L-electron-Rare/mlx
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[CUDA] Fix half type matmul in cutlass kernels (#3469)

Browse Source
This commit is contained in:
Cheng
2026-05-06 08:35:53 +09:00
committed by GitHub
parent 1fdd4e23c2
commit 80bcd1c658
8 changed files with 28 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/tests/cuda_skip.py
-17
View File
@@ -1,22 +1,5 @@
cuda_skip = {
# Lapack ops NYI
"TestLinalg.test_cholesky",
"TestLinalg.test_cholesky_inv",
"TestLinalg.test_eig",
"TestLinalg.test_eigh",
"TestLinalg.test_inverse",
"TestVmap.test_vmap_inverse",
"TestLinalg.test_lu",
"TestLinalg.test_lu_factor",
"TestLinalg.test_pseudo_inverse",
"TestLinalg.test_qr_factorization",
"TestInit.test_orthogonal",
"TestLinalg.test_svd_decomposition",
"TestVmap.test_vmap_svd",
"TestLinalg.test_tri_inverse",
# Quantization NYI
"TestQuantized.test_gather_matmul_grad",
"TestQuantized.test_gather_qmm",
"TestQuantized.test_gather_qmm_sorted",
"TestQuantized.test_gather_qmm_grad",
}
python/tests/test_quantized.py
+1 -1
View File
@@ -1046,7 +1046,7 @@ class TestQuantized(mlx_tests.MLXTestCase):
y3 = scatter_unsort(y3, inv_order, indices.shape)
y4 = scatter_unsort(y4, inv_order, indices.shape)
tol = 1.5e-5 if (dtype == mx.float32) else 2.5e-4
tol = 1.5e-5 if (dtype == mx.float32) else 1e-3
self.assertLess((y1 - y2).abs().max(), tol)
self.assertLess((y1 - y3).abs().max(), tol)