obscura/vllm
2
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2026-08-10 07:48:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CI] Initialize DeepEP FP8 test weights (#49912)

Browse Source 
Signed-off-by: Andreas Karatzas <[email protected]>
This commit is contained in:
Andreas Karatzas
2026-07-28 05:29:08 +08:00
committed by GitHub
parent 1053e248f0
commit 99115fcdcd
1 changed files with 8 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/kernels/moe/test_deepep_moe.py
+8 -2
View File
@@ -66,8 +66,14 @@ def make_weights(
# per-out-channel weight quantization
assert dtype == current_platform.fp8_dtype()
w1 = torch.empty((e, 2 * n, k), device="cuda", dtype=torch.float16)
w2 = torch.empty((e, k, n), device="cuda", dtype=torch.float16)
# Keep FP8 inputs finite and bounded. torch.empty made this test depend on
# 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
k_b_scales = k