P4: Fix dequantize_nvfp4 bridge — handle float8_e4m3fn dtype

dsv4/ops/quantize.py

@@ -367,6 +367,11 @@ def dequantize_nvfp4(x_fp4, x_sf, gsa, shape=None):
     mod = get_cuda_module("dequant_nvfp4", ["dequant_nvfp4.cu"])
     if gsa.dim() == 2:
         gsa = gsa.squeeze(1)  # (M, 1) → (M,)
+    # dequant kernel expects uint8 for both fp4 and sf
+    if x_fp4.dtype != torch.uint8:
+        x_fp4 = x_fp4.view(torch.uint8)
+    if x_sf.dtype != torch.uint8:
+        x_sf = x_sf.view(torch.uint8)
     return mod.dequant_nvfp4(x_fp4, x_sf, gsa)