Fix/training error publish

README.md

@@ -55,6 +55,7 @@ After each memory step, Metis selects informative hidden states and updates the
 ```bash
 conda env create -f environment.yml
 conda activate metis
+python -m pip install --no-build-isolation causal-conv1d==1.6.0 flash-attn==2.7.3
 ```
 
 ### 2. Run inference

README_zh.md

@@ -61,6 +61,7 @@ Metis 仍是一个早期研究系统，并非对外部记忆机制的完整替
 ```bash
 conda env create -f environment.yml
 conda activate metis
+python -m pip install --no-build-isolation causal-conv1d==1.6.0 flash-attn==2.7.3
 ```
 
 仓库中的环境文件是研究环境锁定文件，基于 Python 3.10、PyTorch 2.4.1 + CUDA 11.8、Transformers 5.4.0、DeepSpeed 和 Flash Linear Attention。CUDA 扩展可能需要根据具体硬件与平台进行调整。

environment.yml

@@ -1,17 +1,15 @@
 name: metis
 channels:
-  - conda-forge
   - nvidia
+  - conda-forge
 dependencies:
   - _openmp_mutex=4.5
   - bzip2=1.0.8
   - ca-certificates=2026.6.17
-  - cuda-cccl=12.9.27
-  - cuda-cccl_linux-64=12.9.27
-  - cuda-cudart=11.8.89
-  - cuda-cudart-dev=11.8.89
-  - cuda-nvcc=11.8.89
-  - cuda-version=12.9
+  - nvidia::cuda-cccl=11.8.89
+  - nvidia::cuda-cudart=11.8.89
+  - nvidia::cuda-cudart-dev=11.8.89
+  - nvidia::cuda-nvcc=11.8.89
   - ld_impl_linux-64=2.45.1
   - libexpat=2.8.1
   - libffi=3.5.2
@@ -36,6 +34,7 @@ dependencies:
   - zstd=1.5.7
   - pip:
       - --extra-index-url=https://download.pytorch.org/whl/cu118
+      # CUDA extensions must be built after torch is installed; see the README.
       - accelerate==1.14.0
       - aiohappyeyeballs==2.6.2
       - aiohttp==3.14.1
@@ -45,7 +44,6 @@ dependencies:
       - anyio==4.14.0
       - async-timeout==5.0.1
       - attrs==26.1.0
-      - causal-conv1d==1.6.2.post1
       - certifi==2026.6.17
       - charset-normalizer==3.4.7
       - click==8.4.1
@@ -57,7 +55,6 @@ dependencies:
       - einops==0.8.2
       - exceptiongroup==1.3.1
       - filelock==3.29.0
-      - flash-attn==2.6.3
       - flash-linear-attention==0.2.2
       - fonttools==4.63.0
       - frozenlist==1.8.0
@@ -96,7 +93,7 @@ dependencies:
       - nvidia-cusolver-cu11==11.4.1.48
       - nvidia-cusparse-cu11==11.7.5.86
       - nvidia-ml-py==13.595.45
-      - nvidia-nccl-cu11==2.20.5
+      - nvidia-nccl-cu11==2.21.5
       - nvidia-nvtx-cu11==11.8.86
       - nvitop==1.7.0
       - pandas==2.3.3
@@ -124,12 +121,12 @@ dependencies:
       - six==1.17.0
       - smmap==5.0.3
       - socksio==1.0.0
-      - sympy==1.14.0
+      - sympy==1.13.1
       - tokenizers==0.22.2
-      - torch==2.4.1+cu118
+      - torch==2.5.1+cu118
       - tqdm==4.68.3
       - transformers==5.4.0
-      - triton==3.0.0
+      - triton==3.1.0
       - typer==0.25.1
       - typing-extensions==4.15.0
       - typing-inspection==0.4.2

eval/configs/assets.json

@@ -1,6 +1,6 @@
 {
   "schema_version": 1,
-  "note": "Public model IDs may be replaced by local mirrors. Relative paths resolve from the Metis_dev repository root.",
+  "note": "Public model IDs may be replaced by local mirrors. Relative paths resolve from the repository root.",
   "assets": {
     "qwen3_5_4b": "Qwen/Qwen3.5-4B",
     "qwen3_5_9b": "Qwen/Qwen3.5-9B",

eval/experiments/ablation/run.py

@@ -13,6 +13,7 @@
 import shutil
 import subprocess
 import sys
+import tempfile
 from collections import defaultdict
 from pathlib import Path
 from typing import Any, Iterable
@@ -21,8 +22,9 @@
 
 
 REPO_ROOT = Path(__file__).resolve().parents[3]
-METIS_DEV_ROOT = Path(os.environ.get("METIS_MODEL_REPO_ROOT", str(REPO_ROOT)))
+MODEL_REPO_ROOT = Path(os.environ.get("METIS_MODEL_REPO_ROOT", str(REPO_ROOT)))
 DEFAULT_MATRIX = Path(__file__).resolve().parent / "configs" / "ablation_matrix.json"
+DEFAULT_JUDGE_LOCK = Path(tempfile.gettempdir()) / "metis_eval_judge32.lock"
 READY_STATUSES = {"ready", "reference_ready"}
 REQUIRED_CHECKPOINT_FILES = {
     "config.json",
@@ -562,7 +564,7 @@ def run_raw_stage(args: argparse.Namespace, matrix: dict[str, Any]) -> dict[str,
         "matrix": str(args.matrix),
         "matrix_sha256": sha256(args.matrix),
         "git": git_snapshot(),
-        "metis_dev_git": git_snapshot(METIS_DEV_ROOT),
+        "model_repo_git": git_snapshot(MODEL_REPO_ROOT),
         "python": args.python,
         "checkpoints": selected,
         "benchmarks": benchmarks,
@@ -915,7 +917,7 @@ def build_parser() -> argparse.ArgumentParser:
     parser.add_argument("--judge-concurrency", type=int, default=0)
     parser.add_argument("--judge-base-url", default="https://api.openai.com")
     parser.add_argument("--api-key-env", default="OPENAI_API_KEY")
-    parser.add_argument("--judge-lock", type=Path, default=Path("/tmp/metis_eval_judge32.lock"))
+    parser.add_argument("--judge-lock", type=Path, default=DEFAULT_JUDGE_LOCK)
     parser.add_argument("--judge-max-attempts", type=int, default=18)
     parser.add_argument("--judge-retry-sleep", type=float, default=2.0)
     parser.add_argument("--judge-retry-backoff", type=float, default=1.5)

