pytorch-rocm-llms: unify transformers pin and disable mmap for halo gpt-oss-20b - #659
Merged
Conversation
…mmap on halo Windows halo loaded gpt-oss-20b via safetensors mmap, which exhausts the Windows paged pool on the ~40GB bf16 model (OSError 1450). Add disable_mmap to read weights into RAM instead. Also unify the transformers pin at 5.10.1 across all devices (halo-windows was stale at 4.57.1, which lacks disable_mmap; halo-linux and others-linux were unpinned). Possibly fixes #655.
lucbruni-amd
marked this pull request as draft
July 29, 2026 15:03
lucbruni-amd
marked this pull request as ready for review
July 29, 2026 15:15
Collaborator
|
have we seen perf improvements by disabling mmap? As well, there were some previous errors with the later versions of transformers. lets keep an eye on the CI in case anything starts to fail. |
adamlam2-amd
approved these changes
Aug 10, 2026
Collaborator
|
!orc |
Collaborator
Author
No, if anything it may have a slight model loading cost. Let's merge and monitor CI if you approve. |
OrchestrAI results — ❌ failures (6 passed, 4 failed)
tested |
Collaborator
|
!orc |
Same OSError 1450 mmap failure as run-model when loading gpt-oss-20b on Windows/halo; run-llm-simple runs run_llm.py, so pin it to a full read.
run-summarizer loads gpt-oss-20b on windows/halo and crashed on the mmap'd load; align it with run-model and run_llm.py.
Collaborator
Author
|
Yes. |
Collaborator
|
lgtm |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
On Windows, halo loads
gpt-oss-20b(dequantized to ~40GB bf16 since MXFP4/Triton kernels aren't available on Windows ROCm).safetensorsmemory-maps the shards, and a mapping that large intermittently exhausts the Windows paged pool:OSError 1450 (ERROR_NO_SYSTEM_RESOURCES). Not a RAM limit (halo has 128GB).Fix: pass
disable_mmap=Trueto the halorun-modelload so weights are read into RAM instead of mmapped. Available in transformers 5.x.Also unify the transformers pin at
5.10.1across all devices/OSes: halo-windows was stale at4.57.1(which lacksdisable_mmap), halo-linux and others-linux were floating.Possibly fixes #655.