Upload folder using huggingface_hub

.gitattributes

@@ -1,35 +1,45 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+###############################################################################
+# Set default behavior to automatically normalize line endings.
+###############################################################################
+* text=auto
+
+###############################################################################
+# behavior for image files
+# image files are treated as binary by default.
+###############################################################################
+*.jpg   binary
+*.png   binary
+*.gif   binary
+*.bmp   binary
+*.ico   binary
+*.jpeg  binary
+*.jfif  binary
+
+###############################################################################
+# diff behavior for common document formats
+# Convert binary document formats to text before diffing them.
+###############################################################################
+*.doc   diff=astextplain
+*.DOC   diff=astextplain
+*.docx  diff=astextplain
+*.DOCX  diff=astextplain
+*.dot   diff=astextplain
+*.DOT   diff=astextplain
+*.pdf   diff=astextplain
+*.PDF   diff=astextplain
+*.rtf   diff=astextplain
+*.RTF   diff=astextplain
+
+###############################################################################
+# Force Windows cmd and batch scripts to always use crlf line endings so that if
+# a repo is accessed in Windows via a file share from Unix, the scripts will
+# work.
+###############################################################################
+*.cmd text eol=crlf
+*.bat text eol=crlf
+
+###############################################################################
+# Force bash scripts to always use lf line endings so that if a repo is accessed
+# in Unix via a file share from Windows, the scripts will work.
+###############################################################################
+*.sh text eol=lfexamples/sample_audio.wav filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,37 @@
+# Checkpoints (download with install.sh)
+checkpoints/
+
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+*.egg-info/
+dist/
+build/
+
+# Environments
+.env
+.venv
+venv/
+ENV/
+
+# IDE
+.idea/
+.vscode/
+*.swp
+*.swo
+
+# OS
+.DS_Store
+Thumbs.db
+
+# Logs
+*.log
+logs/
+
+# Temp files
+*.tmp
+*.temp
+uploads/

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright Tsinghua University
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

README.md

@@ -0,0 +1,144 @@
+# SALMONN Inference Server
+
+A ready-to-use inference server for [SALMONN](https://github.com/bytedance/SALMONN) (Speech Audio Language Music Open Neural Network) - a multimodal LLM that can understand speech, audio events, and music.
+
+## Features
+
+- Audio transcription
+- Question answering about audio content
+- Audio description and analysis
+- FastAPI server with REST API
+- Simple Python API
+
+## Requirements
+
+- **GPU**: NVIDIA GPU with 24GB+ VRAM (L4, A100, RTX 4090, etc.)
+- **CUDA**: 11.8 or higher
+- **Python**: 3.10+
+- **Storage**: ~25GB for model checkpoints
+
+## Quick Start
+
+### 1. Clone and Install
+
+```bash
+git clone https://huggingface.co/marcosremar2/salmonn-inference
+cd salmonn-inference
+
+# Create virtual environment
+python -m venv venv
+source venv/bin/activate
+
+# Install dependencies
+pip install -r requirements.txt
+```
+
+### 2. Download Models
+
+```bash
+./install.sh
+```
+
+This downloads (~20GB):
+- Vicuna 7B v1.5 (LLM backbone)
+- Whisper Large v2 (speech encoder)
+- BEATs (audio encoder)
+- SALMONN checkpoint (adapter weights)
+
+### 3. Start Server
+
+```bash
+python server.py
+```
+
+Server runs at `http://localhost:8000`
+
+## API Usage
+
+### Transcribe Audio
+
+```bash
+curl -X POST "http://localhost:8000/transcribe" \
+  -F "audio=@your_audio.wav"
+```
+
+### Ask Questions
+
+```bash
+curl -X POST "http://localhost:8000/chat" \
+  -F "audio=@your_audio.wav" \
+  -F "question=What is being said in this audio?"
+```
+
+### Python API
+
+```python
+from inference import SALMONNInference
+
+model = SALMONNInference()
+model.load()
+
+# Transcribe
+text = model.transcribe("audio.wav")
+
+# Ask questions
+answer = model.chat("audio.wav", "What language is being spoken?")
+
+# Describe audio
+description = model.describe("audio.wav")
+```
+
+## API Endpoints
+
+| Endpoint | Method | Description |
+|----------|--------|-------------|
+| `/` | GET | API info |
+| `/health` | GET | Health check |
+| `/transcribe` | POST | Transcribe audio to text |
+| `/chat` | POST | Ask questions about audio |
+| `/describe` | POST | Get audio description |
+
+## Configuration
+
+Edit `config.yaml` to customize:
+
+```yaml
+model:
+  device: "cuda:0"  # GPU device
+
+server:
+  host: "0.0.0.0"
+  port: 8000
+
+generation:
+  max_new_tokens: 200
+  temperature: 1.0
+```
+
+## Performance
+
+Tested on NVIDIA L4 (24GB):
+
+| Metric | Value |
+|--------|-------|
+| Model Load Time | ~20s |
+| Audio Encode | ~250ms |
+| Time to First Token | ~150ms |
+| Tokens/second | ~18 |
+| GPU Memory | ~16GB |
+
+## Important Note
+
+This repository uses **Vicuna 7B v1.5** (not v1.1). The original SALMONN checkpoint was trained with v1.5, and using v1.1 will result in broken outputs (`<unk>` tokens).
+
+## License
+
+- SALMONN: Apache 2.0
+- Vicuna: Llama 2 Community License
+- Whisper: MIT
+
+## Credits
+
+- [SALMONN](https://github.com/bytedance/SALMONN) by Tsinghua University & ByteDance
+- [Vicuna](https://huggingface.co/lmsys/vicuna-7b-v1.5) by LMSYS
+- [Whisper](https://huggingface.co/openai/whisper-large-v2) by OpenAI

config.yaml

@@ -0,0 +1,37 @@
+# SALMONN Configuration
+
+model:
+  # Model paths (relative to project root)
+  vicuna_path: "checkpoints/vicuna-7b-v1.5"
+  whisper_path: "checkpoints/whisper-large-v2"
+  beats_path: "checkpoints/BEATs_iter3_plus_AS2M.pt"
+  salmonn_ckpt: "checkpoints/salmonn-7b/salmonn_7b_v0.pth"
+  
+  # Model settings
+  device: "cuda:0"
+  low_resource: false  # Set true for 8-bit quantization
+  
+  # LoRA settings (do not change)
+  lora: true
+  lora_rank: 8
+  lora_alpha: 32
+  lora_dropout: 0.1
+  
+  # Prompt settings
+  prompt_template: "USER: {}\nASSISTANT:"
+  prompt_path: "prompts/train_prompt.json"
+
+server:
+  host: "0.0.0.0"
+  port: 8000
+  reload: false
+
+generation:
+  max_new_tokens: 200
+  num_beams: 4
+  do_sample: false
+  min_length: 1
+  temperature: 1.0
+  top_p: 0.9
+  repetition_penalty: 1.0
+  length_penalty: 1.0

examples/example_usage.py

@@ -0,0 +1,53 @@
+"""
+SALMONN Example Usage
+
+This script demonstrates how to use the SALMONN inference API.
+"""
+
+import sys
+sys.path.insert(0, '..')
+
+from inference import SALMONNInference
+
+def main():
+    # Initialize model
+    print("Loading SALMONN model...")
+    model = SALMONNInference(config_path="../config.yaml")
+    model.load()
+    
+    # Example audio file
+    audio_file = "sample_audio.wav"
+    
+    print("\n" + "="*50)
+    print("SALMONN Example Usage")
+    print("="*50)
+    
+    # 1. Transcribe audio
+    print("\n1. Transcription:")
+    text = model.transcribe(audio_file)
+    print(f"   {text}")
+    
+    # 2. Ask questions
+    print("\n2. Question Answering:")
+    questions = [
+        "What language is being spoken?",
+        "What is the tone of the speaker?",
+        "Is this audio about weather?",
+    ]
+    
+    for q in questions:
+        answer = model.chat(audio_file, q)
+        print(f"   Q: {q}")
+        print(f"   A: {answer}")
+        print()
+    
+    # 3. Describe audio
+    print("3. Audio Description:")
+    description = model.describe(audio_file)
+    print(f"   {description}")
+    
+    print("\n" + "="*50)
+    print("Done!")
+
+if __name__ == "__main__":
+    main()

