[AMD][MI35X] 0618 DSV4 sglang

.github/configs/amd-master.yaml

@@ -1874,7 +1874,7 @@ dsr1-fp4-mi355x-sglang-disagg-8k1k-mtp:
           - "DECODE_MTP_SIZE=1"
 
 dsv4-fp4-mi355x-sglang:
-  image: lmsysorg/sglang-rocm:v0.5.13-rocm720-mi35x-20260612
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260618
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: mi355x

benchmarks/single_node/fixed_seq_len/dsv4_fp4_mi355x_sglang.sh

@@ -29,6 +29,7 @@ export SGLANG_DSV4_REASONING_EFFORT=max
 export SGLANG_OPT_DEEPGEMM_HC_PRENORM=false
 export SGLANG_USE_AITER=1
 export SGLANG_USE_ROCM700A=0
+export SGLANG_DP_USE_GATHERV=1
 export SGLANG_OPT_USE_FUSED_COMPRESS=true
 export SGLANG_HACK_FLASHMLA_BACKEND=unified_kv_triton
 export SGLANG_OPT_FP8_WO_A_GEMM=false
@@ -60,9 +61,9 @@ start_gpu_monitor
 PARALLEL_ARGS=(
     --tensor-parallel-size "$TP"
 )
-CHUNKED_PREFILL_SIZE=8192
+CHUNKED_PREFILL_SIZE=$ISL
 if [ "${DP_ATTENTION}" = "true" ]; then
-    CHUNKED_PREFILL_SIZE=$((8192 * TP))
+    CHUNKED_PREFILL_SIZE=$((ISL * TP))
     PARALLEL_ARGS+=(
         --dp "$TP"
         --enable-dp-attention

perf-changelog.yaml

@@ -4014,3 +4014,10 @@
     - "1k/1k: 1p1d-dep4-dep8 (conc 4096,12288), 1p4d-dep4-tp8 (conc 4-128), 1p1d-dep4-dep16 (conc 4096,6144)"
     - "8k/1k: 1p4d-dep4-tep4 (conc 128), 1p4d-dep4-tp8 (conc 4-256), 3p1d-dep4-dep16 (conc 1024), 6p1d-dep4-dep16 (conc 3072), 8p1d-dep4-dep16 (conc 6144)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1862
+
+- config-keys:
+    - dsv4-fp4-mi355x-sglang
+  description:
+    - "Bump image to lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260618."
+    - "Enable SGLANG_DP_USE_GATHERV=1. Change to use all_gatherv + reduce_scatterv for tp + dp config."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1824