From 684dff276ccd4b7071f0d06c023a1df4d7a8e5d6 Mon Sep 17 00:00:00 2001 From: khluu Date: Tue, 3 Nov 2026 18:17:26 -0800 Subject: [PATCH] [CI][DSv4.1] Give the MXFP8 linear test stub a params_dtype #64111 made FlashInferCutedslMxfp8LinearKernel.apply_weights take the output dtype from layer.params_dtype for unquantized inputs, which real linear layers set. The test_kernels.py stand-in never set it, so test_engram_wkv_prefetch_matches_inline fails with AttributeError on every nightly 4xB200 run since. bfloat16 matches the inputs the test feeds, so the output dtype is unchanged from before. Co-Authored-By: Claude Opus 5.5 Signed-off-by: khluu --- .../deepseek_v41/nvidia/specialized_decode/test_kernels.py | 1 + 1 file changed, 1 insertion(+) diff --git a/tests/models/deepseek_v41/nvidia/specialized_decode/test_kernels.py b/tests/models/deepseek_v41/nvidia/specialized_decode/test_kernels.py index c3d1deb741b4..514b305a37dc 100644 --- a/tests/models/deepseek_v41/nvidia/specialized_decode/test_kernels.py +++ b/tests/models/deepseek_v41/nvidia/specialized_decode/test_kernels.py @@ -495,6 +495,7 @@ class _Mxfp8Linear(nn.Module): def __init__(self, n: int, k: int): super().__init__() + self.params_dtype = torch.bfloat16 w = torch.randn(n, k, device=DEVICE) / k**0.5 w_q, w_s = mxfp8_e4m3_quantize( w.to(torch.bfloat16), is_sf_swizzled_layout=False