biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix: use 1-dim tensors for gate_ws2 and gate_input_scale

Browse Source
This commit is contained in:
biondizzle
2026-06-01 11:05:09 +00:00
parent ec8f292112
commit 5f38430423
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
single_shot_inference.py
View File

@@ -723,8 +723,8 @@ def main():
router.load_nvfp4_fused_gate(
gate_weight=gw_fp4,
gate_weight_scale=gw_sf,
gate_ws2=torch.tensor(gw_gs, device=dev), # gsb = weight global scale
gate_input_scale=torch.tensor(1.0 / (6.0 * 448.0), device=dev), # gsa = activation global scale
gate_ws2=torch.tensor([gw_gs], device=dev, dtype=torch.float32), # gsb = weight global scale
gate_input_scale=torch.tensor([1.0 / (6.0 * 448.0)], device=dev, dtype=torch.float32), # gsa = activation global scale
)
else:
router.load_weights(e_bias=eb.to(dev, torch.float32))