transformers integration

README.md

@@ -51,6 +51,79 @@ The predicted text token timestamps can be recovered by subtracting the model's
 
 See the [GitHub repository](https://github.com/kyutai-labs/delayed-streams-modeling/).
 
+### Use with transformers
+
+Install transformers from source:
+```bash
+pip install git+https://github.com/huggingface/transformers
+```
+
+Inference:
+```python
+import torch
+from datasets import load_dataset, Audio
+from transformers import KyutaiSpeechToTextProcessor, KyutaiSpeechToTextForConditionalGeneration
+
+# 1. load the model and the processor
+torch_device = "cuda" if torch.cuda.is_available() else "cpu"
+model_id = "kyutai/stt-2.6b-en_fr"
+
+processor = KyutaiSpeechToTextProcessor.from_pretrained(model_id)
+model = KyutaiSpeechToTextForConditionalGeneration.from_pretrained(model_id, device_map=torch_device)
+
+# 2. load audio samples
+ds = load_dataset(
+    "hf-internal-testing/librispeech_asr_dummy", "clean", split="validation"
+)
+ds = ds.cast_column("audio", Audio(sampling_rate=24000))
+
+# 3. prepare the model inputs
+inputs = processor(
+    ds[0]["audio"]["array"],
+)
+inputs.to(torch_device)
+
+# 4. infer the model
+output_tokens = model.generate(**inputs)
+
+# 5. decode the generated tokens
+print(processor.batch_decode(output_tokens, skip_special_tokens=True))
+```
+
+Batched inference:
+```python
+import torch
+from datasets import load_dataset, Audio
+from transformers import KyutaiSpeechToTextProcessor, KyutaiSpeechToTextForConditionalGeneration
+
+# 1. load the model and the processor
+torch_device = "cuda" if torch.cuda.is_available() else "cpu"
+model_id = "kyutai/stt-2.6b-en_fr"
+
+processor = KyutaiSpeechToTextProcessor.from_pretrained(model_id)
+model = KyutaiSpeechToTextForConditionalGeneration.from_pretrained(model_id, device_map=torch_device)
+
+# 2. load audio samples
+ds = load_dataset(
+    "hf-internal-testing/librispeech_asr_dummy", "clean", split="validation"
+)
+ds = ds.cast_column("audio", Audio(sampling_rate=24000))
+
+# 3. prepare the model inputs
+audio_arrays = [ds[i]["audio"]["array"] for i in range(4)]
+inputs = processor(audio_arrays, return_tensors="pt", padding=True)
+inputs = inputs.to(torch_device)
+
+# 4. infer the model
+output_tokens = model.generate(**inputs)
+
+# 5. decode the generated tokens
+decoded_outputs = processor.batch_decode(output_tokens, skip_special_tokens=True)
+for output in decoded_outputs:
+    print(output)
+```
+####
+
 ## Training Details
 
 ### Training Data

config.json

@@ -75,5 +75,80 @@
   },
   "model_type": "stt",
   "mimi_name": "[email protected]",
-  "tokenizer_name": "tokenizer_en_fr_audio_8000.model"
+  "tokenizer_name": "tokenizer_en_fr_audio_8000.model",
+  "architectures": [
+    "KyutaiSpeechToTextForConditionalGeneration"
+  ],
+  "attention_dropout": 0.0,
+  "audio_bos_token_id": 2048,
+  "audio_pad_token_id": 69569,
+  "bos_token_id": 48000,
+  "codebook_vocab_size": 2049,
+  "codec_config": {
+    "_frame_rate": null,
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "audio_channels": 1,
+    "codebook_dim": 256,
+    "codebook_size": 2048,
+    "compress": 2,
+    "dilation_growth_rate": 2,
+    "head_dim": 64,
+    "hidden_act": "gelu",
+    "hidden_size": 512,
+    "initializer_range": 0.02,
+    "intermediate_size": 2048,
+    "kernel_size": 7,
+    "last_kernel_size": 3,
+    "layer_scale_initial_scale": 0.01,
+    "max_position_embeddings": 8000,
+    "model_type": "mimi",
+    "norm_eps": 1e-05,
+    "num_attention_heads": 8,
+    "num_filters": 64,
+    "num_hidden_layers": 8,
+    "num_key_value_heads": 8,
+    "num_quantizers": 32,
+    "num_residual_layers": 1,
+    "num_semantic_quantizers": 1,
+    "pad_mode": "constant",
+    "residual_kernel_size": 3,
+    "rope_theta": 10000.0,
+    "sampling_rate": 24000,
+    "sliding_window": 250,
+    "trim_right_ratio": 1.0,
+    "upsample_groups": 512,
+    "upsampling_ratios": [
+      8,
+      6,
+      5,
+      4
+    ],
+    "use_cache": false,
+    "use_causal_conv": true,
+    "use_conv_shortcut": false,
+    "use_streaming": false,
+    "vector_quantization_hidden_dimension": 256
+  },
+  "ffn_dim": 11264,
+  "frame_size": 1920,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "max_position_embeddings": 375,
+  "num_attention_heads": 16,
+  "num_codebooks": 32,
+  "num_hidden_layers": 16,
+  "num_key_value_heads": 16,
+  "pad_token_id": 3,
+  "rms_norm_eps": 1e-08,
+  "rope_theta": 100000.0,
+  "sliding_window": 375,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.53.0.dev0",
+  "use_cache": true,
+  "vocab_size": 8001,
+  "transformers_weights": "transformers.safetensors"
 }

generation_config.json

@@ -0,0 +1,8 @@
+{
+  "audio_window_size": 1,
+  "bos_token_id": 48000,
+  "cache_implementation": "sliding_window",
+  "codec_cache_implementation": "sliding_window",
+  "pad_token_id": 3,
+  "transformers_version": "4.53.0.dev0"
+}

preprocessor_config.json

@@ -0,0 +1,13 @@
+{
+  "audio_delay_seconds": 0.5,
+  "audio_silence_prefix_seconds": 0.0,
+  "chunk_length_s": null,
+  "feature_extractor_type": "KyutaiSpeechToTextFeatureExtractor",
+  "feature_size": 1,
+  "overlap": null,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "processor_class": "KyutaiSpeechToTextProcessor",
+  "return_attention_mask": true,
+  "sampling_rate": 24000
+}

special_tokens_map.json

@@ -0,0 +1,3 @@
+{
+  "unk_token": "<unk>"
+}

tokenizer_config.json

@@ -0,0 +1,50 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token_id": null,
+  "chat_template": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token_id": null,
+  "extra_special_tokens": {},
+  "model_input_names": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token_id": null,
+  "processor_class": "KyutaiSpeechToTextProcessor",
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "unk_token": "<unk>"
+}

transformers.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59a6da960020ea4e1436118c63d7cd73f013e0c62466de12d7ad4b64454fd035
+size 2697201444