Currently, opinfo generates random inputs for _scaled_mm but does not enforce type saturation, unlike the upstream test implementation which explicitly saturates both fp8 data types.

Problem:
The current random input generation in sample_inputs_scaled_mm may produce values outside the valid range for the input types, potentially missing edge cases that the CUDA tests intentionally cover.

Solution:
Modify the sample_inputs_scaled_mm implementation to:

1. Apply the same input saturation logic used in CUDA _scaled_mm tests

   pytorch/test/test_matmul_cuda.py

   Lines 1007 to 1052 in 52e4e41

   	def test_scaled_mm_vs_emulated(self, base_dtype):
   	torch.manual_seed(42)
   	input_dtype = e4m3_type
   	output_dtype = base_dtype
   	compare_type = torch.float32
   	x = torch.randn(16, 16, device="cuda", dtype=base_dtype)
   	y = torch.randn(32, 16, device="cuda", dtype=base_dtype).t()
   	x_scale = tensor_to_scale(x, input_dtype).float()
   	y_scale = tensor_to_scale(y, input_dtype).float()
   	x_fp8 = to_fp8_saturated(x * x_scale, input_dtype)
   	y_fp8 = to_fp8_saturated(y * y_scale, input_dtype)
   	# Calculate actual F8 mm
   	out_scaled_mm = mm_float8(
   	x_fp8,
   	y_fp8,
   	a_scale=x_scale,
   	b_scale=y_scale,
   	output_dtype=output_dtype
   	)
   	# Calculate emulated F8 mm
   	out_emulated = mm_float8_emulated(
   	x_fp8,
   	x_scale,
   	y_fp8,
   	y_scale,
   	output_dtype
   	)
   	if output_dtype != base_dtype:
   	out_scaled_mm = out_scaled_mm.to(compare_type)
   	out_scaled_mm = out_scaled_mm / tensor_to_scale(out_scaled_mm, input_dtype)
   	out_emulated = out_emulated.to(compare_type)
   	out_emulated = out_emulated / tensor_to_scale(out_emulated, input_dtype)
   	if base_dtype in {torch.bfloat16, torch.float16}:
   	atol, rtol = 7e-2, 7e-2
   	else:
   	atol, rtol = 3e-3, 3e-3
   	torch.testing.assert_close(out_scaled_mm, out_emulated, atol=atol, rtol=rtol)

2. Maintain the existing random generation approach but clamp values to type-appropriate ranges.

This required a bit of alteration to adapt to the late realization of the inputs. I think I have done this correctly, but open to suggestions.