examples/sample_audio.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa8e0712e9dd12f400579d2c0d15400259c544b12b18fe0508d3b21c79ab0263
+size 142926

inference.py

@@ -0,0 +1,214 @@
+"""
+SALMONN Inference Module
+Simple interface for audio understanding and transcription.
+"""
+
+import torch
+import yaml
+import os
+from pathlib import Path
+from typing import Optional, Generator
+from transformers import WhisperFeatureExtractor
+from omegaconf import OmegaConf
+
+from models.salmonn import SALMONN
+
+
+class SALMONNInference:
+    """
+    Simple inference class for SALMONN model.
+    
+    Usage:
+        model = SALMONNInference()
+        model.load()
+        
+        # Transcribe audio
+        text = model.transcribe("audio.wav")
+        
+        # Ask questions about audio
+        answer = model.chat("audio.wav", "What is being said?")
+    """
+    
+    def __init__(self, config_path: str = "config.yaml"):
+        self.config_path = config_path
+        self.model = None
+        self.wav_processor = None
+        self.config = None
+        self.device = None
+        self._loaded = False
+        
+    def load(self) -> None:
+        """Load the model and processors."""
+        if self._loaded:
+            return
+            
+        # Load config
+        with open(self.config_path, "r") as f:
+            self.config = OmegaConf.create(yaml.safe_load(f))
+        
+        self.device = self.config.model.device
+        
+        print("Loading SALMONN model...")
+        
+        # Build model config for SALMONN
+        model_config = {
+            "llama_path": self.config.model.vicuna_path,
+            "whisper_path": self.config.model.whisper_path,
+            "beats_path": self.config.model.beats_path,
+            "ckpt": self.config.model.salmonn_ckpt,
+            "lora": self.config.model.lora,
+            "lora_rank": self.config.model.lora_rank,
+            "lora_alpha": self.config.model.lora_alpha,
+            "lora_dropout": self.config.model.lora_dropout,
+            "low_resource": self.config.model.get("low_resource", False),
+            "prompt_template": self.config.model.prompt_template,
+            "prompt_path": self.config.model.prompt_path,
+            "freeze_whisper": True,
+            "freeze_beats": True,
+            "use_speech_Qformer": True,
+            "num_speech_query_token": 1,
+            "freeze_speech_QFormer": False,
+            "window_level_Qformer": True,
+            "second_per_window": 0.333333,
+            "second_stride": 0.333333,
+            "multi_prompt": True,
+        }
+        
+        self.model = SALMONN.from_config(model_config)
+        self.model.to(self.device)
+        self.model.eval()
+        
+        self.wav_processor = WhisperFeatureExtractor.from_pretrained(
+            self.config.model.whisper_path
+        )
+        
+        self._loaded = True
+        print("Model loaded successfully!")
+    
+    def _prepare_audio(self, audio_path: str) -> dict:
+        """Prepare audio file for inference."""
+        import soundfile as sf
+        
+        audio, sr = sf.read(audio_path)
+        
+        # Convert to mono if stereo
+        if len(audio.shape) > 1:
+            audio = audio.mean(axis=1)
+        
+        # Resample to 16kHz if needed
+        if sr != 16000:
+            import librosa
+            audio = librosa.resample(audio, orig_sr=sr, target_sr=16000)
+            sr = 16000
+        
+        # Process audio
+        spectrogram = self.wav_processor(
+            audio, 
+            sampling_rate=sr, 
+            return_tensors="pt"
+        )["input_features"].to(self.device)
+        
+        return {
+            "spectrogram": spectrogram,
+            "raw_wav": torch.from_numpy(audio).unsqueeze(0).to(self.device),
+        }
+    
+    def _generate(self, samples: dict, prompt: str) -> str:
+        """Generate response from the model."""
+        prompt_formatted = self.config.model.prompt_template.format(
+            "<Speech><SpeechHere></Speech> " + prompt
+        )
+        
+        gen_config = {
+            "max_new_tokens": self.config.generation.max_new_tokens,
+            "num_beams": self.config.generation.num_beams,
+            "do_sample": self.config.generation.do_sample,
+            "min_length": self.config.generation.min_length,
+            "temperature": self.config.generation.temperature,
+            "top_p": self.config.generation.top_p,
+            "repetition_penalty": self.config.generation.repetition_penalty,
+            "length_penalty": self.config.generation.length_penalty,
+        }
+        
+        with torch.cuda.amp.autocast(dtype=torch.float16):
+            output = self.model.generate(
+                samples,
+                gen_config,
+                prompts=[prompt_formatted]
+            )[0]
+        
+        # Clean output
+        return output.replace("<s>", "").replace("</s>", "").strip()
+    
+    def transcribe(self, audio_path: str) -> str:
+        """
+        Transcribe audio file to text.
+        
+        Args:
+            audio_path: Path to audio file (wav, mp3, etc.)
+            
+        Returns:
+            Transcribed text
+        """
+        if not self._loaded:
+            self.load()
+            
+        samples = self._prepare_audio(audio_path)
+        return self._generate(samples, "Transcribe the speech.")
+    
+    def chat(self, audio_path: str, question: str) -> str:
+        """
+        Ask a question about an audio file.
+        
+        Args:
+            audio_path: Path to audio file
+            question: Question about the audio content
+            
+        Returns:
+            Model response
+        """
+        if not self._loaded:
+            self.load()
+            
+        samples = self._prepare_audio(audio_path)
+        return self._generate(samples, question)
+    
+    def describe(self, audio_path: str) -> str:
+        """
+        Get a description of the audio content.
+        
+        Args:
+            audio_path: Path to audio file
+            
+        Returns:
+            Description of the audio
+        """
+        if not self._loaded:
+            self.load()
+            
+        samples = self._prepare_audio(audio_path)
+        return self._generate(samples, "Describe the audio content in detail.")
+
+
+if __name__ == "__main__":
+    import sys
+    
+    if len(sys.argv) < 2:
+        print("Usage: python inference.py <audio_file> [question]")
+        print("  If question is omitted, transcribes the audio.")
+        sys.exit(1)
+    
+    audio_file = sys.argv[1]
+    question = sys.argv[2] if len(sys.argv) > 2 else None
+    
+    model = SALMONNInference()
+    model.load()
+    
+    if question:
+        print(f"Question: {question}")
+        result = model.chat(audio_file, question)
+    else:
+        print("Transcribing...")
+        result = model.transcribe(audio_file)
+    
+    print(f"Result: {result}")

