fix: _fold_global_scale — remove broken logical_widths branch

The logical_widths branch took expert 0 and 1's global scales and
applied them to ALL experts. For L1 with logical_widths=[3072,3072],
every expert got expert-0's scale on its gate half and expert-1's
scale on its up half. All other experts' global scales were discarded.

The else branch correctly broadcasts each expert's own (E,1) global
scale across (E, N, K//16). Removed the dead logical_widths code.

src/nvfp4_megamoe_kernel/weight_transform.py

@@ -22,25 +22,19 @@ import torch
 
 def _fold_global_scale(
     weight_scale: torch.Tensor,      # (E, N, K//16) float8_e4m3fn
-    weight_scale_2: torch.Tensor,     # (E, num_logical) or (E, 1) or scalar float32
-    logical_widths: list[int] = None,
+    weight_scale_2: torch.Tensor,     # (E, 1) or (E,) or scalar float32
 ) -> torch.Tensor:
-    """Fold global scale into block scales: UE4M3 * FP32 → float32."""
+    """Fold global scale into block scales: UE4M3 * FP32 → float32.
+    
+    Each expert has its own global scale. Broadcasts (E,1,1) → (E, N, K//16).
+    """
     sf_f32 = weight_scale.to(torch.float32)
     gs = weight_scale_2.to(torch.float32)
 
     if gs.numel() == 1:
         sf_f32 = sf_f32 * gs
-    elif gs.numel() > 1 and logical_widths is not None:
-        expanded = []
-        for i, w in enumerate(logical_widths):
-            if i < len(gs.flatten()):
-                expanded.append(gs.flatten()[i].expand(w))
-        if expanded:
-            global_scale = torch.cat(expanded).unsqueeze(1)
-            sf_f32 = sf_f32 * global_scale
     else:
-        # Per-expert or per-row global scale — broadcast multiply
+        # Per-expert global scale — broadcast multiply
         while gs.dim() < sf_f32.dim():
             gs = gs.unsqueeze(-1)
         sf_f32 = sf_f32 * gs.expand_as(sf_f32)
@@ -82,9 +76,12 @@ def transform_nvfp4_weights_for_mega_moe(
     l2_weight, l2_weight_scale = l2_tuple
 
     # Fold global scales into block scales
+    # Both L1 and L2 use per-expert global scales (shape (E,1) or (E,)).
+    # The logical_widths branch was wrong: it treated gs as per-projection
+    # scalars and only used experts 0 and 1's scales for ALL experts.
+    # The else branch correctly broadcasts each expert's own global scale.
     if l1_weight_scale_2 is not None:
-        l1_sf_folded = _fold_global_scale(l1_weight_scale, l1_weight_scale_2,
-                                           logical_widths=[3072, 3072])
+        l1_sf_folded = _fold_global_scale(l1_weight_scale, l1_weight_scale_2)
     else:
         l1_sf_folded = l1_weight_scale.to(torch.float32)