[ROCm] Enable fused AR draft metadata updates for DeepSeek V4

[tpopp]

Builds on vLLM PR #46849, which restores fused multi-step draft decode graphs for autoregressive speculative decoding when every attention backend can update step-dependent draft metadata in place. That PR left DeepSeek V4 ROCm sparse SWA disabled because the AMD path also materializes ROCm-specific ragged SWA indices and indptrs.

Enable the ROCm DeepSeek V4 AITER SparseSWA metadata builder for fused draft decode by refreshing its ragged SWA representation after the generic dense SWA update. The refresh writes directly into the persistent metadata buffers returned by build(), avoiding temporary ragged flat/indptr allocations so captured HIPGraph replay keeps stable tensor storage addresses. Add focused tests for the direct-to-buffer helper and in-place metadata update.

Validation used TP4 on MI355X against patched and control images built from vllm/main@c4e9692. Runtime logs confirmed the patched image used the fused speculator graph path, while control logged fallback because DEEPSEEK_SPARSE_SWA did not support fused multi-step draft decode metadata updates. Smoke, GSM8K, c16, and c128 runs completed with no HIPGraph/cudagraph replay errors, illegal memory access, stale metadata symptoms, or failed benchmark requests.

Serve command shape:
VLLM_ROCM_USE_AITER=1 SAFETENSORS_FAST_GPU=1
vllm serve deepseek-ai/DeepSeek-V4-Flash
--tensor-parallel-size 4
--attention_backend ROCM_AITER_UNIFIED_ATTN
--compilation-config '{"mode":3,"cudagraph_mode":"FULL_DECODE_ONLY"}'
--speculative-config '{"method":"mtp","num_speculative_tokens":3,"attention_backend":"ROCM_AITER_UNIFIED_ATTN"}'
--kv-cache-dtype fp8
--distributed-executor-backend mp
--max_model_len 4096
--max-num-batched-tokens 8192
--max_num_seqs 256
--tokenizer-mode deepseek_v4
--tool-call-parser deepseek_v4
--trust-remote-code

Accuracy command shape:
lm_eval --model local-completions
--tasks gsm8k
--num_fewshot 8
--model_args model=deepseek-ai/DeepSeek-V4-Flash,base_url=http://localhost:/v1/completions,num_concurrent=16,tokenized_requests=False

Accuracy results, full GSM8K 8-shot, no limit:
patched flexible/strict: 0.9492 / 0.9500
control flexible/strict: 0.9416 / 0.9424

Benchmark command shapes:
vllm bench serve --model deepseek-ai/DeepSeek-V4-Flash
--dataset-name random
--random-input-len 512
--random-output-len 1024
--temperature 0
--ignore-eos
--max-concurrency 16
--num-prompts 160
--num-warmups 16

vllm bench serve --model deepseek-ai/DeepSeek-V4-Flash
--dataset-name random
--random-input-len 512
--random-output-len 1024
--temperature 0
--ignore-eos
--max-concurrency 128
--num-prompts 1024
--num-warmups 128

Performance results, output throughput:
c16 patched/control: 2000.58 / 1958.82 tok/s (+2.13%)
c128 patched/control: 7852.55 / 7751.43 tok/s (+1.30%)

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[maeehart]

pre-commit failed on ruff in vllm/platforms/cpu.py. Lines 184 and 190 are E501 (90 and 98 cols, limit 88), and ruff format wants to rewrite that same file. Can you wrap those two lines, run ruff format on vllm/platforms/cpu.py, and push?

[tpopp]

Branch is updated and pre-commit is fixed.