feat(transformers): Add native Transformers weights and usage
What does this PR do?
→ Adds the Transformers-format checkpoint (model.safetensors, config.json, generation_config.json, processor, tokenizer and chat template files) alongside the existing canary-1b-v2.nemo, so the model loads directly with AutoModelForSpeechSeq2Seq / AutoProcessor 🤗
→ Adds a ### 2) Transformers 🤗 usage section to the model card (pipeline, transcription, translation, batch inference and training examples) and the Transformers tag. This mirrors what was done for nvidia/parakeet-tdt-0.6b-v3 in https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3/discussions/39 and https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3/discussions/40
→ Transformers support is being added in https://github.com/huggingface/transformers/pull/46825
Produced by?
→ Weights are exported from canary-1b-v2.nemo with the conversion script that ships with the Transformers PR (src/transformers/models/canary/convert_canary_nemo_to_hf.py); the FastConformer encoder is loaded through the existing ParakeetEncoder, and the decoder weights are permuted to the Transformers concatenated sin/cos positional layout.
→ Verified against NeMo baselines: transcription, translation and batch generations are reproduced exactly, and the slow integration tests in the Transformers PR run against these exact files.
LGTM. Thanks!