[Bugfix] Fix Triton stream capture error on A100 in GDN attention with MTP speculative decoding

[jacob-crux]

Purpose

Fix a Triton "operation not permitted when stream is capturing" error during FULL CUDA graph capture for models using GDN (Gated Delta Net) attention with MTP speculative decoding.

Reproduced with Qwen/Qwen3.5-35B-A3B-GPTQ-Int4 (Qwen3.5 MoE architecture, GPTQ Int4 quantized) on NVIDIA A100-SXM4-80GB.

When --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}' is enabled, the engine fails to start during CUDA graph capture with:

RuntimeError: Triton Error [CUDA]: operation not permitted when stream is capturing
  at vllm/model_executor/layers/mamba/ops/causal_conv1d.py → _causal_conv1d_update_kernel
RuntimeError: Engine core initialization failed.

The spec-decode Triton kernels (causal_conv1d_update with IS_SPEC_DECODING=True, fused_sigmoid_gating_delta_rule_update, fused_gdn_gating, l2norm_fwd_kernel2, etc.) attempt to JIT-compile during the actual CUDA graph capture, calling Triton's load_binary while a CUDA stream capture is active — which CUDA forbids.

Root cause: During _warmup_and_capture, the warmup _dummy_run calls do not pass is_graph_capturing=True. As a result, attention metadata builders use build() instead of build_for_cudagraph_capture(). For GDN attention, build() receives dummy num_decode_draft_tokens=0, so spec_sequence_masks=None and the non-spec decode path is taken during warmup. The spec-decode Triton kernels are therefore never JIT-compiled during warmup. When the actual capture then runs (which always uses is_graph_capturing=True), build_for_cudagraph_capture() auto-generates spec-decode metadata from query_start_loc and triggers the spec-decode code path — at which point Triton tries to JIT-compile the kernels for the first time inside an active stream capture and crashes.

Fix: Pass is_graph_capturing=force_attention to the warmup _dummy_run so that FULL-mode warmup exercises the same build_for_cudagraph_capture() path as the actual capture, ensuring all required Triton kernels are pre-compiled before capture begins. Only affects FULL cudagraph warmup; PIECEWISE mode is unchanged because GDN attention runs as a splitting op outside the captured graph there.

Test Plan

• Server startup with FULL CUDA graph + MTP — Verify the engine initializes and serves requests with qwen3_next_mtp speculative decoding enabled:

# Start server
vllm serve Qwen/Qwen3.5-35B-A3B-GPTQ-Int4 \
  --gpu-memory-utilization 0.85 \
  --language-model-only \
  --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}' \
  --port 8000

# Smoke test
curl -s http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3.5-35B-A3B-GPTQ-Int4",
    "messages": [{"role":"user","content":"What is the capital of South Korea?"}],
    "max_tokens": 100
  }'

Test Result

• Environment: NVIDIA A100-SXM4-80GB x 1

Before fix — Engine fails to start during CUDA graph capture:

(EngineCore_DP0) ERROR [core.py:1108] EngineCore failed to start.
(EngineCore_DP0) ERROR [core.py:1108] RuntimeError: Triton Error [CUDA]: operation not permitted when stream is capturing
(EngineCore_DP0) ERROR [core.py:1108]   at vllm/model_executor/layers/mamba/ops/causal_conv1d.py
(EngineCore_DP0) ERROR [core.py:1108]      → _causal_conv1d_update_kernel
RuntimeError: Engine core initialization failed.

After fix — Engine starts successfully and serves requests normally with MTP speculative decoding enabled.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Code Review

This pull request updates the _warmup_and_capture method in vllm/v1/worker/gpu_model_runner.py to pass the is_graph_capturing argument, set to the value of force_attention, during the model warmup phase. I have no feedback to provide as there are no review comments to evaluate.

[JaheimLee]

any update?