Skip to content

Raise stable ABI minimum to PyTorch 2.11 and fix locality allocation lifetime - #34

Merged
Harry-Chen merged 1 commit into
vllm-project:devfrom
cleonard530:migrate_to_stable_abi_cleanup
Oct 13, 2026
Merged

Harry-Chen merged 1 commit into
vllm-project:devfrom
cleonard530:migrate_to_stable_abi_cleanup

Conversation

@cleonard530

Copy link
Copy Markdown

The stable ABI migration (#21) retained localized GPU allocations until process exit because its PyTorch 2.10 from_blob call had no deleter. This follow-up raises the minimum to PyTorch 2.11 and releases those allocations when their tensors are freed. It also restores the pre-migration stride_c_m = 0 setting in two SM120 BF16 GEMM paths (csrc/jit_kernels/impls/sm120_bf16_gemm.hpp).

Updated the build, wheel metadata, CI, and documentation for the 2.11 minimum. The extension builds successfully, imports with the expected locality operators, and passes torch-abi-audit with no unstable symbols.

cc @Harry-Chen

…locations when their tensors are freed. Restore the original SM120 C-stride setting.

Co-authored-by: codex

Signed-off-by: Chris Leonard <chleonar@redhat.com>
@Harry-Chen
Harry-Chen merged commit bd2e200 into vllm-project:dev Oct 13, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants