight": 0.1672954112291336 + }, + { + "expert": 17, + "weight": 0.11855437606573105 + }, + { + "expert": 20, + "weight": 0.05668004974722862 + }, + { + "expert": 58, + "weight": 0.03611814230680466 + }, + { + "expert": 57, + "weight": 0.03381320834159851 + }, + { + "expert": 48, + "weight": 0.03321889415383339 + } + ], + "selected_weight_sum": 0.44568008184432983, + "top1_top2_margin": 0.04874103516340256 + }, + { + "position": 13, + "token_id": 1525, + "token": "Ġlist", + "decoded_piece": " list", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.11332352459430695 + }, + { + "expert": 24, + "weight": 0.11224807053804398 + }, + { + "expert": 35, + "weight": 0.07935135066509247 + }, + { + "expert": 55, + "weight": 0.07445082813501358 + }, + { + "expert": 28, + "weight": 0.03989847004413605 + }, + { + "expert": 32, + "weight": 0.030583657324314117 + } + ], + "selected_weight_sum": 0.44985590130090714, + "top1_top2_margin": 0.00107545405626297 + }, + { + "position": 14, + "token_id": 285, + "token": "Ġand", + "decoded_piece": " and", + "experts_by_weight": [ + { + "expert": 16, + "weight": 0.15972866117954254 + }, + { + "expert": 22, + "weight": 0.11525166779756546 + }, + { + "expert": 52, + "weight": 0.08245518058538437 + }, + { + "expert": 39, + "weight": 0.06731696426868439 + }, + { + "expert": 29, + "weight": 0.05268394947052002 + }, + { + "expert": 57, + "weight": 0.04680750146508217 + } + ], + "selected_weight_sum": 0.524243924766779, + "top1_top2_margin": 0.04447699338197708 + }, + { + "position": 15, + "token_id": 6716, + "token": "Ġexplain", + "decoded_piece": " explain", + "experts_by_weight": [ + { + "expert": 9, + "weight": 0.16870331764221191 + }, + { + "expert": 41, + "weight": 0.1038370430469513 + }, + { + "expert": 50, + "weight": 0.07737299054861069 + }, + { + "expert": 29, + "weight": 0.052502091974020004 + }, + { + "expert": 48, + "weight": 0.043238840997219086 + }, + { + "expert": 6, + "weight": 0.04061140865087509 + } + ], + "selected_weight_sum": 0.4862656928598881, + "top1_top2_margin": 0.06486627459526062 + }, + { + "position": 16, + "token_id": 895, + "token": "Ġits", + "decoded_piece": " its", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.16002975404262543 + }, + { + "expert": 9, + "weight": 0.13303902745246887 + }, + { + "expert": 61, + "weight": 0.060442566871643066 + }, + { + "expert": 30, + "weight": 0.05143991485238075 + }, + { + "expert": 41, + "weight": 0.0367891788482666 + }, + { + "expert": 47, + "weight": 0.030291805043816566 + } + ], + "selected_weight_sum": 0.4720322471112013, + "top1_top2_margin": 0.026990726590156555 + }, + { + "position": 17, + "token_id": 13913, + "token": "Ġcomplexity", + "decoded_piece": " complexity", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.12802357971668243 + }, + { + "expert": 48, + "weight": 0.08918403834104538 + }, + { + "expert": 54, + "weight": 0.08422041684389114 + }, + { + "expert": 56, + "weight": 0.06852987408638 + }, + { + "expert": 55, + "weight": 0.06726609915494919 + }, + { + "expert": 24, + "weight": 0.04570431262254715 + } + ], + "selected_weight_sum": 0.4829283207654953, + "top1_top2_margin": 0.038839541375637054 + }, + { + "position": 18, + "token_id": 13, + "token": ".", + "decoded_piece": ".", + "experts_by_weight": [ + { + "expert": 22, + "weight": 0.13413840532302856 + }, + { + "expert": 52, + "weight": 0.10561732202768326 + }, + { + "expert": 34, + "weight": 0.07496768981218338 + }, + { + "expert": 3, + "weight": 0.07266590744256973 + }, + { + "expert": 29, + "weight": 0.07079854607582092 + }, + { + "expert": 39, + "weight": 0.0666135922074318 + } + ], + "selected_weight_sum": 0.5248014628887177, + "top1_top2_margin": 0.028521083295345306 + } + ] + }, + { + "id": "math", + "label": "数学推理", + "tokens": 27, + "load": [ + 1, + 2, + 0, + 5, + 2, + 1, + 1, + 8, + 2, + 1, + 4, + 6, + 0, + 0, + 2, + 0, + 2, + 3, + 2, + 2, + 4, + 3, + 4, + 4, + 3, + 1, + 4, + 1, + 0, + 9, + 3, + 7, + 0, + 2, + 2, + 3, + 1, + 3, + 2, + 2, + 6, + 4, + 5, + 0, + 0, + 2, + 3, + 2, + 7, + 4, + 0, + 1, + 2, + 0, + 5, + 4, + 0, + 4, + 1, + 3, + 0, + 4, + 0, + 3 + ], + "metrics": { + "routes": 162, + "used_experts": 51, + "zero_experts": 13, + "mean": 2.53125, + "std": 2.130420207977295, + "cv": 0.8416475057601929, + "gini": 0.4552469253540039, + "entropy_nats": 3.7677295207977295, + "effective_experts": 43.281683030034465, + "min": 0.0, + "p50": 2.0, + "p90": 5.0, + "max": 9.0 + }, + "top_experts": [ + { + "expert": 29, + "routes": 9 + }, + { + "expert": 7, + "routes": 8 + }, + { + "expert": 31, + "routes": 7 + }, + { + "expert": 48, + "routes": 7 + }, + { + "expert": 11, + "routes": 6 + }, + { + "expert": 40, + "routes": 6 + }, + { + "expert": 3, + "routes": 5 + }, + { + "expert": 42, + "routes": 5 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 21, + "weight": 0.2854343056678772 + }, + { + "expert": 49, + "weight": 0.028558792546391487 + }, + { + "expert": 33, + "weight": 0.026019850745797157 + }, + { + "expert": 31, + "weight": 0.018602484837174416 + }, + { + "expert": 14, + "weight": 0.01662612520158291 + }, + { + "expert": 37, + "weight": 0.016511129215359688 + } + ], + "selected_weight_sum": 0.39175268821418285, + "top1_top2_margin": 0.2568755131214857 + }, + { + "position": 1, + "token_id": 8246, + "token": "èĭ¥", + "decoded_piece": "若", + "experts_by_weight": [ + { + "expert": 29, + "weight": 0.32871076464653015 + }, + { + "expert": 25, + "weight": 0.08848361670970917 + }, + { + "expert": 38, + "weight": 0.0565631277859211 + }, + { + "expert": 30, + "weight": 0.05635221675038338 + }, + { + "expert": 49, + "weight": 0.0430927611887455 + }, + { + "expert": 37, + "weight": 0.028256550431251526 + } + ], + "selected_weight_sum": 0.6014590375125408, + "top1_top2_margin": 0.24022714793682098 + }, + { + "position": 2, + "token_id": 2206, + "token": "æŃ£", + "decoded_piece": "正", + "experts_by_weight": [ + { + "expert": 33, + "weight": 0.08287578076124191 + }, + { + "expert": 21, + "weight": 0.0791545957326889 + }, + { + "expert": 49, + "weight": 0.07821590453386307 + }, + { + "expert": 11, + "weight": 0.06597484648227692 + }, + { + "expert": 31, + "weight": 0.05521147698163986 + }, + { + "expert": 37, + "weight": 0.05128243565559387 + } + ], + "selected_weight_sum": 0.41271504014730453, + "top1_top2_margin": 0.003721185028553009 + }, + { + "position": 3, + "token_id": 2266, + "token": "æķ°", + "decoded_piece": "数", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.2757400870323181 + }, + { + "expert": 11, + "weight": 0.07084589451551437 + }, + { + "expert": 54, + "weight": 0.05401374399662018 + }, + { + "expert": 24, + "weight": 0.04900094121694565 + }, + { + "expert": 55, + "weight": 0.04531301185488701 + }, + { + "expert": 31, + "weight": 0.03455324470996857 + } + ], + "selected_weight_sum": 0.5294669233262539, + "top1_top2_margin": 0.20489419251680374 + }, + { + "position": 4, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 40, + "weight": 0.18691080808639526 + }, + { + "expert": 29, + "weight": 0.06510350108146667 + }, + { + "expert": 42, + "weight": 0.06282351911067963 + }, + { + "expert": 46, + "weight": 0.061096590012311935 + }, + { + "expert": 10, + "weight": 0.058895643800497055 + }, + { + "expert": 48, + "weight": 0.04298863559961319 + } + ], + "selected_weight_sum": 0.47781869769096375, + "top1_top2_margin": 0.12180730700492859 + }, + { + "position": 5, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 38, + "weight": 0.9915505647659302 + }, + { + "expert": 41, + "weight": 0.0008498544339090586 + }, + { + "expert": 29, + "weight": 0.0008058093371801078 + }, + { + "expert": 45, + "weight": 0.0007542146486230195 + }, + { + "expert": 34, + "weight": 0.0006039484869688749 + }, + { + "expert": 3, + "weight": 0.0004561223613563925 + } + ], + "selected_weight_sum": 0.9950205140339676, + "top1_top2_margin": 0.9907007103320211 + }, + { + "position": 6, + "token_id": 13211, + "token": "满足", + "decoded_piece": "满足", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.11553625762462616 + }, + { + "expert": 61, + "weight": 0.084024578332901 + }, + { + "expert": 57, + "weight": 0.06242275983095169 + }, + { + "expert": 23, + "weight": 0.056678250432014465 + }, + { + "expert": 17, + "weight": 0.040774986147880554 + }, + { + "expert": 20, + "weight": 0.039831630885601044 + } + ], + "selected_weight_sum": 0.3992684632539749, + "top1_top2_margin": 0.03151167929172516 + }, + { + "position": 7, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 40, + "weight": 0.1632770299911499 + }, + { + "expert": 35, + "weight": 0.06992333382368088 + }, + { + "expert": 59, + "weight": 0.046842269599437714 + }, + { + "expert": 48, + "weight": 0.043937139213085175 + }, + { + "expert": 10, + "weight": 0.037441857159137726 + }, + { + "expert": 7, + "weight": 0.03594948351383209 + } + ], + "selected_weight_sum": 0.3973711133003235, + "top1_top2_margin": 0.09335369616746902 + }, + { + "position": 8, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 26, + "weight": 0.19643105566501617 + }, + { + "expert": 3, + "weight": 0.09501662850379944 + }, + { + "expert": 29, + "weight": 0.08256134390830994 + }, + { + "expert": 41, + "weight": 0.06957186758518219 + }, + { + "expert": 63, + "weight": 0.06131204590201378 + }, + { + "expert": 22, + "weight": 0.05051179975271225 + } + ], + "selected_weight_sum": 0.5554047413170338, + "top1_top2_margin": 0.10141442716121674 + }, + { + "position": 9, + "token_id": 14855, + "token": "åĬłä¸Ĭ", + "decoded_piece": "加上", + "experts_by_weight": [ + { + "expert": 42, + "weight": 0.24526570737361908 + }, + { + "expert": 20, + "weight": 0.060395147651433945 + }, + { + "expert": 31, + "weight": 0.0489298477768898 + }, + { + "expert": 61, + "weight": 0.043506477028131485 + }, + { + "expert": 23, + "weight": 0.03604341298341751 + }, + { + "expert": 57, + "weight": 0.035008423030376434 + } + ], + "selected_weight_sum": 0.46914901584386826, + "top1_top2_margin": 0.18487055972218513 + }, + { + "position": 10, + "token_id": 12985, + "token": "å®ĥçļĦ", + "decoded_piece": "它的", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.18621590733528137 + }, + { + "expert": 30, + "weight": 0.07778092473745346 + }, + { + "expert": 5, + "weight": 0.045964907854795456 + }, + { + "expert": 61, + "weight": 0.03565226122736931 + }, + { + "expert": 47, + "weight": 0.03342017903923988 + }, + { + "expert": 42, + "weight": 0.02921866998076439 + } + ], + "selected_weight_sum": 0.40825285017490387, + "top1_top2_margin": 0.10843498259782791 + }, + { + "position": 11, + "token_id": 8270, + "token": "åĢĴ", + "decoded_piece": "倒", + "experts_by_weight": [ + { + "expert": 54, + "weight": 0.09323352575302124 + }, + { + "expert": 31, + "weight": 0.09294472634792328 + }, + { + "expert": 48, + "weight": 0.07912302017211914 + }, + { + "expert": 7, + "weight": 0.044920433312654495 + }, + { + "expert": 47, + "weight": 0.041864652186632156 + }, + { + "expert": 19, + "weight": 0.035855911672115326 + } + ], + "selected_weight_sum": 0.38794226944446564, + "top1_top2_margin": 0.0002887994050979614 + }, + { + "position": 12, + "token_id": 2266, + "token": "æķ°", + "decoded_piece": "数", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.20481090247631073 + }, + { + "expert": 55, + "weight": 0.0889652669429779 + }, + { + "expert": 11, + "weight": 0.07102048397064209 + }, + { + "expert": 24, + "weight": 0.055527132004499435 + }, + { + "expert": 35, + "weight": 0.048422686755657196 + }, + { + "expert": 54, + "weight": 0.04809562489390373 + } + ], + "selected_weight_sum": 0.5168420970439911, + "top1_top2_margin": 0.11584563553333282 + }, + { + "position": 13, + "token_id": 37360, + "token": "çŃīäºİ", + "decoded_piece": "等于", + "experts_by_weight": [ + { + "expert": 42, + "weight": 0.15381541848182678 + }, + { + "expert": 8, + "weight": 0.11250929534435272 + }, + { + "expert": 23, + "weight": 0.042643386870622635 + }, + { + "expert": 27, + "weight": 0.0421457402408123 + }, + { + "expert": 57, + "weight": 0.038802195340394974 + }, + { + "expert": 17, + "weight": 0.033083997666835785 + } + ], + "selected_weight_sum": 0.4230000339448452, + "top1_top2_margin": 0.04130612313747406 + }, + { + "position": 14, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 3, + "weight": 0.133352592587471 + }, + { + "expert": 59, + "weight": 0.12619337439537048 + }, + { + "expert": 22, + "weight": 0.1030726283788681 + }, + { + "expert": 41, + "weight": 0.07036352902650833 + }, + { + "expert": 18, + "weight": 0.05759592354297638 + }, + { + "expert": 29, + "weight": 0.055169083178043365 + } + ], + "selected_weight_sum": 0.5457471311092377, + "top1_top2_margin": 0.007159218192100525 + }, + { + "position": 15, + "token_id": 18, + "token": "3", + "decoded_piece": "3", + "experts_by_weight": [ + { + "expert": 40, + "weight": 0.23435364663600922 + }, + { + "expert": 10, + "weight": 0.11713644117116928 + }, + { + "expert": 36, + "weight": 0.03138922527432442 + }, + { + "expert": 45, + "weight": 0.0267313364893198 + }, + { + "expert": 1, + "weight": 0.026693860068917274 + }, + { + "expert": 58, + "weight": 0.020750872790813446 + } + ], + "selected_weight_sum": 0.45705538243055344, + "top1_top2_margin": 0.11721720546483994 + }, + { + "position": 16, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 22, + "weight": 0.14716902375221252 + }, + { + "expert": 39, + "weight": 0.14222267270088196 + }, + { + "expert": 52, + "weight": 0.10041153430938721 + }, + { + "expert": 16, + "weight": 0.09503325819969177 + }, + { + "expert": 26, + "weight": 0.05392856523394585 + }, + { + "expert": 29, + "weight": 0.04694347456097603 + } + ], + "selected_weight_sum": 0.5857085287570953, + "top1_top2_margin": 0.004946351051330566 + }, + { + "position": 17, + "token_id": 2989, + "token": "æ±Ĥ", + "decoded_piece": "求", + "experts_by_weight": [ + { + "expert": 9, + "weight": 0.14854924380779266 + }, + { + "expert": 31, + "weight": 0.0715351402759552 + }, + { + "expert": 6, + "weight": 0.05261973291635513 + }, + { + "expert": 29, + "weight": 0.04394381120800972 + }, + { + "expert": 30, + "weight": 0.03829782083630562 + }, + { + "expert": 20, + "weight": 0.03392321988940239 + } + ], + "selected_weight_sum": 0.3888689689338207, + "top1_top2_margin": 0.07701410353183746 + }, + { + "position": 18, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 40, + "weight": 0.13323214650154114 + }, + { + "expert": 7, + "weight": 0.05016558617353439 + }, + { + "expert": 10, + "weight": 0.043240778148174286 + }, + { + "expert": 48, + "weight": 0.04313679784536362 + }, + { + "expert": 35, + "weight": 0.0390276238322258 + }, + { + "expert": 59, + "weight": 0.03554553538560867 + } + ], + "selected_weight_sum": 0.3443484678864479, + "top1_top2_margin": 0.08306656032800674 + }, + { + "position": 19, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 26, + "weight": 0.14701107144355774 + }, + { + "expert": 41, + "weight": 0.08602465689182281 + }, + { + "expert": 63, + "weight": 0.08064025640487671 + }, + { + "expert": 29, + "weight": 0.07543258368968964 + }, + { + "expert": 3, + "weight": 0.07340933382511139 + }, + { + "expert": 34, + "weight": 0.04117288440465927 + } + ], + "selected_weight_sum": 0.5036907866597176, + "top1_top2_margin": 0.060986414551734924 + }, + { + "position": 20, + "token_id": 337, + "token": "çļĦ", + "decoded_piece": "的", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.16039715707302094 + }, + { + "expert": 63, + "weight": 0.05364707112312317 + }, + { + "expert": 57, + "weight": 0.05302438884973526 + }, + { + "expert": 51, + "weight": 0.03915613889694214 + }, + { + "expert": 26, + "weight": 0.032873403280973434 + }, + { + "expert": 14, + "weight": 0.03032870404422283 + } + ], + "selected_weight_sum": 0.36942686326801777, + "top1_top2_margin": 0.10675008594989777 + }, + { + "position": 21, + "token_id": 18718, + "token": "å¹³æĸ¹", + "decoded_piece": "平方", + "experts_by_weight": [ + { + "expert": 40, + "weight": 0.10204068571329117 + }, + { + "expert": 7, + "weight": 0.05241134762763977 + }, + { + "expert": 11, + "weight": 0.051519397646188736 + }, + { + "expert": 46, + "weight": 0.04217739403247833 + }, + { + "expert": 48, + "weight": 0.030644165351986885 + }, + { + "expert": 1, + "weight": 0.030136745423078537 + } + ], + "selected_weight_sum": 0.30892973579466343, + "top1_top2_margin": 0.0496293380856514 + }, + { + "position": 22, + "token_id": 14855, + "token": "åĬłä¸Ĭ", + "decoded_piece": "加上", + "experts_by_weight": [ + { + "expert": 42, + "weight": 0.21430017054080963 + }, + { + "expert": 20, + "weight": 0.04553015157580376 + }, + { + "expert": 61, + "weight": 0.04181747883558273 + }, + { + "expert": 23, + "weight": 0.04179762303829193 + }, + { + "expert": 17, + "weight": 0.03857774659991264 + }, + { + "expert": 16, + "weight": 0.03725086897611618 + } + ], + "selected_weight_sum": 0.4192740395665169, + "top1_top2_margin": 0.16877001896500587 + }, + { + "position": 23, + "token_id": 8270, + "token": "åĢĴ", + "decoded_piece": "倒", + "experts_by_weight": [ + { + "expert": 54, + "weight": 0.07403554767370224 + }, + { + "expert": 31, + "weight": 0.0634426474571228 + }, + { + "expert": 19, + "weight": 0.05088704824447632 + }, + { + "expert": 48, + "weight": 0.05061543732881546 + }, + { + "expert": 21, + "weight": 0.038594309240579605 + }, + { + "expert": 49, + "weight": 0.029449019581079483 + } + ], + "selected_weight_sum": 0.3070240095257759, + "top1_top2_margin": 0.010592900216579437 + }, + { + "position": 24, + "token_id": 51971, + "token": "æķ°çļĦ", + "decoded_piece": "数的", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.20678895711898804 + }, + { + "expert": 11, + "weight": 0.06979629397392273 + }, + { + "expert": 24, + "weight": 0.06282166391611099 + }, + { + "expert": 55, + "weight": 0.06192285567522049 + }, + { + "expert": 54, + "weight": 0.048118963837623596 + }, + { + "expert": 48, + "weight": 0.04148270562291145 + } + ], + "selected_weight_sum": 0.4909314401447773, + "top1_top2_margin": 0.1369926631450653 + }, + { + "position": 25, + "token_id": 18718, + "token": "å¹³æĸ¹", + "decoded_piece": "平方", + "experts_by_weight": [ + { + "expert": 40, + "weight": 0.08547789603471756 + }, + { + "expert": 11, + "weight": 0.05890219286084175 + }, + { + "expert": 7, + "weight": 0.058209121227264404 + }, + { + "expert": 55, + "weight": 0.042358338832855225 + }, + { + "expert": 0, + "weight": 0.032743848860263824 + }, + { + "expert": 46, + "weight": 0.032649461179971695 + } + ], + "selected_weight_sum": 0.31034085899591446, + "top1_top2_margin": 0.02657570317387581 + }, + { + "position": 26, + "token_id": 398, + "token": "ãĢĤ", + "decoded_piece": "。", + "experts_by_weight": [ + { + "expert": 22, + "weight": 0.14948272705078125 + }, + { + "expert": 3, + "weight": 0.11477416008710861 + }, + { + "expert": 52, + "weight": 0.0903598740696907 + }, + { + "expert": 18, + "weight": 0.08863123506307602 + }, + { + "expert": 39, + "weight": 0.08091817051172256 + }, + { + "expert": 29, + "weight": 0.07229965180158615 + } + ], + "selected_weight_sum": 0.5964658185839653, + "top1_top2_margin": 0.03470856696367264 + } + ] + } + ], + "prompt_pair_jaccard": [ + { + "left": "zh_explanation", + "right": "en_architecture", + "used_expert_jaccard": 0.7407407407407407, + "shared_experts": 40, + "union_experts": 54 + }, + { + "left": "zh_explanation", + "right": "code", + "used_expert_jaccard": 