fix(allocator): reuse pinned free blocks

[voltjia]

Summary

• Reuse free allocator blocks even while a captured graph retains their underlying storage.
• Keep graph pins responsible for preventing premature trim or free, rather than making every pinned address exclusive to one graph.
• Update the runtime contract test to cover both size-class and large-block reuse.

Motivation

The modern-stack PagedCompiler captures decode graphs for many batch sizes. The allocator previously refused to reuse a free block whose pin_count was nonzero, so each captured graph duplicated its scratch storage. At the documented NVIDIA capacity boundary, this caused PinnableBlockAllocator allocation failures during P01 warmup and P04 reset_cache.

The same two-line allocator change was validated on NVIDIA A100-SXM4-80GB:

Workload	Before	With pinned-free reuse
P01, 9G-8B, BF16, batch 64, input 4096, output 256, TP1	Allocator error code 2 during warmup	Success, total_time=33219.85 ms
P04, FM9G-70B, BF16, batch 64, input 4096, output 256, TP8	Allocator error code 2 during reset_cache	Success, total_time=46100.18 ms

Allocator tracing for P01 showed underlying allocation events falling from 17331 to 5097 with this change.

Type of Change

• fix — bug fix

Test Results of Involved Models on Supported Platforms

Platform	Result
NVIDIA A100-SXM4-80GB	P01 and P04 capacity cases completed with the allocator reuse change; logs were retained from the diagnostic run.
Local static contracts	python -m unittest test.static.test_infinicore_runtime_contracts -v: 55/55 passed.
Local static suite	python -m unittest discover -s test/static -p 'test_*.py' -v: 104/104 passed.

Benchmark / Performance Impact

This is primarily a capacity and correctness fix. The successful diagnostic runs are listed under Motivation. P01 used an instrumented diagnostic library, so its wall time is not used for performance attribution; P04 was uninstrumented and completed in 46100.18 ms.

Notes for Reviewers

The important invariant is that graph destruction still releases its PinLease, and trim still refuses to free blocks with a nonzero pin_count. Reuse is allowed only after tensor owners mark the block !in_use; this restores sequential same-stream graph behavior without releasing addresses that captured graphs reference.

The A100 diagnostic validation used the historical modern-stack checkout that reproduced the failures plus this allocator change. Static tests and formatting were run on the exact PR head.

CI / ChatOps

CI will be triggered manually after the PR is created.

---

Checklist

Title, Branch, and Commits

• PR title follows Conventional Commits.
• Branch name follows <type>/xxx-yyyy-zzzz.
• Commit message follows Conventional Commits.
• Small PR is a single squashable commit.
• No stray merge commits from main.
• No fixup! / squash! / wip commits remain.
• Existing PR/branch/commit that followed the legacy issue format.

Scope and Design

• Changes are minimal.
• No dead code, debug prints, or ownerless TODOs.
• No unrelated formatting churn.
• No public API changes.

General Code Hygiene

• Comments explain the non-obvious ownership invariant.
• Modified files end with a newline.
• No trailing whitespace, mixed indentation, or BOMs.
• Code references in comments and messages use backticks where appropriate.
• Comments and error messages are English.
• Comments are complete sentences.

C++ Specific

• Code follows the repository formatting and style conventions.
• RAII and existing allocator ownership are preserved.
• Changed files are formatted with the repository formatter configuration.
• No changes reference csrc/models/llama_legacy/.

Python Specific

• Changed Python test code is PEP 8 compliant and formatted.
• No changes reference python/infinilm/auto_config.py.

Testing

• NVIDIA A100 P01/P04 diagnostic capacity runs passed.
• Target runtime contract test passed.
• Full local static suite passed.
• Single request test not rerun on the exact PR head; allocator behavior was covered by the two capacity workloads and static contracts.
• Offline performance test passed for the affected P01/P04 capacity cases in the diagnostic validation.
• Sanity test not rerun; this allocator fix is covered by targeted runtime contracts.
• Service test not rerun; no service API behavior was changed.

Build, CI, and Tooling

• Fresh full build and CI are pending on the PR head.

Documentation

• No user-facing API, build flag, or workflow changed.

Security and Safety

• No secrets, private hosts, customer data, or hardware identifiers are committed.
• No new unsafe ownership or pointer lifetime is introduced.