From 3d9f4d5261e45bf84c2ce3c37d36b603437bd6ee Mon Sep 17 00:00:00 2001 From: "jacky.cheng" Date: Thu, 10 Sep 2026 14:03:55 +0000 Subject: [PATCH 1/5] [AMD][Qwen3.5] Pin MI355X MXFP4 AgentX to sglang-rocm 20260910 Co-Authored-By: ChangLiu0709 <190883410+ChangLiu0709@users.noreply.github.com> --- configs/amd-master.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml index 3d9a3dcca..1a18f05b6 100644 --- a/configs/amd-master.yaml +++ b/configs/amd-master.yaml @@ -361,7 +361,7 @@ qwen3.5-fp4-mi355x-sglang-mtp: - { tp: 4, conc-start: 4, conc-end: 16, spec-decoding: mtp } qwen3.5-fp4-mi355x-sglang-agentic-mtp: - image: lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260908 + image: lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260910 model: amd/Qwen3.5-397B-A17B-MXFP4 model-prefix: qwen3.5 runner: cluster:mi355x-amds From a7447217cb02aae100c5a1406c5b88aa6b8a1ece Mon Sep 17 00:00:00 2001 From: "jacky.cheng" Date: Fri, 11 Sep 2026 10:05:15 +0000 Subject: [PATCH 2/5] [AMD][Qwen3.5] Fix AgentX launch on sglang-rocm 0.5.19: cuda-graph-max-bs-decode v0.5.19 splits --cuda-graph-max-bs into --cuda-graph-max-bs-decode and --cuda-graph-max-bs-prefill and removes the unsuffixed name, so argparse rejects it as ambiguous. Every agentic arm in run 34486742261 died with: sglang serve: error: ambiguous option: --cuda-graph-max-bs could match --cuda-graph-max-bs-decode, --cuda-graph-max-bs-prefill Process died before http://0.0.0.0:8894/health became ready. Use --cuda-graph-max-bs-decode, matching the Qwen3.5 RTX PRO 6000 recipe. The captured decode batch size is unchanged: min(2*CONC, 128). --- benchmarks/single_node/agentic/qwen3.5_fp4_mi355x_sglang_mtp.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/benchmarks/single_node/agentic/qwen3.5_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/qwen3.5_fp4_mi355x_sglang_mtp.sh index e9e3d5eb5..91d4e7b93 100644 --- a/benchmarks/single_node/agentic/qwen3.5_fp4_mi355x_sglang_mtp.sh +++ b/benchmarks/single_node/agentic/qwen3.5_fp4_mi355x_sglang_mtp.sh @@ -113,7 +113,7 @@ SGLANG_CMD=( --watchdog-timeout 1200 --page-size 16 --kv-cache-dtype fp8_e4m3 - --cuda-graph-max-bs "$CUDA_GRAPH_MAX_BS" + --cuda-graph-max-bs-decode "$CUDA_GRAPH_MAX_BS" --max-running-requests "$MAX_RUNNING_REQUESTS" --max-prefill-tokens 16384 --chunked-prefill-size 16384 From 1ce26e1654b02023be891457921b4a51a3940075 Mon Sep 17 00:00:00 2001 From: "jacky.cheng" Date: Fri, 11 Sep 2026 14:27:08 +0000 Subject: [PATCH 3/5] [AMD][Qwen3.5] AgentX INT4 prefill allreduce and recv-interval 60 Switch ROCM_QUICK_REDUCE_QUANTIZATION from INT8 to INT4 and raise the scheduler-recv-interval default from 30 to 60. Local TP2 AgentX vs the previous INT8/recv=30 defaults improved hicache conc28/32 output tok/s/gpu and cut P90 E2E/TPOT; KV-none throughput stayed flat. --- .../single_node/agentic/qwen3.5_fp4_mi355x_sglang_mtp.sh | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/benchmarks/single_node/agentic/qwen3.5_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/qwen3.5_fp4_mi355x_sglang_mtp.sh index 91d4e7b93..9a09392e9 100644 --- a/benchmarks/single_node/agentic/qwen3.5_fp4_mi355x_sglang_mtp.sh +++ b/benchmarks/single_node/agentic/qwen3.5_fp4_mi355x_sglang_mtp.sh @@ -14,7 +14,7 @@ check_env_vars \ MODEL TP CONC EP_SIZE KV_OFFLOADING \ TOTAL_CPU_DRAM_GB RESULT_DIR DURATION -SCHEDULER_RECV_INTERVAL=${SCHEDULER_RECV_INTERVAL:-30} +SCHEDULER_RECV_INTERVAL=${SCHEDULER_RECV_INTERVAL:-60} if [[ -n "${SLURM_JOB_ID:-}" ]]; then echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}" @@ -90,7 +90,7 @@ export SGLANG_USE_AITER=1 export SGLANG_USE_AITER_UNIFIED_ATTN=1 export AITER_FLYDSL_FORCE=1 export SGLANG_MAMBA_SSM_DTYPE=bfloat16 -export ROCM_QUICK_REDUCE_QUANTIZATION=INT8 +export ROCM_QUICK_REDUCE_QUANTIZATION=INT4 export SGLANG_TIMEOUT_KEEP_ALIVE=1800 if [ "${EVAL_ONLY:-false}" != "true" ]; then From ef7191bc6c10beb5bc857cf3e2ff61afa25ac076 Mon Sep 17 00:00:00 2001 From: "jacky.cheng" Date: Fri, 11 Sep 2026 15:23:20 +0000 Subject: [PATCH 4/5] [AMD][Qwen3.5] Bump MI355X MXFP4 AgentX image to sglang-rocm 20260911 --- configs/amd-master.yaml | 2 +- perf-changelog.yaml | 8 ++++++++ 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml index 1a18f05b6..e4bd59ccc 100644 --- a/configs/amd-master.yaml +++ b/configs/amd-master.yaml @@ -361,7 +361,7 @@ qwen3.5-fp4-mi355x-sglang-mtp: - { tp: 4, conc-start: 4, conc-end: 16, spec-decoding: mtp } qwen3.5-fp4-mi355x-sglang-agentic-mtp: - image: lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260910 + image: lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260911 model: amd/Qwen3.5-397B-A17B-MXFP4 model-prefix: qwen3.5 runner: cluster:mi355x-amds diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 1f90deb80..a759b57c8 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -7408,3 +7408,11 @@ - "Enable required power for the seven existing B200 Kimi-K3 recipes, use the pinned AgentX producer and shared collector, fix the exporter import URI, and record DCP8 with offload disabled to match the serving commands." - "启用 Kimi-K3 B200 实测功耗并修正 DCP 与 Offload 元数据。" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3043 + +- config-keys: + - qwen3.5-fp4-mi355x-sglang-agentic-mtp + scenario-type: + - agentic-coding + description: + - "Bump the AgentX image to lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260911 (Docker Hub 200). main moved this key to 20260908 while the PR was open, so the net change against main is 20260908 -> 20260911." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2972 From 2fa9a88e440efa97d455c401751f21f427a821d0 Mon Sep 17 00:00:00 2001 From: "jacky.cheng" Date: Sat, 12 Sep 2026 06:27:22 +0000 Subject: [PATCH 5/5] [AMD][Qwen3.5] Collapse AgentX changelog into one short #2972 entry --- perf-changelog.yaml | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index a759b57c8..abd79ea4d 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -7414,5 +7414,7 @@ scenario-type: - agentic-coding description: - - "Bump the AgentX image to lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260911 (Docker Hub 200). main moved this key to 20260908 while the PR was open, so the net change against main is 20260908 -> 20260911." + - "Bump image to lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260911." + - "Rename --cuda-graph-max-bs to --cuda-graph-max-bs-decode for sglang 0.5.19." + - "Prefill allreduce INT4; scheduler-recv-interval default 60." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2972