Cant get torch compile to work on windows

[TingTingin]

On windows 11 with triton, sageattention 2 and cuda 2.5.1 installed and can't seem to get torch compile to work i get the below error

torch.dynamo.exc.BackendCompilerFailed: backend='inductor' raised:
CompilationError: at 8:11:
def triton(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 56623104
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = tl.full([XBLOCK], True, tl.int1)
x0 = xindex
tmp0 = tl.load(in_ptr0 + (x0), None)
tmp1 = tmp0.to(tl.float32)

[eas125]

Are you getting a "file exists" error?

If so, there is a bug with the windows version. You need to replace some code. From Kijai for reference:

#189 (comment)

[TingTingin]

Are you getting a "file exists" error?

If so, there is a bug with the windows version. You need to replace some code. From Kijai for reference:

#189 (comment)

I got that error initially i fixed it already after fixing that i got this new error

[kijai]

This could be the error you get when trying to compile fp8 on non-40xx series GPU. I'm not sure if there's a way around it, seems to be a Triton limitation. Someone claimed it worked on Triton 3.0, but I can't confirm that. You can try using the bf16 weights and disabling quantization, and enabling auto_cpu_offload to see if that's the issue.

[OperationTwelve]

I got this working without changing any code just by setting dynamic to true

[AILova]

I got this working without changing any code just by setting dynamic to true

Holy cow, this actually fixed mine, too. Assuming, of course, that dynamic doesn't mean "if it doesn't work, just pass it thru"... ;)