0.7272727272727273, + "shared_experts": 40, + "union_experts": 55 + }, + { + "left": "zh_explanation", + "right": "math", + "used_expert_jaccard": 0.7241379310344828, + "shared_experts": 42, + "union_experts": 58 + }, + { + "left": "en_architecture", + "right": "code", + "used_expert_jaccard": 0.6851851851851852, + "shared_experts": 37, + "union_experts": 54 + }, + { + "left": "en_architecture", + "right": "math", + "used_expert_jaccard": 0.5737704918032787, + "shared_experts": 35, + "union_experts": 61 + }, + { + "left": "code", + "right": "math", + "used_expert_jaccard": 0.7321428571428571, + "shared_experts": 41, + "union_experts": 56 + } + ] + } + }, + { + "layer": 5, + "ffn": "moe", + "state_tensors": 203, + "state_numel": 584847872, + "state_bytes": 1169695744, + "timing_ms": { + "load_cpu": 24.87539401045069, + "transfer_to_device": 122.59778601583093, + "forward": 4.010354983620346 + }, + "peak_device_allocated_mib": 1177.7021484375, + "hidden_before": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 3.085932731628418, + "p10": 0.19619415700435638, + "p50": 0.23260724544525146, + "p90": 0.2969282567501068 + }, + "l2_norm": { + "mean": 139.65338134765625, + "p10": 8.878734588623047, + "p50": 10.526601791381836, + "p90": 13.43743896484375 + }, + "finite": true, + "sha256_fp32": "5d5d80c17f47a78f419f6ec9bf16668392a8246393ebfe2b0233841be4875b6e" + }, + "hidden_after": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 3.2100954055786133, + "p10": 0.22690242528915405, + "p50": 0.26399680972099304, + "p90": 0.33678531646728516 + }, + "l2_norm": { + "mean": 145.2723388671875, + "p10": 10.268431663513184, + "p50": 11.947131156921387, + "p90": 15.24116325378418 + }, + "finite": true, + "sha256_fp32": "88f158b68ecb980c55b1b77eebfeb2276d7820443732aa10effca9c85c653481" + }, + "mla": { + "compressed_projection_shape": [ + 4, + 27, + 576 + ], + "latent_shape": [ + 4, + 27, + 512 + ], + "rope_key_shape": [ + 4, + 27, + 64 + ], + "latent_rms": 2.0517733097076416, + "rope_key_rms": 2.469125509262085, + "eager_key_cache_shape": [ + 4, + 16, + 27, + 192 + ], + "eager_value_cache_shape": [ + 4, + 16, + 27, + 128 + ], + "eager_cache_bytes": 1105920 + }, + "routing": { + "aggregate_load": [ + 6, + 12, + 10, + 13, + 13, + 0, + 10, + 19, + 8, + 0, + 7, + 15, + 9, + 3, + 32, + 6, + 6, + 4, + 9, + 14, + 7, + 3, + 8, + 15, + 7, + 7, + 4, + 9, + 11, + 13, + 4, + 0, + 3, + 5, + 22, + 10, + 12, + 3, + 9, + 5, + 4, + 6, + 11, + 9, + 11, + 4, + 11, + 4, + 11, + 14, + 0, + 9, + 7, + 9, + 3, + 10, + 10, + 7, + 3, + 15, + 9, + 3, + 9, + 8 + ], + "aggregate_metrics": { + "routes": 540, + "used_experts": 60, + "zero_experts": 4, + "mean": 8.4375, + "std": 5.415659427642822, + "cv": 0.6418559551239014, + "gini": 0.3323495388031006, + "entropy_nats": 3.9539036750793457, + "effective_experts": 52.13850186069128, + "min": 0.0, + "p50": 8.5, + "p90": 14.0, + "max": 32.0 + }, + "top_experts": [ + { + "expert": 14, + "routes": 32 + }, + { + "expert": 34, + "routes": 22 + }, + { + "expert": 7, + "routes": 19 + }, + { + "expert": 11, + "routes": 15 + }, + { + "expert": 23, + "routes": 15 + }, + { + "expert": 59, + "routes": 15 + }, + { + "expert": 19, + "routes": 14 + }, + { + "expert": 49, + "routes": 14 + }, + { + "expert": 3, + "routes": 13 + }, + { + "expert": 4, + "routes": 13 + }, + { + "expert": 29, + "routes": 13 + }, + { + "expert": 1, + "routes": 12 + } + ], + "selected_weight_sum": { + "mean": 0.4831911623477936, + "p10": 0.37563803791999817, + "p50": 0.4518430829048157, + "p90": 0.6178299784660339 + }, + "top1_top2_margin": { + "mean": 0.09345383197069168, + "p50": 0.027145538479089737, + "p90": 0.16611436009407043 + }, + "per_prompt": [ + { + "id": "zh_explanation", + "label": "中文解释", + "tokens": 26, + "load": [ + 3, + 2, + 2, + 2, + 1, + 0, + 4, + 6, + 3, + 0, + 0, + 2, + 3, + 0, + 7, + 1, + 1, + 3, + 3, + 4, + 2, + 2, + 2, + 4, + 2, + 0, + 3, + 4, + 1, + 5, + 2, + 0, + 2, + 2, + 4, + 5, + 6, + 1, + 3, + 1, + 2, + 1, + 3, + 4, + 2, + 1, + 4, + 1, + 3, + 5, + 0, + 4, + 2, + 0, + 1, + 1, + 2, + 5, + 0, + 9, + 1, + 1, + 5, + 1 + ], + "metrics": { + "routes": 156, + "used_experts": 55, + "zero_experts": 9, + "mean": 2.4375, + "std": 1.9096662998199463, + "cv": 0.7834528684616089, + "gini": 0.4216746687889099, + "entropy_nats": 3.832216262817383, + "effective_experts": 46.16473813406393, + "min": 0.0, + "p50": 2.0, + "p90": 5.0, + "max": 9.0 + }, + "top_experts": [ + { + "expert": 59, + "routes": 9 + }, + { + "expert": 14, + "routes": 7 + }, + { + "expert": 7, + "routes": 6 + }, + { + "expert": 36, + "routes": 6 + }, + { + "expert": 29, + "routes": 5 + }, + { + "expert": 35, + "routes": 5 + }, + { + "expert": 49, + "routes": 5 + }, + { + "expert": 57, + "routes": 5 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.6668257713317871 + }, + { + "expert": 47, + "weight": 0.015490565448999405 + }, + { + "expert": 19, + "weight": 0.012896407395601273 + }, + { + "expert": 20, + "weight": 0.01244935393333435 + }, + { + "expert": 55, + "weight": 0.012159599922597408 + }, + { + "expert": 49, + "weight": 0.011999384500086308 + } + ], + "selected_weight_sum": 0.7318210825324059, + "top1_top2_margin": 0.6513352058827877 + }, + { + "position": 1, + "token_id": 1082, + "token": "ç͍", + "decoded_piece": "用", + "experts_by_weight": [ + { + "expert": 35, + "weight": 0.26780611276626587 + }, + { + "expert": 38, + "weight": 0.17193788290023804 + }, + { + "expert": 36, + "weight": 0.054537083953619 + }, + { + "expert": 48, + "weight": 0.034212712198495865 + }, + { + "expert": 41, + "weight": 0.03277919813990593 + }, + { + "expert": 23, + "weight": 0.02963159792125225 + } + ], + "selected_weight_sum": 0.590904587879777, + "top1_top2_margin": 0.09586822986602783 + }, + { + "position": 2, + "token_id": 83983, + "token": "éĢļä¿Ĺ", + "decoded_piece": "通俗", + "experts_by_weight": [ + { + "expert": 35, + "weight": 0.13310463726520538 + }, + { + "expert": 38, + "weight": 0.10224568098783493 + }, + { + "expert": 36, + "weight": 0.08373391628265381 + }, + { + "expert": 6, + "weight": 0.07357697933912277 + }, + { + "expert": 19, + "weight": 0.06322353333234787 + }, + { + "expert": 49, + "weight": 0.056075744330883026 + } + ], + "selected_weight_sum": 0.5119604915380478, + "top1_top2_margin": 0.030858956277370453 + }, + { + "position": 3, + "token_id": 91476, + "token": "çļĦè¯Ńè¨Ģ", + "decoded_piece": "的语言", + "experts_by_weight": [ + { + "expert": 19, + "weight": 0.237516850233078 + }, + { + "expert": 1, + "weight": 0.16881954669952393 + }, + { + "expert": 49, + "weight": 0.1347130984067917 + }, + { + "expert": 6, + "weight": 0.12524957954883575 + }, + { + "expert": 38, + "weight": 0.051083795726299286 + }, + { + "expert": 32, + "weight": 0.022646525874733925 + } + ], + "selected_weight_sum": 0.7400293964892626, + "top1_top2_margin": 0.06869730353355408 + }, + { + "position": 4, + "token_id": 17882, + "token": "è§£éĩĬ", + "decoded_piece": "解释", + "experts_by_weight": [ + { + "expert": 6, + "weight": 0.1273176521062851 + }, + { + "expert": 23, + "weight": 0.10280857235193253 + }, + { + "expert": 19, + "weight": 0.07830130308866501 + }, + { + "expert": 3, + "weight": 0.07459865510463715 + }, + { + "expert": 49, + "weight": 0.06277203559875488 + }, + { + "expert": 36, + "weight": 0.05869143456220627 + } + ], + "selected_weight_sum": 0.5044896528124809, + "top1_top2_margin": 0.02450907975435257 + }, + { + "position": 5, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 27, + "weight": 0.1488560438156128 + }, + { + "expert": 43, + "weight": 0.08283150941133499 + }, + { + "expert": 17, + "weight": 0.08140339702367783 + }, + { + "expert": 12, + "weight": 0.07671284675598145 + }, + { + "expert": 18, + "weight": 0.058525532484054565 + }, + { + "expert": 0, + "weight": 0.044425301253795624 + } + ], + "selected_weight_sum": 0.49275463074445724, + "top1_top2_margin": 0.0660245344042778 + }, + { + "position": 6, + "token_id": 10285, + "token": "为ä»Ģä¹Ī", + "decoded_piece": "为什么", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.08674228936433792 + }, + { + "expert": 60, + "weight": 0.07026661932468414 + }, + { + "expert": 36, + "weight": 0.06282687932252884 + }, + { + "expert": 35, + "weight": 0.057485852390527725 + }, + { + "expert": 3, + "weight": 0.047155700623989105 + }, + { + "expert": 6, + "weight": 0.04366069287061691 + } + ], + "selected_weight_sum": 0.36813803389668465, + "top1_top2_margin": 0.016475670039653778 + }, + { + "position": 7, + "token_id": 21844, + "token": "ç¨Ģ", + "decoded_piece": "稀", + "experts_by_weight": [ + { + "expert": 11, + "weight": 0.10328876972198486 + }, + { + "expert": 44, + "weight": 0.08392632752656937 + }, + { + "expert": 35, + "weight": 0.0741342306137085 + }, + { + "expert": 37, + "weight": 0.050005458295345306 + }, + { + "expert": 42, + "weight": 0.03275921195745468 + }, + { + "expert": 36, + "weight": 0.028308894485235214 + } + ], + "selected_weight_sum": 0.37242289260029793, + "top1_top2_margin": 0.019362442195415497 + }, + { + "position": 8, + "token_id": 20923, + "token": "çĸı", + "decoded_piece": "疏", + "experts_by_weight": [ + { + "expert": 11, + "weight": 0.21587729454040527 + }, + { + "expert": 7, + "weight": 0.09414123743772507 + }, + { + "expert": 28, + "weight": 0.04408562555909157 + }, + { + "expert": 35, + "weight": 0.03777015209197998 + }, + { + "expert": 34, + "weight": 0.032802462577819824 + }, + { + "expert": 52, + "weight": 0.031176554039120674 + } + ], + "selected_weight_sum": 0.4558533262461424, + "top1_top2_margin": 0.1217360571026802 + }, + { + "position": 9, + "token_id": 10900, + "token": "ä¸ĵå®¶", + "decoded_piece": "专家", + "experts_by_weight": [ + { + "expert": 2, + "weight": 0.08871850371360779 + }, + { + "expert": 7, + "weight": 0.07161764800548553 + }, + { + "expert": 46, + "weight": 0.06334264576435089 + }, + { + "expert": 30, + "weight": 0.059080276638269424 + }, + { + "expert": 8, + "weight": 0.05416818708181381 + }, + { + "expert": 16, + "weight": 0.05293413996696472 + } + ], + "selected_weight_sum": 0.38986140117049217, + "top1_top2_margin": 0.017100855708122253 + }, + { + "position": 10, + "token_id": 39322, + "token": "模åŀĭ", + "decoded_piece": "模型", + "experts_by_weight": [ + { + "expert": 2, + "weight": 0.17880289256572723 + }, + { + "expert": 46, + "weight": 0.1402723491191864 + }, + { + "expert": 61, + "weight": 0.13747455179691315 + }, + { + "expert": 7, + "weight": 0.09143506735563278 + }, + { + "expert": 34, + "weight": 0.08162415027618408 + }, + { + "expert": 8, + "weight": 0.037828776985406876 + } + ], + "selected_weight_sum": 0.6674377880990505, + "top1_top2_margin": 0.03853054344654083 + }, + { + "position": 11, + "token_id": 1876, + "token": "åı¯ä»¥", + "decoded_piece": "可以", + "experts_by_weight": [ + { + "expert": 56, + "weight": 0.12182933837175369 + }, + { + "expert": 51, + "weight": 0.10752717405557632 + }, + { + "expert": 62, + "weight": 0.08328860998153687 + }, + { + "expert": 33, + "weight": 0.0359572097659111 + }, + { + "expert": 45, + "weight": 0.03254731744527817 + }, + { + "expert": 49, + "weight": 0.03201925754547119 + } + ], + "selected_weight_sum": 0.41316890716552734, + "top1_top2_margin": 0.014302164316177368 + }, + { + "position": 12, + "token_id": 9476, + "token": "æĭ¥æľī", + "decoded_piece": "拥有", + "experts_by_weight": [ + { + "expert": 51, + "weight": 0.1414400339126587 + }, + { + "expert": 62, + "weight": 0.10390999913215637 + }, + { + "expert": 23, + "weight": 0.05466355383396149 + }, + { + "expert": 33, + "weight": 0.05069989338517189 + }, + { + "expert": 22, + "weight": 0.043549034744501114 + }, + { + "expert": 0, + "weight": 0.03670540824532509 + } + ], + "selected_weight_sum": 0.43096792325377464, + "top1_top2_margin": 0.03753003478050232 + }, + { + "position": 13, + "token_id": 3599, + "token": "å¾Īå¤ļ", + "decoded_piece": "很多", + "experts_by_weight": [ + { + "expert": 59, + "weight": 0.10780058056116104 + }, + { + "expert": 51, + "weight": 0.09324204176664352 + }, + { + "expert": 62, + "weight": 0.08552185446023941 + }, + { + "expert": 29, + "weight": 0.04821997880935669 + }, + { + "expert": 32, + "weight": 0.03283743932843208 + }, + { + "expert": 14, + "weight": 0.03034812957048416 + } + ], + "selected_weight_sum": 0.3979700244963169, + "top1_top2_margin": 0.014558538794517517 + }, + { + "position": 14, + "token_id": 21273, + "token": "åıĤæķ°", + "decoded_piece": "参数", + "experts_by_weight": [ + { + "expert": 59, + "weight": 0.10011549293994904 + }, + { + "expert": 46, + "weight": 0.09130356460809708 + }, + { + "expert": 34, + "weight": 0.08848028630018234 + }, + { + "expert": 7, + "weight": 0.07407595962285995 + }, + { + "expert": 30, + "weight": 0.04830920696258545 + }, + { + "expert": 40, + "weight": 0.03818826377391815 + } + ], + "selected_weight_sum": 0.440472774207592, + "top1_top2_margin": 0.00881192833185196 + }, + { + "position": 15, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 43, + "weight": 0.12919080257415771 + }, + { + "expert": 12, + "weight": 0.10803435742855072 + }, + { + "expert": 27, + "weight": 0.0981711596250534 + }, + { + "expert": 42, + "weight": 0.05666274577379227 + }, + { + "expert": 18, + "weight": 0.055799245834350586 + }, + { + "expert": 17, + "weight": 0.05242020636796951 + } + ], + "selected_weight_sum": 0.5002785176038742, + "top1_top2_margin": 0.021156445145606995 + }, + { + "position": 16, + "token_id": 1705, + "token": "ä½Ĩ", + "decoded_piece": "但", + "experts_by_weight": [ + { + "expert": 43, + "weight": 0.0960824117064476 + }, + { + "expert": 24, + "weight": 0.0716608390212059 + }, + { + "expert": 14, + "weight": 0.06996157020330429 + }, + { + "expert": 21, + "weight": 0.06649202108383179 + }, + { + "expert": 17, + "weight": 0.06584745645523071 + }, + { + "expert": 27, + "weight": 0.04214527830481529 + } + ], + "selected_weight_sum": 0.4121895767748356, + "top1_top2_margin": 0.0244215726852417 + }, + { + "position": 17, + "token_id": 13712, + "token": "æ¯ı个", + "decoded_piece": "每个", + "experts_by_weight": [ + { + "expert": 24, + "weight": 0.06841839849948883 + }, + { + "expert": 21, + "weight": 0.06105462834239006 + }, + { + "expert": 48, + "weight": 0.052708983421325684 + }, + { + "expert": 36, + "weight": 0.04746530205011368 + }, + { + "expert": 59, + "weight": 0.04617984592914581 + }, + { + "expert": 29, + "weight": 0.037067096680402756 + } + ], + "selected_weight_sum": 0.3128942549228668, + "top1_top2_margin": 0.00736377015709877 + }, + { + "position": 18, + "token_id": 10728, + "token": "Ġtoken", + "decoded_piece": " token", + "experts_by_weight": [ + { + "expert": 34, + "weight": 0.11643253266811371 + }, + { + "expert": 59, + "weight": 0.10292921215295792 + }, + { + "expert": 44, + "weight": 0.06338663399219513 + }, + { + "expert": 7, + "weight": 0.04246857762336731 + }, + { + "expert": 46, + "weight": 0.04223647341132164 + }, + { + "expert": 4, + "weight": 0.03386823087930679 + } + ], + "selected_weight_sum": 0.4013216607272625, + "top1_top2_margin": 0.013503320515155792 + }, + { + "position": 19, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 63, + "weight": 0.9029439091682434 + }, + { + "expert": 15, + "weight": 0.004053776618093252 + }, + { + "expert": 8, + "weight": 0.003351924940943718 + }, + { + "expert": 1, + "weight": 0.0033151512034237385 + }, + { + "expert": 7, + "weight": 0.0029632176738232374 + }, + { + "expert": 54, + "weight": 0.0028055852744728327 + } + ], + "selected_weight_sum": 0.9194335648790002, + "top1_top2_margin": 0.8988901325501502 + }, + { + "position": 20, + "token_id": 2141, + "token": "åıª", + "decoded_piece": "只", + "experts_by_weight": [ + { + "expert": 59, + "weight": 0.09061706066131592 + }, + { + "expert": 14, + "weight": 0.07269977033138275 + }, + { + "expert": 57, + "weight": 0.06583838164806366 + }, + { + "expert": 62, + "weight": 0.05064796283841133 + }, + { + "expert": 56, + "weight": 0.0499817356467247 + }, + { + "expert": 51, + "weight": 0.04805685207247734 + } + ], + "selected_weight_sum": 0.3778417631983757, + "top1_top2_margin": 0.017917290329933167 + }, + { + "position": 21, + "token_id": 5118, + "token": "使ç͍", + "decoded_piece": "使用", + "experts_by_weight": [ + { + "expert": 23, + "weight": 0.09294749051332474 + }, + { + "expert": 57, + "weight": 0.09201069921255112 + }, + { + "expert": 22, + "weight": 0.0672682672739029 + }, + { + "expert": 14, + "weight": 0.06171053647994995 + }, + { + "expert": 59, + "weight": 0.052574750036001205 + }, + { + "expert": 62, + "weight": 0.04288235679268837 + } + ], + "selected_weight_sum": 0.4093941003084183, + "top1_top2_margin": 0.0009367913007736206 + }, + { + "position": 22, + "token_id": 6802, + "token": "åħ¶ä¸Ń", + "decoded_piece": "其中", + "experts_by_weight": [ + { + "expert": 26, + "weight": 0.07986949384212494 + }, + { + "expert": 14, + "weight": 0.06633445620536804 + }, + { + "expert": 0, + "weight": 0.05573779717087746 + }, + { + "expert": 57, + "weight": 0.0536210760474205 + }, + { + "expert": 29, + "weight": 0.05267202854156494 + }, + { + "expert": 59, + "weight": 0.0491383858025074 + } + ], + "selected_weight_sum": 0.3573732376098633, + "top1_top2_margin": 0.013535037636756897 + }, + { + "position": 23, + "token_id": 51905, + "token": "ä¸Ģå°ı", + "decoded_piece": "一小", + "experts_by_weight": [ + { + "expert": 59, + "weight": 0.1389620155096054 + }, + { + "expert": 29, + "weight": 0.07460430264472961 + }, + { + "expert": 26, + "weight": 0.05769191309809685 + }, + { + "expert": 57, + "weight": 0.048479143530130386 + }, + { + "expert": 14, + "weight": 0.03461878374218941 + }, + { + "expert": 40, + "weight": 0.03341403976082802 + } + ], + "selected_weight_sum": 0.3877701982855797, + "top1_top2_margin": 0.0643577128648758 + }, + { + "position": 24, + "token_id": 5781, + "token": "éĥ¨åĪĨ", + "decoded_piece": "部分", + "experts_by_weight": [ + { + "expert": 59, + "weight": 0.09375614672899246 + }, + { + "expert": 57, + "weight": 0.08696839958429337 + }, + { + "expert": 29, + "weight": 0.06338794529438019 + }, + { + "expert": 26, + "weight": 0.059497132897377014 + }, + { + "expert": 14, + "weight": 0.05874354764819145 + }, + { + "expert": 20, + "weight": 0.053002435714006424 + } + ], + "selected_weight_sum": 0.4153556078672409, + "top1_top2_margin": 0.006787747144699097 + }, + { + "position": 25, + "token_id": 398, + "token": "ãĢĤ", + "decoded_piece": "。", + "experts_by_weight": [ + { + "expert": 12, + "weight": 0.14567047357559204 + }, + { + "expert": 43, + "weight": 0.11958866566419601 + }, + { + "expert": 39, + "weight": 0.10240183025598526 + }, + { + "expert": 18, + "weight": 0.07829892635345459 + }, + { + "expert": 42, + "weight": 0.05260668322443962 + }, + { + "expert": 27, + "weight": 0.05196297913789749 + } + ], + "selected_weight_sum": 0.550529558211565, + "top1_top2_margin": 0.026081807911396027 + } + ] + }, + { + "id": "en_architecture", + "label": "English architecture", + "tokens": 18, + "load": [ + 0, + 0, + 6, + 6, + 5, + 0, + 2, + 8, + 1, + 0, + 1, + 6, + 1, + 0, + 1, + 0, + 1, + 0, + 1, + 1, + 1, + 0, + 2, + 2, + 0, + 1, + 0, + 1, + 3, + 3, + 0, + 0, + 0, + 1, + 7, + 1, + 1, + 1, + 2, + 1, + 0, + 2, + 1, + 1, + 5, + 2, + 6, + 1, + 4, + 1, + 0, + 1, + 2, + 0, + 0, + 2, + 2, + 2, + 0, + 1, + 3, + 2, + 2, + 1 + ], + "metrics": { + "routes": 108, + "used_experts": 46, + "zero_experts": 18, + "mean": 1.6875, + "std": 1.9426295757293701, + "cv": 