transformers

df04b012 - [DeepseekV3] OOM cascade root-cause investigation (generator ref leak in conversion_mapping) (#48720)

[DeepseekV3] OOM cascade root-cause investigation (generator ref leak in conversion_mapping) (#48720) * [DeepseekV3] investigation placeholder: OOM cascade from generator ref leak in conversion_mapping.py Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [DeepseekV3] free GPU memory after tests loading real checkpoint to prevent OOM cascade `test_eager_matches_sdpa_generate` loads two copies of `bzantium/tiny-deepseek-v3` (~10 GB each). When it OOMs during generation, the models are not freed before the next test runs, causing a cascade of spurious OOM failures in the suite. - Wrap the generate loop in try/finally so `del model_eager, model_sdpa` + `cleanup(gc_collect=True)` always runs, even when the test fails mid-way. - Add `del model` + `cleanup(gc_collect=True)` at the end of `test_compile_static_cache`. - Change `DeepseekV3IntegrationTest.tearDown` from `gc_collect=False` to `gc_collect=True` so reference cycles are broken between integration tests. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [DeepseekV3] add inline comments explaining cleanup rationale (see PR #48720) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * revert integration test changes (will use MemoryCleanupMixin in a follow-up) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [DeepseekV3] update stale golden values in test_compile_static_cache The expected strings went stale due to torch/env drift on CI runners. The model (bzantium/tiny-deepseek-v3) is untrained so output is gibberish that changes across environments. Values refreshed from latest main CI run (2026-09-11). See PR #48720. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [DeepseekV3] add TODO on compiled static cache path (never passed since added) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>