Issue Station
Sign In
Tools hereToolbox

pytorch - 💡(How to fix) Fix [vllm] [triton 3.7] "Pointer argument cannot be accessed from Triton (cpu tensor?)" on multimodal image preprocessing

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…

With triton==3.7.0 (torch 2.12), a Triton kernel invoked in vLLM's multimodal image preprocessing path now raises:

ValueError: Pointer argument cannot be accessed from Triton (cpu tensor?)

Same path passed on triton 3.6 (torch 2.11). Blocking the torch 2.12 upgrade for vLLM (vllm-project/vllm#40077).

Error Message

E   ValueError: Pointer argument cannot be accessed from Triton (cpu tensor?)

Root Cause

With triton==3.7.0 (torch 2.12), a Triton kernel invoked in vLLM's multimodal image preprocessing path now raises:

ValueError: Pointer argument cannot be accessed from Triton (cpu tensor?)

Same path passed on triton 3.6 (torch 2.11). Blocking the torch 2.12 upgrade for vLLM (vllm-project/vllm#40077).

Code Example

E   ValueError: Pointer argument cannot be accessed from Triton (cpu tensor?)
RAW_BUFFERClick to expand / collapse

Summary

With triton==3.7.0 (torch 2.12), a Triton kernel invoked in vLLM's multimodal image preprocessing path now raises:

ValueError: Pointer argument cannot be accessed from Triton (cpu tensor?)

Same path passed on triton 3.6 (torch 2.11). Blocking the torch 2.12 upgrade for vLLM (vllm-project/vllm#40077).

Environment

  • torch: 2.12.0+cu130
  • triton: 3.7.0
  • CUDA: 13.0
  • Python: 3.12
  • GPU: 2× H100

Affected tests (all same ValueError)

From tests/models/multimodal/generation/test_common.py:

  • test_single_image_models_heavy[chameleon-broadcast-test_case0]
  • test_single_image_models_heavy[chameleon-broadcast-test_case1]
  • test_single_image_models_heavy[llava-broadcast-test_case2]
  • test_single_image_models_heavy[llava-broadcast-test_case3]
  • test_single_image_models_heavy[llava_next-broadcast-test_case4]
  • test_single_image_models_heavy[llava_next-broadcast-test_case5]

Error

E   ValueError: Pointer argument cannot be accessed from Triton (cpu tensor?)

Question / diagnosis

Did triton 3.7 tighten the check that all pointer arguments live on the launch device? If so, the check is catching a real bug on vLLM's side (a CPU tensor reaching a GPU kernel) that triton 3.6 silently tolerated — OR triton 3.7 is mis-classifying a device tensor. A CPU-only repro (small triton kernel + torch.empty(..., device='cpu')) would tell us which — if there's a known reference repro for this check, it'd be useful to link.

Links

cc @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @aakhundov @coconutruben @jataylo

extent analysis

TL;DR

The issue can be resolved by ensuring that all pointer arguments live on the launch device, potentially by moving CPU tensors to the GPU before invoking the Triton kernel.

Guidance

  • Verify if the error is due to a real bug on vLLM's side by creating a CPU-only repro with a small Triton kernel and torch.empty(..., device='cpu').
  • Check if Triton 3.7 is mis-classifying a device tensor by comparing the behavior with Triton 3.6.
  • Ensure that all tensors passed to the Triton kernel are on the same device (either CPU or GPU) to avoid the ValueError.
  • Review the code changes between Triton 3.6 and 3.7 to understand the tightened check on pointer arguments.

Example

No code snippet is provided as it is not clearly supported by the issue.

Notes

The issue seems to be related to the change in Triton 3.7, and resolving it requires understanding the new behavior and ensuring that the code complies with it.

Recommendation

Apply a workaround by moving CPU tensors to the GPU before invoking the Triton kernel, as this is likely to resolve the issue without requiring an upgrade to a fixed version of Triton.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…
#api#installation#tensor shape#autograd error#model save/load

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING

pytorch - 💡(How to fix) Fix [vllm] [triton 3.7] "Pointer argument cannot be accessed from Triton (cpu tensor?)" on multimodal image preprocessing