2026-08-06 06:44:31 +00:00
|
|
|
#!/bin/bash
|
2026-08-08 05:56:58 +00:00
|
|
|
# ==========================================================================
|
|
|
|
|
# SERVING-LAYER-ONLY PATCHES
|
|
|
|
|
#
|
|
|
|
|
# EVIDENCE FROM SUB168 DOCKER LOG (07-23, competition reference):
|
|
|
|
|
# - corex_gdn.py:56 "Loaded fused CoreX GDN decode operator" ✓
|
|
|
|
|
# - corex_moe.py:339 "Using CoreX fused MoE prefill operator" ✓
|
|
|
|
|
# - model_runner.py:1074 (base image's line number)
|
|
|
|
|
# - "Loading model weights took 17.3529 GB"
|
|
|
|
|
# - ZERO NaN warnings
|
|
|
|
|
# - d01: 8.49s, d03_tool_call: PASS in 2.12s
|
|
|
|
|
#
|
|
|
|
|
# EVIDENCE FROM OUR SUB508 DOCKER LOG (08-07):
|
|
|
|
|
# - NO corex_gdn loading
|
|
|
|
|
# - model_runner.py:1119 (our custom code)
|
|
|
|
|
# - "Loading model weights took 16.2303 GB" (1.1GB MISSING)
|
|
|
|
|
# - 16 NaN in prefill, 19 FusedMoE failures
|
|
|
|
|
# - d01: 95.87s, d03_tool_call: FAIL in 49s
|
|
|
|
|
#
|
|
|
|
|
# CONCLUSION: Sub168 succeeds by using BASE IMAGE native model code.
|
|
|
|
|
# Our custom qwen3_5.py/model_runner/etc BREAKS CoreX acceleration.
|
|
|
|
|
#
|
|
|
|
|
# DO NOT deploy: qwen3_5.py, model_runner.py, _custom_ops.py,
|
|
|
|
|
# sampler.py, scheduler.py, sequence.py, xformers.py, paged_attn.py,
|
|
|
|
|
# prefix_prefill.py, logits_processor.py, mamba_cache.py, arg_utils.py
|
|
|
|
|
# ==========================================================================
|
2026-08-05 08:24:38 +00:00
|
|
|
|
2026-08-07 06:20:02 +00:00
|
|
|
cd "$(dirname "$0")"
|
2026-08-07 10:37:41 +00:00
|
|
|
echo "[patch_ops] START — working directory: $(pwd)"
|
2026-08-07 06:20:02 +00:00
|
|
|
|
2026-08-07 10:37:41 +00:00
|
|
|
# Find vllm installation
|
|
|
|
|
VLLM=""
|
|
|
|
|
for P in /usr/local/corex/lib/python3/dist-packages/vllm \
|
|
|
|
|
/usr/local/corex/lib64/python3/dist-packages/vllm; do
|
|
|
|
|
if [ -d "$P" ]; then
|
|
|
|
|
VLLM="$P"
|
|
|
|
|
echo "[patch_ops] Found vllm at: $VLLM"
|
|
|
|
|
break
|
|
|
|
|
fi
|
|
|
|
|
done
|
2026-08-07 04:51:27 +00:00
|
|
|
|
2026-08-07 10:37:41 +00:00
|
|
|
if [ -z "$VLLM" ]; then
|
|
|
|
|
echo "[patch_ops] ERROR: vllm not found"
|
2026-08-05 08:24:38 +00:00
|
|
|
exit 1
|
|
|
|
|
fi
|
|
|
|
|
|
2026-08-08 05:56:58 +00:00
|
|
|
# 1. Transformers config registration (config only, NOT model code)
|
2026-08-07 10:37:41 +00:00
|
|
|
TMODELS=""
|
fix(critical): remove pip install transformers — was breaking corex kernel loading
ROOT CAUSE FOUND from competitor sub168 docker log comparison:
Sub168 (competitor, works):
- corex_gdn.py:56] Loaded fused CoreX GDN decode operator ✓
- corex_moe.py:339] Using CoreX fused MoE prefill operator ✓
- corex_fa2.py:333] Using CoreX FA2 packed prefill ✓
- NO NaN warnings, NO MoE fallback
- max_model_len=256000, gpu_mem=0.95, max_num_seqs=2 (yaml params work)
Sub509 (ours, broken):
- NaN in prefill GatedDeltaNet layer 0 (frac=0.9998) ✗
- FusedMoE native kernel failed, falling back to PyTorch ✗
- NO corex_gdn/corex_moe/corex_fa2 loading logs at all
- max_model_len=100000, gpu_mem=0.9, max_num_seqs=1 (yaml params ignored)
The pip install transformers==4.55.3 in patch_ops.sh was the likely cause:
it changed dependencies that broke corex kernel loading paths.
