[ROCm] fix dtype (#4510)

python/sglang/srt/layers/quantization/w8a8_fp8.py

@@ -108,10 +108,15 @@ class W8A8Fp8LinearMethod(LinearMethodBase):
                     layer.weight, layer.weight.shape[-1]
                 )
                 weight_scale = weight_scale.t().contiguous()
+                if _is_hip:
+                    weight, weight_scale, _ = normalize_e4m3fn_to_e4m3fnuz(
+                        weight=weight, weight_scale=weight_scale
+                    )
             else:
                 # if cutlass not supported, we fall back to use torch._scaled_mm
                 # which requires per tensor quantization on weight
-                qweight, weight_scale = input_to_float8(layer.weight)
+                fp8_dtype = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn
+                qweight, weight_scale = input_to_float8(layer.weight, dtype=fp8_dtype)
 
             # Update the layer with the new values.
             layer.weight = Parameter(qweight.t(), requires_grad=False)