Review cloud MIC baselines fcbf7cf0b35b43b08bea08c7da6d42e3
Measured and independently validated on CPU and CUDA. Review reasons: {"cpu/control/Wav2Vec2/mic-wav2vec2-base-960h-slim/bfloat16": {"reason": "New or unchanged measured reference", "sha256": "f38140ed3e7f3508b0335a953d54880cc0521946ba958900393dae5947bd8ca2", "verdict": "new"}, "cpu/control/Wav2Vec2/mic-wav2vec2-base-960h-slim/default": {"reason": "New or unchanged measured reference", "sha256": "79b34e4278035b20d4c685c97517f3325cf23dfb71dfba95426678cc96c6359f", "verdict": "unchanged"}, "cpu/control/Wav2Vec2/mic-wav2vec2-base-960h-slim/eager": {"reason": "New or unchanged measured reference", "sha256": "18767c5e69accd95d52d042a91c652fc13a16d7e1042652d8d395f3b4fb3f2a4", "verdict": "unchanged"}, "cpu/prototypes/mic-wav2vec2/mic-wav2vec2-slim/bfloat16": {"reason": "New or unchanged measured reference", "sha256": "344f37bc6fdcad4da692fd12d3c94d6d980922ae2a06849ee186ea75236b2455", "verdict": "unchanged"}, "cpu/prototypes/mic-wav2vec2/mic-wav2vec2-slim/default": {"reason": "New or unchanged measured reference", "sha256": "98d3c9214a02c648ddefb9af091489e15ac2fd29d7b480eb72eddab55f1312e4", "verdict": "unchanged"}, "cpu/prototypes/mic-wav2vec2/mic-wav2vec2-slim/eager": {"reason": "New or unchanged measured reference", "sha256": "6ddd79e860174f46dd18c9624f12d008fb322bb93dd1df8485628a54ccf74e0d", "verdict": "unchanged"}, "cuda/control/Wav2Vec2/mic-wav2vec2-base-960h-slim/bfloat16": {"reason": "wav2vec2 fixtures re-recorded in the Transformers CI images (torch 2.14) at transformers e487eaf4, which includes #49386 (masked_spec_embed initialised, fixes #49385). Training loss and gradients now finite in every case, including CPU bf16 (control bf16 baseline is new). CPU float32 drift <=4e-7; CUDA float32 logits <=6.2e-4 and loss +0.01 (first CUDA re-record since torch 2.11); CUDA bf16 loss 565.8->560.0 (1.5 bf16 ulp at that magnitude).", "sha256": "b43ea5225b0a223b164c787065acdb6a2340367cd8651bfeaf03ca9c36c924fe", "verdict": "changed"}, "cuda/control/Wav2Vec2/mic-wav2vec2-base-960h-slim/default": {"reason": "wav2vec2 fixtures re-recorded in the Transformers CI images (torch 2.14) at transformers e487eaf4, which includes #49386 (masked_spec_embed initialised, fixes #49385). Training loss and gradients now finite in every case, including CPU bf16 (control bf16 baseline is new). CPU float32 drift <=4e-7; CUDA float32 logits <=6.2e-4 and loss +0.01 (first CUDA re-record since torch 2.11); CUDA bf16 loss 565.8->560.0 (1.5 bf16 ulp at that magnitude).", "sha256": "3a208ac8c42d6e110d7fcc4205d31b06939bdacf5e51c21dcbaa96ae761ca22c", "verdict": "changed"}, "cuda/control/Wav2Vec2/mic-wav2vec2-base-960h-slim/eager": {"reason": "wav2vec2 fixtures re-recorded in the Transformers CI images (torch 2.14) at transformers e487eaf4, which includes #49386 (masked_spec_embed initialised, fixes #49385). Training loss and gradients now finite in every case, including CPU bf16 (control bf16 baseline is new). CPU float32 drift <=4e-7; CUDA float32 logits <=6.2e-4 and loss +0.01 (first CUDA re-record since torch 2.11); CUDA bf16 loss 565.8->560.0 (1.5 bf16 ulp at that magnitude).", "sha256": "5dea814a2e47014490ab16028272c6eb64d8b4b2e9cd4936debcc3869fe823d8", "verdict": "changed"}, "cuda/prototypes/mic-wav2vec2/mic-wav2vec2-slim/bfloat16": {"reason": "wav2vec2 fixtures re-recorded in the Transformers CI images (torch 2.14) at transformers e487eaf4, which includes #49386 (masked_spec_embed initialised, fixes #49385). Training loss and gradients now finite in every case, including CPU bf16 (control bf16 baseline is new). CPU float32 drift <=4e-7; CUDA float32 logits <=6.2e-4 and loss +0.01 (first CUDA re-record since torch 2.11); CUDA bf16 loss 565.8->560.0 (1.5 bf16 ulp at that magnitude).", "sha256": "90537b9c5d6df87c71ae8ac5d072b24bcd10469155e6378587d2d1c98ebd29b4", "verdict": "changed"}, "cuda/prototypes/mic-wav2vec2/mic-wav2vec2-slim/default": {"reason": "wav2vec2 fixtures re-recorded in the Transformers CI images (torch 2.14) at transformers e487eaf4, which includes #49386 (masked_spec_embed initialised, fixes #49385). Training loss and gradients now finite in every case, including CPU bf16 (control bf16 baseline is new). CPU float32 drift <=4e-7; CUDA float32 logits <=6.2e-4 and loss +0.01 (first CUDA re-record since torch 2.11); CUDA bf16 loss 565.8->560.0 (1.5 bf16 ulp at that magnitude).", "sha256": "b191424d9a5c43b806615f08973166dfdd0c45a3a95cc8bb5617c3ec0a8ff448", "verdict": "changed"}, "cuda/prototypes/mic-wav2vec2/mic-wav2vec2-slim/eager": {"reason": "wav2vec2 fixtures re-recorded in the Transformers CI images (torch 2.14) at transformers e487eaf4, which includes #49386 (masked_spec_embed initialised, fixes #49385). Training loss and gradients now finite in every case, including CPU bf16 (control bf16 baseline is new). CPU float32 drift <=4e-7; CUDA float32 logits <=6.2e-4 and loss +0.01 (first CUDA re-record since torch 2.11); CUDA bf16 loss 565.8->560.0 (1.5 bf16 ulp at that magnitude).", "sha256": "77e179a196ab4ff11e395759cb4e93092eafcb4dca8d824f59f885dc821f18cf", "verdict": "changed"}}
Fixture baselines re-measured in the Transformers CI images (torch 2.14).