fix(CRITICAL): stop overwriting base image model layer — match comp 168 strategy
Root cause of ALL failures: we overwrite base image's production code with our inferior versions, breaking multimodal, killing C++ kernel performance, and causing engine death. Comp 168 evidence (48/52 pass, score=60194): - Uses base image qwen3_5.py (81706B) with full multimodal + CoreX integration - Uses base image corex_gdn/moe/fa2.py with real C++ kernels (libcorex_gdn.so) - Uses base image _custom_ops.py (ERROR spam is harmless) - d01: 8.49s, d05 multimodal: PASS, t13 base64 image: PASS Our sub 508 (21/52 pass, score=0): - Overwrites qwen3_5.py → NO multimodal → engine death on image request - Overwrites corex_*.py → Python fallback → d01: 95.87s (11x slower) - Overwrites _custom_ops.py → may break base fallback chain Changes: 1. patch_ops.sh: qwen3_5.py — KEEP base if >1000 bytes (was: ALWAYS overwrite) 2. patch_ops.sh: corex_*.py — KEEP base if >500 bytes (was: ALWAYS overwrite) 3. patch_ops.sh: _custom_ops.py — KEEP base always (was: ALWAYS overwrite) 4. computility-run.yaml: match comp 168 exactly: - max_model_len: 80000 → 256000 - gpu_memory_utilization: 0.95 → 0.9 - max_num_seqs: 2 → 1 - REMOVE chunked_prefill + batched_tokens - REMOVE limit-mm-per-prompt (base image handles it)
This commit is contained in:
project6-dev
@@ -8,17 +8,14 @@ command:
|
||||
- --served-model-name
|
||||
- llm
|
||||
- --max-model-len
|
||||
- '80000'
|
||||
- '256000'
|
||||
- --gpu-memory-utilization
|
||||
- '0.95'
|
||||
- '0.9'
|
||||
- --trust-remote-code
|
||||
- -tp
|
||||
- '4'
|
||||
- --max-num-seqs
|
||||
- '2'
|
||||
- --max-num-batched-tokens
|
||||
- '4096'
|
||||
- --enable-chunked-prefill
|
||||
- '1'
|
||||
- --disable-log-requests
|
||||
- --disable-frontend-multiprocessing
|
||||
- --enforce-eager
|
||||
@@ -32,8 +29,6 @@ command:
|
||||
- '8192'
|
||||
- --dtype
|
||||
- half
|
||||
- --limit-mm-per-prompt
|
||||
- image=5
|
||||
env:
|
||||
- name: VLLM_ENGINE_ITERATION_TIMEOUT_S
|
||||
value: '3600'
|
||||
|
||||
@@ -90,14 +90,23 @@ find /usr/local/corex/lib64/ -name "*.so" 2>/dev/null | head -20 || echo "[probe
|
||||
echo "[probe] ==========================="
|
||||
|
||||
# 2. Model module — qwen3_5.py
|
||||
# Base image qwen3_5.py (81706 bytes) has NaN in GDN:
|
||||
# CoreXGDN.__init__() got unexpected keyword argument 'num_v_heads'
|
||||
# → all GDN layers fallback to base's PyTorch GDN → NaN frac=0.5000
|
||||
# Our version fixes the GDN math (xllm-aligned cumsum + difference form).
|
||||
# ALWAYS deploy ours.
|
||||
# EVIDENCE: comp 168 uses base image qwen3_5.py (81706 bytes) → 48/52 pass, no NaN, 8.49s d01
|
||||
# Our qwen3_5.py LACKS multimodal support → engine death on image request (d05/t13 FAIL)
|
||||
# Our qwen3_5.py LACKS proper CoreX GDN/MoE/FA2 integration → 95s d01 (11x slower)
|
||||
# KEEP base image version. Only deploy ours if base has no qwen3_5.py.
|
||||
_NATIVE_QW="$VLLM/model_executor/models/qwen3_5.py"
|
||||
cp ./qwen3_5.py "$_NATIVE_QW" && \
|
||||
echo "[patch_ops] qwen3_5.py deployed (replaces base — fixes GDN NaN)"
|
||||
if [ -f "$_NATIVE_QW" ]; then
|
||||
_NATIVE_SIZE=$(stat -c%s "$_NATIVE_QW" 2>/dev/null || echo 0)
|
||||
if [ "$_NATIVE_SIZE" -gt 1000 ]; then
|
||||
echo "[patch_ops] KEEP base image qwen3_5.py ($_NATIVE_SIZE bytes) — proven by comp 168 (48/52 pass)"
|
||||
else
|
||||
cp ./qwen3_5.py "$_NATIVE_QW" && \
|
||||
echo "[patch_ops] qwen3_5.py deployed (base was stub: $_NATIVE_SIZE bytes)"
|
||||
fi
|
||||
else
|
||||
cp ./qwen3_5.py "$_NATIVE_QW" && \
|
||||
echo "[patch_ops] qwen3_5.py deployed (base had no qwen3_5.py)"
|
||||
fi
|
||||
|
||||
# 2b. Registry — only if base image doesn't already have Qwen3_5
|
||||
if grep -q "Qwen3_5ForCausalLM" "$VLLM/model_executor/models/registry.py" 2>/dev/null; then
