[CI] Initialize DeepEP FP8 test weights (#49912)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Jul 27, 2026 at 16:29 UTC
99115fcdcde9736d4097ea15e7bbab368c405edf
1 file changed
+8
-2
tests/kernels/moe/test_deepep_moe.py
+8
-2
@@ -66,8 +66,14 @@ def make_weights(
66
67
# per-out-channel weight quantization
68
assert dtype == current_platform.fp8_dtype()
69
- w1 = torch.empty((e, 2 * n, k), device="cuda", dtype=torch.float16)
70
- w2 = torch.empty((e, k, n), device="cuda", dtype=torch.float16)
69
+ # Keep FP8 inputs finite and bounded. torch.empty made this test depend on
70
+ # allocator contents, while larger values exceed its fixed W8A8 tolerance.
71
+ w1 = torch.randn(
72
+ (e, 2 * n, k), device=current_platform.device_type, dtype=torch.float16
73
+ ).div_(100)
74
+ w2 = torch.randn(
75
+ (e, k, n), device=current_platform.device_type, dtype=torch.float16
76
+ ).div_(100)
77
78
n_b_scales = 2 * n
79
k_b_scales = k