1.1511878967285156, + "gini": 0.5627893805503845, + "entropy_nats": 3.557856321334839, + "effective_experts": 35.087899285630286, + "min": 0.0, + "p50": 1.0, + "p90": 5.0, + "max": 8.0 + }, + "top_experts": [ + { + "expert": 7, + "routes": 8 + }, + { + "expert": 34, + "routes": 7 + }, + { + "expert": 2, + "routes": 6 + }, + { + "expert": 3, + "routes": 6 + }, + { + "expert": 11, + "routes": 6 + }, + { + "expert": 46, + "routes": 6 + }, + { + "expert": 4, + "routes": 5 + }, + { + "expert": 44, + "routes": 5 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.6668257713317871 + }, + { + "expert": 47, + "weight": 0.015490565448999405 + }, + { + "expert": 19, + "weight": 0.012896407395601273 + }, + { + "expert": 20, + "weight": 0.01244935393333435 + }, + { + "expert": 55, + "weight": 0.012159599922597408 + }, + { + "expert": 49, + "weight": 0.011999384500086308 + } + ], + "selected_weight_sum": 0.7318210825324059, + "top1_top2_margin": 0.6513352058827877 + }, + { + "position": 1, + "token_id": 99297, + "token": "Explain", + "decoded_piece": "Explain", + "experts_by_weight": [ + { + "expert": 35, + "weight": 0.17686711251735687 + }, + { + "expert": 38, + "weight": 0.137218177318573 + }, + { + "expert": 41, + "weight": 0.11987470090389252 + }, + { + "expert": 18, + "weight": 0.057543590664863586 + }, + { + "expert": 3, + "weight": 0.04869488254189491 + }, + { + "expert": 44, + "weight": 0.04298888146877289 + } + ], + "selected_weight_sum": 0.5831873454153538, + "top1_top2_margin": 0.039648935198783875 + }, + { + "position": 2, + "token_id": 946, + "token": "Ġhow", + "decoded_piece": " how", + "experts_by_weight": [ + { + "expert": 60, + "weight": 0.11648083478212357 + }, + { + "expert": 3, + "weight": 0.10566136986017227 + }, + { + "expert": 38, + "weight": 0.06881820410490036 + }, + { + "expert": 41, + "weight": 0.059229787439107895 + }, + { + "expert": 6, + "weight": 0.048904579132795334 + }, + { + "expert": 48, + "weight": 0.03643996641039848 + } + ], + "selected_weight_sum": 0.4355347417294979, + "top1_top2_margin": 0.010819464921951294 + }, + { + "position": 3, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.23619119822978973 + }, + { + "expert": 60, + "weight": 0.07677339762449265 + }, + { + "expert": 51, + "weight": 0.03644350543618202 + }, + { + "expert": 36, + "weight": 0.03643375635147095 + }, + { + "expert": 57, + "weight": 0.034974295645952225 + }, + { + "expert": 29, + "weight": 0.03322301805019379 + } + ], + "selected_weight_sum": 0.45403917133808136, + "top1_top2_margin": 0.1594178006052971 + }, + { + "position": 4, + "token_id": 29640, + "token": "Ġcompressed", + "decoded_piece": " compressed", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.14481888711452484 + }, + { + "expert": 11, + "weight": 0.12344990670681 + }, + { + "expert": 44, + "weight": 0.056090760976076126 + }, + { + "expert": 34, + "weight": 0.04771822690963745 + }, + { + "expert": 29, + "weight": 0.040006957948207855 + }, + { + "expert": 28, + "weight": 0.037562284618616104 + } + ], + "selected_weight_sum": 0.4496470242738724, + "top1_top2_margin": 0.021368980407714844 + }, + { + "position": 5, + "token_id": 2131, + "token": "Ġkey", + "decoded_piece": " key", + "experts_by_weight": [ + { + "expert": 34, + "weight": 0.15713807940483093 + }, + { + "expert": 2, + "weight": 0.12892881035804749 + }, + { + "expert": 46, + "weight": 0.0764535441994667 + }, + { + "expert": 61, + "weight": 0.06165243312716484 + }, + { + "expert": 4, + "weight": 0.04296897351741791 + }, + { + "expert": 44, + "weight": 0.042255908250808716 + } + ], + "selected_weight_sum": 0.5093977488577366, + "top1_top2_margin": 0.028209269046783447 + }, + { + "position": 6, + "token_id": 12, + "token": "-", + "decoded_piece": "-", + "experts_by_weight": [ + { + "expert": 63, + "weight": 0.22240972518920898 + }, + { + "expert": 61, + "weight": 0.04973777011036873 + }, + { + "expert": 11, + "weight": 0.03715226426720619 + }, + { + "expert": 10, + "weight": 0.03665049001574516 + }, + { + "expert": 55, + "weight": 0.033358726650476456 + }, + { + "expert": 4, + "weight": 0.03200354427099228 + } + ], + "selected_weight_sum": 0.4113125205039978, + "top1_top2_margin": 0.17267195507884026 + }, + { + "position": 7, + "token_id": 2834, + "token": "value", + "decoded_piece": "value", + "experts_by_weight": [ + { + "expert": 34, + "weight": 0.15645183622837067 + }, + { + "expert": 46, + "weight": 0.09938643872737885 + }, + { + "expert": 2, + "weight": 0.0829453244805336 + }, + { + "expert": 4, + "weight": 0.052371591329574585 + }, + { + "expert": 25, + "weight": 0.050064388662576675 + }, + { + "expert": 52, + "weight": 0.036176759749650955 + } + ], + "selected_weight_sum": 0.4773963391780853, + "top1_top2_margin": 0.05706539750099182 + }, + { + "position": 8, + "token_id": 28937, + "token": "Ġlatent", + "decoded_piece": " latent", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.16698865592479706 + }, + { + "expert": 44, + "weight": 0.09657170623540878 + }, + { + "expert": 11, + "weight": 0.09212639927864075 + }, + { + "expert": 46, + "weight": 0.06852614879608154 + }, + { + "expert": 34, + "weight": 0.06448085606098175 + }, + { + "expert": 2, + "weight": 0.060119882225990295 + } + ], + "selected_weight_sum": 0.5488136485219002, + "top1_top2_margin": 0.07041694968938828 + }, + { + "position": 9, + "token_id": 4183, + "token": "Ġchanges", + "decoded_piece": " changes", + "experts_by_weight": [ + { + "expert": 45, + "weight": 0.09613549709320068 + }, + { + "expert": 7, + "weight": 0.07573243230581284 + }, + { + "expert": 56, + "weight": 0.06894052773714066 + }, + { + "expert": 62, + "weight": 0.06686387956142426 + }, + { + "expert": 3, + "weight": 0.05422895401716232 + }, + { + "expert": 22, + "weight": 0.04179750010371208 + } + ], + "selected_weight_sum": 0.40369879081845284, + "top1_top2_margin": 0.020403064787387848 + }, + { + "position": 10, + "token_id": 254, + "token": "Ġthe", + "decoded_piece": " the", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.1237715482711792 + }, + { + "expert": 3, + "weight": 0.0828896015882492 + }, + { + "expert": 45, + "weight": 0.06915713101625443 + }, + { + "expert": 23, + "weight": 0.05710538104176521 + }, + { + "expert": 62, + "weight": 0.04857157543301582 + }, + { + "expert": 33, + "weight": 0.04292594641447067 + } + ], + "selected_weight_sum": 0.42442118376493454, + "top1_top2_margin": 0.04088194668292999 + }, + { + "position": 11, + "token_id": 4887, + "token": "Ġmemory", + "decoded_piece": " memory", + "experts_by_weight": [ + { + "expert": 34, + "weight": 0.10212866216897964 + }, + { + "expert": 7, + "weight": 0.08982594311237335 + }, + { + "expert": 46, + "weight": 0.05629178136587143 + }, + { + "expert": 4, + "weight": 0.045767851173877716 + }, + { + "expert": 6, + "weight": 0.045069701969623566 + }, + { + "expert": 11, + "weight": 0.040722765028476715 + } + ], + "selected_weight_sum": 0.3798067048192024, + "top1_top2_margin": 0.012302719056606293 + }, + { + "position": 12, + "token_id": 2295, + "token": "Ġcost", + "decoded_piece": " cost", + "experts_by_weight": [ + { + "expert": 46, + "weight": 0.1462714523077011 + }, + { + "expert": 37, + "weight": 0.08453454077243805 + }, + { + "expert": 4, + "weight": 0.08272839337587357 + }, + { + "expert": 7, + "weight": 0.06534243375062943 + }, + { + "expert": 59, + "weight": 0.05198255553841591 + }, + { + "expert": 2, + "weight": 0.04178442433476448 + } + ], + "selected_weight_sum": 0.47264380007982254, + "top1_top2_margin": 0.06173691153526306 + }, + { + "position": 13, + "token_id": 280, + "token": "Ġof", + "decoded_piece": " of", + "experts_by_weight": [ + { + "expert": 60, + "weight": 0.14627379179000854 + }, + { + "expert": 14, + "weight": 0.09714733064174652 + }, + { + "expert": 48, + "weight": 0.09307599067687988 + }, + { + "expert": 23, + "weight": 0.061775561422109604 + }, + { + "expert": 3, + "weight": 0.03882799297571182 + }, + { + "expert": 57, + "weight": 0.037830740213394165 + } + ], + "selected_weight_sum": 0.47493140771985054, + "top1_top2_margin": 0.049126461148262024 + }, + { + "position": 14, + "token_id": 77939, + "token": "Ġautore", + "decoded_piece": " autore", + "experts_by_weight": [ + { + "expert": 44, + "weight": 0.0866532251238823 + }, + { + "expert": 7, + "weight": 0.08156659454107285 + }, + { + "expert": 11, + "weight": 0.0580621175467968 + }, + { + "expert": 16, + "weight": 0.05063927173614502 + }, + { + "expert": 34, + "weight": 0.04073431342840195 + }, + { + "expert": 29, + "weight": 0.03586898744106293 + } + ], + "selected_weight_sum": 0.35352450981736183, + "top1_top2_margin": 0.005086630582809448 + }, + { + "position": 15, + "token_id": 79184, + "token": "gressive", + "decoded_piece": "gressive", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.20286299288272858 + }, + { + "expert": 11, + "weight": 0.060117512941360474 + }, + { + "expert": 2, + "weight": 0.059362683445215225 + }, + { + "expert": 8, + "weight": 0.044468119740486145 + }, + { + "expert": 34, + "weight": 0.044006701558828354 + }, + { + "expert": 28, + "weight": 0.0434880256652832 + } + ], + "selected_weight_sum": 0.454306036233902, + "top1_top2_margin": 0.1427454799413681 + }, + { + "position": 16, + "token_id": 39671, + "token": "Ġdecoding", + "decoded_piece": " decoding", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.22519521415233612 + }, + { + "expert": 46, + "weight": 0.10506051033735275 + }, + { + "expert": 2, + "weight": 0.10295503586530685 + }, + { + "expert": 3, + "weight": 0.0685448870062828 + }, + { + "expert": 22, + "weight": 0.06022454798221588 + }, + { + "expert": 28, + "weight": 0.04068287834525108 + } + ], + "selected_weight_sum": 0.6026630736887455, + "top1_top2_margin": 0.12013470381498337 + }, + { + "position": 17, + "token_id": 13, + "token": ".", + "decoded_piece": ".", + "experts_by_weight": [ + { + "expert": 12, + "weight": 0.10836317390203476 + }, + { + "expert": 43, + "weight": 0.09578173607587814 + }, + { + "expert": 42, + "weight": 0.08955655246973038 + }, + { + "expert": 27, + "weight": 0.08248470723628998 + }, + { + "expert": 56, + "weight": 0.07762836664915085 + }, + { + "expert": 39, + "weight": 0.06929574906826019 + } + ], + "selected_weight_sum": 0.5231102854013443, + "top1_top2_margin": 0.012581437826156616 + } + ] + }, + { + "id": "code", + "label": "Python code", + "tokens": 19, + "load": [ + 1, + 1, + 2, + 4, + 3, + 0, + 4, + 2, + 0, + 0, + 0, + 4, + 2, + 1, + 5, + 1, + 1, + 0, + 2, + 5, + 2, + 0, + 3, + 5, + 3, + 0, + 0, + 2, + 3, + 5, + 2, + 0, + 1, + 0, + 4, + 2, + 1, + 1, + 3, + 1, + 0, + 2, + 2, + 2, + 2, + 1, + 1, + 1, + 3, + 6, + 0, + 2, + 1, + 1, + 0, + 2, + 3, + 0, + 3, + 1, + 5, + 0, + 0, + 0 + ], + "metrics": { + "routes": 114, + "used_experts": 47, + "zero_experts": 17, + "mean": 1.78125, + "std": 1.618040680885315, + "cv": 0.9083737134933472, + "gini": 0.4936951696872711, + "entropy_nats": 3.69284725189209, + "effective_experts": 40.15902719636842, + "min": 0.0, + "p50": 1.5, + "p90": 4.0, + "max": 6.0 + }, + "top_experts": [ + { + "expert": 49, + "routes": 6 + }, + { + "expert": 14, + "routes": 5 + }, + { + "expert": 19, + "routes": 5 + }, + { + "expert": 23, + "routes": 5 + }, + { + "expert": 29, + "routes": 5 + }, + { + "expert": 60, + "routes": 5 + }, + { + "expert": 3, + "routes": 4 + }, + { + "expert": 6, + "routes": 4 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.6668257713317871 + }, + { + "expert": 47, + "weight": 0.015490565448999405 + }, + { + "expert": 19, + "weight": 0.012896407395601273 + }, + { + "expert": 20, + "weight": 0.01244935393333435 + }, + { + "expert": 55, + "weight": 0.012159599922597408 + }, + { + "expert": 49, + "weight": 0.011999384500086308 + } + ], + "selected_weight_sum": 0.7318210825324059, + "top1_top2_margin": 0.6513352058827877 + }, + { + "position": 1, + "token_id": 9081, + "token": "Write", + "decoded_piece": "Write", + "experts_by_weight": [ + { + "expert": 38, + "weight": 0.2189263552427292 + }, + { + "expert": 35, + "weight": 0.1511840671300888 + }, + { + "expert": 41, + "weight": 0.05854736641049385 + }, + { + "expert": 18, + "weight": 0.05501477047801018 + }, + { + "expert": 44, + "weight": 0.052550721913576126 + }, + { + "expert": 15, + "weight": 0.04810547083616257 + } + ], + "selected_weight_sum": 0.5843287520110607, + "top1_top2_margin": 0.06774228811264038 + }, + { + "position": 2, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 38, + "weight": 0.2076040804386139 + }, + { + "expert": 22, + "weight": 0.10240605473518372 + }, + { + "expert": 49, + "weight": 0.08799075335264206 + }, + { + "expert": 35, + "weight": 0.04287009313702583 + }, + { + "expert": 19, + "weight": 0.040391046553850174 + }, + { + "expert": 41, + "weight": 0.03840652480721474 + } + ], + "selected_weight_sum": 0.5196685530245304, + "top1_top2_margin": 0.10519802570343018 + }, + { + "position": 3, + "token_id": 12974, + "token": "ĠPython", + "decoded_piece": " Python", + "experts_by_weight": [ + { + "expert": 49, + "weight": 0.20095491409301758 + }, + { + "expert": 42, + "weight": 0.12679241597652435 + }, + { + "expert": 34, + "weight": 0.056457847356796265 + }, + { + "expert": 44, + "weight": 0.05235559493303299 + }, + { + "expert": 13, + "weight": 0.04632683843374252 + }, + { + "expert": 55, + "weight": 0.04559082165360451 + } + ], + "selected_weight_sum": 0.5284784324467182, + "top1_top2_margin": 0.07416249811649323 + }, + { + "position": 4, + "token_id": 1157, + "token": "Ġfunction", + "decoded_piece": " function", + "experts_by_weight": [ + { + "expert": 49, + "weight": 0.3816191852092743 + }, + { + "expert": 34, + "weight": 0.07868051528930664 + }, + { + "expert": 19, + "weight": 0.05729304254055023 + }, + { + "expert": 30, + "weight": 0.043323855847120285 + }, + { + "expert": 6, + "weight": 0.04145321995019913 + }, + { + "expert": 38, + "weight": 0.026589658111333847 + } + ], + "selected_weight_sum": 0.6289594769477844, + "top1_top2_margin": 0.30293866991996765 + }, + { + "position": 5, + "token_id": 344, + "token": "Ġthat", + "decoded_piece": " that", + "experts_by_weight": [ + { + "expert": 60, + "weight": 0.2202168107032776 + }, + { + "expert": 49, + "weight": 0.11744972318410873 + }, + { + "expert": 51, + "weight": 0.09239882975816727 + }, + { + "expert": 56, + "weight": 0.03130319342017174 + }, + { + "expert": 27, + "weight": 0.031124742701649666 + }, + { + "expert": 36, + "weight": 0.02125764638185501 + } + ], + "selected_weight_sum": 0.51375094614923, + "top1_top2_margin": 0.10276708751916885 + }, + { + "position": 6, + "token_id": 7578, + "token": "Ġreturns", + "decoded_piece": " returns", + "experts_by_weight": [ + { + "expert": 3, + "weight": 0.12485012412071228 + }, + { + "expert": 60, + "weight": 0.10983602702617645 + }, + { + "expert": 45, + "weight": 0.09923596680164337 + }, + { + "expert": 23, + "weight": 0.07021182030439377 + }, + { + "expert": 49, + "weight": 0.0490649938583374 + }, + { + "expert": 51, + "weight": 0.042169809341430664 + } + ], + "selected_weight_sum": 0.49536874145269394, + "top1_top2_margin": 0.015014097094535828 + }, + { + "position": 7, + "token_id": 254, + "token": "Ġthe", + "decoded_piece": " the", + "experts_by_weight": [ + { + "expert": 3, + "weight": 0.14295123517513275 + }, + { + "expert": 48, + "weight": 0.13394472002983093 + }, + { + "expert": 60, + "weight": 0.04427902773022652 + }, + { + "expert": 20, + "weight": 0.0420200414955616 + }, + { + "expert": 23, + "weight": 0.03992121294140816 + }, + { + "expert": 22, + "weight": 0.03541431576013565 + } + ], + "selected_weight_sum": 0.4385305531322956, + "top1_top2_margin": 0.009006515145301819 + }, + { + "position": 8, + "token_id": 1022, + "token": "Ġfirst", + "decoded_piece": " first", + "experts_by_weight": [ + { + "expert": 29, + "weight": 0.10442701727151871 + }, + { + "expert": 11, + "weight": 0.08264623582363129 + }, + { + "expert": 14, + "weight": 0.049039799720048904 + }, + { + "expert": 28, + "weight": 0.046509917825460434 + }, + { + "expert": 59, + "weight": 0.029447149485349655 + }, + { + "expert": 58, + "weight": 0.027899539098143578 + } + ], + "selected_weight_sum": 0.33996965922415257, + "top1_top2_margin": 0.02178078144788742 + }, + { + "position": 9, + "token_id": 9842, + "token": "Ġrepeated", + "decoded_piece": " repeated", + "experts_by_weight": [ + { + "expert": 11, + "weight": 0.23813141882419586 + }, + { + "expert": 28, + "weight": 0.16236260533332825 + }, + { + "expert": 29, + "weight": 0.07772964239120483 + }, + { + "expert": 7, + "weight": 0.04792751371860504 + }, + { + "expert": 14, + "weight": 0.04581587761640549 + }, + { + "expert": 4, + "weight": 0.04462626948952675 + } + ], + "selected_weight_sum": 0.6165933273732662, + "top1_top2_margin": 0.07576881349086761 + }, + { + "position": 10, + "token_id": 3699, + "token": "Ġelement", + "decoded_piece": " element", + "experts_by_weight": [ + { + "expert": 28, + "weight": 0.10746581852436066 + }, + { + "expert": 4, + "weight": 0.081493079662323 + }, + { + "expert": 2, + "weight": 0.07607961446046829 + }, + { + "expert": 34, + "weight": 0.07128653675317764 + }, + { + "expert": 11, + "weight": 0.05328736826777458 + }, + { + "expert": 29, + "weight": 0.05184241756796837 + } + ], + "selected_weight_sum": 0.44145483523607254, + "top1_top2_margin": 0.02597273886203766 + }, + { + "position": 11, + "token_id": 279, + "token": "Ġin", + "decoded_piece": " in", + "experts_by_weight": [ + { + "expert": 60, + "weight": 0.1786193996667862 + }, + { + "expert": 14, + "weight": 0.17806856334209442 + }, + { + "expert": 58, + "weight": 0.09450752288103104 + }, + { + "expert": 23, + "weight": 0.07455217838287354 + }, + { + "expert": 32, + "weight": 0.02587931603193283 + }, + { + "expert": 56, + "weight": 0.025224294513463974 + } + ], + "selected_weight_sum": 0.576851274818182, + "top1_top2_margin": 0.0005508363246917725 + }, + { + "position": 12, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 60, + "weight": 0.1739369034767151 + }, + { + "expert": 29, + "weight": 0.11336424946784973 + }, + { + "expert": 58, + "weight": 0.07465675473213196 + }, + { + "expert": 48, + "weight": 0.06730871647596359 + }, + { + "expert": 23, + "weight": 0.052558671683073044 + }, + { + "expert": 14, + "weight": 0.0453951470553875 + } + ], + "selected_weight_sum": 0.5272204428911209, + "top1_top2_margin": 0.060572654008865356 + }, + { + "position": 13, + "token_id": 1525, + "token": "Ġlist", + "decoded_piece": " list", + "experts_by_weight": [ + { + "expert": 29, + "weight": 0.12000925838947296 + }, + { + "expert": 34, + "weight": 0.11664143949747086 + }, + { + "expert": 4, + "weight": 0.04033638536930084 + }, + { + "expert": 30, + "weight": 0.038349755108356476 + }, + { + "expert": 2, + "weight": 0.03568904474377632 + }, + { + "expert": 1, + "weight": 0.031670890748500824 + } + ], + "selected_weight_sum": 0.3826967738568783, + "top1_top2_margin": 0.0033678188920021057 + }, + { + "position": 14, + "token_id": 285, + "token": "Ġand", + "decoded_piece": " and", + "experts_by_weight": [ + { + "expert": 24, + "weight": 0.09494847059249878 + }, + { + "expert": 56, + "weight": 0.07503247261047363 + }, + { + "expert": 12, + "weight": 0.07015768438577652 + }, + { + "expert": 43, + "weight": 0.06664929538965225 + }, + { + "expert": 14, + "weight": 0.06461586058139801 + }, + { + "expert": 53, + "weight": 0.05943237245082855 + } + ], + "selected_weight_sum": 0.43083615601062775, + "top1_top2_margin": 0.019915997982025146 + }, + { + "position": 15, + "token_id": 6716, + "token": "Ġexplain", + "decoded_piece": " explain", + "experts_by_weight": [ + { + "expert": 3, + "weight": 0.1462646871805191 + }, + { + "expert": 24, + "weight": 0.12913043797016144 + }, + { + "expert": 6, + "weight": 0.06748204678297043 + }, + { + "expert": 19, + "weight": 0.062301188707351685 + }, + { + "expert": 23, + "weight": 0.04884091764688492 + }, + { + "expert": 22, + "weight": 0.045381322503089905 + } + ], + "selected_weight_sum": 0.4994006007909775, + "top1_top2_margin": 0.017134249210357666 + }, + { + "position": 16, + "token_id": 895, + "token": "Ġits", + "decoded_piece": " its", + "experts_by_weight": [ + { + "expert": 48, + "weight": 0.10237778723239899 + }, + { + "expert": 3, + "weight": 0.08468203991651535 + }, + { + "expert": 6, + "weight": 0.07570596784353256 + }, + { + "expert": 16, + "weight": 0.0703137144446373 + }, + { + "expert": 0, + "weight": 0.038703132420778275 + }, + { + "expert": 24, + "weight": 0.0378028005361557 + } + ], + "selected_weight_sum": 0.4095854423940182, + "top1_top2_margin": 0.017695747315883636 + }, + { + "position": 17, + "token_id": 13913, + "token": "Ġcomplexity", + "decoded_piece": " complexity", + "experts_by_weight": [ + { + "expert": 37, + "weight": 0.1097944900393486 + }, + { + "expert": 7, + "weight": 0.0862223356962204 + }, + { + "expert": 46, + "weight": 0.07457230985164642 + }, + { + "expert": 6, + "weight": 0.05728883668780327 + }, + { + "expert": 11, + "weight": 0.05647817999124527 + }, + { + "expert": 19, + "weight": 0.037518858909606934 + } + ], + "selected_weight_sum": 0.4218750111758709, + "top1_top2_margin": 0.023572154343128204 + }, + { + "position": 18, + "token_id": 13, + "token": ".", + "decoded_piece": ".", + "experts_by_weight": [ + { + "expert": 12, + "weight": 0.12260974943637848 + }, + { + "expert": 42, + "weight": 0.11430303007364273 + }, + { + "expert": 27, + "weight": 0.09481319785118103 + }, + { + "expert": 43, + "weight": 0.0880577489733696 + }, + { + "expert": 39, + "weight": 0.07388219237327576 + }, + { + "expert": 18, + "weight": 0.05371580645442009 + } + ], + "selected_weight_sum": 0.5473817251622677, + "top1_top2_margin": 0.008306719362735748 + } + ] + }, + { + "id": "math", + "label": "数学推理", + "tokens": 27, + "load": [ + 2, + 9, + 0, + 1, + 4, + 0, + 0, + 3, + 4, + 0, + 6, + 3, + 3, + 2, + 19, + 4, + 3, + 1, + 3, + 4, + 2, + 1, + 1, + 4, + 2, + 6, + 1, + 2, + 4, + 0, + 0, + 0, + 0, + 2, + 7, + 2, + 4, + 0, + 1, + 2, + 2, + 1, + 5, + 2, + 2, + 0, + 0, + 1, + 1, + 2, + 0, + 2, + 2, + 8, + 2, + 5, + 3, + 0, + 0, + 4, + 0, + 0, + 2, + 6 + ], + "metrics": { + "routes": 162, + "used_experts": 48, + "zero_experts": 16, + "mean": 2.53125, + "std": 2.975928544998169, + "cv": 1.1756755113601685, + "gini": 0.5358796119689941, + "entropy_nats": 3.6021676063537598, + "effective_experts": 36.67765105006978, + "min": 0.0, + "p50": 2.0, + "p90": 5.6999969482421875, + "max": 19.0 + }, + "top_experts": [ + { + "expert": 14, + "routes": 19 + }, + { + "expert": 1, + "routes": 9 + }, + { + "expert": 53, + "routes": 8 + }, + { + "expert": 34, + "routes": 7 + }, + { + "expert": 10, + "routes": 6 + }, + { + "expert": 25, + "routes": 6 + }, + { + "expert": 63, + "routes": 6 + }, + { + "expert": 42, + "routes": 5 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.6668257713317871 + }, + { + "expert": 47, + "weight": 0.015490565448999405 + }, + { + "expert": 19, + "weight": 0.012896407395601273 + }, + { + "expert": 20, + "weight": 0.01244935393333435 + }, + { + "expert": 55, + "weight": 0.012159599922597408 + }, + { + "expert": 49, + "weight": 0.011999384500086308 + } + ], + "selected_weight_sum": 0.7318210825324059, + "top1_top2_margin": 0.6513352058827877 + }, + { + "position": 1, + "token_id": 8246, + "token": "èĭ¥", + "decoded_piece": "若", + "experts_by_weight": [ + { + "expert": 35, + "weight": 0.2157912403345108 + }, + { + "expert": 38, + "weight": 0.20943520963191986 + }, + { + "expert": 36, + "weight": 0.05833398550748825 + }, + { + "expert": 41, + "weight": 0.043179191648960114 + }, + { + "expert": 49, + "weight": 0.03460429981350899 + }, + { + "expert": 15, + "weight": 0.03122684173285961 + } + ], + "selected_weight_sum": 0.5925707686692476, + "top1_top2_margin": 0.006356030702590942 + }, + { + "position": 2, + "token_id": 2206, + "token": "æŃ£", + "decoded_piece": "正", + "experts_by_weight": [ + { + "expert": 36, + "weight": 0.2386857569217682 + }, + { + "expert": 35, + "weight": 0.07329999655485153 + }, + { + "expert": 11, + "weight": 0.07202289253473282 + }, + { + "expert": 52, + "weight": 0.038597580045461655 + }, + { + "expert": 44, + "weight": 0.026481110602617264 + }, + { + "expert": 51, + "weight": 0.022812768816947937 + } + ], + "selected_weight_sum": 0.4719001054763794, + "top1_top2_margin": 0.16538576036691666 + }, + { + "position": 3, + "token_id": 2266, + "token": "æķ°", + "decoded_piece": "数", + "experts_by_weight": [ + { + "expert": 34, + "weight": 0.09990332275629044 + }, + { + "expert": 20, + "weight": 0.09463457763195038 + }, + { + "expert": 14, + "weight": 0.07485338300466537 + }, + { + "expert": 4, + "weight": 0.06988027691841125 + }, + { + "expert": 59, + "weight": 0.04555794224143028 + }, + { + "expert": 8, + "weight": 0.044183291494846344 + } + ], + "selected_weight_sum": 0.42901279404759407, + "top1_top2_margin": 0.005268745124340057 + }, + { + "position": 4, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 63, + "weight": 0.10638757795095444 + }, + { + "expert": 42, + "weight": 0.09196751564741135 + }, + { + "expert": 59, + "weight": 0.051636818796396255 + }, + { + "expert": 14, + "weight": 0.05064014717936516 + }, + { + "expert": 25, + "weight": 0.04116741195321083 + }, + { + "expert": 10, + "weight": 0.03419582173228264 + } + ], + "selected_weight_sum": 0.37599529325962067, + "top1_top2_margin": 0.01442006230354309 + }, + { + "position": 5, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 63, + "weight": 0.9029526710510254 + }, + { + "expert": 15, + "weight": 0.004224724601954222 + }, + { + "expert": 8, + "weight": 0.0033379215747117996 + }, + { + "expert": 1, + "weight": 0.003321886295452714 + }, + { + "expert": 7, + "weight": 0.0029146261513233185 + }, + { + "expert": 54, + "weight": 0.0027777589857578278 + } + ], + "selected_weight_sum": 0.9195295886602253, + "top1_top2_margin": 0.8987279464490712 + }, + { + "position": 6, + "token_id": 13211, + "token": "满足", + "decoded_piece": "满足", + "experts_by_weight": [ + { + "expert": 14, + "weight": 0.28209275007247925 + }, + { + "expert": 23, + "weight": 0.07677952200174332 + }, + { + "expert": 62, + "weight": 0.061268191784620285 + }, + { + "expert": 56, + "weight": 0.04783352091908455 + }, + { + "expert": 36, + "weight": 0.0477428138256073 + }, + { + "expert": 10, + "weight": 0.03472767025232315 + } + ], + "selected_weight_sum": 0.5504444688558578, + "top1_top2_margin": 0.20531322807073593 + }, + { + "position": 7, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 42, + "weight": 0.15821927785873413 + }, + { + "expert": 25, + "weight": 0.049939896911382675 + }, + { + "expert": 34, + "weight": 0.0359053835272789 + }, + { + "expert": 13, + "weight": 0.035540301352739334 + }, + { + "expert": 15, + "weight": 0.03167761489748955 + }, + { + "expert": 16, + "weight": 0.029953069984912872 + } + ], + "selected_weight_sum": 0.34123554453253746, + "top1_top2_margin": 0.10827938094735146 + }, + { + "position": 8, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 18, + "weight": 0.13474272191524506 + }, + { + "expert": 53, + "weight": 0.08470243215560913 + }, + { + "expert": 14, + "weight": 0.06641152501106262 + }, + { + "expert": 55, + "weight": 0.049692537635564804 + }, + { + "expert": 63, + "weight": 0.04948724806308746 + }, + { + "expert": 10, + "weight": 0.04123137518763542 + } + ], + "selected_weight_sum": 0.4262678399682045, + "top1_top2_margin": 0.050040289759635925 + }, + { + "position": 9, + "token_id": 14855, + "token": "åĬłä¸Ĭ", + "decoded_piece": "加上", + "experts_by_weight": [ + { + "expert": 14, + "weight": 0.15503190457820892 + }, + { + "expert": 10, + "weight": 0.11396767944097519 + }, + { + "expert": 53, + "weight": 0.0618465431034565 + }, + { + "expert": 55, + "weight": 0.03268251195549965 + }, + { + "expert": 23, + "weight": 0.03129611164331436 + }, + { + "expert": 22, + "weight": 0.02996181882917881 + } + ], + "selected_weight_sum": 0.42478656955063343, + "top1_top2_margin": 0.041064225137233734 + }, + { + "position": 10, + "token_id": 12985, + "token": "å®ĥçļĦ", + "decoded_piece": "它的", + "experts_by_weight": [ + { + "expert": 14, + "weight": 0.12337606400251389 + }, + { + "expert": 1, + "weight": 0.10133391618728638 + }, + { + "expert": 16, + "weight": 0.07631479203701019 + }, + { + "expert": 0, + "weight": 0.06342805176973343 + }, + { + "expert": 48, + "weight": 0.06302573531866074 + }, + { + "expert": 36, + "weight": 0.046065572649240494 + } + ], + "selected_weight_sum": 0.4735441319644451, + "top1_top2_margin": 0.02204214781522751 + }, + { + "position": 11, + "token_id": 8270, + "token": "åĢĴ", + "decoded_piece": "倒", + "experts_by_weight": [ + { + "expert": 14, + "weight": 0.1234094426035881 + }, + { + "expert": 1, + "weight": 0.0897468701004982 + }, + { + "expert": 11, + "weight": 0.06823419779539108 + }, + { + "expert": 7, + "weight": 0.06477421522140503 + }, + { + "expert": 28, + "weight": 0.04059724509716034 + }, + { + "expert": 33, + "weight": 0.03771129250526428 + } + ], + "selected_weight_sum": 0.42447326332330704, + "top1_top2_margin": 0.033662572503089905 + }, + { + "position": 12, + "token_id": 2266, + "token": "æķ°", + "decoded_piece": "数", + "experts_by_weight": [ + { + "expert": 19, + "weight": 0.09373485296964645 + }, + { + "expert": 4, + "weight": 0.08521629124879837 + }, + { + "expert": 34, + "weight": 0.079887256026268 + }, + { + "expert": 14, + "weight": 0.07751046121120453 + }, + { + "expert": 8, + "weight": 0.07490264624357224 + }, + { + "expert": 1, + "weight": 0.05471431463956833 + } + ], + "selected_weight_sum": 0.4659658223390579, + "top1_top2_margin": 0.008518561720848083 + }, + { + "position": 13, + "token_id": 37360, + "token": "çŃīäºİ", + "decoded_piece": "等于", + "experts_by_weight": [ + { + "expert": 14, + "weight": 0.180343896150589 + }, + { + "expert": 10, + "weight": 0.11404763162136078 + }, + { + "expert": 0, + "weight": 0.06184849888086319 + }, + { + "expert": 62, + "weight": 0.06134380027651787 + }, + { + "expert": 53, + "weight": 0.04147454723715782 + }, + { + "expert": 17, + "weight": 0.029968442395329475 + } + ], + "selected_weight_sum": 0.4890268165618181, + "top1_top2_margin": 0.06629626452922821 + }, + { + "position": 14, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 18, + "weight": 0.14621861279010773 + }, + { + "expert": 42, + "weight": 0.13439302146434784 + }, + { + "expert": 15, + "weight": 0.07689664512872696 + }, + { + "expert": 63, + "weight": 0.05354950577020645 + }, + { + "expert": 12, + "weight": 0.05216725543141365 + }, + { + "expert": 39, + "weight": 0.0497673936188221 + } + ], + "selected_weight_sum": 0.5129924342036247, + "top1_top2_margin": 0.011825591325759888 + }, + { + "position": 15, + "token_id": 18, + "token": "3", + "decoded_piece": "3", + "experts_by_weight": [ + { + "expert": 1, + "weight": 0.12946932017803192 + }, + { + "expert": 25, + "weight": 0.10507825016975403 + }, + { + "expert": 59, + "weight": 0.07482244819402695 + }, + { + "expert": 40, + "weight": 0.042674217373132706 + }, + { + "expert": 28, + "weight": 0.02928566187620163 + }, + { + "expert": 4, + "weight": 0.027979303151369095 + } + ], + "selected_weight_sum": 0.4093092009425163, + "top1_top2_margin": 0.024391070008277893 + }, + { + "position": 16, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 14, + "weight": 0.17965057492256165 + }, + { + "expert": 27, + "weight": 0.10295615345239639 + }, + { + "expert": 12, + "weight": 0.088253453373909 + }, + { + "expert": 43, + "weight": 0.07324301451444626 + }, + { + "expert": 53, + "weight": 0.058590613305568695 + }, + { + "expert": 40, + "weight": 0.044678375124931335 + } + ], + "selected_weight_sum": 0.5473721846938133, + "top1_top2_margin": 0.07669442147016525 + }, + { + "position": 17, + "token_id": 2989, + "token": "æ±Ĥ", + "decoded_piece": "求", + "experts_by_weight": [ + { + "expert": 14, + "weight": 0.15167519450187683 + }, + { + "expert": 3, + "weight": 0.07489567995071411 + }, + { + "expert": 19, + "weight": 0.05545367673039436 + }, + { + "expert": 51, + "weight": 0.05460234358906746 + }, + { + "expert": 21, + "weight": 0.05284403637051582 + }, + { + "expert": 23, + "weight": 0.04110542684793472 + } + ], + "selected_weight_sum": 0.4305763579905033, + "top1_top2_margin": 0.07677951455116272 + }, + { + "position": 18, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 42, + "weight": 0.1560184359550476 + }, + { + "expert": 16, + "weight": 0.039979103952646255 + }, + { + "expert": 34, + "weight": 0.039141543209552765 + }, + { + "expert": 25, + "weight": 0.03823362663388252 + }, + { + "expert": 13, + "weight": 0.0362531840801239 + }, + { + "expert": 63, + "weight": 0.034050628542900085 + } + ], + "selected_weight_sum": 0.34367652237415314, + "top1_top2_margin": 0.11603933200240135 + }, + { + "position": 19, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 18, + "weight": 0.08407339453697205 + }, + { + "expert": 63, + "weight": 0.08105730265378952 + }, + { + "expert": 53, + "weight": 0.07007438689470291 + }, + { + "expert": 55, + "weight": 0.06364663690328598 + }, + { + "expert": 14, + "weight": 0.045173339545726776 + }, + { + "expert": 56, + "weight": 0.037717703729867935 + } + ], + "selected_weight_sum": 0.38174276426434517, + "top1_top2_margin": 0.0030160918831825256 + }, + { + "position": 20, + "token_id": 337, + "token": "çļĦ", + "decoded_piece": "的", + "experts_by_weight": [ + { + "expert": 14, + "weight": 0.1197085976600647 + }, + { + "expert": 1, + "weight": 0.0795862078666687 + }, + { + "expert": 55, + "weight": 0.06330224871635437 + }, + { + "expert": 33, + "weight": 0.057824112474918365 + }, + { + "expert": 54, + "weight": 0.03476481884717941 + }, + { + "expert": 56, + "weight": 0.030369188636541367 + } + ], + "selected_weight_sum": 0.3855551742017269, + "top1_top2_margin": 0.040122389793395996 + }, + { + "position": 21, + "token_id": 18718, + "token": "å¹³æĸ¹", + "decoded_piece": "平方", + "experts_by_weight": [ + { + "expert": 53, + "weight": 0.11117389798164368 + }, + { + "expert": 1, + "weight": 0.07823505997657776 + }, + { + "expert": 14, + "weight": 0.06898786872625351 + }, + { + "expert": 59, + "weight": 0.049189817160367966 + }, + { + "expert": 25, + "weight": 0.0485650971531868 + }, + { + "expert": 34, + "weight": 0.036493945866823196 + } + ], + "selected_weight_sum": 0.3926456868648529, + "top1_top2_margin": 0.03293883800506592 + }, + { + "position": 22, + "token_id": 14855, + "token": "åĬłä¸Ĭ", + "decoded_piece": "加上", + "experts_by_weight": [ + { + "expert": 14, + "weight": 0.1784549057483673 + }, + { + "expert": 10, + "weight": 0.06671992689371109 + }, + { + "expert": 53, + "weight": 0.05699232220649719 + }, + { + "expert": 23, + "weight": 0.04226744547486305 + }, + { + "expert": 24, + "weight": 0.03548656776547432 + }, + { + "expert": 26, + "weight": 0.03004826046526432 + } + ], + "selected_weight_sum": 0.4099694285541773, + "top1_top2_margin": 0.11173497885465622 + }, + { + "position": 23, + "token_id": 8270, + "token": "åĢĴ", + "decoded_piece": "倒", + "experts_by_weight": [ + { + "expert": 14, + "weight": 0.07740753144025803 + }, + { + "expert": 7, + "weight": 0.06655178219079971 + }, + { + "expert": 11, + "weight": 0.0630197674036026 + }, + { + "expert": 24, + "weight": 0.05319470912218094 + }, + { + "expert": 44, + "weight": 0.05203474313020706 + }, + { + "expert": 28, + "weight": 0.03889422491192818 + } + ], + "selected_weight_sum": 0.3511027581989765, + "top1_top2_margin": 0.010855749249458313 + }, + { + "position": 24, + "token_id": 51971, + "token": "æķ°çļĦ", + "decoded_piece": "数的", + "experts_by_weight": [ + { + "expert": 1, + "weight": 0.08456672728061676 + }, + { + "expert": 8, + "weight": 0.08250632137060165 + }, + { + "expert": 14, + "weight": 0.06846188753843307 + }, + { + "expert": 34, + "weight": 0.06562276184558868 + }, + { + "expert": 19, + "weight": 0.062368445098400116 + }, + { + "expert": 4, + "weight": 0.04836544767022133 + } + ], + "selected_weight_sum": 0.4118915908038616, + "top1_top2_margin": 0.002060405910015106 + }, + { + "position": 25, + "token_id": 18718, + "token": "å¹³æĸ¹", + "decoded_piece": "平方", + "experts_by_weight": [ + { + "expert": 1, + "weight": 0.10623366385698318 + }, + { + "expert": 53, + "weight": 0.08463753014802933 + }, + { + "expert": 28, + "weight": 0.05453135445713997 + }, + { + "expert": 14, + "weight": 0.051909394562244415 + }, + { + "expert": 25, + "weight": 0.04595055803656578 + }, + { + "expert": 34, + "weight": 0.040663786232471466 + } + ], + "selected_weight_sum": 0.38392628729343414, + "top1_top2_margin": 0.021596133708953857 + }, + { + "position": 26, + "token_id": 398, + "token": "ãĢĤ", + "decoded_piece": "。", + "experts_by_weight": [ + { + "expert": 12, + "weight": 0.12392057478427887 + }, + { + "expert": 42, + "weight": 0.1172504797577858 + }, + { + "expert": 14, + "weight": 0.0955967903137207 + }, + { + "expert": 43, + "weight": 0.08561467379331589 + }, + { + "expert": 27, + "weight": 0.05499466508626938 + }, + { + "expert": 39, + "weight": 0.05483439192175865 + } + ], + "selected_weight_sum": 0.5322115756571293, + "top1_top2_margin": 0.0066700950264930725 + } + ] + } + ], + "prompt_pair_jaccard": [ + { + "left": "zh_explanation", + "right": "en_architecture", + "used_expert_jaccard": 0.7719298245614035, + "shared_experts": 44, + "union_experts": 57 + }, + { + "left": "zh_explanation", + "right": "code", + "used_expert_jaccard": 0.7586206896551724, + "shared_experts": 44, + "union_experts": 58 + }, + { + "left": "zh_explanation", + "right": "math", + "used_expert_jaccard": 0.7457627118644068, + "shared_experts": 44, + "union_experts": 59 + }, + { + "left": "en_architecture", + "right": "code", + "used_expert_jaccard": 0.6909090909090909, + "shared_experts": 38, + "union_experts": 55 + }, + { + "left": "en_architecture", + "right": "math", + "used_expert_jaccard": 0.6491228070175439, + "shared_experts": 37, + "union_experts": 57 + }, + { + "left": "code", + "right": "math", + "used_expert_jaccard": 0.6379310344827587, + "shared_experts": 37, + "union_experts": 58 + } + ] + } + }, + { + "layer": 6, + "ffn": "moe", + "state_tensors": 203, + "state_numel": 584847872, + "state_bytes": 1169695744, + "timing_ms": { + "load_cpu": 24.91675701458007, + "transfer_to_device": 123.45660501159728, + "forward": 3.9505639870185405 + }, + "peak_device_allocated_mib": 1179.1787109375, + "hidden_before": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 3.2100954055786133, + "p10": 0.22690242528915405, + "p50": 0.26399680972099304, + "p90": 0.33678531646728516 + }, + "l2_norm": { + "mean": 145.2723388671875, + "p10": 10.268431663513184, + "p50": 11.947131156921387, + "p90": 15.24116325378418 + }, + "finite": true, + "sha256_fp32": "88f158b68ecb980c55b1b77eebfeb2276d7820443732aa10effca9c85c653481" + }, + "hidden_after": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 