diff --git a/inferencex-e2e/configs/amd-master.yaml b/inferencex-e2e/configs/amd-master.yaml index b9ddea12e8..59732b57f7 100644 --- a/inferencex-e2e/configs/amd-master.yaml +++ b/inferencex-e2e/configs/amd-master.yaml @@ -181,7 +181,7 @@ dsr1-fp8-mi355x-sglang-mtp: # checkpoint leaves too little KV headroom there for the 256k-capped AgentX # corpus. qwen3.5-fp8-mi355x-sglang-agentic-mtp: - image: lmsysorg/sglang-rocm:v0.5.20-rocm720-mi35x-20260929 + image: lmsysorg/sglang:v0.5.22-rocm720-mi35x@sha256:c4b61360b251d62bb17da63f175d571ab22479eb46a144548d79f3ed3998ea11 model: Qwen/Qwen3.5-397B-A17B-FP8 model-prefix: qwen3.5 runner: cluster:mi355x-amds diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml index 2bd80acd9f..52823b5def 100644 --- a/inferencex-e2e/perf-changelog.yaml +++ b/inferencex-e2e/perf-changelog.yaml @@ -12551,3 +12551,10 @@ - "Add 1P1D TP16 DCP16 points at conc 64, 76 and 96 with an EP16 decode on DeepGEMM MegaMoE (MXFP8 dispatch, BF16 combine), 96 max running requests, CUDA graphs up to batch 96 and 8 draft tokens. Their prefill uses page size 64 and 8 draft tokens. The image and the existing c16, c32 and c48 points are unchanged." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4929 append-only: true + +- config-keys: + - qwen3.5-fp8-mi355x-sglang-agentic-mtp + description: + - "Update SGLang image from v0.5.20 ROCm 7.2 MI35x build 20260929 to v0.5.22 ROCm 7.2 MI35x." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4519 + scenario-type: [agentic-coding]