comfyui problem

[TheMich01]

Hello, i have some problems with the workflow on comfyui:

triton.compiler.errors.CompilationError: at 8:11:
def triton_(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 26214400
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)
^

Set TORCH_LOGS="+dynamo" and TORCHDYNAMO_VERBOSE=1 for more information

You can suppress this exception and fall back to eager by setting:
import torch._dynamo
torch._dynamo.config.suppress_errors = True

Prompt executed in 12.35 seconds

I tried different models, and different library, but it end up alwaiys like this, i'm using ubuntu with a 3090
thanks for your help

[TheMich01]

I made more check using google ai studio, grok, and chagpt the made me do some test, and then i made them wirite all the explanation:

Title: Triton CompilationError when running this node (PyTorch 2.5.1 + Triton 3.1.0, CUDA 12.1)

Body:

Hi,

I’m running into a Triton compilation error when using this node inside ComfyUI.
Here is the error I get:

triton.compiler.errors.CompilationError: at 8:11:
def triton_(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
    xnumel = 26214400
    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)
           ^

PyTorch suggests that I can suppress this by enabling eager mode with:

import torch._dynamo
torch._dynamo.config.suppress_errors = True

but that would just bypass the issue rather than fixing it.

---

My environment

• Python: 3.12 (Conda env)
• Torch: 2.5.1+cu121
• Triton: 3.1.0
• CUDA runtime: 12.1
• GPU: [please add your GPU model + driver version from nvidia-smi]

---

My verification tests

To check if Triton itself was working correctly, I ran the following test script:

import torch
import triton
import triton.language as tl

@triton.jit
def add_kernel(x_ptr, y_ptr, output_ptr, n_elements, BLOCK: tl.constexpr):
    pid = tl.program_id(0)
    block_start = pid * BLOCK
    offsets = block_start + tl.arange(0, BLOCK)
    mask = offsets < n_elements
    x = tl.load(x_ptr + offsets, mask=mask)
    y = tl.load(y_ptr + offsets, mask=mask)
    tl.store(output_ptr + offsets, x + y, mask=mask)

x = torch.ones(1024, device="cuda", dtype=torch.float32)
y = torch.ones(1024, device="cuda", dtype=torch.float32)
z = torch.empty_like(x)

add_kernel[(1024 // 256,)](x, y, z, 1024, BLOCK=256)
print("OK Triton, result:", z[:5])

The output is:

OK Triton, result: tensor([2., 2., 2., 2., 2.], device='cuda:0')

So Triton works fine in my setup. This suggests the issue is specific to how this node generates kernels (likely via torch.compile), rather than a general PyTorch/Triton/driver problem.

---

Question

Could you check if this node uses an unsupported pattern for Triton?
Maybe some type casting (.to(tl.float32)) inside the generated kernel is not being handled properly.

Thanks for looking into this!

[StrongerXi]

PyTorch 2.5.1 + Triton 3.1.0

I feel like I've seen this issue before, for older PyTorch version. Could you try upgrading your pytorch version, e.g., to 2.8?

[TheMich01]

I solved the problem bypassing the wanvideo torch compile settings, disabling this node i can use the workflow, i'm still trying different settings, i also upgrade the pythorch, but i do not find any solution with this node active.
sorry for the delay in answering, i made some test and i had some issues to solve.
thanks for your attention, I apreciate a lot your job