Re-add vllm-project/vllm#58671's ROCm paged MXFP4 sparse indexer

Stacked on #3555. Re-adds the --attention-config flag enabling #58671's
ROCm paged MXFP4 sparse-logits indexer and the --block-size 128 workaround
its active fp4 indexer needs (vLLM's block-size auto-resolution otherwise
picks 64 across this model's 4+ attention backends). Draft until #58671
merges upstream and lands in a nightly image, so #3555 can run now and this
PR can re-sweep immediately once it's available.

Signed-off-by: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

Author: Fangzhou-Ai

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/mi355x-fp4-mtp/agentic.yaml

@@ -6,9 +6,12 @@ base:
   name: dsv41flash-fp4-mi355x-vllm-agentic
   model:
     path: hf:deepseek-ai/DeepSeek-V4.1-Flash
-    # Repinned to a nightly build that includes vllm-project/vllm#58655
-    # (mHC fused Triton seams) and #53492 (sparse MLA Gluon kernel).
-    container: vllm/vllm-openai-rocm:nightly-rocm100-36768d1bfd39094681cdbc8cb37d4b31c0729c89
+    # TODO: repin once a nightly image is released that includes
+    # vllm-project/vllm#58671 (paged MXFP4 sparse indexer, merged upstream),
+    # in addition to #58655 (mHC fused Triton seams) and #53492 (sparse MLA
+    # Gluon kernel) already carried by the previous pin
+    # (nightly-rocm100-36768d1b).
+    container: vllm/vllm-openai-rocm:nightly-rocm100-TBD
     precision: fp4
   resources:
     gpu_type: mi355x
@@ -42,6 +45,16 @@ base:
         # aiter_triton_mxfp4_bf16 would force the W4A16 Triton kernel.
         moe-backend: aiter
         gpu-memory-utilization: 0.9
+        # Enables vllm-project/vllm#58671's ROCm paged MXFP4 sparse-logits
+        # indexer kernel instead of the dense fp8 indexer path.
+        attention-config: '{"indexer_kv_dtype":"mxfp4","indexer_sparse_logits":true}'
+        # With 4+ attention backends active (SparseMLA, SparseSWA, the MXFP4
+        # indexer, Compressor), vLLM's block-size auto-resolution picks the
+        # smallest size every backend merely supports (64), not the 128 the
+        # indexer and SparseMLA backends actually prefer for this fp4 indexer
+        # path. Force 128 explicitly; see vllm-project/vllm's
+        # Platform._preferred_block_size_for_backends.
+        block-size: 128
         # Five-token DSpark with probabilistic drafting. Throughput runs replace
         # block rejection with the golden acceptance length. Adaptive
         # verification trims requests on device, which the ROCm DeepSeek-V4

inferencex-e2e/configs/amd-master.yaml

@@ -1391,9 +1391,10 @@ dsv4-fp4-mi355x-sglang-agentic-mtp:
 # Two GPUs per server doubles the servers per node and is the layout that
 # decides whether DSv4.1-Flash is throughput- or interactivity-bound here.
 dsv41flash-fp4-mi355x-vllm-agentic-dspark:
-  # Repinned to a nightly build that includes vllm-project/vllm#58655 and
-  # #53492 (see agentic.yaml).
-  image: vllm/vllm-openai-rocm:nightly-rocm100-36768d1bfd39094681cdbc8cb37d4b31c0729c89
+  # TODO: repin once a nightly image is released that includes
+  # vllm-project/vllm#58671 (already merged upstream), in addition to
+  # #58655 and #53492 (see agentic.yaml).
+  image: vllm/vllm-openai-rocm:nightly-rocm100-TBD
   model: deepseek-ai/DeepSeek-V4.1-Flash
   model-prefix: dsv41flash
   runner: cluster:mi355x-amds

inferencex-e2e/perf-changelog.yaml

@@ -9127,3 +9127,10 @@
     - "Force VLLM_ROCM_USE_AITER_TRITON_SPARSE_MLA=True to enable vllm-project/vllm#53492's gfx950 Gluon sparse-MLA kernel. Repin the recipe and master config to vllm/vllm-openai-rocm:nightly-rocm100-36768d1bfd39094681cdbc8cb37d4b31c0729c89, which includes #53492 and vllm-project/vllm#58655 (fused mHC Triton kernel). vllm-project/vllm#58208 (DSA candidate topk) was reverted by vllm-project/vllm#59125 before this build, and vllm-project/vllm#58671's MXFP4 sparse indexer landed after it; the indexer is left to #3571."
     - "设置 VLLM_ROCM_USE_AITER_TRITON_SPARSE_MLA=True，启用 vllm-project/vllm#53492 的 gfx950 Gluon sparse-MLA kernel。将配方与主配置固定到 vllm/vllm-openai-rocm:nightly-rocm100-36768d1bfd39094681cdbc8cb37d4b31c0729c89，该镜像包含 #53492 与 vllm-project/vllm#58655（融合 mHC Triton kernel）。vllm-project/vllm#58208（DSA 候选块 topk）在该构建之前已被 vllm-project/vllm#59125 回滚；vllm-project/vllm#58671 的 MXFP4 稀疏 indexer 在该构建之后才合入，留待 #3571。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3555
+
+- config-keys:
+    - dsv41flash-fp4-mi355x-vllm-agentic-dspark
+  description:
+    - "Stacked on #3555. Re-add --attention-config '{\"indexer_kv_dtype\":\"mxfp4\",\"indexer_sparse_logits\":true}' and --block-size 128 to enable vllm-project/vllm#58671's ROCm paged MXFP4 sparse-logits indexer once it merges upstream, replacing the dense fp8 indexer path. A live A/B test (TP2 c16, matched 900s window, #58208 reverted via #59125 so the dense fallback doesn't crash) measured #58671 giving +14.7/+14.9% p50/p90 interactivity and -9.5/-10.8% p50/p90 e2e latency over the dense path, with throughput/GPU unchanged. Draft until #58671 merges and lands in a nightly image; kept separate from #3555 so that PR can run now and this one can re-sweep immediately once #58671 is available."
+    - "基于 #3555 之上。重新添加 --attention-config '{\"indexer_kv_dtype\":\"mxfp4\",\"indexer_sparse_logits\":true}' 与 --block-size 128，以便在 vllm-project/vllm#58671 合并上游后启用其 ROCm 分页 MXFP4 稀疏 logits indexer，取代稠密 fp8 indexer 路径。一次真实的 A/B 测试（TP2、并发 16，匹配 900 秒窗口，并通过 #59125 回滚 #58208 以避免稠密回退路径崩溃）测得 #58671 相较稠密路径带来 interactivity p50/p90 提升 14.7%/14.9%，e2e 延迟 p50/p90 降低 9.5%/10.8%，而 throughput/GPU 无变化。本 PR 保持 draft 状态，直到 #58671 合并并随 nightly 镜像发布；与 #3555 分开提交，以便该 PR 可先行运行，本 PR 待 #58671 可用后立即重新扫描。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3571