eval/tests/test_protocols.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import json
+import re
 import subprocess
 import sys
 import tempfile
@@ -160,15 +161,31 @@ def test_ablation_checkpoints_use_the_shared_asset_registry() -> None:
     assert {item["id"] for item in matrix["checkpoints"]} <= set(assets)
 
 
-def test_release_text_has_no_private_server_paths_or_usernames() -> None:
+def test_release_text_has_no_private_paths_or_development_repo_names() -> None:
     root = Path(__file__).resolve().parents[2]
-    forbidden = (
-        "/" + "mnt/afs/",
+    forbidden_literals = (
+        "/" + "mnt/",
         "/" + "home/",
+        "/" + "root/",
         "/" + "Users/",
+        "/" + "tmp/",
         "api-" + "int.",
     )
-    suffixes = {".py", ".json", ".md", ".toml", ".yml", ".txt"}
+    forbidden_repo_names = (
+        "Metis_" + "dev",
+        "Metis_" + "open",
+    )
+    suffixes = {
+        ".cff",
+        ".json",
+        ".md",
+        ".py",
+        ".sh",
+        ".toml",
+        ".txt",
+        ".yaml",
+        ".yml",
+    }
     candidates = subprocess.run(
         [
             "git",
@@ -177,7 +194,6 @@ def test_release_text_has_no_private_server_paths_or_usernames() -> None:
             "--cached",
             "--others",
             "--exclude-standard",
-            "eval",
         ],
         cwd=root,
         check=True,
@@ -189,7 +205,14 @@ def test_release_text_has_no_private_server_paths_or_usernames() -> None:
         path = root / relative.decode()
         if path.is_file() and path.suffix in suffixes:
             text = path.read_text(encoding="utf-8", errors="replace").lower()
-            assert not any(value.lower() in text for value in forbidden), path
+            assert not any(value.lower() in text for value in forbidden_literals), path
+            assert not any(
+                re.search(
+                    rf"(?<![a-z0-9]){re.escape(value.lower())}(?![a-z0-9])",
+                    text,
+                )
+                for value in forbidden_repo_names
+            ), path
 
 
 def test_generated_result_metadata_is_git_ignored() -> None:

metis/backbone_wrappers/Qwen3_5_wrapper.py

@@ -23,6 +23,26 @@
 from transformers.modeling_outputs import BaseModelOutputWithPast
 
 
+def _zero3_safe_checkpoint(function, hidden_states):
+    """Checkpoint frozen-backbone computation safely under ZeRO-3.
+
+    Reentrant checkpoint avoids the PyTorch 2.5 non-reentrant saved-tensor
+    metadata conflict with ZeRO-3 parameter re-partitioning.
+
+    If the input does not require gradients, the enclosed backbone operation
+    is fully frozen, so checkpointing it provides no backward-memory benefit.
+    """
+    if not torch.is_grad_enabled() or not hidden_states.requires_grad:
+        return function(hidden_states)
+
+    return _ckpt(
+        function,
+        hidden_states,
+        use_reentrant=True,
+        preserve_rng_state=True,
+    )
+
+
 class Qwen3_5DecoderLayerForMetis(DecoderLayerWrapperForMetis):
     def __init__(self, config, raw_decoder):
         super().__init__(config, raw_decoder)
@@ -137,7 +157,10 @@ def _attn(hidden_states):
             attn_output = self_attn.o_proj(attn_output)
             return memory_for_query, attn_output, residual
 
-        memory_for_query, attn_output, residual = _ckpt(_attn, hidden_states, use_reentrant=False)
+        memory_for_query, attn_output, residual = _zero3_safe_checkpoint(
+            _attn,
+            hidden_states,
+        )
 
         return memory_for_query, attn_output, {'residual': residual}, self_attn.o_proj
 
@@ -183,7 +206,7 @@ def _mix(hidden_states):
                 )
                 return residual + mixed
 
-            hidden_states = _ckpt(_mix, hidden_states, use_reentrant=False)
+            hidden_states = _zero3_safe_checkpoint(_mix, hidden_states)
 
         elif self.raw_decoder.layer_type == "full_attention":
             residual = cache_dict['residual']
@@ -194,7 +217,7 @@ def _mlp(hidden_states):
             normed = self.raw_decoder.post_attention_layernorm(hidden_states)
             return residual + self.raw_decoder.mlp(normed)
 
-        hidden_states = _ckpt(_mlp, hidden_states, use_reentrant=False)
+        hidden_states = _zero3_safe_checkpoint(_mlp, hidden_states)
 
         return hidden_states