Without corex_gdn, GatedDeltaNet falls back to Python → NaN.
Without corex_moe, MoE falls back to PyTorch → 10x slower.
Fix: Remove pip install, use base image's transformers version.
Only register qwen3_5 config files without upgrading the package.
2026-08-07 10:02:09 +00:00
|
|
|
for P in /usr/local/lib/python3.10/site-packages/transformers/models \
|
|
|
|
|
/usr/local/corex/lib/python3/dist-packages/transformers/models \
|
|
|
|
|
/usr/local/corex/lib64/python3/dist-packages/transformers/models; do
|
|
|
|
|
if [ -d "$P" ]; then
|
2026-08-07 10:37:41 +00:00
|
|
|
TMODELS="$P"
|
fix(critical): remove pip install transformers — was breaking corex kernel loading
ROOT CAUSE FOUND from competitor sub168 docker log comparison:
Sub168 (competitor, works):
- corex_gdn.py:56] Loaded fused CoreX GDN decode operator ✓
- corex_moe.py:339] Using CoreX fused MoE prefill operator ✓
- corex_fa2.py:333] Using CoreX FA2 packed prefill ✓
- NO NaN warnings, NO MoE fallback
- max_model_len=256000, gpu_mem=0.95, max_num_seqs=2 (yaml params work)
Sub509 (ours, broken):
- NaN in prefill GatedDeltaNet layer 0 (frac=0.9998) ✗
- FusedMoE native kernel failed, falling back to PyTorch ✗
- NO corex_gdn/corex_moe/corex_fa2 loading logs at all
- max_model_len=100000, gpu_mem=0.9, max_num_seqs=1 (yaml params ignored)
The pip install transformers==4.55.3 in patch_ops.sh was the likely cause:
it changed dependencies that broke corex kernel loading paths.
Without corex_gdn, GatedDeltaNet falls back to Python → NaN.
Without corex_moe, MoE falls back to PyTorch → 10x slower.
Fix: Remove pip install, use base image's transformers version.
Only register qwen3_5 config files without upgrading the package.
2026-08-07 10:02:09 +00:00
|
|
|
break
|
|
|
|
|
fi
|
|
|
|
|
done
|
2026-08-07 10:37:41 +00:00
|
|
|
if [ -n "$TMODELS" ]; then
|
|
|
|
|
cp -r ./qwen3_5 "$TMODELS/" 2>/dev/null && echo "[patch_ops] qwen3_5 config copied" || true
|
|
|
|
|
cp -r ./qwen3_5_moe "$TMODELS/" 2>/dev/null && echo "[patch_ops] qwen3_5_moe config copied" || true
|
|
|
|
|
python3 ./patch_transformers_qwen3_5.py 2>&1 || echo "[patch_ops] WARNING: transformers patch failed (non-fatal)"
|
fix(critical): remove pip install transformers — was breaking corex kernel loading
ROOT CAUSE FOUND from competitor sub168 docker log comparison:
Sub168 (competitor, works):
- corex_gdn.py:56] Loaded fused CoreX GDN decode operator ✓
- corex_moe.py:339] Using CoreX fused MoE prefill operator ✓
- corex_fa2.py:333] Using CoreX FA2 packed prefill ✓
- NO NaN warnings, NO MoE fallback
- max_model_len=256000, gpu_mem=0.95, max_num_seqs=2 (yaml params work)
Sub509 (ours, broken):
- NaN in prefill GatedDeltaNet layer 0 (frac=0.9998) ✗
- FusedMoE native kernel failed, falling back to PyTorch ✗
- NO corex_gdn/corex_moe/corex_fa2 loading logs at all
- max_model_len=100000, gpu_mem=0.9, max_num_seqs=1 (yaml params ignored)
The pip install transformers==4.55.3 in patch_ops.sh was the likely cause:
it changed dependencies that broke corex kernel loading paths.
