fix(cpp): add cross-process file lock for JIT kernel builds

[Ndgandhi23]

Summary

Two quantize() processes on one host share the same JIT build cache, but the lock guarding it is process-local. torch serializes builds with a lock file that has no timeout or owner check. If a process is killed mid-build, that file is left behind and every later load() waits on it forever at 0% cpu. Reproduced locally without a GPU. Surfaced while investigating #3015.

The fix is an fcntl.flock around the build step. The kernel releases the lock automatically when the holding process dies, so a killed process can never leave the cache locked. Once the lock is held, any leftover torch lock file is removed. The wait is also bounded, so a stuck holder degrades the load into the existing non-JIT fallback instead of hanging a worker.

What Changed

• gptqmodel/utils/cpp.py:
  • New _cross_process_build_lock context manager using stdlib fcntl, with a no-op fallback where fcntl is unavailable. The lock file lives outside the build dir so cache clears never delete it.
  • TorchOpsJitExtension.load() runs its cache-probe and compile section under this lock, and removes any stale torch lock file after acquiring it.
  • The lock wait is bounded by GPTQMODEL_TORCH_OPS_LOCK_TIMEOUT (default 600s, or 5x the kernel's compile baseline if larger). On timeout the load fails cleanly into the non-JIT path.
  • clear_cache() and force-rebuild delete the cache under the same lock, so a wipe cannot land under another process's in-flight build.
  • Note: most of the diff line count is the existing load() tail reindented into the new with block. It is byte-identical apart from the stale-lock removal call.
• tests/test_torch_ops_jit_extension.py: three new tests — a stale torch lock file is removed before compiling, a SIGKILLed lock holder releases the lock so load proceeds, and a live holder hits the bounded wait instead of hanging. The subprocess tests skip where fcntl is unavailable.
• README.md: new "JIT kernel cache and multi-process quantization" section documenting the cache dir and both env vars.

Tests

• I added a new simple/fast unit test for this change, or documented why that is not applicable.
• I ran the new targeted test locally before opening this PR.
• I ran any other directly relevant local tests.

Paste the exact test commands and results here:

pytest tests/test_torch_ops_jit_extension.py tests/test_extension_load_api.py tests/test_cpp_jit_progress.py tests/test_cpp_nvcc_flags.py
# 57 passed in 4.18s

cd format && ruff check --config ruff.toml \
  ../gptqmodel/utils/cpp.py \
  ../tests/test_torch_ops_jit_extension.py
# All checks passed!

git diff --check
# no output

Review Requirements

• I personally reviewed every file in this diff.
• I checked that the code matches existing project structure, APIs, and conventions.
• I avoided unnecessary monkeypatching and used the project's normal extension points where possible.

Notes

flock on NFS-mounted cache dirs has weaker advisory semantics; worst case is today's behavior.

Assisted with Claude Code; every change personally designed and reviewed.

[Qubitium]

@Ndgandhi23 LGTM. Thank you!