Triton error on a GFX1100 while using TheRock nightly wheels on Windows.

[alexsarmiento]

I am getting "Pointer argument (at 0) cannot be accessed from Triton (cpu tensor?)" error. I tried various versions of triton from https://github.com/lshqqytiger/triton and pytorch from https://d2awnip2yjpvqn.cloudfront.net/v2/gfx110X-dgpu/.

Also tried with my local builds of these, in python 3.11 and 3.12, micromamba environments. I tried: disabling the integrated gpu in the BIOS, setting HIP_VISIBLE_DEVICES=1, and using torch.cuda.set_device(). Tried with amdclang, clang, clang-cl.

I am not getting that error with the pytorch 2.7 from scottt/rocm-TheRock builds for my gfx1100

I wonder if anyone else can reproduce this problem to rule out an environment issue on my part.

For instance:

$ python minimal.py

import torch

@torch.compile()
def f(x, y):
    return x + y.sum()

f(torch.randn(3, device='cuda'), torch.randn(3))

Results:

HIP Library Path: C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\_rocm_sdk_core\bin\amdhip64_7.dll
Traceback (most recent call last):
  File "C:\Users\sarmi\repo\triton\python\tutorials\minimal.py", line 7, in <module>
    f(torch.randn(3, device='cuda:0'), torch.randn(3))
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_dynamo\eval_frame.py", line 804, in compile_wrapper
    return fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^
  File "C:\Users\sarmi\repo\triton\python\tutorials\minimal.py", line 3, in f
    @torch.compile()
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_dynamo\eval_frame.py", line 1005, in _fn
    return fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_functorch\aot_autograd.py", line 1130, in forward
    return compiled_fn(full_args)
           ^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_functorch\_aot_autograd\runtime_wrappers.py", line 353, in runtime_wrapper
    all_outs = call_func_at_runtime_with_args(
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_functorch\_aot_autograd\utils.py", line 129, in call_func_at_runtime_with_args
    out = normalize_as_list(f(args))
                            ^^^^^^^
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_functorch\_aot_autograd\runtime_wrappers.py", line 526, in wrapper
    return compiled_fn(runtime_args)
           ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_inductor\output_code.py", line 588, in __call__
    return self.current_callable(inputs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_inductor\utils.py", line 2897, in run
    out = model(new_inputs)
          ^^^^^^^^^^^^^^^^^
  File "C:\Users\sarmi\AppData\Local\Temp\torchinductor_sarmi\st\cstd7ee6t4lfuid2v7smlwa4qat3zqkhx4vwd7uk42evomqn3tzt.py", line 129, in call
    triton_poi_fused_add_1.run(arg1_1, buf0.item(), buf1, 3, stream=stream0)
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_inductor\runtime\triton_heuristics.py", line 1242, in run
    return launcher(
           ^^^^^^^^^
  File "<string>", line 5, in launcher
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\triton\backends\amd\driver.py", line 548, in __call__
    self.launch(self.launch_cooperative_grid, *args)
ValueError: Pointer argument (at 0) cannot be accessed from Triton (cpu tensor?)

Same problem with the scripts at "triton/python/tutorials"

[jammm]

f(torch.randn(3, device='cuda'), torch.randn(3))
The first argument is on GPU while the second one is on the CPU. Shouldn't both be on the GPU?

[alexsarmiento]

@jammm

pip list
...
rocm                            7.0.0rc20250820
rocm-sdk-core                   7.0.0rc20250820
rocm-sdk-devel                  7.0.0rc20250820
rocm-sdk-libraries-gfx110X-dgpu 7.0.0rc20250820
...
torch                           2.9.0a0+rocmsdk20250820
triton                          3.4.0+gita9c80202
...

$env:HIP_PATH = "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\_rocm_sdk_core"
$env:CC = "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\_rocm_sdk_core\lib\llvm\bin\clang.exe"
$env:HIP_VISIBLE_DEVICES = "1"

python.exe .\minimal.py

import torch

@torch.compile()
def f(x, y):
    return x + y.sum()

f(torch.randn(3, device='cuda'), torch.randn(3,device="cuda"))

Same error.