Without corex_gdn, GatedDeltaNet falls back to Python → NaN.
Without corex_moe, MoE falls back to PyTorch → 10x slower.
Fix: Remove pip install, use base image's transformers version.
Only register qwen3_5 config files without upgrading the package.
2026-08-07 10:02:09 +00:00
|
|
|
else
|
2026-08-07 10:37:41 +00:00
|
|
|
echo "[patch_ops] WARNING: transformers/models not found"
|
fix(critical): remove pip install transformers — was breaking corex kernel loading
ROOT CAUSE FOUND from competitor sub168 docker log comparison:
Sub168 (competitor, works):
- corex_gdn.py:56] Loaded fused CoreX GDN decode operator ✓
- corex_moe.py:339] Using CoreX fused MoE prefill operator ✓
- corex_fa2.py:333] Using CoreX FA2 packed prefill ✓
- NO NaN warnings, NO MoE fallback
- max_model_len=256000, gpu_mem=0.95, max_num_seqs=2 (yaml params work)
Sub509 (ours, broken):
- NaN in prefill GatedDeltaNet layer 0 (frac=0.9998) ✗
- FusedMoE native kernel failed, falling back to PyTorch ✗
- NO corex_gdn/corex_moe/corex_fa2 loading logs at all
- max_model_len=100000, gpu_mem=0.9, max_num_seqs=1 (yaml params ignored)
The pip install transformers==4.55.3 in patch_ops.sh was the likely cause:
it changed dependencies that broke corex kernel loading paths.
Without corex_gdn, GatedDeltaNet falls back to Python → NaN.
Without corex_moe, MoE falls back to PyTorch → 10x slower.
Fix: Remove pip install, use base image's transformers version.
Only register qwen3_5 config files without upgrading the package.
2026-08-07 10:02:09 +00:00
|
|
|
fi
|
2026-07-30 16:06:20 +00:00
|
|
|
|
2026-08-08 05:56:58 +00:00
|
|
|
# 2. 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
|
|
|
|
|
echo "[patch_ops] registry already has Qwen3_5 — NOT overwriting"
|
|
|
|
|
else
|
2026-08-07 10:37:41 +00:00
|
|
|
cp ./registry.py "$VLLM/model_executor/models/registry.py" 2>/dev/null && \
|
2026-08-08 05:56:58 +00:00
|
|
|
echo "[patch_ops] registry.py deployed" || true
|
arch(critical): deploy ALL customized files to container — qwen3_5.py was NEVER running
ROOT CAUSE FOUND: patch_ops.sh only deployed serving-layer files
(tool_parser, reasoning, protocol, serving_chat) but NEVER deployed:
- qwen3_5.py (1712 lines of NaN-safe DeltaNet + CCCL patterns)
- _custom_ops.py (MoE kernel fallback for BI-V100)
- model_runner.py (has_inner_state for DeltaNet MambaCacheManager)
- sampler.py, sequence.py, scheduler.py, arg_utils.py
- xformers.py, paged_attn.py, prefix_prefill.py
- logits_processor.py, mamba_cache.py
The container was running the BASE IMAGE's original qwen3_5.py which has:
- NO NaN clamping (g.clamp, cumsum.clamp, state.clamp)
- NO overflow_cast protection (CCCL pattern)
- NO forward substitution fallback (cuSOLVER unavailable on BI-V100)
- NO batched GEMM MoE decode (3 launches vs 16)
- NO sorted-segment MoE prefill (CCCL histogram pattern)
- NO GDN prefix-cache state save/restore
This explains why Docker logs showed 99.98% NaN in EVERY DeltaNet layer
despite our qwen3_5.py having comprehensive numerical guards.