install.sh

@@ -0,0 +1,94 @@
+#!/bin/bash
+# SALMONN Installation Script
+# Downloads all required models (~20GB total)
+
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+CHECKPOINTS_DIR="$SCRIPT_DIR/checkpoints"
+
+echo "=============================================="
+echo "SALMONN Model Installation"
+echo "=============================================="
+echo ""
+echo "This will download approximately 20GB of model files."
+echo "Make sure you have enough disk space and a stable internet connection."
+echo ""
+
+# Create checkpoints directory
+mkdir -p "$CHECKPOINTS_DIR"
+cd "$CHECKPOINTS_DIR"
+
+# Check if git-lfs is installed
+if ! command -v git-lfs &> /dev/null; then
+    echo "Installing git-lfs..."
+    if command -v apt-get &> /dev/null; then
+        sudo apt-get update && sudo apt-get install -y git-lfs
+    elif command -v yum &> /dev/null; then
+        sudo yum install -y git-lfs
+    elif command -v brew &> /dev/null; then
+        brew install git-lfs
+    else
+        echo "Please install git-lfs manually: https://git-lfs.github.com/"
+        exit 1
+    fi
+fi
+git lfs install
+
+# 1. Download Vicuna 7B v1.5 (~13GB)
+echo ""
+echo "[1/4] Downloading Vicuna 7B v1.5 (~13GB)..."
+if [ -d "vicuna-7b-v1.5" ] && [ -f "vicuna-7b-v1.5/pytorch_model-00001-of-00002.bin" ]; then
+    echo "  -> Already exists, skipping..."
+else
+    rm -rf vicuna-7b-v1.5
+    git clone https://huggingface.co/lmsys/vicuna-7b-v1.5
+    echo "  -> Done!"
+fi
+
+# 2. Download Whisper Large v2 (~3GB)
+echo ""
+echo "[2/4] Downloading Whisper Large v2 (~3GB)..."
+if [ -d "whisper-large-v2" ] && [ -f "whisper-large-v2/model.safetensors" ]; then
+    echo "  -> Already exists, skipping..."
+else
+    rm -rf whisper-large-v2
+    git clone https://huggingface.co/openai/whisper-large-v2
+    echo "  -> Done!"
+fi
+
+# 3. Download BEATs (~350MB)
+echo ""
+echo "[3/4] Downloading BEATs checkpoint (~350MB)..."
+if [ -f "BEATs_iter3_plus_AS2M.pt" ]; then
+    echo "  -> Already exists, skipping..."
+else
+    wget -q --show-progress https://valle.blob.core.windows.net/share/BEATs/BEATs_iter3_plus_AS2M.pt
+    echo "  -> Done!"
+fi
+
+# 4. Download SALMONN checkpoint (~350MB)
+echo ""
+echo "[4/4] Downloading SALMONN 7B checkpoint (~350MB)..."
+if [ -d "salmonn-7b" ] && [ -f "salmonn-7b/salmonn_7b_v0.pth" ]; then
+    echo "  -> Already exists, skipping..."
+else
+    rm -rf salmonn-7b
+    mkdir -p salmonn-7b
+    cd salmonn-7b
+    wget -q --show-progress https://huggingface.co/tsinghua-ee/SALMONN-7B/resolve/main/salmonn_7b_v0.pth
+    cd ..
+    echo "  -> Done!"
+fi
+
+echo ""
+echo "=============================================="
+echo "Installation Complete!"
+echo "=============================================="
+echo ""
+echo "Downloaded models:"
+du -sh "$CHECKPOINTS_DIR"/*
+echo ""
+echo "To start the server, run:"
+echo "  python server.py"
+echo ""

