transformers

5bed6296 - [MusicgenMelody] Fix conditioning silently dropped at generation step 0 (#48679)

[MusicgenMelody] Fix conditioning silently dropped at generation step 0 (#48679) * [MusicgenMelody] Fix conditioning silently dropped at generation step 0 PR #43679 (`_prepare_cache_for_generation`) pre-creates a DynamicCache (or EncoderDecoderCache) before the first generation step, so `past_key_values` is no longer None at step 0. The guard if past_key_values is not None: encoder_hidden_states = None # drop conditioning in `MusicgenMelodyForConditionalGeneration.prepare_inputs_for_generation` then silently drops all conditioning at step 0, making every generation call effectively unconditioned regardless of the audio/text inputs. Fix: only suppress `encoder_hidden_states` once the self-attention cache actually contains tokens (`past_length > 0`). An empty pre-created cache at step 0 must pass conditioning through. Also apply the symmetric fix to `MusicgenMelodyForCausalLM` and `MusicgenForCausalLM` (`input_ids` trimming guard), and update the two stale sampling test golden values that reflected the broken unconditioned outputs. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Revert unnecessary tolerance widening in sampling tests The two sampling tests use set_seed so they are deterministic on a given runner; the captured golden values should match within the original 1e-4 / 2e-4 tolerances. The looser 1e-3 was overly permissive and could mask real future regressions. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [MusicgenMelody] Update all golden values to current CI environment (torch 2.11) Update expected values for all 10 integration tests to match outputs from the fixed model on the current runner (torch 2.11 / post-April-5 environment). The 8 deterministic tests already passed with the code fix alone (max_abs_diff well within atol), but are updated here for consistency. The 2 sampling tests (unconditional_sampling, text_prompt_sampling) were failing since 2026-04-05 due to the unrelated torch 2.10 -> 2.11 numerical drift and required new values. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Revert modeling_musicgen.py change; fix stereo right golden value to match CI runner - Revert the defensive fix to MusicgenForCausalLM (base Musicgen is unaffected by the conditioning bug; melody does not use that model) - Revert EXPECTED_VALUES_RIGHT_SECOND_SAMPLE to original values — our test runner produced slightly different outputs than the CI runner for this slice Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * up-to-date code -> local slow CI run green --------- Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> Co-authored-by: raushan <raushan@huggingface.co>