Also fixes:
- serving_chat.py: n>1 returns 400 instead of clamping (prevents OOM cascade)
- serving_chat.py: improved d07 content fallback (multi-layer extraction)
2026-08-08 05:37:46 +00:00
|
|
|
fi
|
|
|
|
|
|
2026-08-07 10:37:41 +00:00
|
|
|
# 3. Tool parser
|
|
|
|
|
mkdir -p "$VLLM/entrypoints/openai/tool_parsers" 2>/dev/null || true
|
|
|
|
|
cp ./qwen3coder_tool_parser.py "$VLLM/entrypoints/openai/tool_parsers/" 2>/dev/null || true
|
|
|
|
|
cp ./tool_parsers_init.py "$VLLM/entrypoints/openai/tool_parsers/__init__.py" 2>/dev/null || true
|
|
|
|
|
echo "[patch_ops] tool parser deployed"
|
2026-07-30 16:06:20 +00:00
|
|
|
|
2026-08-07 10:37:41 +00:00
|
|
|
# 4. Reasoning parser
|
|
|
|
|
cp -r ./reasoning "$VLLM/" 2>/dev/null || true
|
|
|
|
|
echo "[patch_ops] reasoning parser deployed"
|
2026-07-30 16:06:20 +00:00
|
|
|
|
2026-08-08 05:56:58 +00:00
|
|
|
# 5. Serving layer ONLY
|
2026-08-07 10:37:41 +00:00
|
|
|
cp ./protocol.py "$VLLM/entrypoints/openai/protocol.py" 2>/dev/null || true
|
|
|
|
|
cp ./cli_args.py "$VLLM/entrypoints/openai/cli_args.py" 2>/dev/null || true
|
|
|
|
|
cp ./serving_chat.py "$VLLM/entrypoints/openai/serving_chat.py" 2>/dev/null || true
|
|
|
|
|
cp ./api_server.py "$VLLM/entrypoints/openai/api_server.py" 2>/dev/null || true
|
|
|
|
|
cp ./chat_utils.py "$VLLM/entrypoints/chat_utils.py" 2>/dev/null || true
|
|
|
|
|
echo "[patch_ops] serving layer deployed"
|
fix(critical): stop replacing base image compute files — use corex native kernels
ROOT CAUSE OF ALL FAILURES:
patch_ops.sh was replacing qwen3_5.py, _custom_ops.py, model_runner.py,
xformers.py, paged_attn.py, prefix_prefill.py, logits_processor.py,
sampler.py, arg_utils.py — killing base image's CoreX fused kernels.
Evidence from competitor sub168 docker logs (d03 PASS in 2.12s):
- 'Using fused CoreX GDN decode operator' (DeltaNet)
- 'Using CoreX fused MoE prefill operator: tokens=4096, kernel=expert-grouped-wmma'
- 'Using CoreX FA2 packed prefill: B=2 Hq=4 Hkv=1 D=256'
- ZERO NaN warnings
- Model weights: 17.35GB (full)
Our sub509 (d03 FAIL in 49s):
- 'NaN in prefill GatedDeltaNet layer 0 (frac=0.9998)' — 99.98% NaN!
- 'FusedMoE native kernel failed, falling back to pure PyTorch'
- No CoreX FA2
- Model weights: 16.23GB (incomplete — 1.1GB missing)
CCCL design principle (dispatch_reduce_deterministic.cuh, transform.cu):
Let the framework's policy_selector choose optimal kernel config per
hardware — never hand-replace the dispatch layer.
Now patch_ops.sh ONLY patches serving layer:
- protocol.py, serving_chat.py, api_server.py, chat_utils.py, cli_args.py
- qwen3coder_tool_parser.py (tool call XML parsing)
- reasoning/ (think tag parsing)
- registry.py (register Qwen3_5 model type)
- transformers models (qwen3_5 config)
Base image compute files PRESERVED:
qwen3_5.py, _custom_ops.py, model_runner.py, xformers.py,
paged_attn.py, prefix_prefill.py, logits_processor.py, sampler.py,
arg_utils.py, sequence.py, scheduler.py
2026-08-07 09:21:43 +00:00
|
|
|
|
2026-08-08 05:56:58 +00:00
|
|
|
# 6. Mirror to second vllm path if exists
|
2026-08-07 10:37:41 +00:00
|
|
|
VLLM2=""
|
|
|
|
|
for P in /usr/local/corex/lib/python3/dist-packages/vllm \
|
|
|
|
|
/usr/local/corex/lib64/python3/dist-packages/vllm; do
