Claude
|
16f0b30d2e
|
fix(critical): 3 bugs causing 0.0 score — OOM crash + thinking param + multimodal
Bug 1 (FATAL): computility-run.yaml max-model-len 256000 → 131072,
gpu-memory-utilization 0.95 → 0.90, max-num-batched-tokens 4096 → 8192.
Server OOM'd on t2_n_2, killed all subsequent modules (replay=0, opencompass=0).
Bug 2 (functional): protocol.py thinking={enable:true/false} was accepted
but NEVER mapped to chat_template_kwargs.enable_thinking. Qwen3 template
never received the parameter → t1a, t1c, d07, d10 all FAIL.
Bug 3 (functional): chat_utils.py _placeholder_str didn't handle qwen3_5
model_type for multimodal → d05_multimodal HTTP 400 TypeError.
Expected: functional pass rate 0.41 → 0.90+, server stays alive for all
4 modules, total score 0.0 → 60000+ (matching reference sub 168).
|
2026-08-07 07:05:40 +00:00 |
|
dylanyunlon
|
ef6abf3dc7
|
[DEPLOY] Complete submission: baseline + all optimizations
Adds ALL files needed for Dockerfile build:
- qwen3_6_scripts/ (baseline patches + our optimizations)
- vllm/ (full vllm package)
- paged_attention_v2_pytorch.py (V2 with single-bmm optimization)
- Dockerfile + computility-run.yaml
Our optimizations vs baseline:
1. paged_attn.py: pre-gathered context KV (eliminates 194 gather calls),
Triton try/fallback, V2 heuristic, threshold 32K→64K
2. paged_attention_v2_pytorch.py: fills NotImplementedError,
single-bmm Phase 1 (195 launches → 3)
3. patch_enable_triton.py: HAS_TRITON=True with safety fallback
4. patch_triton_tuning.py: BLOCK=64, NUM_WARPS=4 for BI-V100
5. computility-run.yaml: gpu-memory-utilization 0.9→0.95,
max-num-batched-tokens 8192→16384
This repo can now be submitted to dev.modelhub.org.cn as-is.
|
2026-07-30 16:06:20 +00:00 |
|