minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[AMD] Bump amd/Kimi-K2.5-MXFP4 revision to align with shared-experts fusion (#25740)

Browse Source
This commit is contained in:
YC Yen-Ching Tseng
2026-05-18 22:47:05 -07:00
committed by GitHub
parent 1d19721394
commit 7c3f614e23
1 changed files with 7 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/amd/test_kimi_k25_mxfp4.py
+7 -1
View File
@@ -27,7 +27,13 @@ from sglang.test.test_utils import (
register_amd_ci(est_time=3600, suite="stage-c-test-large-8-gpu-amd-mi35x")
KIMI_K25_MXFP4_MODEL_PATH = "amd/Kimi-K2.5-MXFP4"
KIMI_K25_MXFP4_REVISION = "b071bc6f8eb042e093e14f3b8bdbad71c18e09d3"
# Bumped from b071bc6f -> 419004c8 (HF main HEAD as of 2026-05-18). The pinned
# b071bc6f revision keeps shared_experts unquantized (bf16), which is
# incompatible with the shared-experts fusion path enabled for Kimi-K2.5
# (n_routed_experts=384) in #25390. Revisions from 94d8c1bd onward quantize
# shared_experts to MXFP4 so the fusion can copy weights between routed and
# shared experts without a dtype/shape mismatch.
KIMI_K25_MXFP4_REVISION = "419004c8716cf22c929aa15d39b85e09a8a2091a"
SERVER_LAUNCH_TIMEOUT = 3600