From 3c480b6dc3fe30c98da4c0e662be631f65def609 Mon Sep 17 00:00:00 2001 From: JohnQinAMD Date: Tue, 2 Jun 2026 20:01:24 +0000 Subject: [PATCH 1/2] dsr1-fp4-mi355x-sglang: bump ROCm 7.0->7.2 image + add TP4 search-space Bump image lmsysorg/sglang:v0.5.12-rocm700-mi35x -> lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260519 (ROCm 7.0->7.2; same image validated multi-node in PR #1566). Add tp:4 search-space. Recipe unchanged (already QR-INT4 + range-ratio 0.8). fp8-mi355x untouched. Co-Authored-By: Claude Opus 4.8 --- .github/configs/amd-master.yaml | 7 +++++-- perf-changelog.yaml | 9 +++++++++ 2 files changed, 14 insertions(+), 2 deletions(-) diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml index fb3966ce6..5947724cf 100644 --- a/.github/configs/amd-master.yaml +++ b/.github/configs/amd-master.yaml @@ -1,5 +1,5 @@ dsr1-fp4-mi355x-sglang: - image: lmsysorg/sglang:v0.5.12-rocm700-mi35x + image: lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260519 model: amd/DeepSeek-R1-0528-MXFP4-Preview model-prefix: dsr1 runner: mi355x @@ -16,6 +16,7 @@ dsr1-fp4-mi355x-sglang: - isl: 8192 osl: 1024 search-space: + - { tp: 4, conc-start: 4, conc-end: 64 } - { tp: 8, conc-start: 4, conc-end: 64 } # Agentic-coding sweep commented out for this image-bump PR — the # 10-conc agentic matrix amplifies sweep cost and the bump validation @@ -27,7 +28,7 @@ dsr1-fp4-mi355x-sglang: # - { tp: 8, offloading: none, conc-list: [1, 2, 4, 8, 12, 16, 32, 64, 128, 256] } dsr1-fp4-mi355x-sglang-mtp: - image: lmsysorg/sglang:v0.5.12-rocm700-mi35x + image: lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260519 model: amd/DeepSeek-R1-0528-MXFP4 model-prefix: dsr1 runner: mi355x @@ -39,10 +40,12 @@ dsr1-fp4-mi355x-sglang-mtp: - isl: 1024 osl: 1024 search-space: + - { tp: 4, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp } - { tp: 8, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp } - isl: 8192 osl: 1024 search-space: + - { tp: 4, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp } - { tp: 8, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp } dsr1-fp4-mi355x-atom: diff --git a/perf-changelog.yaml b/perf-changelog.yaml index d201e9f3b..5900818fc 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3395,3 +3395,12 @@ description: - "Add DeepSeek-V4-Pro FP4 MI355X ATOM MTP3 benchmark; image rocm/atom:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.3" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1627 + +- config-keys: + - dsr1-fp4-mi355x-sglang + - dsr1-fp4-mi355x-sglang-mtp + description: + - "Image: lmsysorg/sglang:v0.5.12-rocm700-mi35x -> lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260519 (ROCm 7.0 -> 7.2; same image already validated multi-node in PR #1566)" + - "Add tp:4 to search-space (8k1k for non-mtp; 1k1k+8k1k for mtp) to surface the TP4 throughput-per-GPU frontier" + - "Local MI355X bench (matched cfg: range-ratio 0.8 + ROCM_QUICK_REDUCE_QUANTIZATION=INT4): 8k1k decode TPOT 9-25% faster at every conc vs the rocm700 CI; c64 now leads both axes (TPOT 16.5 vs 18.1 ms, tput 2954 vs 2841 tok/s/gpu). TP4 c64 8k1k = 4254 tok/s/gpu (> TP8 2954)" + pr-link: PR_LINK_PENDING From 4f52374a583ee523b477a94511c13867b6573c81 Mon Sep 17 00:00:00 2001 From: JohnQinAMD Date: Tue, 2 Jun 2026 20:05:12 +0000 Subject: [PATCH 2/2] perf-changelog: fill pr-link for dsr1-fp4-mi355x-sglang rocm720 bump (#1645) Co-Authored-By: Claude Opus 4.8 --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 5900818fc..03564a5e3 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3403,4 +3403,4 @@ - "Image: lmsysorg/sglang:v0.5.12-rocm700-mi35x -> lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260519 (ROCm 7.0 -> 7.2; same image already validated multi-node in PR #1566)" - "Add tp:4 to search-space (8k1k for non-mtp; 1k1k+8k1k for mtp) to surface the TP4 throughput-per-GPU frontier" - "Local MI355X bench (matched cfg: range-ratio 0.8 + ROCM_QUICK_REDUCE_QUANTIZATION=INT4): 8k1k decode TPOT 9-25% faster at every conc vs the rocm700 CI; c64 now leads both axes (TPOT 16.5 vs 18.1 ms, tput 2954 vs 2841 tok/s/gpu). TP4 c64 8k1k = 4254 tok/s/gpu (> TP8 2954)" - pr-link: PR_LINK_PENDING + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1645