[DeepSeek][ROCm] V4.1-Flash: enable sparse-MLA Gluon kernel + MXFP4 sparse indexer on MI355X

[Fangzhou-Ai]

Purpose

Add to the MI355X single_node_tp override in DeepSeek-V4.1-Flash.yaml:

• VLLM_ROCM_USE_AITER_TRITON_SPARSE_MLA: "1" — forces
  vllm-project/vllm#53492's
  gfx950-only AITER Gluon sparse-MLA kernel. That PR's own test plan
  validates it on DSv4.1-Flash at TP4/MI355X: faster prefill/decode
  TTFT/TPOT than the default Triton sparse-MLA path, unchanged GSM8K
  accuracy.
• --attention-config '{"indexer_kv_dtype":"mxfp4","indexer_sparse_logits":true}'
  — enables vllm-project/vllm#58671's
  ROCm paged MXFP4 sparse-logits indexer kernel, replacing the dense fp8
  indexer path.
• --block-size 128 — the fp4 indexer and SparseMLA backend prefer this
  once active; vLLM's block-size auto-resolution otherwise falls back to
  64, the smallest size any of this model's 4+ attention backends merely
  supports.

Both vLLM PRs are merged and gfx950-scoped. This mirrors
SemiAnalysisAI/InferenceX#3555
(sparse-MLA env var) and
SemiAnalysisAI/InferenceX#3571
(mxfp4 indexer + block-size), so the recipe and the InferenceX benchmark
config agree on MI355X.

Scoped to mi355x only — the shared hardware_overrides.amd block also
covers MI325X (gfx942), which doesn't have this kernel support, and this
file has no MI350X-specific override today; I have no MI350X validation of
my own to add one.

Duplicate-work check

Searched open/closed PRs and issues on this repo for
VLLM_ROCM_USE_AITER_TRITON_SPARSE_MLA, indexer_kv_dtype, mxfp4 indexer,
block-size 128, and DeepSeek-V4.1-Flash mi355x. No existing PR adds these
to this recipe. PR #962 (same author, open) touches the same file but a
disjoint, older section (image pin + --moe-backend) predating the current
file structure.

Test Plan

node scripts/build-recipes-api.mjs

Test Result

✓ JSON API: 195 models (169 with recommended_command, 819 default-hw
  alternatives, 1697 per-hw renderings), 227 promoted variants, 10
  strategies, 2 kv-store deployments, 2 platforms (7 variant collisions
  skipped)

No new warnings or errors versus a clean main run of the same script.

AI assistance disclosure

Investigated with Claude Sonnet 5 (Cursor): confirmed vLLM PRs #53492 and
#58671 are merged and gfx950-scoped, pulled their DSv4.1-Flash/MI355X
validation numbers, cross-checked against the two InferenceX PRs, and ran
this repo's own YAML validator. I reviewed the diff and am responsible for
it.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
vllm-recipes	Ready	Preview	Sep 30, 2026 12:15am UTC

[gemini-code-assist]

Code Review

This pull request updates the metadata date and adds the VLLM_ROCM_USE_AITER_TRITON_SPARSE_MLA environment variable to enable ROCm sparse-MLA kernel optimizations. The reviewer suggests moving this environment variable from the strategy-specific overrides to the top-level hardware overrides block to ensure the optimization is applied across all compatible serving strategies on MI355X hardware.

[gemini-code-assist]

The environment variable VLLM_ROCM_USE_AITER_TRITON_SPARSE_MLA: "1" is a hardware-specific optimization for gfx950 (MI355X) and is independent of the serving strategy. Currently, it is defined under strategy_overrides.single_node_tp.hardware_overrides.mi355x, which means it will only be applied when using the single_node_tp strategy. To ensure this optimization is applied across all compatible strategies (such as multi_node_tp or pd_cluster) when running on mi355x, it should be moved to the top-level hardware_overrides.mi355x block instead.