From c9dbedd5825a3367b74b539d5f5a8f48685a7477 Mon Sep 17 00:00:00 2001 From: "jacky.cheng" Date: Fri, 31 Jul 2026 00:57:31 +0000 Subject: [PATCH 1/4] [AMD][MI35X] Bump qwen3.5-mxfp4 MI355X SGLang single-node image to v0.5.16 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Bump the image for qwen3.5-fp4-mi355x-sglang and qwen3.5-fp4-mi355x-sglang-mtp from lmsysorg/sglang-rocm:v0.5.15-rocm720-mi35x-20260713 to lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260730, and append a perf-changelog entry triggering a re-benchmark for the two config-keys. 中文：将 qwen3.5-fp4-mi355x-sglang 与 qwen3.5-fp4-mi355x-sglang-mtp 两个配置的镜像从 lmsysorg/sglang-rocm:v0.5.15-rocm720-mi35x-20260713 升级到 lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260730，并在 perf-changelog.yaml 追加一条记录，触发这两个 config-key 的重新基准测试。 --- configs/amd-master.yaml | 4 ++-- perf-changelog.yaml | 7 +++++++ 2 files changed, 9 insertions(+), 2 deletions(-) diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml index b8f6f4fc31..200e45ce44 100644 --- a/configs/amd-master.yaml +++ b/configs/amd-master.yaml @@ -366,7 +366,7 @@ qwen3.5-fp8-mi355x-sglang-disagg: - "DECODE_MTP_SIZE=0" qwen3.5-fp4-mi355x-sglang: - image: lmsysorg/sglang-rocm:v0.5.15-rocm720-mi35x-20260713 + image: lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260730 model: amd/Qwen3.5-397B-A17B-MXFP4 model-prefix: qwen3.5 runner: mi355x @@ -398,7 +398,7 @@ qwen3.5-fp4-mi355x-atom: - { tp: 4, conc-start: 4, conc-end: 16 } qwen3.5-fp4-mi355x-sglang-mtp: - image: lmsysorg/sglang-rocm:v0.5.15-rocm720-mi35x-20260713 + image: lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260730 model: amd/Qwen3.5-397B-A17B-MXFP4 model-prefix: qwen3.5 runner: mi355x diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 6eebd97070..6f6dcbc48d 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5355,3 +5355,10 @@ - "Apply the accuracy-gated Kimi-K2.5 MXFP4 settings: tuned AITER MXFP4 MoE, fused shared experts, FP8 KV cache, block size 16, 16384 batched tokens, 512 sequences, async scheduling, gpu-memory-utilization 0.85 (headroom for CUDA-graph capture on MI355X), and the AITER BF16 GEMM path" - "Extend the TP4 and TP8 8k1k concurrency sweep from 64 to 128 (1k1k deprecated per #2263)" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2213 + +- config-keys: + - qwen3.5-fp4-mi355x-sglang + - qwen3.5-fp4-mi355x-sglang-mtp + description: + - "Bump image from lmsysorg/sglang-rocm:v0.5.15-rocm720-mi35x-20260713 to lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260730" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2201 From 760e7d3d700b720a2f6d5ae3c858cbab55922cb5 Mon Sep 17 00:00:00 2001 From: "jacky.cheng" Date: Fri, 31 Jul 2026 01:00:29 +0000 Subject: [PATCH 2/4] [AMD][MI35X] Point qwen3.5-mxfp4 changelog entry to PR #2433 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Update the newly-added qwen3.5-mxfp4-mi355x-sglang changelog entry's pr-link from the #2201 reference placeholder to this PR (#2433). 中文：将新增的 qwen3.5-mxfp4-mi355x-sglang changelog 记录的 pr-link 从参考占位 #2201 更新为本 PR（#2433）。 --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 6f6dcbc48d..828296abb7 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5361,4 +5361,4 @@ - qwen3.5-fp4-mi355x-sglang-mtp description: - "Bump image from lmsysorg/sglang-rocm:v0.5.15-rocm720-mi35x-20260713 to lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260730" - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2201 + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2433 From 8eb57da588d1d4903bfba974b467911d61d7cafe Mon Sep 17 00:00:00 2001 From: "jacky.cheng" Date: Tue, 4 Aug 2026 02:33:22 +0000 Subject: [PATCH 3/4] [AMD][MI35X] Sync perf-changelog with main