@@ -6,7 +6,10 @@ base:
   name: dsv41flash-fp4-mi355x-vllm-agentic
   model:
     path: hf:deepseek-ai/DeepSeek-V4.1-Flash
-    container: vllm/vllm-openai-rocm:nightly-rocm100-7f1a5398e9610d96c473931a26c0e12bbe0d0423
+    # Repinned to a nightly build that includes vllm-project/vllm#58671
+    # (paged MXFP4 sparse indexer), #58655 (mHC fused Triton seams) and
+    # #53492 (sparse MLA Gluon kernel).
+    container: vllm/vllm-openai-rocm:nightly-rocm100-ac9126e58aa7bbab1856ba6593ba4d5003fea516
     precision: fp4
   resources:
     gpu_type: mi355x
@@ -40,6 +43,12 @@ 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}'
+        # Auto-resolution picks 64 (min across backends); force 128, which
+        # the indexer and SparseMLA actually prefer.
+        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
@@ -54,6 +63,7 @@ base:
         VLLM_ROCM_USE_AITER: '1'
         VLLM_ROCM_USE_AITER_MOE: '1'
         AITER_TRITON_LOG_LEVEL: ERROR
+        VLLM_ROCM_USE_AITER_TRITON_SPARSE_MLA: 'True'
         # DeepseekV41ForCausalLM is not torch-compiled upstream; breakable
         # graphs keep FULL_AND_PIECEWISE capture working.
         VLLM_USE_BREAKABLE_CUDAGRAPH: '1'
@@ -70,11 +80,11 @@ base:
       WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
       AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'vllm:'
 
-# One variant per point. Engram tables move to pinned host memory at TP2 and
-# at c128, where resident tables leave too little KV. Graph capture covers
-# 2x CONC (at least 128 sequences) x (1 + 5 drafts), up to 2048 tokens. The
-# prefill chunk shrinks at TP2 from c64 and at c128 to buy KV room, and where
-# it drops below 1024 x 6 slots the scheduler is capped at the captured shape.
+# One variant per point (c128 excluded for both TP2 and TP4). Engram tables
+# move to pinned host memory at TP2, where resident tables leave too little
+# KV. Graph capture covers 2x CONC x (1 + 5 drafts). The prefill chunk shrinks
+# at TP2 from c64 to buy KV room, and where it drops below 1024 x 6 slots the
+# scheduler is capped at the captured shape.
 override_tp4_c1:
   roles:
     agg:
@@ -166,19 +176,6 @@ override_tp4_c64:
     env:
       CONC: '64'
 
-override_tp4_c128:
-  roles:
-    agg:
-      gpus: 4
-      args:
-        tensor-parallel-size: 4
-        engram-config: '{"cpu_offload":true}'
-        max-cudagraph-capture-size: 2048
-        max-num-batched-tokens: 8192
-  benchmark:
-    env:
-      CONC: '128'
-
 override_tp2_c1:
   roles:
     agg:
@@ -269,17 +266,3 @@ override_tp2_c64:
   benchmark:
     env:
       CONC: '64'
-
-override_tp2_c128:
-  roles:
-    agg:
-      gpus: 2
-      args:
-        tensor-parallel-size: 2
-        engram-config: '{"cpu_offload":true}'
-        max-cudagraph-capture-size: 2048
-        max-num-batched-tokens: 4096
-        max-num-seqs: 256
-  benchmark:
-    env:
-      CONC: '128'
@@ -1457,8 +1457,9 @@ 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:
-  # ROCm 10.0 nightly channel, shared with kimik3-fp4-mi355x-vllm-agentic-mtp.
-  image: vllm/vllm-openai-rocm:nightly-rocm100-29468dde8b515031dc6d4d9d06bf0a2fa0442098
+  # Repinned to a nightly build that includes vllm-project/vllm#58671,
+  # #58655 and #53492 (see agentic.yaml).
+  image: vllm/vllm-openai-rocm:nightly-rocm100-ac9126e58aa7bbab1856ba6593ba4d5003fea516
   model: deepseek-ai/DeepSeek-V4.1-Flash
   model-prefix: dsv41flash
   runner: cluster:mi355x-amds
@@ -1470,12 +1471,12 @@ dsv41flash-fp4-mi355x-vllm-agentic-dspark:
     - dram-utilization: 0.60
       search-space:
       # The recipe only changes the memory split where a point ran short of KV:
-      # TP=4 offloads the engram tables at c128, and the prefill chunk shrinks at
-      # TP=2 c64, TP=4 c128 and TP=2 c128. Every point is measured here rather
-      # than combined from the earlier run: the image move leaves concurrency
-      # 1-32 measured only on the superseded nightly-eed1f3d0 pin.
-      - { tp: 4, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/mi355x-fp4-mtp/agentic.yaml }
-      - { tp: 2, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/mi355x-fp4-mtp/agentic.yaml }
+      # the prefill chunk shrinks at TP=2 c64. c128 is excluded for both TP=2
+      # and TP=4. Every point is measured here rather than combined from the
+      # earlier run: the image move leaves concurrency 1-32 measured only on
+      # the superseded nightly-eed1f3d0 pin.
+      - { tp: 4, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/mi355x-fp4-mtp/agentic.yaml }
+      - { tp: 2, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/mi355x-fp4-mtp/agentic.yaml }
 
 # The golden acceptance curve remains the provisional GLM-5.2-derived curve
 # documented in infx/golden_al_distribution/glm5.3_mtp.yaml.
 
@@ -9168,3 +9168,10 @@
     - "Relevant ATOM changes in the range: V4 decode reuses the sparse prefill ASM (ROCm/ATOM#2271), greedy sampler picks via aiter.topk_select (ROCm/ATOM#2244), and FP8 block scales declared scale_fmt ue8m0 are now stored as E8M0 on gfx950 by default (ROCm/ATOM#2419; previously FP32 unless ATOM_FP8_BLOCKSCALE_USE_E8M0_SCALE=1). The upstream DeepSeek-V4 recipes are unchanged across the range, and every recipe flag and choice (all2all-backend rccl, dp-load-balance least_tokens, moe-backend standard) remains valid."
     - "No data-type or precision change to the DeepSeek-V4-Pro-0813 DSpark draft: no online quantization is configured, so it keeps its checkpoint precision; the E8M0 scale storage represents the checkpoint's power-of-two block scales exactly. kv-cache-dtype and index-cache-dtype touch cache storage only."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3605
+
+- config-keys:
+    - dsv41flash-fp4-mi355x-vllm-agentic-dspark
+  description:
+    - "Repin to vllm/vllm-openai-rocm:nightly-rocm100-ac9126e58aa7bbab1856ba6593ba4d5003fea516, which includes vllm-project/vllm#58671, #58655 and #53492. Re-add --attention-config '{\"indexer_kv_dtype\":\"mxfp4\",\"indexer_sparse_logits\":true}' and --block-size 128 to enable #58671's ROCm paged MXFP4 sparse-logits indexer, 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. Drops c128 from both TP2 and TP4 conc-lists for this recipe."
+    - "固定到 vllm/vllm-openai-rocm:nightly-rocm100-ac9126e58aa7bbab1856ba6593ba4d5003fea516，该镜像包含 vllm-project/vllm#58671、#58655 与 #53492。重新添加 --attention-config '{\"indexer_kv_dtype\":\"mxfp4\",\"indexer_sparse_logits\":true}' 与 --block-size 128，以启用 #58671 的 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。此配方的 TP2 与 TP4 conc-list 均移除 c128。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3571