models/Qformer.py

@@ -0,0 +1,1217 @@
+"""
+Adapted from salesforce@LAVIS. Below is the original copyright:
+ * Copyright (c) 2023, salesforce.com, inc.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ * For full license text, see LICENSE.txt file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+ * By Junnan Li
+ * Based on huggingface code base
+ * https://github.com/huggingface/transformers/blob/v4.15.0/src/transformers/models/bert
+"""
+
+import math
+import os
+import warnings
+from dataclasses import dataclass
+from typing import Optional, Tuple, Dict, Any
+
+import torch
+from torch import Tensor, device, dtype, nn
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+import torch.nn.functional as F
+
+from transformers.activations import ACT2FN
+from transformers.file_utils import (
+    ModelOutput,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    NextSentencePredictorOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import (
+    PreTrainedModel,
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    prune_linear_layer,
+)
+from transformers.utils import logging
+from transformers.models.bert.configuration_bert import BertConfig
+
+logger = logging.get_logger(__name__)
+
+
+class BertEmbeddings(nn.Module):
+    """Construct the embeddings from word and position embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(
+            config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id
+        )
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings, config.hidden_size
+        )
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer(
+            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1))
+        )
+        self.position_embedding_type = getattr(
+            config, "position_embedding_type", "absolute"
+        )
+
+        self.config = config
+
+    def forward(
+        self,
+        input_ids=None,
+        position_ids=None,
+        query_embeds=None,
+        past_key_values_length=0,
+    ):
+        if input_ids is not None:
+            seq_length = input_ids.size()[1]
+        else:
+            seq_length = 0
+
+        if position_ids is None:
+            position_ids = self.position_ids[
+                :, past_key_values_length : seq_length + past_key_values_length
+            ].clone()
+
+        if input_ids is not None:
+            embeddings = self.word_embeddings(input_ids)
+            if self.position_embedding_type == "absolute":
+                position_embeddings = self.position_embeddings(position_ids)
+                embeddings = embeddings + position_embeddings
+
+            if query_embeds is not None:
+                embeddings = torch.cat((query_embeds, embeddings), dim=1)
+        else:
+            embeddings = query_embeds
+
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class BertSelfAttention(nn.Module):
+    def __init__(self, config, is_cross_attention):
+        super().__init__()
+        self.config = config
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(
+            config, "embedding_size"
+        ):
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention "
+                "heads (%d)" % (config.hidden_size, config.num_attention_heads)
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        if is_cross_attention:
+            self.key = nn.Linear(config.encoder_width, self.all_head_size)
+            self.value = nn.Linear(config.encoder_width, self.all_head_size)
+        else:
+            self.key = nn.Linear(config.hidden_size, self.all_head_size)
+            self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        if (
+            self.position_embedding_type == "relative_key"
+            or self.position_embedding_type == "relative_key_query"
+        ):
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(
+                2 * config.max_position_embeddings - 1, self.attention_head_size
+            )
+        self.save_attention = False
+
+    def save_attn_gradients(self, attn_gradients):
+        self.attn_gradients = attn_gradients
+
+    def get_attn_gradients(self):
+        return self.attn_gradients
+
+    def save_attention_map(self, attention_map):
+        self.attention_map = attention_map
+
+    def get_attention_map(self):
+        return self.attention_map
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (
+            self.num_attention_heads,
+            self.attention_head_size,
+        )
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        mixed_query_layer = self.query(hidden_states)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if (
+            self.position_embedding_type == "relative_key"
+            or self.position_embedding_type == "relative_key_query"
+        ):
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(
+                seq_length, dtype=torch.long, device=hidden_states.device
+            ).view(-1, 1)
+            position_ids_r = torch.arange(
+                seq_length, dtype=torch.long, device=hidden_states.device
+            ).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(
+                distance + self.max_position_embeddings - 1
+            )
+            positional_embedding = positional_embedding.to(
+                dtype=query_layer.dtype
+            )  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding
+                )
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding
+                )
+                relative_position_scores_key = torch.einsum(
+                    "bhrd,lrd->bhlr", key_layer, positional_embedding
+                )
+                attention_scores = (
+                    attention_scores
+                    + relative_position_scores_query
+                    + relative_position_scores_key
+                )
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+
+        if is_cross_attention and self.save_attention:
+            self.save_attention_map(attention_probs)
+            attention_probs.register_hook(self.save_attn_gradients)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs_dropped = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs_dropped = attention_probs_dropped * head_mask
+
+        context_layer = torch.matmul(attention_probs_dropped, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = (
+            (context_layer, attention_probs) if output_attentions else (context_layer,)
+        )
+
+        outputs = outputs + (past_key_value,)
+        return outputs
+
+
+class BertSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertAttention(nn.Module):
+    def __init__(self, config, is_cross_attention=False):
+        super().__init__()
+        self.self = BertSelfAttention(config, is_cross_attention)
+        self.output = BertSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads,
+            self.self.num_attention_heads,
+            self.self.attention_head_size,
+            self.pruned_heads,
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = (
+            self.self.attention_head_size * self.self.num_attention_heads
+        )
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+
+        outputs = (attention_output,) + self_outputs[
+            1:
+        ]  # add attentions if we output them
+        return outputs
+
+
+class BertIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+class BertOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertLayer(nn.Module):
+    def __init__(self, config, layer_num):
+        super().__init__()
+        self.config = config
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = BertAttention(config)
+        self.layer_num = layer_num
+        if (
+            self.config.add_cross_attention
+            and layer_num % self.config.cross_attention_freq == 0
+        ):
+            self.crossattention = BertAttention(
+                config, is_cross_attention=self.config.add_cross_attention
+            )
+            self.has_cross_attention = True
+        else:
+            self.has_cross_attention = False
+        self.intermediate = BertIntermediate(config)
+        self.output = BertOutput(config)
+
+        self.intermediate_query = BertIntermediate(config)
+        self.output_query = BertOutput(config)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+        query_length=0,
+    ):
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = (
+            past_key_value[:2] if past_key_value is not None else None
+        )
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+        )
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:-1]
+
+        present_key_value = self_attention_outputs[-1]
+
+        if query_length > 0:
+            query_attention_output = attention_output[:, :query_length, :]
+
+            if self.has_cross_attention:
+                assert (
+                    encoder_hidden_states is not None
+                ), "encoder_hidden_states must be given for cross-attention layers"
+                cross_attention_outputs = self.crossattention(
+                    query_attention_output,
+                    attention_mask,
+                    head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    output_attentions=output_attentions,
+                )
+                query_attention_output = cross_attention_outputs[0]
+                outputs = (
+                    outputs + cross_attention_outputs[1:-1]
+                )  # add cross attentions if we output attention weights
+
+            layer_output = apply_chunking_to_forward(
+                self.feed_forward_chunk_query,
+                self.chunk_size_feed_forward,
+                self.seq_len_dim,
+                query_attention_output,
+            )
+            if attention_output.shape[1] > query_length:
+                layer_output_text = apply_chunking_to_forward(
+                    self.feed_forward_chunk,
+                    self.chunk_size_feed_forward,
+                    self.seq_len_dim,
+                    attention_output[:, query_length:, :],
+                )
+                layer_output = torch.cat([layer_output, layer_output_text], dim=1)
+        else:
+            layer_output = apply_chunking_to_forward(
+                self.feed_forward_chunk,
+                self.chunk_size_feed_forward,
+                self.seq_len_dim,
+                attention_output,
+            )
+        outputs = (layer_output,) + outputs
+
+        outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+    def feed_forward_chunk_query(self, attention_output):
+        intermediate_output = self.intermediate_query(attention_output)
+        layer_output = self.output_query(intermediate_output, attention_output)
+        return layer_output
+
+
+class BertEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList(
+            [BertLayer(config, i) for i in range(config.num_hidden_layers)]
+        )
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=False,
+        output_hidden_states=False,
+        return_dict=True,
+        query_length=0,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = (
+            () if output_attentions and self.config.add_cross_attention else None
+        )
+
+        next_decoder_cache = () if use_cache else None
+
+        for i in range(self.config.num_hidden_layers):
+            layer_module = self.layer[i]
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if getattr(self.config, "gradient_checkpointing", False) and self.training:
+
+                if use_cache:
+                    logger.warn(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(
+                            *inputs, past_key_value, output_attentions, query_length
+                        )
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                    query_length,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+class BertPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class BertPredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+class BertLMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = BertPredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+class BertOnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = BertLMPredictionHead(config)
+
+    def forward(self, sequence_output):
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+class BertPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = BertConfig
+    base_model_prefix = "bert"
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        if isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+
+class BertModel(BertPreTrainedModel):
+    """
+    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
