fix(critical): match Sub168 config exactly + disable risky numerical patch

1. computility-run.yaml: restore Sub168's proven params:
   - max-model-len=256000 (not 100000)
   - gpu-memory-utilization=0.95 (not 0.90)
   - max-num-seqs=2 (not 1)
   - max-num-batched-tokens=4096 (restored)
   - enable-chunked-prefill (restored)
   These params worked for Sub168. Now that pip install is removed,
   they should work for us too.

2. patch_ops.sh: disable patch_numerical_stability.py
   If corex_gdn loads (which it should without pip install breaking deps),
   Python GatedDeltaNet fallback never runs, so numerical patches are
   unnecessary. Running regex replacements on qwen3_5.py risks breaking
   corex import conditions.
This commit is contained in:
project6
2026-08-07 10:27:53 +00:00
parent 1ba0dd3966
commit 2680d62ec8
2 changed files with 14 additions and 15 deletions

View File

@@ -8,14 +8,16 @@ command:
- --served-model-name
- llm
- --max-model-len
- '100000'
- '256000'
- --gpu-memory-utilization
- '0.90'
- '0.95'
- --trust-remote-code
- -tp
- '4'
- --max-num-seqs
- '1'
- '2'
- --max-num-batched-tokens
- '4096'
- --disable-log-requests
- --disable-frontend-multiprocessing
- --enforce-eager
@@ -25,6 +27,7 @@ command:
- --reasoning-parser
- qwen3
- --enable-prefix-caching
- --enable-chunked-prefill
- --dtype
- half
env:

View File

@@ -107,19 +107,15 @@ done
echo "[patch_ops] reasoning parser + serving files installed"
# ============================================================
# 4. CCCL Agent-pattern: numerical stability patch for qwen3_5.py
# Sub509 docker logs: 99.98% NaN in every GatedDeltaNet layer.
# Base image has NaN detection + nan_to_num(nan=0.0), but that
# means DeltaNet layers output all-zeros → model "brain dead"
# → can't produce <tool_call> XML → d03 FAIL.
#
# Strategy (CCCL optionally_static): detect what guards exist,
# inject ONLY what's missing. Preserve corex kernel paths.
# Agent flow: Init → Detect → Patch → Verify.
# 4. Numerical stability patch — DISABLED
# If corex_gdn loads (which it should without pip install),
# the Python _torch_chunk_gated_delta_rule is NEVER called.
# Patching qwen3_5.py risks breaking corex import conditions.
# Only enable this if docker logs still show NaN after corex fix.
# ============================================================
python3 ./patch_numerical_stability.py 2>&1 || \
echo "[patch_ops] WARNING: numerical stability patch failed (non-fatal)"
echo "[patch_ops] numerical stability patch complete"
# python3 ./patch_numerical_stability.py 2>&1 || \
# echo "[patch_ops] WARNING: numerical stability patch failed (non-fatal)"
echo "[patch_ops] numerical stability patch SKIPPED (corex_gdn handles DeltaNet)"
# ============================================================
# 5. DO NOT full-replace these files — base image has optimized versions.