python.exe .\minimal.py
HIP Library Path: C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\_rocm_sdk_core\bin\amdhip64_7.dll
Traceback (most recent call last):
  File "C:\Users\sarmi\repo\triton\python\tutorials\minimal.py", line 7, in <module>
    f(torch.randn(3, device='cuda'), torch.randn(3,device="cuda"))
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_dynamo\eval_frame.py", line 804, in compile_wrapper
    return fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^
  File "C:\Users\sarmi\repo\triton\python\tutorials\minimal.py", line 3, in f
    @torch.compile()
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_dynamo\eval_frame.py", line 1005, in _fn
    return fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_functorch\aot_autograd.py", line 1130, in forward
    return compiled_fn(full_args)
           ^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_functorch\_aot_autograd\runtime_wrappers.py", line 353, in runtime_wrapper
    all_outs = call_func_at_runtime_with_args(
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_functorch\_aot_autograd\utils.py", line 129, in call_func_at_runtime_with_args
    out = normalize_as_list(f(args))
                            ^^^^^^^
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_functorch\_aot_autograd\runtime_wrappers.py", line 724, in inner_fn
    outs = compiled_fn(args)
           ^^^^^^^^^^^^^^^^^
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_functorch\_aot_autograd\runtime_wrappers.py", line 526, in wrapper
    return compiled_fn(runtime_args)
           ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_inductor\output_code.py", line 588, in __call__
    return self.current_callable(inputs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_inductor\utils.py", line 2897, in run
    out = model(new_inputs)
          ^^^^^^^^^^^^^^^^^
  File "C:\Users\sarmi\AppData\Local\Temp\torchinductor_sarmi\aj\cajeudapnkvyraymt2ffivmw4luto3mbsflk4frref47v7w7jnfa.py", line 108, in call
    triton_poi_fused_add_sum_0.run(arg1_1, arg0_1, buf0, 3, stream=stream0)
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\torch\_inductor\runtime\triton_heuristics.py", line 1240, in run
    return launcher(
           ^^^^^^^^^
  File "<string>", line 5, in launcher
  File "C:\Users\sarmi\micromamba\envs\py12\Lib\site-packages\triton\backends\amd\driver.py", line 548, in __call__
    self.launch(self.launch_cooperative_grid, *args)
ValueError: Pointer argument (at 0) cannot be accessed from Triton (cpu tensor?)

The same error manifests with the triton\python\tutorials examples, Comfyui, etc.
But torch.compile works with the https://github.com/scottt/rocm-TheRock/releases/tag/v6.5.0rc-pytorch-gfx110x wheels.

My question is whether torch.compile is broken just for me, and if someone can confirm it works for them. Can you try to reproduce with any torch.compile code?

[ScottTodd]

FWIW, I get these errors using the CI-produced wheels (without aotriton) on gfx1100:

(3.12.venv) λ pip freeze
filelock==3.19.1
fsspec==2025.7.0
Jinja2==3.1.6
MarkupSafe==3.0.2
mpmath==1.3.0
networkx==3.5
numpy==2.3.2
rocm==7.0.0rc20250821
rocm-sdk-core==7.0.0rc20250821
rocm-sdk-libraries-gfx110X-dgpu==7.0.0rc20250821
setuptools==80.9.0
sympy==1.14.0
torch==2.9.0a0+rocm7.0.0rc20250821
typing_extensions==4.14.1



(3.12.venv) λ cat D:\scratch\pytorch\minimal.py
import torch

@torch.compile()
def f(x, y):
    return x + y.sum()

f(torch.randn(3, device="cuda"), torch.randn(3, device="cuda"))



(3.12.venv) λ python D:\scratch\pytorch\minimal.py
HIP Library Path: D:\scratch\therock\3.12.venv\Lib\site-packages\_rocm_sdk_core\bin\amdhip64_7.dll
Traceback (most recent call last):
  File "D:\scratch\pytorch\torch_compile_demo.py", line 9, in <module>
    f(torch.randn(3, device="cuda"), torch.randn(3, device="cuda"))
  File "D:\scratch\therock\3.12.venv\Lib\site-packages\torch\_dynamo\eval_frame.py", line 817, in compile_wrapper
    raise e.remove_dynamo_frames() from None  # see TORCHDYNAMO_VERBOSE=1
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\scratch\therock\3.12.venv\Lib\site-packages\torch\_inductor\scheduler.py", line 4284, in create_backend
    raise TritonMissing(inspect.currentframe())
torch._inductor.exc.TritonMissing: Cannot find a working triton installation. Either the package is not installed or it is too old. More information on installing Triton can be found at: https://github.com/triton-lang/triton

Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"

That seems in line with "Running TorchInductor on GPU requires Triton" from https://docs.pytorch.org/tutorials/intermediate/torch_compile_tutorial.html#basic-usage

[alexsarmiento]

Yes, I am using the triton wheels from lee's builds. https://github.com/lshqqytiger/triton/releases

[alexsarmiento]

I have tested the amd's pytorch preview https://gpuopen.com/learn/pytorch-windows-amd-llm-guide/ , and torch compile inductor"works" with triton. And by "works" I mean the error ""Pointer argument (at 0) cannot be accessed from Triton (cpu tensor?)"" is not present with these wheels. There might be other issues but not this one.

[benlann]

I got this error last time I tried ROCm7rc with lshqqytiger's triton, couple minutes into wanvideosampler:

torch._inductor.exc.InductorError: RuntimeError: LLD failed to link hsaco source

[taylding-amd]

Hi @alexsarmiento , do you still see this issue in the nightly PyTorch wheel? (from https://pytorch.org/)

pip3 install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/rocm7.0

[alexsarmiento]

Hi @alexsarmiento , do you still see this issue in the nightly PyTorch wheel? (from https://pytorch.org/)

pip3 install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/rocm7.0

No, I am only testing rocm pytorch builds for windows. That repo is missing such builds.