+    cross-attention is added between the self-attention layers, following the architecture described in `Attention is
+    all you need <https://arxiv.org/abs/1706.03762>`__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
+    Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
+    argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an
+    input to the forward pass.
+    """
+
+    def __init__(self, config, add_pooling_layer=False):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = BertEmbeddings(config)
+
+        self.encoder = BertEncoder(config)
+
+        self.pooler = BertPooler(config) if add_pooling_layer else None
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def get_extended_attention_mask(
+        self,
+        attention_mask: Tensor,
+        input_shape: Tuple[int],
+        device: device,
+        is_decoder: bool,
+        has_query: bool = False,
+    ) -> Tensor:
+        """
+        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
+
+        Arguments:
+            attention_mask (:obj:`torch.Tensor`):
+                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
+            input_shape (:obj:`Tuple[int]`):
+                The shape of the input to the model.
+            device: (:obj:`torch.device`):
+                The device of the input to the model.
+
+        Returns:
+            :obj:`torch.Tensor` The extended attention mask, with a the same dtype as :obj:`attention_mask.dtype`.
+        """
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if attention_mask.dim() == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif attention_mask.dim() == 2:
+            # Provided a padding mask of dimensions [batch_size, seq_length]
+            # - if the model is a decoder, apply a causal mask in addition to the padding mask
+            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            if is_decoder:
+                batch_size, seq_length = input_shape
+
+                seq_ids = torch.arange(seq_length, device=device)
+                causal_mask = (
+                    seq_ids[None, None, :].repeat(batch_size, seq_length, 1)
+                    <= seq_ids[None, :, None]
+                )
+
+                # add a prefix ones mask to the causal mask
+                # causal and attention masks must have same type with pytorch version < 1.3
+                causal_mask = causal_mask.to(attention_mask.dtype)
+
+                if causal_mask.shape[1] < attention_mask.shape[1]:
+                    prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]
+                    if has_query:  # UniLM style attention mask
+                        causal_mask = torch.cat(
+                            [
+                                torch.zeros(
+                                    (batch_size, prefix_seq_len, seq_length),
+                                    device=device,
+                                    dtype=causal_mask.dtype,
+                                ),
+                                causal_mask,
+                            ],
+                            axis=1,
+                        )
+                    causal_mask = torch.cat(
+                        [
+                            torch.ones(
+                                (batch_size, causal_mask.shape[1], prefix_seq_len),
+                                device=device,
+                                dtype=causal_mask.dtype,
+                            ),
+                            causal_mask,
+                        ],
+                        axis=-1,
+                    )
+                extended_attention_mask = (
+                    causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
+                )
+            else:
+                extended_attention_mask = attention_mask[:, None, None, :]
+        else:
+            raise ValueError(
+                "Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(
+                    input_shape, attention_mask.shape
+                )
+            )
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = extended_attention_mask.to(
+            dtype=self.dtype
+        )  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+        return extended_attention_mask
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        head_mask=None,
+        query_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        is_decoder=False,
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        """
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        # use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        if input_ids is None:
+            assert (
+                query_embeds is not None
+            ), "You have to specify query_embeds when input_ids is None"
+
+        # past_key_values_length
+        past_key_values_length = (
+            past_key_values[0][0].shape[2] - self.config.query_length
+            if past_key_values is not None
+            else 0
+        )
+
+        query_length = query_embeds.shape[1] if query_embeds is not None else 0
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            query_embeds=query_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+
+        input_shape = embedding_output.size()[:-1]
+        batch_size, seq_length = input_shape
+        device = embedding_output.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                ((batch_size, seq_length + past_key_values_length)), device=device
+            )
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if is_decoder:
+            extended_attention_mask = self.get_extended_attention_mask(
+                attention_mask,
+                input_ids.shape,
+                device,
+                is_decoder,
+                has_query=(query_embeds is not None),
+            )
+        else:
+            extended_attention_mask = self.get_extended_attention_mask(
+                attention_mask, input_shape, device, is_decoder
+            )
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if encoder_hidden_states is not None:
+            if type(encoder_hidden_states) == list:
+                encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[
+                    0
+                ].size()
+            else:
+                (
+                    encoder_batch_size,
+                    encoder_sequence_length,
+                    _,
+                ) = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+
+            if type(encoder_attention_mask) == list:
+                encoder_extended_attention_mask = [
+                    self.invert_attention_mask(mask) for mask in encoder_attention_mask
+                ]
+            elif encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+                encoder_extended_attention_mask = self.invert_attention_mask(
+                    encoder_attention_mask
+                )
+            else:
+                encoder_extended_attention_mask = self.invert_attention_mask(
+                    encoder_attention_mask
+                )
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            query_length=query_length,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = (
+            self.pooler(sequence_output) if self.pooler is not None else None
+        )
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+class BertLMHeadModel(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        head_mask=None,
+        query_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        past_key_values=None,
+        use_cache=True,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        return_logits=False,
+        is_decoder=True,
+        reduction="mean",
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
+            ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are
+            ignored (masked), the loss is only computed for the tokens with labels n ``[0, ..., config.vocab_size]``
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        Returns:
+        Example::
+            >>> from transformers import BertTokenizer, BertLMHeadModel, BertConfig
+            >>> import torch
+            >>> tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
+            >>> config = BertConfig.from_pretrained("bert-base-cased")
+            >>> model = BertLMHeadModel.from_pretrained('bert-base-cased', config=config)
+            >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+            >>> outputs = model(**inputs)
+            >>> prediction_logits = outputs.logits
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+        if labels is not None:
+            use_cache = False
+        if past_key_values is not None:
+            query_embeds = None
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            query_embeds=query_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            is_decoder=is_decoder,
+        )
+
+        sequence_output = outputs[0]
+        if query_embeds is not None:
+            sequence_output = outputs[0][:, query_embeds.shape[1] :, :]
+
+        prediction_scores = self.cls(sequence_output)
+
+        if return_logits:
+            return prediction_scores[:, :-1, :].contiguous()
+
+        lm_loss = None
+        if labels is not None:
+            # we are doing next-token prediction; shift prediction scores and input ids by one
+            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
+            labels = labels[:, 1:].contiguous()
+            loss_fct = CrossEntropyLoss(reduction=reduction, label_smoothing=0.1)
+            lm_loss = loss_fct(
+                shifted_prediction_scores.view(-1, self.config.vocab_size),
+                labels.view(-1),
+            )
+            if reduction == "none":
+                lm_loss = lm_loss.view(prediction_scores.size(0), -1).sum(1)
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return ((lm_loss,) + output) if lm_loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=lm_loss,
+            logits=prediction_scores,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self, input_ids, query_embeds, past=None, attention_mask=None, **model_kwargs
+    ):
+        # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
+        if attention_mask is None:
+            attention_mask = input_ids.new_ones(input_ids.shape)
+        query_mask = input_ids.new_ones(query_embeds.shape[:-1])
+        attention_mask = torch.cat([query_mask, attention_mask], dim=-1)
+
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            input_ids = input_ids[:, -1:]
+
+        return {
+            "input_ids": input_ids,
+            "query_embeds": query_embeds,
+            "attention_mask": attention_mask,
+            "past_key_values": past,
+            "encoder_hidden_states": model_kwargs.get("encoder_hidden_states", None),
+            "encoder_attention_mask": model_kwargs.get("encoder_attention_mask", None),
+            "is_decoder": True,
+        }
+
+    def _reorder_cache(self, past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            reordered_past += (
+                tuple(
+                    past_state.index_select(0, beam_idx) for past_state in layer_past
+                ),
+            )
+        return reordered_past
+
+
+class BertForMaskedLM(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        head_mask=None,
+        query_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        return_logits=False,
+        is_decoder=False,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for computing the masked language modeling loss. Indices should be in ``[-100, 0, ...,
+            config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are ignored
+            (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``
+        """
+
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            query_embeds=query_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            is_decoder=is_decoder,
+        )
+
+        if query_embeds is not None:
+            sequence_output = outputs[0][:, query_embeds.shape[1] :, :]
+        prediction_scores = self.cls(sequence_output)
+
+        if return_logits:
+            return prediction_scores
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()  # -100 index = padding token
+            masked_lm_loss = loss_fct(
+                prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)
+            )
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return (
+                ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+            )
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