3.3090813159942627, + "p10": 0.2696530222892761, + "p50": 0.31665489077568054, + "p90": 0.3847275674343109 + }, + "l2_norm": { + "mean": 149.75193786621094, + "p10": 12.203103065490723, + "p50": 14.330163955688477, + "p90": 17.410781860351562 + }, + "finite": true, + "sha256_fp32": "1ffd41521b5a32af38695b767a5ecc556ddeae1e1efe5339743e5c8c7644030e" + }, + "mla": { + "compressed_projection_shape": [ + 4, + 27, + 576 + ], + "latent_shape": [ + 4, + 27, + 512 + ], + "rope_key_shape": [ + 4, + 27, + 64 + ], + "latent_rms": 2.3470914363861084, + "rope_key_rms": 2.1612274646759033, + "eager_key_cache_shape": [ + 4, + 16, + 27, + 192 + ], + "eager_value_cache_shape": [ + 4, + 16, + 27, + 128 + ], + "eager_cache_bytes": 1105920 + }, + "routing": { + "aggregate_load": [ + 16, + 3, + 14, + 4, + 25, + 9, + 7, + 14, + 11, + 5, + 0, + 3, + 9, + 5, + 2, + 6, + 10, + 17, + 6, + 10, + 4, + 7, + 6, + 5, + 2, + 40, + 7, + 6, + 16, + 10, + 11, + 9, + 7, + 2, + 6, + 1, + 19, + 16, + 3, + 8, + 10, + 9, + 10, + 14, + 8, + 6, + 10, + 8, + 2, + 5, + 7, + 2, + 11, + 5, + 2, + 6, + 17, + 14, + 11, + 8, + 7, + 0, + 0, + 7 + ], + "aggregate_metrics": { + "routes": 540, + "used_experts": 61, + "zero_experts": 3, + "mean": 8.4375, + "std": 6.458352565765381, + "cv": 0.7654343843460083, + "gini": 0.3778356611728668, + "entropy_nats": 3.9015467166900635, + "effective_experts": 49.47891982211325, + "min": 0.0, + "p50": 7.0, + "p90": 16.0, + "max": 40.0 + }, + "top_experts": [ + { + "expert": 25, + "routes": 40 + }, + { + "expert": 4, + "routes": 25 + }, + { + "expert": 36, + "routes": 19 + }, + { + "expert": 17, + "routes": 17 + }, + { + "expert": 56, + "routes": 17 + }, + { + "expert": 0, + "routes": 16 + }, + { + "expert": 28, + "routes": 16 + }, + { + "expert": 37, + "routes": 16 + }, + { + "expert": 2, + "routes": 14 + }, + { + "expert": 7, + "routes": 14 + }, + { + "expert": 43, + "routes": 14 + }, + { + "expert": 57, + "routes": 14 + } + ], + "selected_weight_sum": { + "mean": 0.44403982162475586, + "p10": 0.332468181848526, + "p50": 0.425670325756073, + "p90": 0.571208655834198 + }, + "top1_top2_margin": { + "mean": 0.053601086139678955, + "p50": 0.026227694004774094, + "p90": 0.07640297710895538 + }, + "per_prompt": [ + { + "id": "zh_explanation", + "label": "中文解释", + "tokens": 26, + "load": [ + 7, + 1, + 6, + 3, + 1, + 3, + 1, + 3, + 6, + 0, + 0, + 0, + 4, + 1, + 0, + 1, + 3, + 4, + 3, + 2, + 0, + 2, + 2, + 1, + 1, + 13, + 2, + 1, + 7, + 5, + 1, + 1, + 1, + 1, + 1, + 0, + 5, + 3, + 0, + 2, + 2, + 2, + 3, + 8, + 2, + 2, + 4, + 4, + 0, + 0, + 1, + 0, + 5, + 3, + 1, + 0, + 6, + 5, + 4, + 0, + 6, + 0, + 0, + 0 + ], + "metrics": { + "routes": 156, + "used_experts": 49, + "zero_experts": 15, + "mean": 2.4375, + "std": 2.5126662254333496, + "cv": 1.0308374166488647, + "gini": 0.5288461446762085, + "entropy_nats": 3.646120071411133, + "effective_experts": 38.325676313953316, + "min": 0.0, + "p50": 2.0, + "p90": 6.0, + "max": 13.0 + }, + "top_experts": [ + { + "expert": 25, + "routes": 13 + }, + { + "expert": 43, + "routes": 8 + }, + { + "expert": 0, + "routes": 7 + }, + { + "expert": 28, + "routes": 7 + }, + { + "expert": 2, + "routes": 6 + }, + { + "expert": 8, + "routes": 6 + }, + { + "expert": 56, + "routes": 6 + }, + { + "expert": 60, + "routes": 6 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 22, + "weight": 0.5437576770782471 + }, + { + "expert": 21, + "weight": 0.02887026220560074 + }, + { + "expert": 13, + "weight": 0.022947292774915695 + }, + { + "expert": 42, + "weight": 0.020739872008562088 + }, + { + "expert": 47, + "weight": 0.019217059016227722 + }, + { + "expert": 32, + "weight": 0.01740080676972866 + } + ], + "selected_weight_sum": 0.652932969853282, + "top1_top2_margin": 0.5148874148726463 + }, + { + "position": 1, + "token_id": 1082, + "token": "ç͍", + "decoded_piece": "用", + "experts_by_weight": [ + { + "expert": 46, + "weight": 0.19940678775310516 + }, + { + "expert": 37, + "weight": 0.08581104874610901 + }, + { + "expert": 56, + "weight": 0.07319452613592148 + }, + { + "expert": 19, + "weight": 0.05708487331867218 + }, + { + "expert": 39, + "weight": 0.04857153818011284 + }, + { + "expert": 26, + "weight": 0.03971230611205101 + } + ], + "selected_weight_sum": 0.5037810802459717, + "top1_top2_margin": 0.11359573900699615 + }, + { + "position": 2, + "token_id": 83983, + "token": "éĢļä¿Ĺ", + "decoded_piece": "通俗", + "experts_by_weight": [ + { + "expert": 57, + "weight": 0.0797773152589798 + }, + { + "expert": 2, + "weight": 0.07066472619771957 + }, + { + "expert": 24, + "weight": 0.05039288476109505 + }, + { + "expert": 53, + "weight": 0.047804661095142365 + }, + { + "expert": 19, + "weight": 0.04701150208711624 + }, + { + "expert": 58, + "weight": 0.03956286236643791 + } + ], + "selected_weight_sum": 0.33521395176649094, + "top1_top2_margin": 0.009112589061260223 + }, + { + "position": 3, + "token_id": 91476, + "token": "çļĦè¯Ńè¨Ģ", + "decoded_piece": "的语言", + "experts_by_weight": [ + { + "expert": 2, + "weight": 0.20149613916873932 + }, + { + "expert": 29, + "weight": 0.06121385097503662 + }, + { + "expert": 53, + "weight": 0.045187439769506454 + }, + { + "expert": 57, + "weight": 0.04054460674524307 + }, + { + "expert": 60, + "weight": 0.04029814898967743 + }, + { + "expert": 3, + "weight": 0.038398370146751404 + } + ], + "selected_weight_sum": 0.4271385557949543, + "top1_top2_margin": 0.1402822881937027 + }, + { + "position": 4, + "token_id": 17882, + "token": "è§£éĩĬ", + "decoded_piece": "解释", + "experts_by_weight": [ + { + "expert": 44, + "weight": 0.08854366093873978 + }, + { + "expert": 2, + "weight": 0.07951278984546661 + }, + { + "expert": 5, + "weight": 0.06590777635574341 + }, + { + "expert": 46, + "weight": 0.06553405523300171 + }, + { + "expert": 7, + "weight": 0.055855643004179 + }, + { + "expert": 4, + "weight": 0.038224026560783386 + } + ], + "selected_weight_sum": 0.3935779519379139, + "top1_top2_margin": 0.009030871093273163 + }, + { + "position": 5, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 16, + "weight": 0.13611505925655365 + }, + { + "expert": 56, + "weight": 0.1061541959643364 + }, + { + "expert": 5, + "weight": 0.10146880894899368 + }, + { + "expert": 46, + "weight": 0.0781978890299797 + }, + { + "expert": 12, + "weight": 0.07308310270309448 + }, + { + "expert": 43, + "weight": 0.058526765555143356 + } + ], + "selected_weight_sum": 0.5535458214581013, + "top1_top2_margin": 0.029960863292217255 + }, + { + "position": 6, + "token_id": 10285, + "token": "为ä»Ģä¹Ī", + "decoded_piece": "为什么", + "experts_by_weight": [ + { + "expert": 5, + "weight": 0.12138903886079788 + }, + { + "expert": 43, + "weight": 0.0896969884634018 + }, + { + "expert": 56, + "weight": 0.07434116303920746 + }, + { + "expert": 6, + "weight": 0.043124958872795105 + }, + { + "expert": 52, + "weight": 0.042676858603954315 + }, + { + "expert": 46, + "weight": 0.04177377000451088 + } + ], + "selected_weight_sum": 0.41300277784466743, + "top1_top2_margin": 0.03169205039739609 + }, + { + "position": 7, + "token_id": 21844, + "token": "ç¨Ģ", + "decoded_piece": "稀", + "experts_by_weight": [ + { + "expert": 25, + "weight": 0.06200406327843666 + }, + { + "expert": 26, + "weight": 0.05429893359541893 + }, + { + "expert": 27, + "weight": 0.04909799247980118 + }, + { + "expert": 41, + "weight": 0.047943904995918274 + }, + { + "expert": 60, + "weight": 0.03833471238613129 + }, + { + "expert": 57, + "weight": 0.03801535442471504 + } + ], + "selected_weight_sum": 0.28969496116042137, + "top1_top2_margin": 0.007705129683017731 + }, + { + "position": 8, + "token_id": 20923, + "token": "çĸı", + "decoded_piece": "疏", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.17407171428203583 + }, + { + "expert": 25, + "weight": 0.1118895411491394 + }, + { + "expert": 57, + "weight": 0.09100628644227982 + }, + { + "expert": 31, + "weight": 0.08462925255298615 + }, + { + "expert": 60, + "weight": 0.04508446529507637 + }, + { + "expert": 30, + "weight": 0.03972841799259186 + } + ], + "selected_weight_sum": 0.5464096777141094, + "top1_top2_margin": 0.06218217313289642 + }, + { + "position": 9, + "token_id": 10900, + "token": "ä¸ĵå®¶", + "decoded_piece": "专家", + "experts_by_weight": [ + { + "expert": 47, + "weight": 0.10310626029968262 + }, + { + "expert": 53, + "weight": 0.08589243143796921 + }, + { + "expert": 29, + "weight": 0.05800008028745651 + }, + { + "expert": 21, + "weight": 0.05487441271543503 + }, + { + "expert": 17, + "weight": 0.052628058940172195 + }, + { + "expert": 2, + "weight": 0.04682231321930885 + } + ], + "selected_weight_sum": 0.4013235569000244, + "top1_top2_margin": 0.01721382886171341 + }, + { + "position": 10, + "token_id": 39322, + "token": "模åŀĭ", + "decoded_piece": "模型", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.14891529083251953 + }, + { + "expert": 47, + "weight": 0.12191086262464523 + }, + { + "expert": 29, + "weight": 0.0684787854552269 + }, + { + "expert": 42, + "weight": 0.06209414824843407 + }, + { + "expert": 25, + "weight": 0.0608627125620842 + }, + { + "expert": 2, + "weight": 0.05374256893992424 + } + ], + "selected_weight_sum": 0.5160043686628342, + "top1_top2_margin": 0.027004428207874298 + }, + { + "position": 11, + "token_id": 1876, + "token": "åı¯ä»¥", + "decoded_piece": "可以", + "experts_by_weight": [ + { + "expert": 43, + "weight": 0.09292622655630112 + }, + { + "expert": 0, + "weight": 0.06740861386060715 + }, + { + "expert": 52, + "weight": 0.06254689395427704 + }, + { + "expert": 33, + "weight": 0.04910411313176155 + }, + { + "expert": 56, + "weight": 0.03792039304971695 + }, + { + "expert": 47, + "weight": 0.03512315824627876 + } + ], + "selected_weight_sum": 0.34502939879894257, + "top1_top2_margin": 0.02551761269569397 + }, + { + "position": 12, + "token_id": 9476, + "token": "æĭ¥æľī", + "decoded_piece": "拥有", + "experts_by_weight": [ + { + "expert": 0, + "weight": 0.10773254930973053 + }, + { + "expert": 7, + "weight": 0.09062737226486206 + }, + { + "expert": 58, + "weight": 0.08368217200040817 + }, + { + "expert": 52, + "weight": 0.07159384340047836 + }, + { + "expert": 43, + "weight": 0.04937493056058884 + }, + { + "expert": 25, + "weight": 0.024813424795866013 + } + ], + "selected_weight_sum": 0.427824292331934, + "top1_top2_margin": 0.01710517704486847 + }, + { + "position": 13, + "token_id": 3599, + "token": "å¾Īå¤ļ", + "decoded_piece": "很多", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.1371069699525833 + }, + { + "expert": 25, + "weight": 0.08427796512842178 + }, + { + "expert": 58, + "weight": 0.08417855203151703 + }, + { + "expert": 36, + "weight": 0.04595799744129181 + }, + { + "expert": 0, + "weight": 0.045302487909793854 + }, + { + "expert": 52, + "weight": 0.041003432124853134 + } + ], + "selected_weight_sum": 0.4378274045884609, + "top1_top2_margin": 0.05282900482416153 + }, + { + "position": 14, + "token_id": 21273, + "token": "åıĤæķ°", + "decoded_piece": "参数", + "experts_by_weight": [ + { + "expert": 25, + "weight": 0.13282836973667145 + }, + { + "expert": 2, + "weight": 0.09014931321144104 + }, + { + "expert": 28, + "weight": 0.0724765732884407 + }, + { + "expert": 29, + "weight": 0.0615108460187912 + }, + { + "expert": 57, + "weight": 0.05170028656721115 + }, + { + "expert": 3, + "weight": 0.03531194105744362 + } + ], + "selected_weight_sum": 0.44397732987999916, + "top1_top2_margin": 0.04267905652523041 + }, + { + "position": 15, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 16, + "weight": 0.11967949569225311 + }, + { + "expert": 18, + "weight": 0.10622544586658478 + }, + { + "expert": 12, + "weight": 0.08323302865028381 + }, + { + "expert": 56, + "weight": 0.07838869094848633 + }, + { + "expert": 43, + "weight": 0.06008601188659668 + }, + { + "expert": 37, + "weight": 0.05834601819515228 + } + ], + "selected_weight_sum": 0.505958691239357, + "top1_top2_margin": 0.013454049825668335 + }, + { + "position": 16, + "token_id": 1705, + "token": "ä½Ĩ", + "decoded_piece": "但", + "experts_by_weight": [ + { + "expert": 18, + "weight": 0.09913211315870285 + }, + { + "expert": 34, + "weight": 0.09904447197914124 + }, + { + "expert": 54, + "weight": 0.054669421166181564 + }, + { + "expert": 12, + "weight": 0.054521091282367706 + }, + { + "expert": 0, + "weight": 0.04755903407931328 + }, + { + "expert": 43, + "weight": 0.03531811386346817 + } + ], + "selected_weight_sum": 0.3902442455291748, + "top1_top2_margin": 8.764117956161499e-05 + }, + { + "position": 17, + "token_id": 13712, + "token": "æ¯ı个", + "decoded_piece": "每个", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.08446930348873138 + }, + { + "expert": 36, + "weight": 0.07750283181667328 + }, + { + "expert": 28, + "weight": 0.06822194904088974 + }, + { + "expert": 25, + "weight": 0.0491325743496418 + }, + { + "expert": 15, + "weight": 0.037747886031866074 + }, + { + "expert": 0, + "weight": 0.0376180075109005 + } + ], + "selected_weight_sum": 0.3546925522387028, + "top1_top2_margin": 0.0069664716720581055 + }, + { + "position": 18, + "token_id": 10728, + "token": "Ġtoken", + "decoded_piece": " token", + "experts_by_weight": [ + { + "expert": 28, + "weight": 0.1324920803308487 + }, + { + "expert": 29, + "weight": 0.1233772411942482 + }, + { + "expert": 25, + "weight": 0.06866113841533661 + }, + { + "expert": 17, + "weight": 0.047023892402648926 + }, + { + "expert": 3, + "weight": 0.044400230050086975 + }, + { + "expert": 41, + "weight": 0.033736228942871094 + } + ], + "selected_weight_sum": 0.4496908113360405, + "top1_top2_margin": 0.009114839136600494 + }, + { + "position": 19, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 39, + "weight": 0.048708073794841766 + }, + { + "expert": 42, + "weight": 0.038140591233968735 + }, + { + "expert": 1, + "weight": 0.03512806445360184 + }, + { + "expert": 23, + "weight": 0.03305252268910408 + }, + { + "expert": 22, + "weight": 0.029746539890766144 + }, + { + "expert": 45, + "weight": 0.029309971258044243 + } + ], + "selected_weight_sum": 0.2140857633203268, + "top1_top2_margin": 0.010567482560873032 + }, + { + "position": 20, + "token_id": 2141, + "token": "åıª", + "decoded_piece": "只", + "experts_by_weight": [ + { + "expert": 28, + "weight": 0.09421364217996597 + }, + { + "expert": 8, + "weight": 0.08210399001836777 + }, + { + "expert": 0, + "weight": 0.0750744491815567 + }, + { + "expert": 25, + "weight": 0.05757107213139534 + }, + { + "expert": 43, + "weight": 0.0511380210518837 + }, + { + "expert": 52, + "weight": 0.031326521188020706 + } + ], + "selected_weight_sum": 0.3914276957511902, + "top1_top2_margin": 0.012109652161598206 + }, + { + "position": 21, + "token_id": 5118, + "token": "使ç͍", + "decoded_piece": "使用", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.13225959241390228 + }, + { + "expert": 0, + "weight": 0.08616293966770172 + }, + { + "expert": 25, + "weight": 0.05643099546432495 + }, + { + "expert": 43, + "weight": 0.0435882993042469 + }, + { + "expert": 44, + "weight": 0.03772665560245514 + }, + { + "expert": 60, + "weight": 0.03494388610124588 + } + ], + "selected_weight_sum": 0.3911123685538769, + "top1_top2_margin": 0.04609665274620056 + }, + { + "position": 22, + "token_id": 6802, + "token": "åħ¶ä¸Ń", + "decoded_piece": "其中", + "experts_by_weight": [ + { + "expert": 36, + "weight": 0.09671742469072342 + }, + { + "expert": 40, + "weight": 0.08142291009426117 + }, + { + "expert": 28, + "weight": 0.08033161610364914 + }, + { + "expert": 25, + "weight": 0.06336305290460587 + }, + { + "expert": 8, + "weight": 0.05350400134921074 + }, + { + "expert": 58, + "weight": 0.04034680873155594 + } + ], + "selected_weight_sum": 0.41568581387400627, + "top1_top2_margin": 0.01529451459646225 + }, + { + "position": 23, + "token_id": 51905, + "token": "ä¸Ģå°ı", + "decoded_piece": "一小", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.1243104413151741 + }, + { + "expert": 25, + "weight": 0.09060382097959518 + }, + { + "expert": 28, + "weight": 0.07431431114673615 + }, + { + "expert": 36, + "weight": 0.06538720428943634 + }, + { + "expert": 45, + "weight": 0.04744257777929306 + }, + { + "expert": 60, + "weight": 0.032568901777267456 + } + ], + "selected_weight_sum": 0.4346272572875023, + "top1_top2_margin": 0.03370662033557892 + }, + { + "position": 24, + "token_id": 5781, + "token": "éĥ¨åĪĨ", + "decoded_piece": "部分", + "experts_by_weight": [ + { + "expert": 25, + "weight": 0.139021098613739 + }, + { + "expert": 28, + "weight": 0.0859280452132225 + }, + { + "expert": 8, + "weight": 0.07490529119968414 + }, + { + "expert": 60, + "weight": 0.07370275259017944 + }, + { + "expert": 36, + "weight": 0.04788759723305702 + }, + { + "expert": 40, + "weight": 0.029776398092508316 + } + ], + "selected_weight_sum": 0.45122118294239044, + "top1_top2_margin": 0.05309305340051651 + }, + { + "position": 25, + "token_id": 398, + "token": "ãĢĤ", + "decoded_piece": "。", + "experts_by_weight": [ + { + "expert": 16, + "weight": 0.12179290503263474 + }, + { + "expert": 37, + "weight": 0.09725924581289291 + }, + { + "expert": 56, + "weight": 0.09423467516899109 + }, + { + "expert": 18, + "weight": 0.09347380697727203 + }, + { + "expert": 12, + "weight": 0.07409657537937164 + }, + { + "expert": 50, + "weight": 0.07257192581892014 + } + ], + "selected_weight_sum": 0.5534291341900826, + "top1_top2_margin": 0.02453365921974182 + } + ] + }, + { + "id": "en_architecture", + "label": "English architecture", + "tokens": 18, + "load": [ + 3, + 0, + 5, + 0, + 7, + 2, + 0, + 4, + 1, + 0, + 0, + 0, + 1, + 1, + 2, + 0, + 1, + 8, + 1, + 2, + 0, + 1, + 1, + 1, + 0, + 12, + 0, + 0, + 0, + 1, + 10, + 3, + 2, + 0, + 0, + 0, + 3, + 3, + 3, + 3, + 0, + 1, + 1, + 2, + 1, + 1, + 1, + 1, + 1, + 1, + 1, + 0, + 2, + 2, + 0, + 1, + 2, + 7, + 1, + 1, + 0, + 0, + 0, + 0 + ], + "metrics": { + "routes": 108, + "used_experts": 42, + "zero_experts": 22, + "mean": 1.6875, + "std": 2.429154872894287, + "cv": 1.4394991397857666, + "gini": 0.6342592835426331, + "entropy_nats": 3.3749022483825684, + "effective_experts": 29.221427199847042, + "min": 0.0, + "p50": 1.0, + "p90": 3.6999969482421875, + "max": 12.0 + }, + "top_experts": [ + { + "expert": 25, + "routes": 12 + }, + { + "expert": 30, + "routes": 10 + }, + { + "expert": 17, + "routes": 8 + }, + { + "expert": 4, + "routes": 7 + }, + { + "expert": 57, + "routes": 7 + }, + { + "expert": 2, + "routes": 5 + }, + { + "expert": 7, + "routes": 4 + }, + { + "expert": 0, + "routes": 3 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 22, + "weight": 0.5437576770782471 + }, + { + "expert": 21, + "weight": 0.02887026220560074 + }, + { + "expert": 13, + "weight": 0.022947292774915695 + }, + { + "expert": 42, + "weight": 0.020739872008562088 + }, + { + "expert": 47, + "weight": 0.019217059016227722 + }, + { + "expert": 32, + "weight": 0.01740080676972866 + } + ], + "selected_weight_sum": 0.652932969853282, + "top1_top2_margin": 0.5148874148726463 + }, + { + "position": 1, + "token_id": 99297, + "token": "Explain", + "decoded_piece": "Explain", + "experts_by_weight": [ + { + "expert": 19, + "weight": 0.10629317164421082 + }, + { + "expert": 46, + "weight": 0.09721477329730988 + }, + { + "expert": 37, + "weight": 0.09492044895887375 + }, + { + "expert": 39, + "weight": 0.05780201405286789 + }, + { + "expert": 4, + "weight": 0.04771720618009567 + }, + { + "expert": 48, + "weight": 0.041242267936468124 + } + ], + "selected_weight_sum": 0.4451898820698261, + "top1_top2_margin": 0.00907839834690094 + }, + { + "position": 2, + "token_id": 946, + "token": "Ġhow", + "decoded_piece": " how", + "experts_by_weight": [ + { + "expert": 19, + "weight": 0.12368354946374893 + }, + { + "expert": 5, + "weight": 0.09157591313123703 + }, + { + "expert": 52, + "weight": 0.05463002249598503 + }, + { + "expert": 7, + "weight": 0.050672709941864014 + }, + { + "expert": 39, + "weight": 0.047128066420555115 + }, + { + "expert": 56, + "weight": 0.03596038371324539 + } + ], + "selected_weight_sum": 0.4036506451666355, + "top1_top2_margin": 0.0321076363325119 + }, + { + "position": 3, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 52, + "weight": 0.0887380987405777 + }, + { + "expert": 8, + "weight": 0.08550800383090973 + }, + { + "expert": 5, + "weight": 0.07985977083444595 + }, + { + "expert": 36, + "weight": 0.06503360718488693 + }, + { + "expert": 58, + "weight": 0.04625506326556206 + }, + { + "expert": 23, + "weight": 0.0421152301132679 + } + ], + "selected_weight_sum": 0.40750977396965027, + "top1_top2_margin": 0.0032300949096679688 + }, + { + "position": 4, + "token_id": 29640, + "token": "Ġcompressed", + "decoded_piece": " compressed", + "experts_by_weight": [ + { + "expert": 57, + "weight": 0.1338948905467987 + }, + { + "expert": 4, + "weight": 0.08762071281671524 + }, + { + "expert": 30, + "weight": 0.0809745118021965 + }, + { + "expert": 25, + "weight": 0.0679486021399498 + }, + { + "expert": 17, + "weight": 0.06139545142650604 + }, + { + "expert": 31, + "weight": 0.05613873153924942 + } + ], + "selected_weight_sum": 0.4879729002714157, + "top1_top2_margin": 0.046274177730083466 + }, + { + "position": 5, + "token_id": 2131, + "token": "Ġkey", + "decoded_piece": " key", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.12676671147346497 + }, + { + "expert": 25, + "weight": 0.11693397909402847 + }, + { + "expert": 17, + "weight": 0.10827948153018951 + }, + { + "expert": 38, + "weight": 0.07470021396875381 + }, + { + "expert": 14, + "weight": 0.05182885378599167 + }, + { + "expert": 30, + "weight": 0.04812060296535492 + } + ], + "selected_weight_sum": 0.5266298428177834, + "top1_top2_margin": 0.009832732379436493 + }, + { + "position": 6, + "token_id": 12, + "token": "-", + "decoded_piece": "-", + "experts_by_weight": [ + { + "expert": 57, + "weight": 0.06554465740919113 + }, + { + "expert": 38, + "weight": 0.05510614812374115 + }, + { + "expert": 55, + "weight": 0.04951103776693344 + }, + { + "expert": 25, + "weight": 0.04858021065592766 + }, + { + "expert": 39, + "weight": 0.0470929853618145 + }, + { + "expert": 4, + "weight": 0.04168501868844032 + } + ], + "selected_weight_sum": 0.3075200580060482, + "top1_top2_margin": 0.010438509285449982 + }, + { + "position": 7, + "token_id": 2834, + "token": "value", + "decoded_piece": "value", + "experts_by_weight": [ + { + "expert": 59, + "weight": 0.1504407376050949 + }, + { + "expert": 4, + "weight": 0.10922424495220184 + }, + { + "expert": 17, + "weight": 0.08334412425756454 + }, + { + "expert": 25, + "weight": 0.06348013132810593 + }, + { + "expert": 57, + "weight": 0.05469481274485588 + }, + { + "expert": 45, + "weight": 0.05172499269247055 + } + ], + "selected_weight_sum": 0.5129090435802937, + "top1_top2_margin": 0.041216492652893066 + }, + { + "position": 8, + "token_id": 28937, + "token": "Ġlatent", + "decoded_piece": " latent", + "experts_by_weight": [ + { + "expert": 57, + "weight": 0.15120594203472137 + }, + { + "expert": 17, + "weight": 0.09751661121845245 + }, + { + "expert": 25, + "weight": 0.07147576659917831 + }, + { + "expert": 2, + "weight": 0.05824965238571167 + }, + { + "expert": 30, + "weight": 0.05139882490038872 + }, + { + "expert": 31, + "weight": 0.04974324256181717 + } + ], + "selected_weight_sum": 0.4795900397002697, + "top1_top2_margin": 0.05368933081626892 + }, + { + "position": 9, + "token_id": 4183, + "token": "Ġchanges", + "decoded_piece": " changes", + "experts_by_weight": [ + { + "expert": 25, + "weight": 0.11196812242269516 + }, + { + "expert": 7, + "weight": 0.09889062494039536 + }, + { + "expert": 2, + "weight": 0.06677774339914322 + }, + { + "expert": 43, + "weight": 0.044559650123119354 + }, + { + "expert": 0, + "weight": 0.04172351956367493 + }, + { + "expert": 53, + "weight": 0.037076983600854874 + } + ], + "selected_weight_sum": 0.4009966440498829, + "top1_top2_margin": 0.013077497482299805 + }, + { + "position": 10, + "token_id": 254, + "token": "Ġthe", + "decoded_piece": " the", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.1058885008096695 + }, + { + "expert": 0, + "weight": 0.07895072549581528 + }, + { + "expert": 43, + "weight": 0.06406665593385696 + }, + { + "expert": 36, + "weight": 0.05847375839948654 + }, + { + "expert": 25, + "weight": 0.054556623101234436 + }, + { + "expert": 30, + "weight": 0.034156013280153275 + } + ], + "selected_weight_sum": 0.396092277020216, + "top1_top2_margin": 0.026937775313854218 + }, + { + "position": 11, + "token_id": 4887, + "token": "Ġmemory", + "decoded_piece": " memory", + "experts_by_weight": [ + { + "expert": 30, + "weight": 0.09737883508205414 + }, + { + "expert": 2, + "weight": 0.08197834342718124 + }, + { + "expert": 4, + "weight": 0.05927024781703949 + }, + { + "expert": 25, + "weight": 0.04981620982289314 + }, + { + "expert": 57, + "weight": 0.04698258265852928 + }, + { + "expert": 38, + "weight": 0.04634204879403114 + } + ], + "selected_weight_sum": 0.38176826760172844, + "top1_top2_margin": 0.015400491654872894 + }, + { + "position": 12, + "token_id": 2295, + "token": "Ġcost", + "decoded_piece": " cost", + "experts_by_weight": [ + { + "expert": 25, + "weight": 0.15634430944919586 + }, + { + "expert": 30, + "weight": 0.12142481654882431 + }, + { + "expert": 2, + "weight": 0.11154589802026749 + }, + { + "expert": 31, + "weight": 0.09600093215703964 + }, + { + "expert": 53, + "weight": 0.028693636879324913 + }, + { + "expert": 17, + "weight": 0.027443602681159973 + } + ], + "selected_weight_sum": 0.5414531957358122, + "top1_top2_margin": 0.03491949290037155 + }, + { + "position": 13, + "token_id": 280, + "token": "Ġof", + "decoded_piece": " of", + "experts_by_weight": [ + { + "expert": 25, + "weight": 0.10794661939144135 + }, + { + "expert": 36, + "weight": 0.0671105906367302 + }, + { + "expert": 49, + "weight": 0.05974861606955528 + }, + { + "expert": 30, + "weight": 0.05525089055299759 + }, + { + "expert": 7, + "weight": 0.04587714374065399 + }, + { + "expert": 0, + "weight": 0.03875768184661865 + } + ], + "selected_weight_sum": 0.37469154223799706, + "top1_top2_margin": 0.04083602875471115 + }, + { + "position": 14, + "token_id": 77939, + "token": "Ġautore", + "decoded_piece": " autore", + "experts_by_weight": [ + { + "expert": 57, + "weight": 0.06619898229837418 + }, + { + "expert": 17, + "weight": 0.05084220692515373 + }, + { + "expert": 14, + "weight": 0.03945222496986389 + }, + { + "expert": 41, + "weight": 0.038507360965013504 + }, + { + "expert": 30, + "weight": 0.03479800373315811 + }, + { + "expert": 37, + "weight": 0.03332655876874924 + } + ], + "selected_weight_sum": 0.26312533766031265, + "top1_top2_margin": 0.015356775373220444 + }, + { + "position": 15, + "token_id": 79184, + "token": "gressive", + "decoded_piece": "gressive", + "experts_by_weight": [ + { + "expert": 17, + "weight": 0.21340444684028625 + }, + { + "expert": 57, + "weight": 0.14111490547657013 + }, + { + "expert": 25, + "weight": 0.07872821390628815 + }, + { + "expert": 30, + "weight": 0.04866563901305199 + }, + { + "expert": 32, + "weight": 0.04710451513528824 + }, + { + "expert": 2, + "weight": 0.02478797361254692 + } + ], + "selected_weight_sum": 0.5538056939840317, + "top1_top2_margin": 0.07228954136371613 + }, + { + "position": 16, + "token_id": 39671, + "token": "Ġdecoding", + "decoded_piece": " decoding", + "experts_by_weight": [ + { + "expert": 30, + "weight": 0.19381695985794067 + }, + { + "expert": 17, + "weight": 0.16847439110279083 + }, + { + "expert": 4, + "weight": 0.10773279517889023 + }, + { + "expert": 29, + "weight": 0.06173955276608467 + }, + { + "expert": 44, + "weight": 0.038706641644239426 + }, + { + "expert": 25, + "weight": 0.027917681261897087 + } + ], + "selected_weight_sum": 0.5983880218118429, + "top1_top2_margin": 0.02534256875514984 + }, + { + "position": 17, + "token_id": 13, + "token": ".", + "decoded_piece": ".", + "experts_by_weight": [ + { + "expert": 16, + "weight": 0.12360352277755737 + }, + { + "expert": 12, + "weight": 0.1027035191655159 + }, + { + "expert": 56, + "weight": 0.09440722316503525 + }, + { + "expert": 18, + "weight": 0.09196682274341583 + }, + { + "expert": 37, + "weight": 0.08447214215993881 + }, + { + "expert": 50, + "weight": 0.08360416442155838 + } + ], + "selected_weight_sum": 0.5807573944330215, + "top1_top2_margin": 0.020900003612041473 + } + ] + }, + { + "id": "code", + "label": "Python code", + "tokens": 19, + "load": [ + 0, + 0, + 3, + 1, + 12, + 3, + 0, + 5, + 4, + 5, + 0, + 1, + 2, + 1, + 0, + 4, + 2, + 2, + 1, + 4, + 0, + 1, + 1, + 2, + 1, + 5, + 1, + 3, + 3, + 4, + 0, + 2, + 2, + 0, + 1, + 1, + 5, + 2, + 0, + 0, + 4, + 1, + 3, + 1, + 4, + 0, + 1, + 1, + 1, + 2, + 2, + 0, + 2, + 0, + 0, + 0, + 3, + 1, + 2, + 0, + 1, + 0, + 0, + 1 + ], + "metrics": { + "routes": 114, + "used_experts": 46, + "zero_experts": 18, + "mean": 1.78125, + "std": 2.0037167072296143, + "cv": 1.1248935461044312, + "gini": 0.5419408082962036, + "entropy_nats": 3.5931899547576904, + "effective_experts": 36.349845541114355, + "min": 0.0, + "p50": 1.0, + "p90": 4.0, + "max": 12.0 + }, + "top_experts": [ + { + "expert": 4, + "routes": 12 + }, + { + "expert": 7, + "routes": 5 + }, + { + "expert": 9, + "routes": 5 + }, + { + "expert": 25, + "routes": 5 + }, + { + "expert": 36, + "routes": 5 + }, + { + "expert": 8, + "routes": 4 + }, + { + "expert": 15, + "routes": 4 + }, + { + "expert": 19, + "routes": 4 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 22, + "weight": 0.5437576770782471 + }, + { + "expert": 21, + "weight": 0.02887026220560074 + }, + { + "expert": 13, + "weight": 0.022947292774915695 + }, + { + "expert": 42, + "weight": 0.020739872008562088 + }, + { + "expert": 47, + "weight": 0.019217059016227722 + }, + { + "expert": 32, + "weight": 0.01740080676972866 + } + ], + "selected_weight_sum": 0.652932969853282, + "top1_top2_margin": 0.5148874148726463 + }, + { + "position": 1, + "token_id": 9081, + "token": "Write", + "decoded_piece": "Write", + "experts_by_weight": [ + { + "expert": 46, + "weight": 0.13753922283649445 + }, + { + "expert": 4, + "weight": 0.09548301994800568 + }, + { + "expert": 37, + "weight": 0.061647929251194 + }, + { + "expert": 9, + "weight": 0.055846862494945526 + }, + { + "expert": 26, + "weight": 0.05239253491163254 + }, + { + "expert": 19, + "weight": 0.047255437821149826 + } + ], + "selected_weight_sum": 0.450165007263422, + "top1_top2_margin": 0.04205620288848877 + }, + { + "position": 2, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 8, + "weight": 0.09976524114608765 + }, + { + "expert": 4, + "weight": 0.09795945137739182 + }, + { + "expert": 19, + "weight": 0.08393261581659317 + }, + { + "expert": 44, + "weight": 0.07459735125303268 + }, + { + "expert": 7, + "weight": 0.05948468670248985 + }, + { + "expert": 23, + "weight": 0.03450009599328041 + } + ], + "selected_weight_sum": 0.4502394422888756, + "top1_top2_margin": 0.0018057897686958313 + }, + { + "position": 3, + "token_id": 12974, + "token": "ĠPython", + "decoded_piece": " Python", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.1781829595565796 + }, + { + "expert": 29, + "weight": 0.11575023084878922 + }, + { + "expert": 27, + "weight": 0.10451281070709229 + }, + { + "expert": 41, + "weight": 0.07266294956207275 + }, + { + "expert": 32, + "weight": 0.051248300820589066 + }, + { + "expert": 48, + "weight": 0.043809834867715836 + } + ], + "selected_weight_sum": 0.5661670863628387, + "top1_top2_margin": 0.062432728707790375 + }, + { + "position": 4, + "token_id": 1157, + "token": "Ġfunction", + "decoded_piece": " function", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.19366487860679626 + }, + { + "expert": 29, + "weight": 0.11906357854604721 + }, + { + "expert": 19, + "weight": 0.03140224888920784 + }, + { + "expert": 28, + "weight": 0.03139836713671684 + }, + { + "expert": 2, + "weight": 0.031194070354104042 + }, + { + "expert": 15, + "weight": 0.02878504991531372 + } + ], + "selected_weight_sum": 0.4355081934481859, + "top1_top2_margin": 0.07460130006074905 + }, + { + "position": 5, + "token_id": 344, + "token": "Ġthat", + "decoded_piece": " that", + "experts_by_weight": [ + { + "expert": 40, + "weight": 0.09220249950885773 + }, + { + "expert": 19, + "weight": 0.06857212632894516 + }, + { + "expert": 52, + "weight": 0.06615664809942245 + }, + { + "expert": 56, + "weight": 0.04282122477889061 + }, + { + "expert": 35, + "weight": 0.04161207005381584 + }, + { + "expert": 4, + "weight": 0.031904879957437515 + } + ], + "selected_weight_sum": 0.3432694487273693, + "top1_top2_margin": 0.023630373179912567 + }, + { + "position": 6, + "token_id": 7578, + "token": "Ġreturns", + "decoded_piece": " returns", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.15439683198928833 + }, + { + "expert": 7, + "weight": 0.11837179958820343 + }, + { + "expert": 9, + "weight": 0.09164318442344666 + }, + { + "expert": 40, + "weight": 0.03915099427103996 + }, + { + "expert": 52, + "weight": 0.03787941113114357 + }, + { + "expert": 44, + "weight": 0.03748919814825058 + } + ], + "selected_weight_sum": 0.47893141955137253, + "top1_top2_margin": 0.0360250324010849 + }, + { + "position": 7, + "token_id": 254, + "token": "Ġthe", + "decoded_piece": " the", + "experts_by_weight": [ + { + "expert": 36, + "weight": 0.1356981247663498 + }, + { + "expert": 7, + "weight": 0.09688761830329895 + }, + { + "expert": 4, + "weight": 0.05860389024019241 + }, + { + "expert": 44, + "weight": 0.04973827674984932 + }, + { + "expert": 58, + "weight": 0.04758499935269356 + }, + { + "expert": 23, + "weight": 0.035689134150743484 + } + ], + "selected_weight_sum": 0.4242020435631275, + "top1_top2_margin": 0.03881050646305084 + }, + { + "position": 8, + "token_id": 1022, + "token": "Ġfirst", + "decoded_piece": " first", + "experts_by_weight": [ + { + "expert": 36, + "weight": 0.18592320382595062 + }, + { + "expert": 8, + "weight": 0.17726802825927734 + }, + { + "expert": 4, + "weight": 0.06434556841850281 + }, + { + "expert": 9, + "weight": 0.058037169277668 + }, + { + "expert": 25, + "weight": 0.04549534246325493 + }, + { + "expert": 15, + "weight": 0.03907840698957443 + } + ], + "selected_weight_sum": 0.5701477192342281, + "top1_top2_margin": 0.008655175566673279 + }, + { + "position": 9, + "token_id": 9842, + "token": "Ġrepeated", + "decoded_piece": " repeated", + "experts_by_weight": [ + { + "expert": 25, + "weight": 0.15881554782390594 + }, + { + "expert": 9, + "weight": 0.1423594057559967 + }, + { + "expert": 31, + "weight": 0.13588689267635345 + }, + { + "expert": 8, + "weight": 0.08650468289852142 + }, + { + "expert": 57, + "weight": 0.03664877638220787 + }, + { + "expert": 36, + "weight": 0.0329442135989666 + } + ], + "selected_weight_sum": 0.593159519135952, + "top1_top2_margin": 0.01645614206790924 + }, + { + "position": 10, + "token_id": 3699, + "token": "Ġelement", + "decoded_piece": " element", + "experts_by_weight": [ + { + "expert": 25, + "weight": 0.15322136878967285 + }, + { + "expert": 17, + "weight": 0.09143341332674026 + }, + { + "expert": 28, + "weight": 0.08360946178436279 + }, + { + "expert": 27, + "weight": 0.05222221836447716 + }, + { + "expert": 4, + "weight": 0.04394563287496567 + }, + { + "expert": 15, + "weight": 0.04124316945672035 + } + ], + "selected_weight_sum": 0.4656752645969391, + "top1_top2_margin": 0.06178795546293259 + }, + { + "position": 11, + "token_id": 279, + "token": "Ġin", + "decoded_piece": " in", + "experts_by_weight": [ + { + "expert": 49, + "weight": 0.16705435514450073 + }, + { + "expert": 28, + "weight": 0.06526098400354385 + }, + { + "expert": 25, + "weight": 0.0634346604347229 + }, + { + "expert": 40, + "weight": 0.03620058670639992 + }, + { + "expert": 42, + "weight": 0.035932786762714386 + }, + { + "expert": 27, + "weight": 0.029904931783676147 + } + ], + "selected_weight_sum": 0.39778830483555794, + "top1_top2_margin": 0.10179337114095688 + }, + { + "position": 12, + "token_id": 245, + "token": "Ġa", + "decoded_piece": " a", + "experts_by_weight": [ + { + "expert": 36, + "weight": 0.14008726179599762 + }, + { + "expert": 49, + "weight": 0.12481290847063065 + }, + { + "expert": 8, + "weight": 0.07218452543020248 + }, + { + "expert": 58, + "weight": 0.05240301042795181 + }, + { + "expert": 4, + "weight": 0.040102530270814896 + }, + { + "expert": 42, + "weight": 0.03472679853439331 + } + ], + "selected_weight_sum": 0.46431703492999077, + "top1_top2_margin": 0.015274353325366974 + }, + { + "position": 13, + "token_id": 1525, + "token": "Ġlist", + "decoded_piece": " list", + "experts_by_weight": [ + { + "expert": 4, + "weight": 0.10503388941287994 + }, + { + "expert": 63, + "weight": 0.09194021672010422 + }, + { + "expert": 29, + "weight": 0.07014487683773041 + }, + { + "expert": 3, + "weight": 0.048768382519483566 + }, + { + "expert": 17, + "weight": 0.04372379556298256 + }, + { + "expert": 15, + "weight": 0.03325546905398369 + } + ], + "selected_weight_sum": 0.3928666301071644, + "top1_top2_margin": 0.013093672692775726 + }, + { + "position": 14, + "token_id": 285, + "token": "Ġand", + "decoded_piece": " and", + "experts_by_weight": [ + { + "expert": 34, + "weight": 0.11566011607646942 + }, + { + "expert": 12, + "weight": 0.08769606053829193 + }, + { + "expert": 16, + "weight": 0.06660442799329758 + }, + { + "expert": 18, + "weight": 0.05762474611401558 + }, + { + "expert": 50, + "weight": 0.04991111159324646 + }, + { + "expert": 56, + "weight": 0.03884124383330345 + } + ], + "selected_weight_sum": 0.4163377061486244, + "top1_top2_margin": 0.02796405553817749 + }, + { + "position": 15, + "token_id": 6716, + "token": "Ġexplain", + "decoded_piece": " explain", + "experts_by_weight": [ + { + "expert": 7, + "weight": 0.13019512593746185 + }, + { + "expert": 5, + "weight": 0.07663741707801819 + }, + { + "expert": 4, + "weight": 0.05870531126856804 + }, + { + "expert": 44, + "weight": 0.05767242982983589 + }, + { + "expert": 9, + "weight": 0.054065629839897156 + }, + { + "expert": 2, + "weight": 0.046069856733083725 + } + ], + "selected_weight_sum": 0.42334577068686485, + "top1_top2_margin": 0.053557708859443665 + }, + { + "position": 16, + "token_id": 895, + "token": "Ġits", + "decoded_piece": " its", + "experts_by_weight": [ + { + "expert": 36, + "weight": 0.1395944356918335 + }, + { + "expert": 43, + "weight": 0.10110253840684891 + }, + { + "expert": 40, + "weight": 0.09792105853557587 + }, + { + "expert": 5, + "weight": 0.09229704737663269 + }, + { + "expert": 7, + "weight": 0.03483238443732262 + }, + { + "expert": 24, + "weight": 0.020935704931616783 + } + ], + "selected_weight_sum": 0.48668316937983036, + "top1_top2_margin": 0.03849189728498459 + }, + { + "position": 17, + "token_id": 13913, + "token": "Ġcomplexity", + "decoded_piece": " complexity", + "experts_by_weight": [ + { + "expert": 2, + "weight": 0.1412113904953003 + }, + { + "expert": 25, + "weight": 0.1274697482585907 + }, + { + "expert": 31, + "weight": 0.08632361143827438 + }, + { + "expert": 29, + "weight": 0.043605998158454895 + }, + { + "expert": 60, + "weight": 0.03586355224251747 + }, + { + "expert": 5, + "weight": 0.03131873905658722 + } + ], + "selected_weight_sum": 0.46579303964972496, + "top1_top2_margin": 0.013741642236709595 + }, + { + "position": 18, + "token_id": 13, + "token": ".", + "decoded_piece": ".", + "experts_by_weight": [ + { + "expert": 50, + "weight": 0.13818378746509552 + }, + { + "expert": 16, + "weight": 0.12461185455322266 + }, + { + "expert": 37, + "weight": 0.11088532954454422 + }, + { + "expert": 12, + "weight": 0.10322917252779007 + }, + { + "expert": 56, + "weight": 0.08050823956727982 + }, + { + "expert": 11, + "weight": 0.059470295906066895 + } + ], + "selected_weight_sum": 0.6168886795639992, + "top1_top2_margin": 0.013571932911872864 + } + ] + }, + { + "id": "math", + "label": "数学推理", + "tokens": 27, + "load": [ + 6, + 2, + 0, + 0, + 5, + 1, + 6, + 2, + 0, + 0, + 0, + 2, + 2, + 2, + 0, + 1, + 4, + 3, + 1, + 2, + 4, + 3, + 2, + 1, + 0, + 10, + 4, + 2, + 6, + 0, + 0, + 3, + 2, + 1, + 4, + 0, + 6, + 8, + 0, + 3, + 4, + 5, + 3, + 3, + 1, + 3, + 4, + 2, + 0, + 2, + 3, + 2, + 2, + 0, + 1, + 5, + 6, + 1, + 4, + 7, + 0, + 0, + 0, + 6 + ], + "metrics": { + "routes": 162, + "used_experts": 48, + "zero_experts": 16, + "mean": 2.53125, + "std": 2.3160500526428223, + "cv": 0.9149827361106873, + "gini": 0.4951774775981903, + "entropy_nats": 3.697890281677246, + "effective_experts": 40.362061891369216, + "min": 0.0, + "p50": 2.0, + "p90": 6.0, + "max": 10.0 + }, + "top_experts": [ + { + "expert": 25, + "routes": 10 + }, + { + "expert": 37, + "routes": 8 + }, + { + "expert": 59, + "routes": 7 + }, + { + "expert": 0, + "routes": 6 + }, + { + "expert": 6, + "routes": 6 + }, + { + "expert": 28, + "routes": 6 + }, + { + "expert": 36, + "routes": 6 + }, + { + "expert": 56, + "routes": 6 + } + ], + "token_routes": [ + { + "position": 0, + "token_id": 100000, + "token": "<|begin▁of▁sentence|>", + "decoded_piece": "<|begin▁of▁sentence|>", + "experts_by_weight": [ + { + "expert": 22, + "weight": 0.5437576770782471 + }, + { + "expert": 21, + "weight": 0.02887026220560074 + }, + { + "expert": 13, + "weight": 0.022947292774915695 + }, + { + "expert": 42, + "weight": 0.020739872008562088 + }, + { + "expert": 47, + "weight": 0.019217059016227722 + }, + { + "expert": 32, + "weight": 0.01740080676972866 + } + ], + "selected_weight_sum": 0.652932969853282, + "top1_top2_margin": 0.5148874148726463 + }, + { + "position": 1, + "token_id": 8246, + "token": "èĭ¥", + "decoded_piece": "若", + "experts_by_weight": [ + { + "expert": 46, + "weight": 0.14609992504119873 + }, + { + "expert": 37, + "weight": 0.1228821724653244 + }, + { + "expert": 19, + "weight": 0.10514180362224579 + }, + { + "expert": 56, + "weight": 0.07545209676027298 + }, + { + "expert": 55, + "weight": 0.054526668041944504 + }, + { + "expert": 39, + "weight": 0.04116235300898552 + } + ], + "selected_weight_sum": 0.5452650189399719, + "top1_top2_margin": 0.02321775257587433 + }, + { + "position": 2, + "token_id": 2206, + "token": "æŃ£", + "decoded_piece": "正", + "experts_by_weight": [ + { + "expert": 26, + "weight": 0.0876239538192749 + }, + { + "expert": 45, + "weight": 0.07643693685531616 + }, + { + "expert": 0, + "weight": 0.057786520570516586 + }, + { + "expert": 52, + "weight": 0.040895964950323105 + }, + { + "expert": 41, + "weight": 0.040521040558815 + }, + { + "expert": 13, + "weight": 0.03683890402317047 + } + ], + "selected_weight_sum": 0.34010332077741623, + "top1_top2_margin": 0.01118701696395874 + }, + { + "position": 3, + "token_id": 2266, + "token": "æķ°", + "decoded_piece": "数", + "experts_by_weight": [ + { + "expert": 25, + "weight": 0.12693962454795837 + }, + { + "expert": 31, + "weight": 0.11401762068271637 + }, + { + "expert": 63, + "weight": 0.05557835474610329 + }, + { + "expert": 28, + "weight": 0.04544328525662422 + }, + { + "expert": 57, + "weight": 0.03626488149166107 + }, + { + "expert": 26, + "weight": 0.0351380929350853 + } + ], + "selected_weight_sum": 0.4133818596601486, + "top1_top2_margin": 0.012922003865242004 + }, + { + "position": 4, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 41, + "weight": 0.15854693949222565 + }, + { + "expert": 6, + "weight": 0.0794890820980072 + }, + { + "expert": 4, + "weight": 0.06637027114629745 + }, + { + "expert": 55, + "weight": 0.03351316973567009 + }, + { + "expert": 37, + "weight": 0.03202974796295166 + }, + { + "expert": 27, + "weight": 0.02984650433063507 + } + ], + "selected_weight_sum": 0.3997957147657871, + "top1_top2_margin": 0.07905785739421844 + }, + { + "position": 5, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 39, + "weight": 0.048063986003398895 + }, + { + "expert": 42, + "weight": 0.03813319653272629 + }, + { + "expert": 1, + "weight": 0.03565631061792374 + }, + { + "expert": 23, + "weight": 0.033526234328746796 + }, + { + "expert": 22, + "weight": 0.02928059734404087 + }, + { + "expert": 45, + "weight": 0.02924760803580284 + } + ], + "selected_weight_sum": 0.21390793286263943, + "top1_top2_margin": 0.009930789470672607 + }, + { + "position": 6, + "token_id": 13211, + "token": "满足", + "decoded_piece": "满足", + "experts_by_weight": [ + { + "expert": 58, + "weight": 0.12237146496772766 + }, + { + "expert": 7, + "weight": 0.07029765099287033 + }, + { + "expert": 43, + "weight": 0.059514883905649185 + }, + { + "expert": 40, + "weight": 0.05232641473412514 + }, + { + "expert": 0, + "weight": 0.05189927667379379 + }, + { + "expert": 5, + "weight": 0.05173082277178764 + } + ], + "selected_weight_sum": 0.40814051404595375, + "top1_top2_margin": 0.05207381397485733 + }, + { + "position": 7, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 41, + "weight": 0.14802685379981995 + }, + { + "expert": 6, + "weight": 0.11230576783418655 + }, + { + "expert": 4, + "weight": 0.045813966542482376 + }, + { + "expert": 37, + "weight": 0.04318483546376228 + }, + { + "expert": 20, + "weight": 0.037408940494060516 + }, + { + "expert": 59, + "weight": 0.03519348427653313 + } + ], + "selected_weight_sum": 0.4219338484108448, + "top1_top2_margin": 0.03572108596563339 + }, + { + "position": 8, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 56, + "weight": 0.15301837027072906 + }, + { + "expert": 46, + "weight": 0.10797952860593796 + }, + { + "expert": 37, + "weight": 0.10473429411649704 + }, + { + "expert": 16, + "weight": 0.06219639629125595 + }, + { + "expert": 6, + "weight": 0.03330612555146217 + }, + { + "expert": 50, + "weight": 0.029627667739987373 + } + ], + "selected_weight_sum": 0.49086238257586956, + "top1_top2_margin": 0.04503884166479111 + }, + { + "position": 9, + "token_id": 14855, + "token": "åĬłä¸Ĭ", + "decoded_piece": "加上", + "experts_by_weight": [ + { + "expert": 34, + "weight": 0.14497965574264526 + }, + { + "expert": 55, + "weight": 0.14415185153484344 + }, + { + "expert": 28, + "weight": 0.03581972047686577 + }, + { + "expert": 45, + "weight": 0.032462701201438904 + }, + { + "expert": 0, + "weight": 0.03084731660783291 + }, + { + "expert": 4, + "weight": 0.02867933362722397 + } + ], + "selected_weight_sum": 0.41694057919085026, + "top1_top2_margin": 0.0008278042078018188 + }, + { + "position": 10, + "token_id": 12985, + "token": "å®ĥçļĦ", + "decoded_piece": "它的", + "experts_by_weight": [ + { + "expert": 36, + "weight": 0.25599992275238037 + }, + { + "expert": 40, + "weight": 0.10082335025072098 + }, + { + "expert": 58, + "weight": 0.06489169597625732 + }, + { + "expert": 34, + "weight": 0.04302974417805672 + }, + { + "expert": 15, + "weight": 0.030839329585433006 + }, + { + "expert": 49, + "weight": 0.02814270369708538 + } + ], + "selected_weight_sum": 0.5237267464399338, + "top1_top2_margin": 0.1551765725016594 + }, + { + "position": 11, + "token_id": 8270, + "token": "åĢĴ", + "decoded_piece": "倒", + "experts_by_weight": [ + { + "expert": 25, + "weight": 0.09931953251361847 + }, + { + "expert": 36, + "weight": 0.05962561443448067 + }, + { + "expert": 26, + "weight": 0.05795827880501747 + }, + { + "expert": 17, + "weight": 0.04644583910703659 + }, + { + "expert": 31, + "weight": 0.03724290803074837 + }, + { + "expert": 40, + "weight": 0.03425101935863495 + } + ], + "selected_weight_sum": 0.3348431922495365, + "top1_top2_margin": 0.0396939180791378 + }, + { + "position": 12, + "token_id": 2266, + "token": "æķ°", + "decoded_piece": "数", + "experts_by_weight": [ + { + "expert": 63, + "weight": 0.15665771067142487 + }, + { + "expert": 25, + "weight": 0.11005081236362457 + }, + { + "expert": 28, + "weight": 0.09605225920677185 + }, + { + "expert": 17, + "weight": 0.06298534572124481 + }, + { + "expert": 59, + "weight": 0.0517939031124115 + }, + { + "expert": 4, + "weight": 0.04924814775586128 + } + ], + "selected_weight_sum": 0.5267881788313389, + "top1_top2_margin": 0.04660689830780029 + }, + { + "position": 13, + "token_id": 37360, + "token": "çŃīäºİ", + "decoded_piece": "等于", + "experts_by_weight": [ + { + "expert": 55, + "weight": 0.15661580860614777 + }, + { + "expert": 34, + "weight": 0.09555711597204208 + }, + { + "expert": 25, + "weight": 0.0472816526889801 + }, + { + "expert": 58, + "weight": 0.0455225333571434 + }, + { + "expert": 0, + "weight": 0.03498144820332527 + }, + { + "expert": 43, + "weight": 0.02906597964465618 + } + ], + "selected_weight_sum": 0.4090245384722948, + "top1_top2_margin": 0.06105869263410568 + }, + { + "position": 14, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 56, + "weight": 0.1745140701532364 + }, + { + "expert": 46, + "weight": 0.1652524471282959 + }, + { + "expert": 37, + "weight": 0.16063380241394043 + }, + { + "expert": 16, + "weight": 0.09852830320596695 + }, + { + "expert": 6, + "weight": 0.07098746299743652 + }, + { + "expert": 50, + "weight": 0.05819505825638771 + } + ], + "selected_weight_sum": 0.7281111441552639, + "top1_top2_margin": 0.00926162302494049 + }, + { + "position": 15, + "token_id": 18, + "token": "3", + "decoded_piece": "3", + "experts_by_weight": [ + { + "expert": 59, + "weight": 0.10042109340429306 + }, + { + "expert": 51, + "weight": 0.08639753609895706 + }, + { + "expert": 6, + "weight": 0.05279490351676941 + }, + { + "expert": 63, + "weight": 0.040343768894672394 + }, + { + "expert": 42, + "weight": 0.0377325676381588 + }, + { + "expert": 25, + "weight": 0.0330699160695076 + } + ], + "selected_weight_sum": 0.3507597856223583, + "top1_top2_margin": 0.014023557305335999 + }, + { + "position": 16, + "token_id": 19304, + "token": "ï¼Į", + "decoded_piece": ",", + "experts_by_weight": [ + { + "expert": 16, + "weight": 0.10927784442901611 + }, + { + "expert": 12, + "weight": 0.10593557357788086 + }, + { + "expert": 18, + "weight": 0.08180209249258041 + }, + { + "expert": 11, + "weight": 0.07350264489650726 + }, + { + "expert": 56, + "weight": 0.06115064397454262 + }, + { + "expert": 54, + "weight": 0.05411747470498085 + } + ], + "selected_weight_sum": 0.4857862740755081, + "top1_top2_margin": 0.003342270851135254 + }, + { + "position": 17, + "token_id": 2989, + "token": "æ±Ĥ", + "decoded_piece": "求", + "experts_by_weight": [ + { + "expert": 0, + "weight": 0.0868334174156189 + }, + { + "expert": 7, + "weight": 0.07825968414545059 + }, + { + "expert": 19, + "weight": 0.04057038575410843 + }, + { + "expert": 44, + "weight": 0.037275321781635284 + }, + { + "expert": 52, + "weight": 0.03410649672150612 + }, + { + "expert": 36, + "weight": 0.032320231199264526 + } + ], + "selected_weight_sum": 0.30936553701758385, + "top1_top2_margin": 0.008573733270168304 + }, + { + "position": 18, + "token_id": 1376, + "token": "Ġx", + "decoded_piece": " x", + "experts_by_weight": [ + { + "expert": 41, + "weight": 0.1231520026922226 + }, + { + "expert": 6, + "weight": 0.06557122617959976 + }, + { + "expert": 20, + "weight": 0.0597735270857811 + }, + { + "expert": 4, + "weight": 0.05677611008286476 + }, + { + "expert": 37, + "weight": 0.03178378567099571 + }, + { + "expert": 59, + "weight": 0.031433671712875366 + } + ], + "selected_weight_sum": 0.3684903234243393, + "top1_top2_margin": 0.05758077651262283 + }, + { + "position": 19, + "token_id": 207, + "token": "Ġ", + "decoded_piece": " ", + "experts_by_weight": [ + { + "expert": 56, + "weight": 0.09447808563709259 + }, + { + "expert": 37, + "weight": 0.07532291114330292 + }, + { + "expert": 46, + "weight": 0.06776554137468338 + }, + { + "expert": 21, + "weight": 0.05218034237623215 + }, + { + "expert": 39, + "weight": 0.03654729574918747 + }, + { + "expert": 33, + "weight": 0.032638877630233765 + } + ], + "selected_weight_sum": 0.35893305391073227, + "top1_top2_margin": 0.019155174493789673 + }, + { + "position": 20, + "token_id": 337, + "token": "çļĦ", + "decoded_piece": "的", + "experts_by_weight": [ + { + "expert": 36, + "weight": 0.1181933805346489 + }, + { + "expert": 32, + "weight": 0.04208538308739662 + }, + { + "expert": 21, + "weight": 0.0391385555267334 + }, + { + "expert": 40, + "weight": 0.03826816380023956 + }, + { + "expert": 43, + "weight": 0.036307066679000854 + }, + { + "expert": 1, + "weight": 0.03165046498179436 + } + ], + "selected_weight_sum": 0.3056430146098137, + "top1_top2_margin": 0.07610799744725227 + }, + { + "position": 21, + "token_id": 18718, + "token": "å¹³æĸ¹", + "decoded_piece": "平方", + "experts_by_weight": [ + { + "expert": 63, + "weight": 0.09191884845495224 + }, + { + "expert": 25, + "weight": 0.0589422807097435 + }, + { + "expert": 28, + "weight": 0.043951794505119324 + }, + { + "expert": 59, + "weight": 0.04311515390872955 + }, + { + "expert": 31, + "weight": 0.04050914943218231 + }, + { + "expert": 51, + "weight": 0.032655391842126846 + } + ], + "selected_weight_sum": 0.3110926188528538, + "top1_top2_margin": 0.03297656774520874 + }, + { + "position": 22, + "token_id": 14855, + "token": "åĬłä¸Ĭ", + "decoded_piece": "加上", + "experts_by_weight": [ + { + "expert": 55, + "weight": 0.14588066935539246 + }, + { + "expert": 34, + "weight": 0.14225462079048157 + }, + { + "expert": 58, + "weight": 0.034157682210206985 + }, + { + "expert": 49, + "weight": 0.029166897758841515 + }, + { + "expert": 0, + "weight": 0.028123507276177406 + }, + { + "expert": 25, + "weight": 0.025089992210268974 + } + ], + "selected_weight_sum": 0.4046733696013689, + "top1_top2_margin": 0.0036260485649108887 + }, + { + "position": 23, + "token_id": 8270, + "token": "åĢĴ", + "decoded_piece": "倒", + "experts_by_weight": [ + { + "expert": 20, + "weight": 0.057385675609111786 + }, + { + "expert": 26, + "weight": 0.054804299026727676 + }, + { + "expert": 59, + "weight": 0.054245784878730774 + }, + { + "expert": 25, + "weight": 0.05022210627794266 + }, + { + "expert": 36, + "weight": 0.039826568216085434 + }, + { + "expert": 41, + "weight": 0.03494826331734657 + } + ], + "selected_weight_sum": 0.2914326973259449, + "top1_top2_margin": 0.0025813765823841095 + }, + { + "position": 24, + "token_id": 51971, + "token": "æķ°çļĦ", + "decoded_piece": "数的", + "experts_by_weight": [ + { + "expert": 28, + "weight": 0.0728067085146904 + }, + { + "expert": 63, + "weight": 0.07093488425016403 + }, + { + "expert": 25, + "weight": 0.0637853667140007 + }, + { + "expert": 17, + "weight": 0.05612759292125702 + }, + { + "expert": 36, + "weight": 0.047605425119400024 + }, + { + "expert": 20, + "weight": 0.0461735725402832 + } + ], + "selected_weight_sum": 0.3574335500597954, + "top1_top2_margin": 0.0018718242645263672 + }, + { + "position": 25, + "token_id": 18718, + "token": "å¹³æĸ¹", + "decoded_piece": "平方", + "experts_by_weight": [ + { + "expert": 59, + "weight": 0.13647541403770447 + }, + { + "expert": 63, + "weight": 0.10184752941131592 + }, + { + "expert": 25, + "weight": 0.059981655329465866 + }, + { + "expert": 28, + "weight": 0.0428638830780983 + }, + { + "expert": 27, + "weight": 0.0374276340007782 + }, + { + "expert": 47, + "weight": 0.031549129635095596 + } + ], + "selected_weight_sum": 0.41014524549245834, + "top1_top2_margin": 0.03462788462638855 + }, + { + "position": 26, + "token_id": 398, + "token": "ãĢĤ", + "decoded_piece": "。", + "experts_by_weight": [ + { + "expert": 16, + "weight": 0.11203551292419434 + }, + { + "expert": 11, + "weight": 0.1093420758843422 + }, + { + "expert": 12, + "weight": 0.08985795080661774 + }, + { + "expert": 37, + "weight": 0.08588015288114548 + }, + { + "expert": 56, + "weight": 0.07133956253528595 + }, + { + "expert": 50, + "weight": 0.06675778329372406 + } + ], + "selected_weight_sum": 0.5352130383253098, + "top1_top2_margin": 0.0026934370398521423 + } + ] + } + ], + "prompt_pair_jaccard": [ + { + "left": "zh_explanation", + "right": "en_architecture", + "used_expert_jaccard": 0.6545454545454545, + "shared_experts": 36, + "union_experts": 55 + }, + { + "left": "zh_explanation", + "right": "code", + "used_expert_jaccard": 0.7272727272727273, + "shared_experts": 40, + "union_experts": 55 + }, + { + "left": "zh_explanation", + "right": "math", + "used_expert_jaccard": 0.7321428571428571, + "shared_experts": 41, + "union_experts": 56 + }, + { + "left": "en_architecture", + "right": "code", + "used_expert_jaccard": 0.6, + "shared_experts": 33, + "union_experts": 55 + }, + { + "left": "en_architecture", + "right": "math", + "used_expert_jaccard": 0.6071428571428571, + "shared_experts": 34, + "union_experts": 56 + }, + { + "left": "code", + "right": "math", + "used_expert_jaccard": 0.6491228070175439, + "shared_experts": 37, + "union_experts": 57 + } + ] + } + } + ], + "final_hidden": { + "valid_tokens": 90, + "hidden_rms": { + "mean": 3.3090813159942627, + "p10": 0.2696530222892761, + "p50": 0.31665489077568054, + "p90": 0.3847275674343109 + }, + "l2_norm": { + "mean": 149.75193786621094, + "p10": 12.203103065490723, + "p50": 14.330163955688477, + "p90": 17.410781860351562 + }, + "finite": true, + "sha256_fp32": "1ffd41521b5a32af38695b767a5ecc556ddeae1e1efe5339743e5c8c7644030e" + } + }, + "cache_accounting": { + "dtype": "BF16", + "latent_elements_per_token_layer": 576, + "latent_bytes_per_token_layer": 1152, + "hf_eager_elements_per_token_layer": 5120, + "hf_eager_bytes_per_token_layer": 10240, + "eager_over_latent_ratio": 8.88888888888889, + "latent_reduction_vs_eager": 0.8875, + "components": { + "latent_content": 512, + "rope_key": 64, + "expanded_key": 3072, + "expanded_value": 2048 + } + } +} diff --git a/src/pages/deepseek/index.astro b/src/pages/deepseek/index.astro index 0b8b795..574ff8c 100644 --- a/src/pages/deepseek/index.astro +++ b/src/pages/deepseek/index.astro @@ -2,6 +2,7 @@ import BaseLayout from "@/layouts/BaseLayout.astro"; import DeepSeekLineage from "@/components/DeepSeekLineage.astro"; import DeepSeekLab from "@/components/DeepSeekLab.astro"; +import DeepSeekArtifactLab from "@/components/DeepSeekArtifactLab.astro"; import { deepseekBranches, deepseekLedgers, deepseekPaperChain, deepseekWaves } from "@/data/deepseek"; const toc = [ @@ -27,21 +28,22 @@ const toc = [ ["19", "v4-state", "V4:异构状态与稳定性"], ["20", "k3", "与 K3 的继承边界"], ["21", "lab", "四联交互实验"], - ["22", "branches", "别漏掉旁支"], - ["23", "audit", "事实、推导与教学模型"], + ["22", "artifact", "真实权重执行"], + ["23", "branches", "别漏掉旁支"], + ["24", "audit", "事实、推导与教学模型"], ["↳", "papers", "六十节点阅读链"], ]; ---
-

