perf(config): match competitor Sub168's proven engine params

From competitor docker log analysis:
- max_model_len: 100000 → 256000 (competitor proven, 19259 GPU blocks)
- gpu_memory_utilization: 0.90 → 0.95 (competitor proven)
- max_num_batched_tokens: None → 4096 (competitor proven)
- enable_chunked_prefill: off → on (competitor proven, critical for 256K context)
- max_num_seqs stays at 2 (matches competitor)

Competitor Sub168 scored 60194 with these exact params before OOM at replay tail.
Our code has OOM-surviving advantages they lack (n>1 clamp, max_completion_tokens).

Docker log evidence: competitor's vLLM started with 19259 GPU blocks at 0.95 util,
ran for ~1h18m before OOM in layernorm.py x.float() at 31.72 GiB GPU.
This commit is contained in:
project6
2026-08-07 08:48:20 +00:00
parent bf6ceb0b12
commit 391866785e

View File

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