models/__init__.py

@@ -0,0 +1,18 @@
+# Copyright (2024) Tsinghua University, Bytedance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .salmonn import SALMONN
+
+def load_model(config):
+    return SALMONN.from_config(config)

models/beats/BEATs.py

@@ -0,0 +1,180 @@
+# --------------------------------------------------------
+# BEATs: Audio Pre-Training with Acoustic Tokenizers (https://arxiv.org/abs/2212.09058)
+# Github source: https://github.com/microsoft/unilm/tree/master/beats
+# Copyright (c) 2022 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Based on fairseq code bases
+# https://github.com/pytorch/fairseq
+# --------------------------------------------------------
+
+
+import torch
+import torch.nn as nn
+from torch.nn import LayerNorm
+import torchaudio.compliance.kaldi as ta_kaldi
+
+from .backbone import (
+    TransformerEncoder,
+)
+
+import logging
+from typing import Optional
+
+logger = logging.getLogger(__name__)
+
+
+class BEATsConfig:
+    def __init__(self, cfg=None):
+        self.input_patch_size: int = -1  # path size of patch embedding
+        self.embed_dim: int = 512  # patch embedding dimension
+        self.conv_bias: bool = False  # include bias in conv encoder
+
+        self.encoder_layers: int = 12  # num encoder layers in the transformer
+        self.encoder_embed_dim: int = 768  # encoder embedding dimension
+        self.encoder_ffn_embed_dim: int = 3072  # encoder embedding dimension for FFN
+        self.encoder_attention_heads: int = 12  # num encoder attention heads
+        self.activation_fn: str = "gelu"  # activation function to use
+
+        self.layer_wise_gradient_decay_ratio: float = 1.0  # ratio for layer-wise gradient decay
+        self.layer_norm_first: bool = False  # apply layernorm first in the transformer
+        self.deep_norm: bool = False  # apply deep_norm first in the transformer
+
+        # dropouts
+        self.dropout: float = 0.1  # dropout probability for the transformer
+        self.attention_dropout: float = 0.1  # dropout probability for attention weights
+        self.activation_dropout: float = 0.0  # dropout probability after activation in FFN
+        self.encoder_layerdrop: float = 0.0  # probability of dropping a tarnsformer layer
+        self.dropout_input: float = 0.0  # dropout to apply to the input (after feat extr)
+
+        # positional embeddings
+        self.conv_pos: int = 128  # number of filters for convolutional positional embeddings
+        self.conv_pos_groups: int = 16  # number of groups for convolutional positional embedding
+
+        # relative position embedding
+        self.relative_position_embedding: bool = False  # apply relative position embedding
+        self.num_buckets: int = 320  # number of buckets for relative position embedding
+        self.max_distance: int = 1280  # maximum distance for relative position embedding
+        self.gru_rel_pos: bool = False  # apply gated relative position embedding
+
+        # label predictor
+        self.finetuned_model: bool = False  # whether the model is a fine-tuned model.
+        self.predictor_dropout: float = 0.1  # dropout probability for the predictor
+        self.predictor_class: int = 527  # target class number for the predictor
+
+        if cfg is not None:
+            self.update(cfg)
+
+    def update(self, cfg: dict):
+        self.__dict__.update(cfg)
+
+
+class BEATs(nn.Module):
+    def __init__(
+            self,
+            cfg: BEATsConfig,
+    ) -> None:
+        super().__init__()
+        logger.info(f"BEATs Config: {cfg.__dict__}")
+
+        self.cfg = cfg
+
+        self.embed = cfg.embed_dim
+        self.post_extract_proj = (
+            nn.Linear(self.embed, cfg.encoder_embed_dim)
+            if self.embed != cfg.encoder_embed_dim
+            else None
+        )
+
+        self.input_patch_size = cfg.input_patch_size
+        self.patch_embedding = nn.Conv2d(1, self.embed, kernel_size=self.input_patch_size, stride=self.input_patch_size,
+                                         bias=cfg.conv_bias)
+
+        self.dropout_input = nn.Dropout(cfg.dropout_input)
+
+        assert not cfg.deep_norm or not cfg.layer_norm_first
+        self.encoder = TransformerEncoder(cfg)
+        self.layer_norm = LayerNorm(self.embed)
+
+        if cfg.finetuned_model:
+            self.predictor_dropout = nn.Dropout(cfg.predictor_dropout)
+            self.predictor = nn.Linear(cfg.encoder_embed_dim, cfg.predictor_class)
+        else:
+            self.predictor = None
+
+    def forward_padding_mask(
+            self,
+            features: torch.Tensor,
+            padding_mask: torch.Tensor,
+    ) -> torch.Tensor:
+        extra = padding_mask.size(1) % features.size(1)
+        if extra > 0:
+            padding_mask = padding_mask[:, :-extra]
+        padding_mask = padding_mask.view(
+            padding_mask.size(0), features.size(1), -1
+        )
+        padding_mask = padding_mask.all(-1)
+        return padding_mask
+
+    def preprocess(
+            self,
+            source: torch.Tensor,
+            fbank_mean: float = 15.41663,
+            fbank_std: float = 6.55582,
+    ) -> torch.Tensor:
+        fbanks = []
+        for waveform in source:
+            waveform = waveform.unsqueeze(0) * 2 ** 15
+            fbank = ta_kaldi.fbank(waveform, num_mel_bins=128, sample_frequency=16000, frame_length=25, frame_shift=10)
+            fbanks.append(fbank)
+        fbank = torch.stack(fbanks, dim=0)
+        fbank = (fbank - fbank_mean) / (2 * fbank_std)
+        return fbank
+
+    def extract_features(
+            self,
+            source: torch.Tensor,
+            padding_mask: Optional[torch.Tensor] = None,
+            fbank_mean: float = 15.41663,
+            fbank_std: float = 6.55582,
+            feature_only=False,
+    ):
+        fbank = self.preprocess(source, fbank_mean=fbank_mean, fbank_std=fbank_std).to(torch.float32)
+
+        if padding_mask is not None:
+            padding_mask = self.forward_padding_mask(fbank, padding_mask)
+
+        fbank = fbank.unsqueeze(1)
+        features = self.patch_embedding(fbank)
+        features = features.reshape(features.shape[0], features.shape[1], -1)
+        features = features.transpose(1, 2)
+        features = self.layer_norm(features)
+
+        if padding_mask is not None:
+            padding_mask = self.forward_padding_mask(features, padding_mask)
+
+        if self.post_extract_proj is not None:
+            features = self.post_extract_proj(features)
+
+        x = self.dropout_input(features)
+
+        x, layer_results = self.encoder(
+            x,
+            padding_mask=padding_mask,
+        )
+
+        if not feature_only and self.predictor is not None:
+            x = self.predictor_dropout(x)
+            logits = self.predictor(x)
+
+            if padding_mask is not None and padding_mask.any():
+                logits[padding_mask] = 0