|
||||
@@ -181,21 +190,33 @@ if [ -n "$VLLM2" ]; then
|
||||
fi
|
||||
|
||||
# Deploy corex_gdn.py + corex_moe.py + corex_fa2.py → vllm model_executor/models/
|
||||
# MUST overwrite: base image's corex_gdn.py produces NaN (GDN frac=0.5000).
|
||||
# Our versions have fixed GDN math (fp32 accumulation, cumsum clamp).
|
||||
if [ -f "/workspace/ex_engine/python/corex_gdn.py" ]; then
|
||||
cp "/workspace/ex_engine/python/corex_gdn.py" "$VLLM/model_executor/models/corex_gdn.py" && \
|
||||
echo "[patch_ops] corex_gdn.py deployed (overwrites base — fixes NaN)"
|
||||
cp "/workspace/ex_engine/python/corex_moe.py" "$VLLM/model_executor/models/corex_moe.py" && \
|
||||
echo "[patch_ops] corex_moe.py deployed"
|
||||
cp "/workspace/ex_engine/python/corex_fa2.py" "$VLLM/model_executor/models/corex_fa2.py" && \
|
||||
echo "[patch_ops] corex_fa2.py deployed"
|
||||
if [ -n "$VLLM2" ]; then
|
||||
cp "/workspace/ex_engine/python/corex_gdn.py" "$VLLM2/model_executor/models/corex_gdn.py" 2>/dev/null || true
|
||||
cp "/workspace/ex_engine/python/corex_moe.py" "$VLLM2/model_executor/models/corex_moe.py" 2>/dev/null || true
|
||||
cp "/workspace/ex_engine/python/corex_fa2.py" "$VLLM2/model_executor/models/corex_fa2.py" 2>/dev/null || true
|
||||
# EVIDENCE: comp 168 uses base image corex modules with real C++ kernels (libcorex_gdn.so)
|
||||
# → d01 in 8.49s, corex_gdn.py:56 "Loaded fused CoreX GDN decode operator"
|
||||
# Our Python fallback versions are 11x slower (d01 in 95.87s).
|
||||
# KEEP base image versions if they exist and are non-trivial.
|
||||
for _COREX_MOD in corex_gdn.py corex_moe.py corex_fa2.py; do
|
||||
_NATIVE="$VLLM/model_executor/models/$_COREX_MOD"
|
||||
_OURS="/workspace/ex_engine/python/$_COREX_MOD"
|
||||
if [ -f "$_NATIVE" ]; then
|
||||
_SZ=$(stat -c%s "$_NATIVE" 2>/dev/null || echo 0)
|
||||
if [ "$_SZ" -gt 500 ]; then
|
||||
echo "[patch_ops] KEEP base $_COREX_MOD ($_SZ bytes) — real C++ kernel dispatch"
|
||||
elif [ -f "$_OURS" ]; then
|
||||
cp "$_OURS" "$_NATIVE" && echo "[patch_ops] $_COREX_MOD deployed (base was stub: $_SZ bytes)"
|
||||
fi
|
||||
elif [ -f "$_OURS" ]; then
|
||||
cp "$_OURS" "$_NATIVE" && echo "[patch_ops] $_COREX_MOD deployed (base had none)"
|
||||
fi
|
||||
fi
|
||||
# Mirror to VLLM2
|
||||
if [ -n "$VLLM2" ]; then
|
||||
_NATIVE2="$VLLM2/model_executor/models/$_COREX_MOD"
|
||||
if [ -f "$_NATIVE2" ]; then
|
||||
_SZ2=$(stat -c%s "$_NATIVE2" 2>/dev/null || echo 0)
|
||||
[ "$_SZ2" -gt 500 ] && continue
|
||||
fi
|
||||
[ -f "$_OURS" ] && cp "$_OURS" "$_NATIVE2" 2>/dev/null || true
|
||||
fi
|
||||
done
|
||||
|
||||
# Deploy EX Engine Python module + C++ bridge into vllm importable path
|
||||
EX_ENGINE_SRC="/workspace/ex_engine"
|
||||
@@ -268,15 +289,9 @@ fi
|
||||
echo "[patch_ops] DONE — EX Engine + SM70 GDN kernel + MoE topk kernel + serving layer deployed"
|
||||
echo "[patch_ops] Deployed: qwen3_5.py, flash_qla_sm70, ex_engine factors, paged_attn.py, mamba_cache.py, sequence.py, scheduler.py, xformers patches, serving layer"
|
||||
echo "[patch_ops] EX factors replace: vllm_moe_topk_softmax (2304 calls/token), gdn_chunk_fwd (NaN fix)"
|
||||
# Deploy patched _custom_ops.py — fixes topk_softmax ERROR log spam
|
||||
# Base image ixf_F.vllm_moe_topk_softmax is missing; our patch tries
|
||||
# ixformer._C.topk_softmax first, then silent PyTorch fallback.
|
||||
cp ./_custom_ops.py "$VLLM/_custom_ops.py" 2>/dev/null && \
|
||||
echo "[patch_ops] _custom_ops.py deployed (topk_softmax fix)" || \
|
||||
echo "[patch_ops] WARNING: _custom_ops.py deploy failed"
|
||||
if [ -n "$VLLM2" ]; then
|
||||
cp ./_custom_ops.py "$VLLM2/_custom_ops.py" 2>/dev/null || true
|
||||
fi
|
||||
# _custom_ops.py — comp 168 has the same topk_softmax ERROR spam but still works (48/52 pass).
|
||||
# Do NOT overwrite. The base image handles it via its own fallback chain.
|
||||
echo "[patch_ops] KEEP base _custom_ops.py — comp 168 proves ERROR spam is harmless"
|
||||
|
||||
echo "[patch_ops] NOT deployed (base image native): model_runner.py, sampler.py, logits_processor.py, arg_utils.py"
|
||||
|
||||
|
||||
Reference in New Issue
Block a user