cf7824313f
Profiling OOM: flash_attn_varlen_func allocates large temp buffers during profile_run with 8192 dummy tokens. patch_worker_profile_override.py already has skip logic when num_gpu_blocks_override is set. 4000 blocks × 16 tokens/block = 64K token KV capacity. With max-model-len=80000 and prefix caching, this is sufficient. Sub 168 reference: ran 262K context on 0.95 util without override because base image profiling doesn't use flash_attn_varlen (uses native xformers).
55 lines
1.2 KiB
YAML
55 lines
1.2 KiB
YAML
concurrency: 1
|
|
command:
|
|
- python3
|
|
- -m
|
|
- vllm.entrypoints.openai.api_server
|
|
- --model
|
|
- /model
|
|
- --served-model-name
|
|
- llm
|
|
- --max-model-len
|
|
- '80000'
|
|
- --gpu-memory-utilization
|
|
- '0.85'
|
|
- --trust-remote-code
|
|
- -tp
|
|
- '4'
|
|
- --max-num-seqs
|
|
- '2'
|
|
- --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
|
|
- --enforce-eager
|
|
- --dtype
|
|
- half
|
|
- --num-gpu-blocks-override
|
|
- '4000'
|
|
env:
|
|
- name: VLLM_ENGINE_ITERATION_TIMEOUT_S
|
|
value: 3600
|
|
- name: BI100_MOE_COREX_DIRECT_ROUTED
|
|
value: 1
|
|
- name: BI100_GDN_COREX_PACKED_DECODE
|
|
value: 1
|
|
- name: BI100_HYBRID_KV_ACCOUNTING
|
|
value: full_attention
|
|
- name: BI100_GDN_CACHE_POLICY
|
|
value: admission64
|
|
- name: BI100_GDN_RESTORE_MODE
|
|
value: hybrid64
|
|
- name: BI100_MOE_COREX_TOPK_SOFTMAX
|
|
value: '1'
|
|
- name: PYTORCH_CUDA_ALLOC_CONF
|
|
value: max_split_size_mb:512
|
|
|