fix: in-place prepack to avoid 2× peak memory

torch.stack(packed) held all expert tensors + final stack (~3.5 GiB).
Now pre-allocate output and fill in-place — only 1 expert tmp + final
tensor in memory at any time.

src/nvfp4_megamoe_kernel/nvfp4_mega_moe.py

@@ -142,11 +142,23 @@ def _prepack_weight_sf(weight_sf, N, K, tag):
     # M=128 — TODO: test with M=1, M=256 to confirm.
     M_for_layout = 128
 
-    packed = []
-    for e in range(E):
-        packed.append(prepack_sfb(weight_sf[e], M_for_layout, N, K))
+    # Pre-allocate output tensor and fill in-place to avoid 2× peak memory
+    # (torch.stack would hold all expert tensors + the final stack = ~3.5 GiB)
+    packed0 = prepack_sfb(weight_sf[0], M_for_layout, N, K)
+    packed = torch.empty(
+        (E, *packed0.shape),
+        dtype=packed0.dtype,
+        device=packed0.device,
+    )
+    packed[0].copy_(packed0)
+    del packed0
 
-    packed = torch.stack(packed, dim=0).contiguous()
+    for e in range(1, E):
+        tmp = prepack_sfb(weight_sf[e], M_for_layout, N, K)
+        packed[e].copy_(tmp)
+        del tmp
+
+    packed = packed.contiguous()
     _prepack_weight_sf._cache[cache_key] = packed
     _prepack_weight_sf._cache_order.append(cache_key)