Update DSV4 Pro MI355X nightly and enable monokernel / 更新 DSV4 Pro MI355X nightly 镜像并启用 monokernel

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/dsv4/vllm/mi355x-fp4-mtp/agentic.yaml

@@ -39,6 +39,7 @@ base:
         data-parallel-size: 1
         gpu-memory-utilization: 0.86
         moe-backend: aiter
+        monokernel: true
         compilation-config: '{"mode":3,"cudagraph_mode":"FULL_AND_PIECEWISE"}'
         # Six-token DSpark with probabilistic drafting. Throughput runs replace
         # verification with the golden acceptance length.
@@ -167,6 +168,8 @@ override_dep8_c64:
       args:
         tensor-parallel-size: 1
         data-parallel-size: 8
+        # Pro's monokernel requires TP8 without data parallelism.
+        monokernel: false
         prefill-schedule-interval: 8
         long-prefill-token-threshold: 16384
         max-num-seqs: 64
@@ -189,6 +192,7 @@ override_dep8_c128:
       args:
         tensor-parallel-size: 1
         data-parallel-size: 8
+        monokernel: false
         prefill-schedule-interval: 8
         long-prefill-token-threshold: 16384
         max-num-seqs: 128
@@ -211,6 +215,7 @@ override_dep8_c192:
       args:
         tensor-parallel-size: 1
         data-parallel-size: 8
+        monokernel: false
         prefill-schedule-interval: 8
         long-prefill-token-threshold: 16384
         max-num-seqs: 192

inferencex-e2e/configs/amd-master.yaml

@@ -988,7 +988,7 @@ qwen3.5-fp8-mi325x-sglang-mtp:
 # DeepSeek-V4-Pro-0813 FP4 AgentX on one MI355X node using vLLM DSpark K6.
 # Keep the historical config key for changelog compatibility.
 dsv4-fp4-mi355x-vllm-agentic-mtp:
-  image: vllm/vllm-openai-rocm:nightly-e9757321527ca1ecd514c07c1418dd2c53da3d19
+  image: vllm/vllm-openai-rocm:nightly-e4a5b5537344b103ed10de7764378ef09604ef3d
   model: deepseek-ai/DeepSeek-V4-Pro-0813
   model-prefix: dsv4
   runner: cluster:mi355x-amds

inferencex-e2e/perf-changelog.yaml

@@ -11604,3 +11604,15 @@
     - "Upstream SGLang: mono-kernel specs are registered by architecture and model type (sgl-project/sglang#46036) on the new plugin kernel backend (#45389); the DSpark draft MoE is kept out of the target's expert placement (#45299); attn_sink is refreshed after weight loading (#41630); the draft and final mHC streams are combined from BF16 without an FP32 copy (#45964, #45940)."
     - "Draft model data type: no change. The DSpark draft (mtp.0-2 in DeepSeek-V4-Pro-0813) runs as shipped, like the target layers: FP8 E4M3 attention, shared-expert and main_proj weights with 128x128 UE8M0 block scales, MXFP4 routed experts, BF16 norms, router gate and Markov head, and wo_a dequantized to BF16 at load. Its KV stays a BF16 ring under SGLANG_DSV4_UNIFIED_KV_FP8, and no sgl-amd-mono spec covers the DeepSeek-V4 DSpark draft, so the draft keeps SGLang's forward."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4596
+
+- config-keys:
+    - dsv4-fp4-mi355x-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Image: vllm/vllm-openai-rocm:nightly-e9757321527ca1ecd514c07c1418dd2c53da3d19 -> nightly-e4a5b5537344b103ed10de7764378ef09604ef3d (vLLM commit of 2026-10-25, 1953 commits ahead)."
+    - "Points: unchanged (TP8 c1/4/8/16/32/40/48/56/64; DP-attention DEP8 c64/128/192 behind vllm-router)."
+    - "TP8 points: add --monokernel. On the target's DeepSeek-V4-Pro layers it runs a whole decoder layer as two persistent FlyDSL launches with the TP all-reduces inside, for decode-only steps of at most 64 rows (up to 9 requests at 7 DSpark tokens each) in eager or FULL CUDA graph mode; prefill, mixed and wider steps keep the standard path. DEP8 points set monokernel: false, because it requires TP8 with no data or expert parallelism and refuses other deployments at startup."
+    - "Upstream vLLM: the DeepSeek-V4-Pro MI355X mono decode layer (vllm-project/vllm#61482); the CSA/HCA default chain is enqueued first at small batch sizes (#61378); the ROCm sparse prefill uses the shared prefill chunk plan (#58405); the C128 compressor skip is no longer baked into CUDA graphs (#59709)."
+    - "Draft model data type: no change. The DSpark draft (mtp.0-2 of DeepSeek-V4-Pro-0813 at revision 72e1d3230f6c) runs as shipped with the checkpoint's FP8 quant config: FP8 E4M3 attention, shared-expert and main_proj weights with 128x128 UE8M0 block scales, MXFP4 routed experts, BF16 norms, router gate, confidence and Markov heads. Its KV cache inherits the target's fp8_ds_mla, and --monokernel skips draft layers, so the draft keeps vLLM's standard forward."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4526