SPOTLIGHT / DEEPSEEK · ROUND 02 ALGORITHM × SYSTEM × EVIDENCE

+

SPOTLIGHT / DEEPSEEK · ROUND 03 ALGORITHM × SYSTEM × REAL WEIGHTS

不要背模型名
要看懂每次为什么转向

这不是七篇报告的摘要,而是一套可追问、可计算、可反驳的技术谱系: @@ -53,9 +55,9 @@ const toc = [

SPAN
2024.01 → 2026.06
LEDGERS
24 张问题账
LINEAGE
10 次技术转向
-
LABS
4 个可操作实验
+
LABS
8 个可操作实验
EVIDENCE
60 个一手 / 官方节点
-
STATUS
重点专题 · 二轮深读
+
STATUS
三轮 · 真实权重执行
@@ -762,8 +764,25 @@ const toc = [ +
+

22 OFFICIAL WEIGHTS / EXECUTED

+

从“MLA 与 MoE 的概念”再往前一步:让官方 V2-Lite 权重真的跑起来

+

+ 前面的四联实验负责建立公式与角色合同;下面的四联工件实验固定官方 revision、tokenizer、 + 模型代码和 checkpoint 第一分片,在 RTX 5090 上连续执行 layer 0–6。它把真实观测、shape 推导、 + 实现差距和未覆盖范围放在同一张证据图里。 +

+
+
X / FORWARD7 / 27 layers

1 个 dense 层 + 6 个 MoE 层;layer 7 因跨分片停止。

+
X / ROUTES3,240

90 个有效 token × 6 层 × top-6 routed experts。

+
X + D / CACHE576 ↔ 5,120

latent 合同与 HF eager 实际展开元素,两张账同时保留。

+
X / RERUN31 / 31 exact

hidden hashes、MLA shapes、loads 与全部 token routes。

+
+ +
+
-

22 THE MAIN LINE IS NOT THE WHOLE TREE

+

23 THE MAIN LINE IS NOT THE WHOLE TREE

如果只读 V2 → V3 → R1 → V4,会漏掉五条反过来影响主线的旁支

{deepseekBranches.map(([name, line, text, url]) => ( @@ -783,7 +802,7 @@ const toc = [
-

23 EVIDENCE AUDIT

+

24 EVIDENCE AUDIT

同一张页面里有三种知识,它们的语气必须不同

@@ -837,6 +856,43 @@ const toc = [ radial-gradient(circle at 65% 38%, rgba(84, 124, 116, 0.13), transparent 24%); } + .artifact-callout { + display: grid; + grid-template-columns: repeat(4, minmax(0, 1fr)); + max-width: 980px; + margin: 30px 0; + border-top: 1px solid var(--line); + border-left: 1px solid var(--line); + } + + .artifact-callout article { + min-height: 155px; + padding: 20px; + border-right: 1px solid var(--line); + border-bottom: 1px solid var(--line); + background: var(--paper-raised); + } + + .artifact-callout span { + display: block; + color: var(--copper); + font: 700 0.58rem/1 var(--mono); + letter-spacing: 0.08em; + } + + .artifact-callout b { + display: block; + margin-top: 18px; + font-size: 1.14rem; + } + + .artifact-callout p { + margin: 11px 0 0; + color: var(--muted); + font-size: 0.68rem; + line-height: 1.55; + } + .moe-compare, .four-layer { display: grid; @@ -1674,6 +1730,7 @@ const toc = [ .followup-grid, .state-machines, .v4-contract, + .artifact-callout, .branch-grid, .audit-grid, .paper-chain.expanded { diff --git a/src/pages/index.astro b/src/pages/index.astro index d782178..c93c809 100644 --- a/src/pages/index.astro +++ b/src/pages/index.astro @@ -145,17 +145,17 @@ const paths = [
-

NEW / DEEPSEEK ROUND 02 CAPACITY · STATE · SYSTEM · REASONING

+

NEW / DEEPSEEK ROUND 03 LINEAGE · REAL WEIGHTS · ROUTES · CACHE

从 Dense 到百万上下文:每次创新都在偿还上一代最贵的一张账

- 用二十四张问题账和十次技术转向,从 DeepSeek LLM、MoE、V2 的 MLA 权重吸收, - 走到 V3 的 FP8 / DualPipe / MTP、R1 与 DAPO / Dr.GRPO 反查、V3.2 Agent 环境和 V4 异构长状态。 + 用二十四张问题账和十次技术转向走完 Dense→V4,再固定官方 V2-Lite 权重执行 7/27 层: + 逐 token 检查 3,240 次专家选择,并把 latent 状态与 HF eager cache 的实现差距摆在同一张账上。

LINEAGE
1991 → 2026 · 10 次转向
NODES
60 个一手 / 官方节点
-
LAB
MoE · MLA · V3 协同 · RL 偏差
+
LAB
4 公式实验 · 4 真实工件实验
diff --git a/src/pages/progress/index.astro b/src/pages/progress/index.astro index 566b8af..c7f1eea 100644 --- a/src/pages/progress/index.astro +++ b/src/pages/progress/index.astro @@ -15,7 +15,7 @@ const workstreams = [ { label: "表示、位置与残差高速公路", value: 81, next: "加入真实 hidden-state / norm traces、长上下文位置外推复现与更多深层稳定性消融" }, { label: "Scaling Laws", value: 74, next: "加入真实拟合复现、置信区间与更多模型族对照" }, { label: "数据工程与预训练配方", value: 73, next: "逐图精读 FineWeb / DCLM,加入真实去重与 mixture traces" }, - { label: "DeepSeek 专题", value: 83, next: "加入真实专家负载、MLA kernel、RL 训练 traces 与独立复现" }, + { label: "DeepSeek 专题", value: 87, next: "真实 latent-cache kernel、更大样本负载、FP8/pipeline 与 R1-like RL 复现" }, { label: "指令微调与人类偏好", value: 75, next: "加入真实偏好分歧样本、RM 长度偏置与 PPO/DPO 小模型复现" }, { label: "推理与测试时扩展", value: 76, next: "真实模型采样曲线、PRM 案例与逐篇图表精读" }, { label: "工具使用与长程 Agent", value: 74, next: "补真实环境 traces、cross-harness 对照、Agent RL 训练曲线与安全案例" }, @@ -50,7 +50,7 @@ const workstreams = [
OVERALL
专题平均 {average}%
READABLE
{published} 个首版可读专题
ACTIVE
{researching} 个研究/写作中
-
UPDATED
2026-07-29 12:40 CST
+
UPDATED
2026-07-29 14:15 CST
MODE
持续迭代,不锁死版本
@@ -97,12 +97,12 @@ const workstreams = [
✓

K3 报告已结构化拆解

47 页报告目录、151 条参考来源和架构/后训练/系统主线已经提取。

✓

17 专题知识图

从语言模型基础到评测安全,包含先修依赖和三条贯穿案例。

✓

编辑式网站系统

响应式导航、章节模板、侧栏、进度、论文链和证据提示组件。

-
✓

七十一个原创交互视图

K3 三轴图、八联报告实验与四联开放工件实验,以及语言模型前史、Transformer、表示深度、DeepSeek、长上下文、MoE、推理、Agent、多模态、训练系统、推理服务、Scaling、数据工程、数值、Alignment 与评测安全专题。

+
✓

七十五个原创交互视图

K3 三轴图、八联报告实验与四联开放工件实验,DeepSeek 四联公式实验与四联真实权重实验,以及语言模型前史、Transformer、表示深度、长上下文、MoE、推理、Agent、多模态、训练系统、推理服务、Scaling、数据工程、数值、Alignment 与评测安全专题。

✓

十七篇首版长文

K3、语言模型前史、Transformer、表示/位置/残差、DeepSeek、Scaling、数据工程、长上下文、MoE、后训练、推理、Agent、原生多模态、训练系统、推理服务、数值优化与评测安全专题。

✓

语言模型前史深度专题

八张独立问题账、33 个正式节点、20 段长文与概率—向量—记忆—对齐四联实验。

✓

Transformer 深度专题

十张独立问题账、40 个正式节点、21 段正文与 QKV—Mask—多头位置—Block 成本四联实验。

✓

表示、位置与残差高速公路深度专题

二十张问题账、66 个一手节点、DeepSeek/Kimi 双谱系,以及 Token—位置—Norm—Residual/FFN 四联实验。

-
✓

DeepSeek 技术谱系二轮深读

二十四张问题账、十次技术转向、60 个一手/官方节点,以及稀疏容量—MLA 缓存—V3 协同—RL 偏差四联实验。

+
✓

DeepSeek 三轮真实权重里程碑

在二十四张问题账、十次转向与四联公式实验上,新增 V2-Lite 7/27 层连续 forward、3,240 次真实专家选择、MLA/eager cache 实现账与 31/31 exact 复跑四联实验。

✓

Kimi K3 技术报告二轮深读

三十二张问题账、Figure 1–16 / Table 1–5 审计、100 节点阅读链,以及 Delta—Decay—AttnRes—LatentMoE—SiTU—QB—MOPD—Cache 八联实验。

✓

Kimi K3 三轮开放工件里程碑

固定官方 revisions,审计 96 个 shards、497,220 个 tensor entries 与真实 KDA / MLA / MoE / MoonViT shapes;四联实验分开显示层型、tensor anatomy、参数范围和复现边界。

✓

FlashKDA RTX 5090 执行闸门

隔离 CUDA 13.0 / glibc 2.39 编译 sm_120a wheel;6/6 官方参考逐元素相等,并完成 fixed / varlen、三种 state mode 的 1,800 个 CUDA Event samples。

@@ -134,7 +134,7 @@ const workstreams = [
优先级专题本轮交付完成闸门
P0K3 三轮

开放权重 traces → FlashKDA / AttnRes / MoE 真实行为 → Figure 1–16 数值重绘与独立复现

运行证据 + 逐图复现
-
P0DeepSeek 三轮

真实 expert load / MLA kernel → FP8 / pipeline traces → R1-like RL 小模型复现

运行证据 + 独立复现
+
P0DeepSeek 三轮

真实 latent-cache kernel / 更大负载样本 → FP8 / pipeline traces → R1-like RL 小模型复现

运行证据 + 独立复现
P0Transformer 二轮

多头电路逐图 → Pre/Post-LN 真实 traces → Flash/KV 配置与 kernel 对照

逐图笔记 + 实测边界
P0表示、位置与残差二轮

真实 hidden-state / norm traces → 长上下文位置外推 → mHC / AttnRes 深层稳定性消融

可复现实验 + 逐图笔记
P0语言模型前史二轮

Kneser–Ney / LSTM / Bahdanau 逐图 → 真实小语料复现 → tokenizer 公平性

可复现实验 + 逐图笔记