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

@@ -11480,3 +11480,10 @@
     - "Update the Muse Spark Open B300 curve to lmsysorg/sglang:nightly-dev-20261021-c1c287a4 and set SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION=0 at every concurrency; run TP8/EP8 at concurrency 1, 4, 8, 10, 14, 16, 20, 24 and TP4/EP4 HiCache at concurrency 16, 28, 32, 40, 44, 48."
     - "Use shortest-prefill-first at TP8 concurrency 10-24, stream interval 20 at TP8 concurrency 4, 20, 24 and TP4 concurrency 28-48, SGLANG_TRTLLM_MHA_DECODE_SEQ_LEN_SPLITS=4 at TP4 concurrency 40-48, and at concurrency 44 and 48 request admission 24 and 26 with FCFS, no prefill-decode interval, HiCache write_back and SWA/full ratio 0.1."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4417
+
+- config-keys:
+    - dsv4-fp4-mi355x-vllm-agentic-mtp
+  description:
+    - "Update DeepSeek-V4-Pro-0813 on MI355X to vllm/vllm-openai-rocm:nightly-e4a5b5537344b103ed10de7764378ef09604ef3d and enable --monokernel for TP8 points; keep it disabled for data-parallel points."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4526
+  scenario-type: [agentic-coding]