gistfile1.txt

Here's the repro:

  import torch
  torch._dynamo.config.capture_scalar_outputs = True

  def fn(x, val_tensor):
      val = val_tensor.item()  # Creates an unbacked float (fp64)
      scaled = val * 2.0       # fp64 * fp64 = fp64 computation
      return x * scaled        # fp32 * fp64 -> needs downcast

  x = torch.ones(4, device='cuda', dtype=torch.float32)
  val_tensor = torch.tensor(0.1, device='cuda', dtype=torch.float64)

  compiled = torch.compile(fn)
  out = compiled(x, val_tensor)
  print(f'Output: {out}')

  Generated Triton kernel shows fp64 computation traced in:

  tmp1 = tl.load(in_ptr1 + (0))          # Load fp64 value from tensor
  tmp2 = tl.broadcast_to(tmp1, [XBLOCK])
  tmp3 = tl.full([1], 2.0, tl.float64)   # fp64 constant  
  tmp4 = tmp3 * tmp2                      # fp64 * fp64 computation
  tmp5 = tmp4.to(tl.float32)              # Downcast to fp32
  tmp6 = tmp0 * tmp5                      # fp32 * fp32

  Run with TORCH_LOGS="+output_code" to see the generated kernel.