Vllm Xpu Profile

vllm-xpu-profile

Profile a vLLM-XPU server with torch.profiler to see scheduler, KV manager, attention backend, and batching alongside XPU op timeline. For pure-PyTorch traces use torch-xpu-profile; for SYCL kernel level use xpu-profile-unitrace.

Modes

Mode A — server + HTTP bracket

Launch the server with profiler flags. Always launch from the official upstream vllm/vllm-openai-xpu:latest image — even if a running container or host process is using a different image (e.g. intel/llm-scaler-vllm), do not reuse that image for the profiling server; those stacks are out of scope (see What this skill does NOT cover).

docker run -d --name vllm-xpu-prof \
    --device /dev/dri \
    -v /dev/dri/by-path:/dev/dri/by-path:ro \
    --group-add "$(getent group render | cut -d: -f3)" \
    --ipc=host \
    -e ZE_AFFINITY_MASK=0 \
    -e VLLM_WORKER_MULTIPROC_METHOD=spawn \
   -e HTTP_PROXY -e HTTPS_PROXY -e NO_PROXY \
   -e http_proxy -e https_proxy -e no_proxy \
   -e HF_TOKEN \
    -v "$HOME/.cache/huggingface:/root/.cache/huggingface" \
    -v "$PWD/traces:/work/traces" \
    -p 8000:8000 \
   vllm/vllm-openai-xpu:latest \
   Qwen/Qwen2.5-1.5B-Instruct \
        --dtype bfloat16 --enforce-eager --max-model-len 4096 \
        --profiler-config.profiler=torch \
        --profiler-config.torch_profiler_dir=/work/traces

Wait for Application startup complete, then bracket your window:

# Warmup (don't profile this) — kernel cache hot
for i in 1 2 3; do
    curl -s http://localhost:8000/v1/chat/completions \
        -H 'Content-Type: application/json' \
        -d '{"model":"Qwen/Qwen2.5-1.5B-Instruct",
             "messages":[{"role":"user","content":"hi"}],"max_tokens":32}' >/dev/null
done

curl -X POST http://localhost:8000/start_profile

# Workload to characterise
for i in $(seq 1 8); do
    curl -s http://localhost:8000/v1/chat/completions \
        -H 'Content-Type: application/json' \
        -d '{"model":"Qwen/Qwen2.5-1.5B-Instruct",
             "messages":[{"role":"user","content":"Write a short paragraph."}],
             "max_tokens":128}' &
done; wait

curl -X POST http://localhost:8000/stop_profile
ls traces/

Qwen/Qwen2.5-1.5B-Instruct is public, so HF_TOKEN is optional; it is forwarded for gated/private replacements and anonymous Hub rate limits. -e HF_TOKEN (no =value) passes through the host's already-exported variable without putting the token in the docker process argv, where ps//proc on a shared host could expose it. The proxy lines are required on hosts whose containers cannot reach the Hub directly. Docker ignores unset proxy variable names.

Cleanup after the trace has flushed:

docker stop vllm-xpu-prof && docker rm vllm-xpu-prof

Three artifacts written per profile (allow a few seconds for flush):

• <timestamp>-rank-<N>.<timestamp>.pt.trace.json.gz — per-worker Chrome trace. Drag-and-drop into https://ui.perfetto.dev; Perfetto handles .json.gz directly.

• <container>_<N>.async_llm.<timestamp>.pt.trace.json.gz — async engine trace (smaller).

• profiler_out_<N>.txt — read this first, sort-by-XPU-time table of top kernels. Most "what's slow" questions are answered by its top 5 rows:

  Name                     Self XPU %  XPU total   # of Calls
  gemm_kernel              28.50%      26.232ms    4753
  aten::mm                 21.79%      20.061ms    3577
  _C::fused_add_rms_norm   14.11%      12.988ms    2352
  _vllm_fa2_C::varlen_fwd   6.70%      17.979ms    1176
  _C::rotary_embedding      6.70%       6.172ms    1176
  Memcpy H2D                2.11%       1.945ms     412
  

Mode B — offline bench with --profile

docker run --rm \
   --entrypoint vllm \
    --device /dev/dri \
    -v /dev/dri/by-path:/dev/dri/by-path:ro \
    --group-add "$(getent group render | cut -d: -f3)" \
    --ipc=host \
    -e ZE_AFFINITY_MASK=0 \
   -e HTTP_PROXY -e HTTPS_PROXY -e NO_PROXY \
   -e http_proxy -e https_proxy -e no_proxy \
   -e HF_TOKEN \
    -v "$HOME/.cache/huggingface:/root/.cache/huggingface" \
    -v "$PWD/traces:/work/traces" \
   vllm/vllm-openai-xpu:latest \
   bench latency \
        --model Qwen/Qwen2.5-1.5B-Instruct \
        --dtype bfloat16 \
        --enforce-eager \
        --num-iters-warmup 2 --num-iters 5 \
        --profile \
        --profiler-config.profiler=torch \
        --profiler-config.torch_profiler_dir=/work/traces

What to look for

1. Step boundary markers — EngineCore.step spans. Time per step / batch size = TPOT contribution per request.
2. Attention path — confirm kernel name matches your --attention-backend. If you passed TRITON_ATTN but trace shows flash_attention_xpu, confirm the flag reached the engine:

   docker logs <name> 2>&1 | grep -E "attention_backend|Using.*backend"
   

3. Gaps between steps — wide white bands at boundaries mean scheduler / KV manager is on the critical path. Normal under --enforce-eager; tighter without it.
4. Quant kernel names — for --quantization fp8 expect fp8_gemm_w8a16; for awq expect int4_gemm_w4a8. Seeing int4_gemm_w4a16 with AWQ -> silent fall-through (vLLM #38064).
5. Per-rank spans — multi-GPU -tp N shows N timelines. Skew between ranks means a slow rank is gating throughput.

Common errors

• torch_profiler_dir is only applicable when profiler is set to 'torch' -> add --profiler-config.profiler=torch.
• /start_profile 404 -> server launched without profiler-config flags; restart.
• Empty trace -> too few requests inside the bracket; send 5–10+.
• Trace doesn't appear after /stop_profile -> flush is async; wait a few seconds, or check server logs.

Env vars

(vLLM profiler config is via CLI flags --profiler-config.*, not env vars.)

What this skill does NOT cover

• Pure-PyTorch profiling (no server) -> torch-xpu-profile.
• SYCL-kernel events / hardware metrics -> xpu-profile-unitrace.
• Fixing a hot kernel — out of scope.
• intel/llm-scaler-vllm images — out of scope; launch the profiling server from upstream vllm/vllm-openai-xpu:latest instead.

References

• vLLM profiling guide: https://docs.vllm.ai/en/latest/contributing/profiling/
• Perfetto trace viewer: https://ui.perfetto.dev