|
|
|
|
|
if [ -d "$P" ] && [ "$P" != "$VLLM" ]; then
|
|
|
|
|
VLLM2="$P"
|
|
|
|
|
break
|
|
|
|
|
fi
|
|
|
|
|
done
|
|
|
|
|
if [ -n "$VLLM2" ]; then
|
2026-08-08 05:56:58 +00:00
|
|
|
echo "[patch_ops] Second vllm at: $VLLM2"
|
|
|
|
|
if ! grep -q "Qwen3_5ForCausalLM" "$VLLM2/model_executor/models/registry.py" 2>/dev/null; then
|
|
|
|
|
cp ./registry.py "$VLLM2/model_executor/models/registry.py" 2>/dev/null || true
|
|
|
|
|
fi
|
2026-08-07 10:37:41 +00:00
|
|
|
mkdir -p "$VLLM2/entrypoints/openai/tool_parsers" 2>/dev/null || true
|
|
|
|
|
cp ./qwen3coder_tool_parser.py "$VLLM2/entrypoints/openai/tool_parsers/" 2>/dev/null || true
|
|
|
|
|
cp ./tool_parsers_init.py "$VLLM2/entrypoints/openai/tool_parsers/__init__.py" 2>/dev/null || true
|
|
|
|
|
cp -r ./reasoning "$VLLM2/" 2>/dev/null || true
|
|
|
|
|
cp ./protocol.py "$VLLM2/entrypoints/openai/protocol.py" 2>/dev/null || true
|
|
|
|
|
cp ./cli_args.py "$VLLM2/entrypoints/openai/cli_args.py" 2>/dev/null || true
|
|
|
|
|
cp ./serving_chat.py "$VLLM2/entrypoints/openai/serving_chat.py" 2>/dev/null || true
|
|
|
|
|
cp ./api_server.py "$VLLM2/entrypoints/openai/api_server.py" 2>/dev/null || true
|
|
|
|
|
cp ./chat_utils.py "$VLLM2/entrypoints/chat_utils.py" 2>/dev/null || true
|
|
|
|
|
fi
|
fix(critical): stop replacing base image compute files — use corex native kernels
ROOT CAUSE OF ALL FAILURES:
patch_ops.sh was replacing qwen3_5.py, _custom_ops.py, model_runner.py,
xformers.py, paged_attn.py, prefix_prefill.py, logits_processor.py,
sampler.py, arg_utils.py — killing base image's CoreX fused kernels.
Evidence from competitor sub168 docker logs (d03 PASS in 2.12s):
- 'Using fused CoreX GDN decode operator' (DeltaNet)
- 'Using CoreX fused MoE prefill operator: tokens=4096, kernel=expert-grouped-wmma'
- 'Using CoreX FA2 packed prefill: B=2 Hq=4 Hkv=1 D=256'
- ZERO NaN warnings
- Model weights: 17.35GB (full)
Our sub509 (d03 FAIL in 49s):
- 'NaN in prefill GatedDeltaNet layer 0 (frac=0.9998)' — 99.98% NaN!
- 'FusedMoE native kernel failed, falling back to pure PyTorch'
- No CoreX FA2
- Model weights: 16.23GB (incomplete — 1.1GB missing)
CCCL design principle (dispatch_reduce_deterministic.cuh, transform.cu):
Let the framework's policy_selector choose optimal kernel config per
hardware — never hand-replace the dispatch layer.
Now patch_ops.sh ONLY patches serving layer:
- protocol.py, serving_chat.py, api_server.py, chat_utils.py, cli_args.py
- qwen3coder_tool_parser.py (tool call XML parsing)
- reasoning/ (think tag parsing)
- registry.py (register Qwen3_5 model type)
- transformers models (qwen3_5 config)
Base image compute files PRESERVED:
qwen3_5.py, _custom_ops.py, model_runner.py, xformers.py,
paged_attn.py, prefix_prefill.py, logits_processor.py, sampler.py,
arg_utils.py, sequence.py, scheduler.py
2026-08-07 09:21:43 +00:00
|
|
|
|
2026-08-08 05:56:58 +00:00
|
|
|
echo "[patch_ops] DONE — serving-only patches, CoreX native model PRESERVED"
|
|
|
|
|
echo "[patch_ops] NOT deployed (base image native): qwen3_5.py, model_runner.py, _custom_ops.py, sampler.py, scheduler.py, sequence.py, xformers.py, paged_attn.py, prefix_prefill.py, logits_processor.py, mamba_cache.py, arg_utils.py"
|