From 65a9952d292a17e47934c61d826458c631caf8b8 Mon Sep 17 00:00:00 2001 From: jiacao-amd Date: Tue, 23 Jun 2026 13:57:12 -0500 Subject: [PATCH 1/2] glm5.1-fp4-mi355x-sglang: bump image to v0.5.13.post1-20260622 + enable aiter allreduce fusion - Bump sglang image v0.5.12.post1-20260529 -> v0.5.13.post1-20260622: the older image's aiter (dev98) does not tolerate the host ROCm 7.1.1 x container 7.2.0 mismatch on MI355X and produces garbled output; the newer aiter (dev240) does. The newer build is also faster at decode. - Add --enable-aiter-allreduce-fusion: fuse TP all-reduce + RMSNorm (atom-style local_device_load_rmsnorm). Marginal perf; gated by CI GSM8K eval. Co-Authored-By: Claude Opus 4.8 --- .github/configs/amd-master.yaml | 2 +- benchmarks/single_node/fixed_seq_len/glm5.1_fp4_mi355x.sh | 1 + 2 files changed, 2 insertions(+), 1 deletion(-) diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml index 5e5452c4c..3b0d908be 100644 --- a/.github/configs/amd-master.yaml +++ b/.github/configs/amd-master.yaml @@ -679,7 +679,7 @@ glm5-fp8-mi355x-atom: - { tp: 8, conc-start: 4, conc-end: 256 } glm5.1-fp4-mi355x-sglang: - image: lmsysorg/sglang-rocm:v0.5.12.post1-rocm720-mi35x-20260529 + image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260622 model: amd/GLM-5.1-MXFP4 model-prefix: glm5.1 runner: mi355x diff --git a/benchmarks/single_node/fixed_seq_len/glm5.1_fp4_mi355x.sh b/benchmarks/single_node/fixed_seq_len/glm5.1_fp4_mi355x.sh index aada63d56..4e0d507c6 100644 --- a/benchmarks/single_node/fixed_seq_len/glm5.1_fp4_mi355x.sh +++ b/benchmarks/single_node/fixed_seq_len/glm5.1_fp4_mi355x.sh @@ -52,6 +52,7 @@ python3 -m sglang.launch_server \ --nsa-decode-backend tilelang $EVAL_CONTEXT_ARGS \ --kv-cache-dtype fp8_e4m3 \ --tokenizer-worker-num $((TP*2)) \ + --enable-aiter-allreduce-fusion \ --disable-radix-cache> $SERVER_LOG 2>&1 & SERVER_PID=$! From 40ba9039a2d08a63810ae6c35818a95ff71ee077 Mon Sep 17 00:00:00 2001 From: Chun Fang Date: Wed, 24 Jun 2026 07:34:40 +0000 Subject: [PATCH 2/2] Update the perf-changelog --- perf-changelog.yaml | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index fd3d0a6e7..5c9eb9d87 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3996,3 +3996,10 @@ - "Recover the skipped official ingest for PR #1797 from validated sweep run 27591355916 (attempt 6)" - "No benchmark configuration change; reuse the exact 25-point fixed-sequence matrix and 2 eval jobs" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1869 + +- config-keys: + - glm5.1-fp4-mi355x-sglang + description: + - "Bump SGLang ROCm image from v0.5.12.post1-rocm720-mi35x-20260529 to v0.5.13.post1-rocm720-mi35x-20260622" + - "Enable aiter allreduce fusion via --enable-aiter-allreduce-fusion in benchmarks/single_node/fixed_seq_len/glm5.1_fp4_mi355x.sh" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1905