+                logits = logits.sum(dim=1)
+                logits = logits / (~padding_mask).sum(dim=1).unsqueeze(-1).expand_as(logits)
+            else:
+                logits = logits.mean(dim=1)
+
+            lprobs = torch.sigmoid(logits)
+
+            return lprobs, padding_mask
+        else:
+            return x, padding_mask

models/beats/Tokenizers.py

@@ -0,0 +1,172 @@
+# --------------------------------------------------------
+# BEATs: Audio Pre-Training with Acoustic Tokenizers (https://arxiv.org/abs/2212.09058)
+# Github source: https://github.com/microsoft/unilm/tree/master/beats
+# Copyright (c) 2022 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Based on fairseq code bases
+# https://github.com/pytorch/fairseq
+# --------------------------------------------------------
+
+
+import torch
+import torch.nn as nn
+from torch.nn import LayerNorm
+import torchaudio.compliance.kaldi as ta_kaldi
+
+from .backbone import (
+    TransformerEncoder,
+)
+from .quantizer import (
+    NormEMAVectorQuantizer,
+)
+
+import logging
+from typing import Optional
+
+logger = logging.getLogger(__name__)
+
+
+class TokenizersConfig:
+    def __init__(self, cfg=None):
+        self.input_patch_size: int = -1  # path size of patch embedding
+        self.embed_dim: int = 512  # patch embedding dimension
+        self.conv_bias: bool = False  # include bias in conv encoder
+
+        self.encoder_layers: int = 12  # num encoder layers in the transformer
+        self.encoder_embed_dim: int = 768  # encoder embedding dimension
+        self.encoder_ffn_embed_dim: int = 3072  # encoder embedding dimension for FFN
+        self.encoder_attention_heads: int = 12  # num encoder attention heads
+        self.activation_fn: str = "gelu"  # activation function to use
+
+        self.layer_norm_first: bool = False  # apply layernorm first in the transformer
+        self.deep_norm: bool = False  # apply deep_norm first in the transformer
+
+        # dropouts
+        self.dropout: float = 0.1  # dropout probability for the transformer
+        self.attention_dropout: float = 0.1  # dropout probability for attention weights
+        self.activation_dropout: float = 0.0  # dropout probability after activation in FFN
+        self.encoder_layerdrop: float = 0.0  # probability of dropping a tarnsformer layer
+        self.dropout_input: float = 0.0  # dropout to apply to the input (after feat extr)
+
+        # positional embeddings
+        self.conv_pos: int = 128  # number of filters for convolutional positional embeddings
+        self.conv_pos_groups: int = 16  # number of groups for convolutional positional embedding
+
+        # relative position embedding
+        self.relative_position_embedding: bool = False  # apply relative position embedding
+        self.num_buckets: int = 320  # number of buckets for relative position embedding
+        self.max_distance: int = 1280  # maximum distance for relative position embedding
+        self.gru_rel_pos: bool = False  # apply gated relative position embedding
+
+        # quantizer
+        self.quant_n: int = 1024 # codebook number in quantizer
+        self.quant_dim: int = 256    # codebook dimension in quantizer
+
+        if cfg is not None:
+            self.update(cfg)
+
+    def update(self, cfg: dict):
+        self.__dict__.update(cfg)
+
+
+class Tokenizers(nn.Module):
+    def __init__(
+            self,
+            cfg: TokenizersConfig,
+    ) -> None:
+        super().__init__()
+        logger.info(f"Tokenizers Config: {cfg.__dict__}")
+
+        self.cfg = cfg
+
+        self.embed = cfg.embed_dim
+        self.post_extract_proj = (
+            nn.Linear(self.embed, cfg.encoder_embed_dim)
+            if self.embed != cfg.encoder_embed_dim
+            else None
+        )
+
+        self.input_patch_size = cfg.input_patch_size
+        self.patch_embedding = nn.Conv2d(1, self.embed, kernel_size=self.input_patch_size, stride=self.input_patch_size,
+                                         bias=cfg.conv_bias)
+
+        self.dropout_input = nn.Dropout(cfg.dropout_input)
+
+        assert not cfg.deep_norm or not cfg.layer_norm_first
+        self.encoder = TransformerEncoder(cfg)
+        self.layer_norm = LayerNorm(self.embed)
+
+        self.quantize = NormEMAVectorQuantizer(
+            n_embed=cfg.quant_n, embedding_dim=cfg.quant_dim, beta=1.0, kmeans_init=True, decay=0.99,
+        )
+        self.quant_n = cfg.quant_n
+        self.quantize_layer = nn.Sequential(
+            nn.Linear(cfg.encoder_embed_dim, cfg.encoder_embed_dim),
+            nn.Tanh(),
+            nn.Linear(cfg.encoder_embed_dim, cfg.quant_dim)  # for quantize
+        )
+
+    def forward_padding_mask(
+            self,
+            features: torch.Tensor,
+            padding_mask: torch.Tensor,
+    ) -> torch.Tensor:
+        extra = padding_mask.size(1) % features.size(1)
+        if extra > 0:
+            padding_mask = padding_mask[:, :-extra]
+        padding_mask = padding_mask.view(
+            padding_mask.size(0), features.size(1), -1
+        )
+        padding_mask = padding_mask.all(-1)
+        return padding_mask
+
+    def preprocess(
+            self,
+            source: torch.Tensor,
+            fbank_mean: float = 15.41663,
+            fbank_std: float = 6.55582,
+    ) -> torch.Tensor:
+        fbanks = []
+        for waveform in source:
+            waveform = waveform.unsqueeze(0) * 2 ** 15
+            fbank = ta_kaldi.fbank(waveform, num_mel_bins=128, sample_frequency=16000, frame_length=25, frame_shift=10)
+            fbanks.append(fbank)
+        fbank = torch.stack(fbanks, dim=0)
+        fbank = (fbank - fbank_mean) / (2 * fbank_std)
+        return fbank
+
+    def extract_labels(
+            self,
+            source: torch.Tensor,
+            padding_mask: Optional[torch.Tensor] = None,
+            fbank_mean: float = 15.41663,
+            fbank_std: float = 6.55582,
+    ):
+        fbank = self.preprocess(source, fbank_mean=fbank_mean, fbank_std=fbank_std)
+
+        if padding_mask is not None:
+            padding_mask = self.forward_padding_mask(fbank, padding_mask)
+
+        fbank = fbank.unsqueeze(1)
+        features = self.patch_embedding(fbank)
+        features = features.reshape(features.shape[0], features.shape[1], -1)
+        features = features.transpose(1, 2)
+        features = self.layer_norm(features)
+
+        if padding_mask is not None:
+            padding_mask = self.forward_padding_mask(features, padding_mask)
+
+        if self.post_extract_proj is not None:
+            features = self.post_extract_proj(features)
+
+        x = self.dropout_input(features)
+
+        x, layer_results = self.encoder(
+            x,
+            padding_mask=padding_mask,
+        )
+
+        quantize_input = self.quantize_layer(x)
+        quantize_feature, embed_loss, embed_ind = self.quantize(quantize_input)
+
+        return embed_ind

