Support vLLM v0.28 accelerator memory APIs

[Joey-gvwal]

Summary

• Route torch.accelerator.get_memory_info() to torch.musa.mem_get_info() only on torch_musa versions before 2.11.0.post2. The version check strips the local +musa... suffix so post2 and later keep the fixed unified accelerator implementation; missing or malformed versions conservatively retain the workaround.
• Exclude torchada-owned stable compatibility headers from in-place CUDA-to-MUSA rewriting during editable/source builds.
• Bump torchada to 0.1.83 across package metadata, runtime version, tests, benchmark history, and documentation.

Validation

S5000 / torch_musa 2.11 post1

• Candidate: 4357d66fb3f4e8128ae800cf28740294c65a69f3
• Image: sh-harbor.mthreads.com/mcctest/vllm:v0.24.0-ph1-5.2.0-torch2.11.0.post1-20260820
• Image digest: sha256:c9b7268be71cbba113696006fd84b139d7a3a0f7d1ac21b141b1631e2831e67a
• Hardware: one MTT S5000, driver 5.2.0-server
• Runtime: Python 3.10.12, torch and torch_musa 2.11.0.post1+musa5.2.0, torchada 0.1.83
• Baseline/candidate A/B: the unpatched post1 torch.accelerator.get_memory_info() reproduced NotImplementedError: getMemoryInfo is not implemented for this allocator yet; the PR candidate passed and resolved it to torch.musa.mem_get_info.
• Runtime inventory confirmed nine post1 memory overrides; empty_host_cache was absent from both modules. A real 16 MiB MUSA allocation was reflected by memory_allocated, memory_stats, peak/reset APIs, empty_cache, synchronize, and get_memory_info.
• get_memory_info passed with None, integer, string, and torch.device arguments. Same-name and remapped fallback behavior also passed when the official accelerator API was absent.
• Focused accelerator tests: 40 passed with no skips.
• Other changed-surface tests (test_inplace_porting.py, test_platform.py): 26 passed.
• Full torchada suite: 465 passed, 19 skipped, 11 warnings in 55.43s.

Other validation

• Post1/post2 version-boundary, local-suffix trimming, malformed-version fallback, post2 native-selection, and post2 missing-API fallback tests passed. The post2 selection checks are simulated until an actual 2.11.0.post2 wheel is available.
• Pre-commit hooks passed for all files changed by the review fix.
• Existing PR validation: vLLM-MUSA v0.28.0 source build/install with editable torchada on an 8-GPU MUSA 5.2.0 environment.

Scope note: this compatibility layer wraps the top-level torch.accelerator.* API used by vLLM. The separate torch.accelerator.memory.* submodule is outside this PR's scope.

[yeahdongcn]

@lijing-mt Please be aware of this issue.

[yeahdongcn]

S5000 merge-readiness validation completed at head 4357d66 on the requested torch2.11.0.post1 image (digest sha256:c9b7268...e67a). The unpatched baseline reproduced the allocator NotImplementedError; the candidate passed the top-level override/remap/fallback A/B, allocation-backed memory checks, 40 focused accelerator tests, 26 other changed-surface tests, and the full suite (465 passed, 19 skipped). The PR body now contains the exact environment and scope.