[CI] Initialize DeepEP FP8 test weights (#49912)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
@@ -66,8 +66,14 @@ def make_weights(
|
|||||||
|
|
||||||
# per-out-channel weight quantization
|
# per-out-channel weight quantization
|
||||||
assert dtype == current_platform.fp8_dtype()
|
assert dtype == current_platform.fp8_dtype()
|
||||||
w1 = torch.empty((e, 2 * n, k), device="cuda", dtype=torch.float16)
|
# Keep FP8 inputs finite and bounded. torch.empty made this test depend on
|
||||||
w2 = torch.empty((e, k, n), device="cuda", dtype=torch.float16)
|
# allocator contents, while larger values exceed its fixed W8A8 tolerance.
|
||||||
|
w1 = torch.randn(
|
||||||
|
(e, 2 * n, k), device=current_platform.device_type, dtype=torch.float16
|
||||||
|
).div_(100)
|
||||||
|
w2 = torch.randn(
|
||||||
|
(e, k, n), device=current_platform.device_type, dtype=torch.float16
|
||||||
|
).div_(100)
|
||||||
|
|
||||||
n_b_scales = 2 * n
|
n_b_scales = 2 * n
|
||||||
k_b_scales = k
|
k_b_scales = k
|
||||||
|
|||||||
Reference in New Issue
Block a user