c5a0d61851
Synced files from EngineX baseline zip (2026-06-30): - ADD paged_attn.py (root): production paged attention with PyTorch fallback - ADD launch_service: BI-V100 server startup script with env configuration - SYNC computility-run.yaml: gpu_memory=0.9, batched_tokens=8192, seq_capture=32768 - SYNC qwen3_6_scripts/paged_attn.py: +311 lines, Triton bypass docs, _forward_decode_pytorch shape docs - SYNC qwen3_6_scripts/qwen3_5.py: -72 lines, revert optimized MoE prefill to baseline (untested on BI-V100) - KEEP Dockerfile: repo version has V2/Triton/head256 optimization patches not in baseline Baseline commit: 1902c81fdd373943f17f5983eb8750758c7f4a69 Source: enginex-vllm-bi100-qwen36-main.zip (dev.modelhub.org.cn)
35 lines
707 B
YAML
35 lines
707 B
YAML
concurrency: 1
|
|
command:
|
|
- python3
|
|
- -m
|
|
- vllm.entrypoints.openai.api_server
|
|
- --model
|
|
- /model
|
|
- --served-model-name
|
|
- llm
|
|
- --max-model-len
|
|
- '100000'
|
|
- --gpu-memory-utilization
|
|
- '0.9'
|
|
- --trust-remote-code
|
|
- -tp
|
|
- '4'
|
|
- --max-num-seqs
|
|
- '1'
|
|
- --disable-log-requests
|
|
- --disable-frontend-multiprocessing
|
|
- --max-num-batched-tokens
|
|
- '8192'
|
|
- --enable-chunked-prefill
|
|
- --max-seq-len-to-capture
|
|
- '32768'
|
|
- --enable-auto-tool-choice
|
|
- --tool-call-parser
|
|
- qwen3_coder
|
|
- --reasoning-parser
|
|
- qwen3
|
|
- --enable-prefix-caching
|
|
env:
|
|
- name: VLLM_ENGINE_ITERATION_TIMEOUT_S
|
|
value: 3600
|