Enhance model training lite workflow

skills/model-training-lite/scripts/render_sft_submission_plan.py

@@ -3,8 +3,10 @@ from __future__ import annotations
 
 import argparse
 import json
+import os
 import re
 import sys
+from pathlib import Path
 from pathlib import PurePosixPath
 from typing import Any
 
@@ -26,14 +28,18 @@ def main() -> int:
 
 
 def render_plan(payload: dict[str, Any]) -> dict[str, Any]:
+    defaults = load_local_defaults()
     missing: list[str] = []
-    jupyter_url = clean(payload.get("jupyter_url"))
+    jupyter_url = clean(payload.get("jupyter_url")) or clean(defaults.get("jupyter_url"))
+    jupyter_auth_type = clean(payload.get("jupyter_auth_type") or payload.get("auth_type")) or clean(defaults.get("auth_type"))
+    jupyter_password = clean(payload.get("jupyter_password") or payload.get("password")) or clean(defaults.get("password"))
+    jupyter_token = clean(payload.get("jupyter_token") or payload.get("token")) or clean(defaults.get("token"))
     data_commit = clean(payload.get("data_commit") or payload.get("commit"))
     data_path = clean(payload.get("data_path"))
     if not jupyter_url:
         missing.append("jupyter_url")
-    if not data_commit and not data_path:
-        missing.append("data_commit or data_path")
+    if jupyter_url and not (jupyter_password or jupyter_token or clean(payload.get("jupyter_cookie")) or clean(defaults.get("cookie"))):
+        missing.append("jupyter auth")
 
     owner = clean(payload.get("owner")) or "wuyang6"
     run_name = clean(payload.get("run_name")) or default_run_name(data_commit or data_path or "manual")
@@ -43,6 +49,12 @@ def render_plan(payload: dict[str, Any]) -> dict[str, Any]:
     data_repo = clean(payload.get("data_repo")) or DEFAULT_DATA_REPO
     data_branch = clean(payload.get("data_branch")) or DEFAULT_BRANCH
     recipe = clean(payload.get("recipe")) or DEFAULT_RECIPE
+    run_eval = bool(payload.get("run_eval") or payload.get("eval_after_train"))
+    model_old = clean(payload.get("model_old")) or "/mnt/wangsenhao/verl_zk/qwen4b_cispo_wokl_add_bvt_2/global_step_5/actor/huggingface"
+    wf_root = clean(payload.get("workflow_root")) or "/mnt/xiaoai-zk-model-train-tj5/workflow5"
+    warnings: list[str] = []
+    if not data_commit and not data_path:
+        warnings.append("data_commit 未提供：计划将使用 data_branch 当前 HEAD，复现性弱于固定 commit。")
 
