Skip to content

v2.12.2

Choose a tag to compare

View all tags
@torchlens-release torchlens-release released this 27 Apr 20:05
· 18075 commits to main since this release
v2.12.2
b9c5a1a

v2.12.2 (2026-04-27)

This release is published under the GPL-3.0-only License.

Bug Fixes

  • perf: Cache bytecode column offsets, fast-skip empty-branch attribution, guard CUDA probes on CPU runs (b61202d)

Bundle three independent performance fixes from the 2026-04-27 profiling audit (.project-context/research/profiling_audit_2026-04-27.md). All three are localized and behaviorally identical to the slow paths they short-circuit.

  1. Cache _get_col_offset() per code object id. The disassembled instruction-offset -> column-offset map is immutable for a given code object, so we build it once and reuse it. On GPT-2 the audit measured ~16.5s self time across dis.* (~50% of profiled runtime); the cache reduces repeated work to a dict lookup. A soft 100K-entry cap with a one-shot warning protects against pathological workloads; real-world models stay well under the cap.

  2. Fast-skip Step 5 branch attribution when no conditional bools were captured. The slow path's only branch-attributing input is ModelLog.internally_terminated_bool_layers; if it is empty the downstream collections all resolve to their empty defaults (already set in ModelLog.init and capture/output_tensors.py). The precondition checks every derived ModelLog conditional collection so any caller that pre-populated them still triggers the slow path. A defensive assert on the slow path will trip if a future bool-detector refactor produces conditional keys without populating internally_terminated_bool_layers, preventing silent fast-path drift.

  3. Gate torch.cuda.empty_cache() calls in postprocess and capture paths behind the cached _is_cuda_available() helper. CUDA availability is process-fixed, so once cached the helper short-circuits without touching the CUDA driver / NVML probes that the audit measured at ~12% overhead on CPU-only ResNet50.

Behavior is unchanged. ModelLog output is byte-identical for the test fixtures. tests/test_perf_bundle.py adds 11 behavioral tests covering each fix; smoke + non-slow suites pass without regression.

Detailed Changes: v2.12.1...v2.12.2

Assets 4
Loading