Context

#46849 added the infrastructure to fuse last N-1 draft steps into a single cudagraph for models with attention backends that support it. Currently DeepSeek V4 satisfies the conditions to enable this, when using MTP > 1.

This PR

It turns out we can remove the first attention metadata build before drafting the last N-1 speculative tokens in a fused loop. In this PR, I make the necessary changes to DeepseekSparseSWAMetadataBuilder.update_draft_decode_metadata and refactor AutoRegressiveSpeculator so that the first draft metadata build can be safely skipped during FULL cudagraph replay.

With these changes, DSV4 + MTP no longer rebuilds metadata during proposal, resulting in a correctness-preserving removal of about 1ms of host work per decode step. No significant throughput improvement was observed from these changes.

Profile

Captured while serving DeepSeek-V4-Flash + MTP=3 on 4xGB200:
Before

After

In the decode step above, propose's CPU time shrinks by an entire millisecond.

Benchmarks / Evals

Setup. DeepSeek-V4-Flash, MTP with num_speculative_tokens=3, 4x GB200 (TP4), --block-size 256 --kv-cache-dtype fp8 --moe-backend deep_gemm --no-enable-prefix-caching, VLLM_USE_V2_MODEL_RUNNER=1. Baseline is main at 34b069080d; PR is cfb93dedca. One server per arm; trace, benchmark and eval share it.

Speed-Bench 2K/2K (throughput_16k, low_entropy, temperature 0)

Prompts scaled 8x concurrency (min 16), seed 0. Greedy decoding keeps acceptance deterministic, so ITL p50 is the step-time metric.

Step time is unchanged at every concurrency. The ~1 ms of host work removed per step is not on the critical path for this model at TP4; throughput deltas track the small acceptance-length differences and are within noise.

AIME25 (MQA, 2 epochs x 30 problems, thinking on, temperature 1.0, top-p 0.95, max 98304 tokens, 64 concurrent)

Mean accuracy identical. pass@2 differs by one problem in 30, which is expected sampling variance at temperature 1.