     commands = {
         "prepare_workspace": [
@@ -53,7 +65,11 @@ def render_plan(payload: dict[str, Any]) -> dict[str, Any]:
             f"[ -d ai-planning/.git ] || git clone -b {data_branch} {data_repo} ai-planning",
             f"git -C ai-planning fetch origin {data_branch}",
             f"git -C ai-planning checkout {data_branch}",
-            f"git -C ai-planning reset --hard {data_commit}" if data_commit else f"# use existing data_path: {data_path}",
+            f"git -C ai-planning reset --hard {data_commit}" if data_commit else (
+                f"# use existing data_path: {data_path}" if data_path else
+                f"git -C ai-planning reset --hard origin/{data_branch}"
+            ),
+            "git -C ai-planning rev-parse HEAD",
         ],
         "submit_sft": [
             "source ~/.cloudml-cli/.profile",
@@ -67,21 +83,42 @@ def render_plan(payload: dict[str, Any]) -> dict[str, Any]:
             "cml custom_train describe <JOB_ID>",
             f"test -f {workspace}/sft_output/_SUCCESS",
         ],
+        "submit_eval": [
+            "source ~/.cloudml-cli/.profile",
+            f"export AUTORESEARCH_CHAT_ROOT={workspace}",
+            'export AUTORESEARCH_ROOT="$AUTORESEARCH_CHAT_ROOT"',
+            f"export WF_ROOT={wf_root}",
+            f"cd {workspace}",
+            'eval "$(./scripts/resolve_run_ids.sh)"',
+            f'./scripts/submit_cml_eval.sh "$EVAL_RUNDIC" "{workspace}/sft_output" "{model_old}"',
+        ],
+        "analyze_eval": [
+            f"python {workspace}/scripts/summarize_eval_result.py --workflow-root {wf_root} --run-dic <EVAL_RUNDIC> --output {workspace}/results/eval_summary_<EVAL_RUNDIC>.md",
+        ],
     }
     return {
         "ok": not missing,
         "missing": missing,
         "plan": {
             "jupyter_url": jupyter_url,
+            "jupyter_auth": {
+                "type": jupyter_auth_type or ("token" if jupyter_token else "password" if jupyter_password else ""),
+                "configured": bool(jupyter_password or jupyter_token or clean(payload.get("jupyter_cookie")) or clean(defaults.get("cookie"))),
+                "source": auth_source(payload, defaults),
+            },
             "recipe": recipe,
             "workspace": workspace,
             "data_repo": data_repo,
             "data_branch": data_branch,
             "data_commit": data_commit,
             "data_path": data_path,
+            "run_eval": run_eval,
+            "model_old": model_old,
+            "workflow_root": wf_root,
             "success_marker": f"{workspace}/sft_output/_SUCCESS",
         },
         "commands": commands,
+        "warnings": warnings,
         "required_outputs": [
             "CloudML JobID",
             "CloudML task URL",
@@ -89,6 +126,8 @@ def render_plan(payload: dict[str, Any]) -> dict[str, Any]:
             "EVAL_RUNDIC",
             "workspace",
             "success_marker",
+            "eval workflow id when run_eval=true",
+            "eval summary path when analysis is requested",
         ],
     }
 
@@ -97,6 +136,35 @@ def clean(value: Any) -> str:
     return str(value or "").strip()
 
 
+def load_local_defaults() -> dict[str, Any]:
+    defaults: dict[str, Any] = {}
+    env_defaults = {
+        "jupyter_url": os.environ.get("MODEL_TRAINING_LITE_JUPYTER_URL"),
+        "password": os.environ.get("MODEL_TRAINING_LITE_JUPYTER_PASSWORD"),
+        "token": os.environ.get("MODEL_TRAINING_LITE_JUPYTER_TOKEN"),
+        "cookie": os.environ.get("MODEL_TRAINING_LITE_JUPYTER_COOKIE"),
+        "auth_type": os.environ.get("MODEL_TRAINING_LITE_JUPYTER_AUTH_TYPE"),
+    }
+    defaults.update({key: value for key, value in env_defaults.items() if value})
+    config_path = Path(__file__).resolve().parents[1] / ".local" / "jupyter_defaults.json"
+    if config_path.exists():
+        try:
+            file_defaults = json.loads(config_path.read_text(encoding="utf-8"))
+            if isinstance(file_defaults, dict):
+                defaults.update({key: value for key, value in file_defaults.items() if value})
+        except json.JSONDecodeError as exc:
+            raise SystemExit(f"invalid local defaults JSON: {config_path}: {exc}") from exc
+    return defaults
+
+
+def auth_source(payload: dict[str, Any], defaults: dict[str, Any]) -> str:
+    if any(clean(payload.get(key)) for key in ("jupyter_password", "password", "jupyter_token", "token", "jupyter_cookie")):
+        return "input"
+    if any(clean(defaults.get(key)) for key in ("password", "token", "cookie")):
+        return "local_default"
+    return ""
+
+
 def default_run_name(seed: str) -> str:
     slug = re.sub(r"[^a-zA-Z0-9_-]+", "_", seed).strip("_")[:32] or "manual"
     return f"manual_sft_{slug}"