diff --git a/test/registered/amd/test_kimi_k25_mxfp4.py b/test/registered/amd/test_kimi_k25_mxfp4.py index 1ce83f8eb..43c40fe79 100644 --- a/test/registered/amd/test_kimi_k25_mxfp4.py +++ b/test/registered/amd/test_kimi_k25_mxfp4.py @@ -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