transformers

39f81b04 - Add integration tests for MuseGlimmerAssistantModel (#48796)

Add integration tests for MuseGlimmerAssistantModel (#48796) * Add integration tests for MuseGlimmerAssistantModel Two tests: 1. test_drafter_forward_output_shape: standalone drafter forward with synthetic inputs — checks output shape [1, block_size, hidden_size] and all-finite values. No main model needed. 2. test_dflash_speculative_generation: end-to-end DFlash pipeline with main model + drafter via generate(..., speculation_type="dflash"). Expected prefix is identical to the greedy-decode reference in MuseGlimmerIntegrationTest (DFlash is lossless by design). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Fix imports: cleanup/slow/require_torch_accelerator from testing_utils Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Use 70% GPU memory cap in get_model() to avoid forward-pass OOM Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Use 80% GPU cap + no disk offload to avoid meta-tensor OOM in DFlash test Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Try 90% GPU memory cap Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Load main model in 4-bit to fit on single 24 GiB accelerator Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * style: ruff line-length fix Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Use MemoryCleanupMixin instead of manual setUp/tearDown cleanup calls Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Drop redundant tearDownClass: MemoryCleanupMixin handles it via __init_subclass__ Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Move expected string into test; add clarifying comments Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Use Expectations class; compare full completion; clarify prompt-strip slice Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Remove stale comment about 70% memory cap (now using 4-bit loading) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * style: ruff line-length fix Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Address vasqu review: trim docstrings and inline comments Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Remove with torch.no_grad() from integration test Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Remove test_drafter_forward_output_shape (unit-test concern, not integration) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Trim docstring; move reference note to inline comment Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Keep Expectations key on one line with fmt: off/on Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Move fmt: off/on around the full Expectations block Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Remove verbose comment from get_model Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>