Skip to content

Commit 786dc46

Browse files
Add MI355X config: qwen3.5-fp4-sglang-mtp
Adds the qwen3.5-fp4-mi355x-sglang-mtp recipe using the latest sglang-rocm nightly v0.5.12-rocm720-mi35x-20260517 (matches PRs #1440-1444 which bump other mi355x recipes to the same image). Replaces the stale #1078 which targeted v0.5.10rc0-rocm700-mi35x-20260417 and was tagged broken; that branch reverted unrelated merged changes when rebased. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
1 parent c07bf5d commit 786dc46

2 files changed

Lines changed: 26 additions & 0 deletions

File tree

.github/configs/amd-master.yaml

Lines changed: 20 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -367,6 +367,26 @@ qwen3.5-fp4-mi355x-atom:
367367
- { tp: 2, conc-start: 4, conc-end: 256 }
368368
- { tp: 4, conc-start: 4, conc-end: 16 }
369369

370+
qwen3.5-fp4-mi355x-sglang-mtp:
371+
image: lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260517
372+
model: amd/Qwen3.5-397B-A17B-MXFP4
373+
model-prefix: qwen3.5
374+
runner: mi355x
375+
precision: fp4
376+
framework: sglang
377+
multinode: false
378+
seq-len-configs:
379+
- isl: 1024
380+
osl: 1024
381+
search-space:
382+
- { tp: 2, conc-start: 4, conc-end: 256, spec-decoding: mtp }
383+
- { tp: 4, conc-start: 4, conc-end: 16, spec-decoding: mtp }
384+
- isl: 8192
385+
osl: 1024
386+
search-space:
387+
- { tp: 2, conc-start: 4, conc-end: 256, spec-decoding: mtp }
388+
- { tp: 4, conc-start: 4, conc-end: 16, spec-decoding: mtp }
389+
370390
qwen3.5-fp8-mi300x-sglang:
371391
image: lmsysorg/sglang:v0.5.10-rocm720-mi30x
372392
model: Qwen/Qwen3.5-397B-A17B-FP8

perf-changelog.yaml

Lines changed: 6 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2629,3 +2629,9 @@
26292629
description:
26302630
- "Update vLLM ROCm image from v0.18.0 to v0.21.0"
26312631
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1404
2632+
2633+
- config-keys:
2634+
- qwen3.5-fp4-mi355x-sglang-mtp
2635+
description:
2636+
- "Add MI355X config: qwen3.5-fp4-sglang-mtp using lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260517"
2637+
pr-link: PLACEHOLDER

0 commit comments

Comments
 (0)