Skip to content

Commit 5c9f96a

Browse files
authored
[AMD][MI35X] 0521 DSV4 (#1548)
* Update config * Update change log
1 parent f210796 commit 5c9f96a

2 files changed

Lines changed: 11 additions & 5 deletions

File tree

.github/configs/amd-master.yaml

Lines changed: 5 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -1774,7 +1774,7 @@ dsr1-fp4-mi355x-sglang-disagg-mtp:
17741774
# image tag, so bumping sglang is just an image tag bump here. Sweeps
17751775
# DP-attention on/off and EP=8.
17761776
dsv4-fp4-mi355x-sglang:
1777-
image: rocm/sgl-dev:rocm720-mi35x-b19052c-20260518-DSv4
1777+
image: rocm/sgl-dev:rocm720-mi35x-8c3b5aa-20260521-DSv4
17781778
model: deepseek-ai/DeepSeek-V4-Pro
17791779
model-prefix: dsv4
17801780
runner: mi355x
@@ -1786,13 +1786,13 @@ dsv4-fp4-mi355x-sglang:
17861786
- isl: 1024
17871787
osl: 1024
17881788
search-space:
1789-
- { tp: 8, dp-attn: true, conc-start: 32, conc-end: 1024 }
1790-
- { tp: 8, dp-attn: false, conc-start: 1 , conc-end: 64 }
1789+
- { tp: 8, dp-attn: true, conc-start: 64, conc-end: 1024 }
1790+
- { tp: 8, dp-attn: false, conc-start: 1 , conc-end: 32 }
17911791
- isl: 8192
17921792
osl: 1024
17931793
search-space:
1794-
- { tp: 8, dp-attn: true, conc-start: 32, conc-end: 512 }
1795-
- { tp: 8, dp-attn: false, conc-start: 1, conc-end: 64 }
1794+
- { tp: 8, dp-attn: true, conc-start: 64, conc-end: 512 }
1795+
- { tp: 8, dp-attn: false, conc-start: 1, conc-end: 32 }
17961796

17971797
# DSv4 on MI355X via vLLM, using the official vllm/vllm-openai-rocm
17981798
# nightly image. DSv4 base ROCm support (vllm-project/vllm#40871) merged

perf-changelog.yaml

Lines changed: 6 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -3062,3 +3062,9 @@
30623062
description:
30633063
- "Bump vLLM ROCm image from nightly-b50646e5effd7cb5884cd96fdff4c53c18521198 to nightly-4f940896a32c9e2a0eba7f50d521bf5f6b4de458"
30643064
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1546
3065+
3066+
- config-keys:
3067+
- dsv4-fp4-mi355x-sglang
3068+
description:
3069+
- "Bump image to rocm/sgl-dev:rocm720-mi35x-8c3b5aa-20260521-DSv4"
3070+
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1548

0 commit comments

Comments
 (0)