and re-append qwen3.5-mxfp4 v0.5.16 entry --- perf-changelog.yaml | 48 +++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 48 insertions(+) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 828296abb7..4140fc5140 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5356,6 +5356,54 @@ - "Extend the TP4 and TP8 8k1k concurrency sweep from 64 to 128 (1k1k deprecated per #2263)" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2213 +- config-keys: + - qwen3.5-fp4-b300-sglang-agentic-mtp + description: + - "Add Qwen3.5-397B-A17B NVFP4 AgentX benchmark on B300 with SGLang native NEXTN MTP" + - "Use the 256k trace dataset and golden synthetic acceptance length 3.39" + - "Use the shared AIPerf watchdog to cap whole-trajectory runtime idle gaps at 300 seconds" + - "Image: lmsysorg/sglang:v0.5.16-cu130" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2421 + +- config-keys: + - qwen3.5-fp4-b200-sglang-agentic-mtp + description: + - "Add Qwen3.5-397B-A17B NVFP4 AgentX benchmark on B200 with SGLang native NEXTN MTP" + - "Use the 256k trace dataset and golden synthetic acceptance length 3.39" + - "Use the shared AIPerf watchdog to cap whole-trajectory runtime idle gaps at 300 seconds" + - "Image: lmsysorg/sglang:v0.5.16-cu130" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2420 + + +- config-keys: + - minimaxm3-fp4-b200-dynamo-vllm + description: + - "Bump vLLM nightly image to nightly-5e35a6f4f9bbc217c599692157ca985c894373f7 (fixed-len MiniMax-M3 fix); update all B200 FP4 8k1k recipe YAML container fields to match" + - "Enable VLLM_MINIMAX_M3_MSA_DECODE_BACKEND=cutlass in prefill and decode environments" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2472 + +- config-keys: + - minimaxm3-fp4-b300-dynamo-vllm-8k1k-tp1 + - minimaxm3-fp4-b300-dynamo-vllm-8k1k-tp1-tep4 + description: + - "Bump vLLM nightly image to nightly-5e35a6f4f9bbc217c599692157ca985c894373f7 (fixed-len MiniMax-M3 fix); update B300 FP4 8k1k tp1 recipe YAML container fields to match" + - "Enable VLLM_MINIMAX_M3_MSA_DECODE_BACKEND=cutlass in prefill and decode environments" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2473 + +- config-keys: + - minimaxm3-fp4-b300-dynamo-vllm-mtp + description: + - "Bump vLLM nightly image to nightly-5e35a6f4f9bbc217c599692157ca985c894373f7 (fixed-len MiniMax-M3 fix); update B300 FP4 MTP recipe YAML container fields to match (excluding 2p1d-dep2-dep4 used by legacy-dep4)" + - "Enable VLLM_MINIMAX_M3_MSA_DECODE_BACKEND=cutlass in prefill and decode environments" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2471 + +- config-keys: + - minimaxm3-fp4-b300-dynamo-vllm-mtp-legacy-dep4 + description: + - "Bump vLLM nightly image to nightly-5e35a6f4f9bbc217c599692157ca985c894373f7 (fixed-len MiniMax-M3 fix); update 2p1d-dep2-dep4 recipe YAML container to match" + - "Enable minimax_m3_msa_decode_backend=cutlass in decode attention-config" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2483 + - config-keys: - qwen3.5-fp4-mi355x-sglang - qwen3.5-fp4-mi355x-sglang-mtp From 72d62d5e7b07133649f4c33afb11bbbdb0ced70a Mon Sep 17 00:00:00 2001 From: "jacky.cheng" Date: Tue, 4 Aug 2026 06:24:50 +0000 Subject: [PATCH 4/4] [AMD][MI35X] Add trailing newline to perf-changelog.yaml Fix CI validation failure: perf-changelog.yaml did not end with a newline. --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index a8f6f00383..83c562957c 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5432,4 +5432,4 @@ description: - "Bump image from lmsysorg/sglang-rocm:v0.5.15-rocm720-mi35x-20260713 to lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260730" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2433 - \ No newline at end of file +