[AMD][MI35X] 0526 DSV4 (#1568)

Author: 1am9trash

.github/configs/amd-master.yaml

@@ -1813,7 +1813,7 @@ dsr1-fp4-mi355x-sglang-disagg-mtp:
 # image tag, so bumping sglang is just an image tag bump here. Sweeps
 # DP-attention on/off and EP=8.
 dsv4-fp4-mi355x-sglang:
-  image: rocm/sgl-dev:rocm720-mi35x-8c3b5aa-20260521-DSv4
+  image: rocm/sgl-dev:rocm720-mi35x-f96ac98-20260526-DSv4
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: mi355x
@@ -1825,12 +1825,12 @@ dsv4-fp4-mi355x-sglang:
     - isl: 1024
       osl: 1024
       search-space:
-      - { tp: 8, dp-attn: true, conc-start: 64, conc-end: 1024 }
+      - { tp: 8, dp-attn: true, conc-start: 64, conc-end: 2048 }
       - { tp: 8, dp-attn: false, conc-start: 1 , conc-end: 32 }
     - isl: 8192
       osl: 1024
       search-space:
-      - { tp: 8, dp-attn: true, conc-start: 64, conc-end: 512 }
+      - { tp: 8, dp-attn: true, conc-start: 64, conc-end: 2048 }
       - { tp: 8, dp-attn: false, conc-start: 1, conc-end: 32 }
 
 # DSv4 on MI355X via vLLM, using the official vllm/vllm-openai-rocm

benchmarks/single_node/dsv4_fp4_mi355x_sglang.sh

@@ -77,7 +77,7 @@ export SGLANG_FORCE_TRITON_MOE_FP8=0
 export SGLANG_HACK_FLASHMLA_BACKEND=triton
 export SGLANG_OPT_USE_TILELANG_INDEXER=true
 export SGLANG_OPT_USE_TRITON_SWA_PREPARE=true
-export AITER_BF16_FP8_MOE_BOUND=1
+export AITER_BF16_FP8_MOE_BOUND=0
 export SGLANG_OPT_FUSE_WQA_WKV=true
 export SGLANG_OPT_USE_FUSED_PAGED_COMPRESS=true
 export SGLANG_OPT_USE_MULTI_STREAM_OVERLAP=0
@@ -116,6 +116,8 @@ python3 -m sglang.launch_server \
     --disable-radix-cache \
     --attention-backend compressed \
     --max-running-requests ${CONC} \
+    --mem-fraction-static 0.90 \
+    --swa-full-tokens-ratio 0.15 \
     --page-size 256 \
     --context-length $MAX_MODEL_LEN \
     --chunked-prefill-size 8192 \

perf-changelog.yaml

@@ -3146,6 +3146,13 @@
     - "Set vLLM serving knobs in benchmarks/single_node/minimaxm2.5_fp8_h200.sh: generated benchmark max-model-len, previous eval max-model-len handling, fp8 KV cache, FlashInfer attention/autotune, Triton MoE, and MiniMax QK norm fusion"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1354
 
+- config-keys:
+    - dsv4-fp4-mi355x-sglang
+  description:
+    - "Bump image to rocm/sgl-dev:rocm720-mi35x-f96ac98-20260526-DSv4"
+    - "Add args to avoid kvcache pool full issue on high conc"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1568
+
 - config-keys:
     - qwen3.5-fp8-h200-sglang
     - dsr1-fp8-mi355x-sglang