[dsv4] Trigger MHC prenorm prewarm at weight-load time with rank sync (#29988)

2026-07-02 23:05:03 -07:00
parent a6ee64d237
commit e81f05cf4f
3 changed files with 68 additions and 180 deletions
@@ -18,7 +18,6 @@ def test_mhc_fused_post_pre_matches_unfused(
        pytest.skip("CUDA is required for TileLang mHC kernels")

    monkeypatch.setattr(mhc, "is_dsa_prefill_cp_round_robin_split", lambda: False)
-    monkeypatch.setattr(mhc, "_mhc_pre_warmed", True)
    torch.manual_seed(0)
    device = torch.device("cuda")
    hc_mult = 4