[AMD][DSV4] Add a DP-attention arm and refresh the MI355X SGLang AgentX key

benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh

@@ -74,6 +74,7 @@ export SGLANG_USE_ROCM700A=0
 export SGLANG_HACK_FLASHMLA_BACKEND=unified_kv_triton
 export AITER_BF16_FP8_MOE_BOUND=0
 export TORCH_BLAS_PREFER_HIPBLASLT=1
+export HSA_NO_SCRATCH_RECLAIM=0
 # aiter batched GEMM for the absorbed MLA projections, carried by the v0.5.18
 # image and off by default in environ.py.
 export SGLANG_OPT_USE_AITER_BATCHED_GEMM=1
@@ -140,12 +141,11 @@ else
     echo "Error: unsupported TP '$TP' (expected: 4 or 8)" >&2
     exit 1
 fi
-# MTP adds a draft KV pool and extra graph captures on top of the spec-none
-# footprint, which ran at 0.90. 0.89 recovers most of that: the DSv4 compressor
-# state pools are sized from the full-attention pool and allocated after it,
-# outside this budget, so the remainder has to stay large enough to cover them.
-MEM_FRACTION_STATIC=0.89
+MEM_FRACTION_STATIC="${MEM_FRACTION_STATIC:-0.86}"
 PARALLEL_ARGS=(--tensor-parallel-size "$TP")
+SHARED_EXPERTS_ARGS=(--enforce-shared-experts-fusion)
+SWA_FULL_TOKENS_RATIO="${SWA_FULL_TOKENS_RATIO:-0.10}"
+export GPU_MAX_HW_QUEUES="${GPU_MAX_HW_QUEUES:-2}"
 if [ "$DP_ATTENTION" = "true" ]; then
     USE_SGLANG_ROUTER=true
     export AIPERF_HTTP_X_SMG_ROUTING_KEY_FROM_CORRELATION_ID=true
@@ -157,14 +157,21 @@ if [ "$DP_ATTENTION" = "true" ]; then
     export SGLANG_DP_SHARED_EXPERT_LOCAL=1
     export SGLANG_DP_USE_GATHERV=1
     export SGLANG_DP_USE_REDUCE_SCATTER=1
-    export GPU_MAX_HW_QUEUES=2
+    export GPU_MAX_HW_QUEUES="${GPU_MAX_HW_QUEUES_DP:-5}"
+    SHARED_EXPERTS_ARGS=(--disable-shared-experts-fusion)
+    SWA_FULL_TOKENS_RATIO="${SWA_FULL_TOKENS_RATIO_DP:-0.15}"
 
     # Chunked prefill is a whole-engine budget, so widen it by the DP degree.
-    CHUNKED_PREFILL_SIZE=$((8192 * TP))
+    CHUNKED_PREFILL_SIZE=$((CHUNKED_PREFILL_SIZE * TP))
     PARALLEL_ARGS+=(
         --dp "$TP"
         --enable-dp-attention
         --enable-prefill-delayer
+        --enable-two-batch-overlap
+        --enable-dp-attention-local-control-broadcast
+        --tokenizer-worker-num "$TP"
+        --stream-interval 20
+        --prefill-decode-interval 10
     )
 fi
 
@@ -223,10 +230,11 @@ SGLANG_CMD=(
     --trust-remote-code
     "${PARALLEL_ARGS[@]}"
     --attention-backend dsv4
+    --enable-deepseek-v4-fp4-indexer
     --page-size 256
-    --swa-full-tokens-ratio 0.10
+    --swa-full-tokens-ratio "$SWA_FULL_TOKENS_RATIO"
     --kv-cache-dtype fp8_e4m3
-    --enforce-shared-experts-fusion
+    "${SHARED_EXPERTS_ARGS[@]}"
     --tool-call-parser deepseekv4
     --reasoning-parser deepseek-v4
     --chunked-prefill-size "$CHUNKED_PREFILL_SIZE"

configs/amd-master.yaml

@@ -1742,7 +1742,7 @@ glm5.2-fp4-mi355x-atom-agentic-mtp:
 
 
 dsv4-fp4-mi355x-sglang-agentic-mtp:
-  image: lmsysorg/sglang-rocm:v0.5.18-rocm720-mi35x-20260829
+  image: lmsysorg/sglang-rocm:v0.5.18-rocm720-mi35x-20260902
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:mi355x-amds
@@ -1753,6 +1753,7 @@ dsv4-fp4-mi355x-sglang-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 4, ep: 1, dp-attn: false, kv-offloading: none, conc-list: [1, 2, 4, 8, 10], spec-decoding: mtp }
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [16, 32, 48], spec-decoding: mtp }
+      - { tp: 4, ep: 1, dp-attn: false, kv-offloading: none, conc-list: [1, 4, 8], spec-decoding: mtp }
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [32, 48], spec-decoding: mtp }
       - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, conc-list: [16], spec-decoding: mtp }
+      - { tp: 8, ep: 1, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 96, 128, 160], spec-decoding: mtp }

perf-changelog.yaml

@@ -6858,3 +6858,18 @@
   description:
     - "Refresh the DeepSeek-V4-Pro FP4 B200 vLLM image and search space, pin the Nscale NVFP4 checkpoint, use the FlashInfer CuTeDSL MoE backend for expert-parallel runs with EPLB disabled, set GPU memory utilization to 0.95, and cap the model length at 12288."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2534
+
+- config-keys:
+    - dsv4-fp4-mi355x-sglang-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Update image from lmsysorg/sglang-rocm:v0.5.18-rocm720-mi35x-20260829 to lmsysorg/sglang-rocm:v0.5.18-rocm720-mi35x-20260902."
+    - "Add a DP-attention arm: tp 8, ep 1, dp-attn true, hicache host KV tier, concurrency [64, 96, 128, 160]. It runs behind sglang-router with consistent hashing on the AIPerf correlation id so multi-turn sessions stay on the DP rank holding their prefix."
+    - "Serve with --enable-deepseek-v4-fp4-indexer and set HSA_NO_SCRATCH_RECLAIM=0 on all arms."
+    - "DP-attention serving flags: --enable-two-batch-overlap, --enable-dp-attention-local-control-broadcast, --tokenizer-worker-num equal to TP, --stream-interval 20 and --prefill-decode-interval 10, alongside the existing --enable-prefill-delayer. Chunked prefill now scales the per-TP base by the DP degree instead of a fixed 8192."
+    - "Shared-experts fusion is now per-arm: kept enforced on the tensor-parallel arms and disabled under DP attention, matching the DP baselines."
+    - "GPU_MAX_HW_QUEUES is now set explicitly per-arm: 2 on the tensor-parallel arms and 5 under DP attention, the documented companion to two-batch overlap."
+    - "Lower mem-fraction-static from 0.89 to 0.86 on every arm: TP4, TP8 and TP8 with DP attention. swa-full-tokens-ratio becomes per-arm: 0.10 on the tensor-parallel arms as before, 0.15 under DP attention."
+    - "Drop concurrency 2 and 10 from the TP4 arm, leaving [1, 4, 8], and drop concurrency 16 from the TP8 hicache arm, leaving [32, 48]. Concurrency 16 remains on the TP8 no-offload arm."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2800