models/beats/backbone.py

@@ -0,0 +1,783 @@
+# --------------------------------------------------------
+# BEATs: Audio Pre-Training with Acoustic Tokenizers (https://arxiv.org/abs/2212.09058)
+# Github source: https://github.com/microsoft/unilm/tree/master/beats
+# Copyright (c) 2022 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Based on fairseq code bases
+# https://github.com/pytorch/fairseq
+# --------------------------------------------------------
+
+import math
+import numpy as np
+from typing import Dict, Optional, Tuple
+import torch
+from torch import Tensor, nn
+import torch.nn.functional as F
+from torch.nn import LayerNorm, Parameter
+from .modules import (
+    GradMultiply,
+    SamePad,
+    get_activation_fn,
+    GLU_Linear,
+    quant_noise,
+)
+
+
+class TransformerEncoder(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+
+        self.dropout = args.dropout
+        self.embedding_dim = args.encoder_embed_dim
+
+        self.pos_conv = nn.Conv1d(
+            self.embedding_dim,
+            self.embedding_dim,
+            kernel_size=args.conv_pos,
+            padding=args.conv_pos // 2,
+            groups=args.conv_pos_groups,
+        )
+        dropout = 0
+        std = math.sqrt((4 * (1.0 - dropout)) / (args.conv_pos * self.embedding_dim))
+        nn.init.normal_(self.pos_conv.weight, mean=0, std=std)
+        nn.init.constant_(self.pos_conv.bias, 0)
+
+        self.pos_conv = nn.utils.weight_norm(self.pos_conv, name="weight", dim=2)
+        self.pos_conv = nn.Sequential(self.pos_conv, SamePad(args.conv_pos), nn.GELU())
+
+        if hasattr(args, "relative_position_embedding"):
+            self.relative_position_embedding = args.relative_position_embedding
+            self.num_buckets = args.num_buckets
+            self.max_distance = args.max_distance
+        else:
+            self.relative_position_embedding = False
+            self.num_buckets = 0
+            self.max_distance = 0
+
+        self.layers = nn.ModuleList(
+            [
+                TransformerSentenceEncoderLayer(
+                    embedding_dim=self.embedding_dim,
+                    ffn_embedding_dim=args.encoder_ffn_embed_dim,
+                    num_attention_heads=args.encoder_attention_heads,
+                    dropout=self.dropout,
+                    attention_dropout=args.attention_dropout,
+                    activation_dropout=args.activation_dropout,
+                    activation_fn=args.activation_fn,
+                    layer_norm_first=args.layer_norm_first,
+                    deep_norm=args.deep_norm,
+                    has_relative_attention_bias=self.relative_position_embedding,
+                    num_buckets=self.num_buckets,
+                    max_distance=self.max_distance,
+                    gru_rel_pos=args.gru_rel_pos,
+                    encoder_layers=args.encoder_layers,
+                )
+                for i in range(args.encoder_layers)
+            ]
+        )
+        if self.relative_position_embedding:
+            for i in range(1, args.encoder_layers):
+                del self.layers[i].self_attn.relative_attention_bias
+                self.layers[i].self_attn.relative_attention_bias = self.layers[0].self_attn.relative_attention_bias
+
+        self.layer_norm_first = args.layer_norm_first
+        self.layer_norm = LayerNorm(self.embedding_dim)
+        self.layerdrop = args.encoder_layerdrop
+
+        self.apply(init_bert_params)
+
+        if args.deep_norm:
+            deep_norm_beta = math.pow(8 * args.encoder_layers, -1 / 4)
+            for i in range(args.encoder_layers):
+                nn.init.xavier_normal_(self.layers[i].self_attn.k_proj.weight, gain=1)
+                nn.init.xavier_normal_(self.layers[i].self_attn.v_proj.weight, gain=deep_norm_beta)
+                nn.init.xavier_normal_(self.layers[i].self_attn.q_proj.weight, gain=1)
+                nn.init.xavier_normal_(self.layers[i].self_attn.out_proj.weight, gain=deep_norm_beta)
+                nn.init.xavier_normal_(self.layers[i].fc1.weight, gain=deep_norm_beta)
+                nn.init.xavier_normal_(self.layers[i].fc2.weight, gain=deep_norm_beta)
+
+        self.layer_wise_gradient_decay_ratio = getattr(args, "layer_wise_gradient_decay_ratio", 1)
+
+    def forward(self, x, padding_mask=None, layer=None):
+        x, layer_results = self.extract_features(x, padding_mask, layer)
+
+        if self.layer_norm_first and layer is None:
+            x = self.layer_norm(x)
+
+        return x, layer_results
+
+    def extract_features(self, x, padding_mask=None, tgt_layer=None):
+
+        if padding_mask is not None:
+            x[padding_mask] = 0
+
+        x_conv = self.pos_conv(x.transpose(1, 2))
+        x_conv = x_conv.transpose(1, 2)
+        x = x + x_conv
+
+        if not self.layer_norm_first:
+            x = self.layer_norm(x)
+
+        x = F.dropout(x, p=self.dropout, training=self.training)
+
+        # B x T x C -> T x B x C
+        x = x.transpose(0, 1)
+
+        layer_results = []
+        z = None
+        if tgt_layer is not None:
+            layer_results.append((x, z))
+        r = None
+        pos_bias = None
+        for i, layer in enumerate(self.layers):
+            if self.layer_wise_gradient_decay_ratio != 1.0:
+                x = GradMultiply.apply(x, self.layer_wise_gradient_decay_ratio)
+            dropout_probability = np.random.random()
+            if not self.training or (dropout_probability > self.layerdrop):
+                x, z, pos_bias = layer(x, self_attn_padding_mask=padding_mask, need_weights=False, pos_bias=pos_bias)
+            if tgt_layer is not None:
+                layer_results.append((x, z))
+            if i == tgt_layer:
+                r = x
+                break
+
+        if r is not None:
+            x = r
+
+        # T x B x C -> B x T x C
+        x = x.transpose(0, 1)
+
+        return x, layer_results
+
+
+class TransformerSentenceEncoderLayer(nn.Module):
+    def __init__(
+            self,
+            embedding_dim: float = 768,
+            ffn_embedding_dim: float = 3072,
+            num_attention_heads: float = 8,
+            dropout: float = 0.1,
+            attention_dropout: float = 0.1,
+            activation_dropout: float = 0.1,
+            activation_fn: str = "relu",
+            layer_norm_first: bool = False,
+            deep_norm: bool = False,
+            has_relative_attention_bias: bool = False,
+            num_buckets: int = 0,
+            max_distance: int = 0,
+            rescale_init: bool = False,
+            gru_rel_pos: bool = False,
+            encoder_layers: int = 0,
+    ) -> None:
+
+        super().__init__()
+        self.embedding_dim = embedding_dim
+        self.dropout = dropout
+        self.activation_dropout = activation_dropout
+
+        self.activation_name = activation_fn
+        self.activation_fn = get_activation_fn(activation_fn)
+        self.self_attn = MultiheadAttention(
+            self.embedding_dim,
+            num_attention_heads,
+            dropout=attention_dropout,
+            self_attention=True,
+            has_relative_attention_bias=has_relative_attention_bias,
+            num_buckets=num_buckets,
+            max_distance=max_distance,
+            rescale_init=rescale_init,
+            gru_rel_pos=gru_rel_pos,
+        )
+
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(self.activation_dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.layer_norm_first = layer_norm_first
+
+        self.self_attn_layer_norm = LayerNorm(self.embedding_dim)
+
+        if self.activation_name == "glu":
+            self.fc1 = GLU_Linear(self.embedding_dim, ffn_embedding_dim, "swish")
+        else:
+            self.fc1 = nn.Linear(self.embedding_dim, ffn_embedding_dim)
+        self.fc2 = nn.Linear(ffn_embedding_dim, self.embedding_dim)
+
+        self.final_layer_norm = LayerNorm(self.embedding_dim)
+
+        self.deep_norm = deep_norm
+        if self.deep_norm:
+            self.deep_norm_alpha = math.pow(2 * encoder_layers, 1 / 4)
+        else:
+            self.deep_norm_alpha = 1
+
+    def forward(
+            self,
+            x: torch.Tensor,
+            self_attn_mask: torch.Tensor = None,
+            self_attn_padding_mask: torch.Tensor = None,
+            need_weights: bool = False,
+            pos_bias=None
+    ):
+        residual = x
+
+        if self.layer_norm_first:
+            x = self.self_attn_layer_norm(x)
+            x, attn, pos_bias = self.self_attn(
+                query=x,
+                key=x,
+                value=x,
+                key_padding_mask=self_attn_padding_mask,
+                need_weights=False,
+                attn_mask=self_attn_mask,
+                position_bias=pos_bias
+            )
+            x = self.dropout1(x)
+            x = residual + x
+
+            residual = x
+            x = self.final_layer_norm(x)
+            if self.activation_name == "glu":
+                x = self.fc1(x)
+            else:
+                x = self.activation_fn(self.fc1(x))
+            x = self.dropout2(x)
+            x = self.fc2(x)
+            x = self.dropout3(x)
+            x = residual + x
+        else:
+            x, attn, pos_bias = self.self_attn(
+                query=x,
+                key=x,
+                value=x,
+                key_padding_mask=self_attn_padding_mask,
+                need_weights=need_weights,
+                attn_mask=self_attn_mask,
+                position_bias=pos_bias
+            )
+
+            x = self.dropout1(x)
+            x = residual * self.deep_norm_alpha + x
+
+            x = self.self_attn_layer_norm(x)
+
+            residual = x
+            if self.activation_name == "glu":
+                x = self.fc1(x)
+            else:
+                x = self.activation_fn(self.fc1(x))
+            x = self.dropout2(x)
+            x = self.fc2(x)
+            x = self.dropout3(x)
+            x = residual * self.deep_norm_alpha + x
+            x = self.final_layer_norm(x)
+
+        return x, attn, pos_bias
+
+
+class MultiheadAttention(nn.Module):
+    """Multi-headed attention.
+
+    See "Attention Is All You Need" for more details.
+    """
+
+    def __init__(
+            self,
+            embed_dim,
+            num_heads,
+            kdim=None,
+            vdim=None,
+            dropout=0.0,
+            bias=True,
+            add_bias_kv=False,
+            add_zero_attn=False,
+            self_attention=False,
+            encoder_decoder_attention=False,
+            q_noise=0.0,
+            qn_block_size=8,
+            has_relative_attention_bias=False,
+            num_buckets=32,
+            max_distance=128,
+            gru_rel_pos=False,
+            rescale_init=False,
+    ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.kdim = kdim if kdim is not None else embed_dim
+        self.vdim = vdim if vdim is not None else embed_dim
+        self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim
+
+        self.num_heads = num_heads
+        self.dropout_module = nn.Dropout(dropout)
+
+        self.has_relative_attention_bias = has_relative_attention_bias
+        self.num_buckets = num_buckets
+        self.max_distance = max_distance
+        if self.has_relative_attention_bias:
+            self.relative_attention_bias = nn.Embedding(num_buckets, num_heads)
+
+        self.head_dim = embed_dim // num_heads
+        self.q_head_dim = self.head_dim
+        self.k_head_dim = self.head_dim
+        assert (
+                self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+        self.scaling = self.head_dim ** -0.5
+
+        self.self_attention = self_attention
+        self.encoder_decoder_attention = encoder_decoder_attention
+
+        assert not self.self_attention or self.qkv_same_dim, (
+            "Self-attention requires query, key and " "value to be of the same size"
+        )
+
+        k_bias = True
+        if rescale_init:
+            k_bias = False
+
+        k_embed_dim = embed_dim
+        q_embed_dim = embed_dim
+
+        self.k_proj = quant_noise(
+            nn.Linear(self.kdim, k_embed_dim, bias=k_bias), q_noise, qn_block_size
+        )
+        self.v_proj = quant_noise(
+            nn.Linear(self.vdim, embed_dim, bias=bias), q_noise, qn_block_size
+        )
+        self.q_proj = quant_noise(
+            nn.Linear(embed_dim, q_embed_dim, bias=bias), q_noise, qn_block_size
+        )
+
+        self.out_proj = quant_noise(
+            nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size
+        )
+
+        if add_bias_kv:
+            self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim))
+            self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim))
+        else:
+            self.bias_k = self.bias_v = None
+
+        self.add_zero_attn = add_zero_attn
+
+        self.gru_rel_pos = gru_rel_pos
+        if self.gru_rel_pos:
+            self.grep_linear = nn.Linear(self.q_head_dim, 8)
+            self.grep_a = nn.Parameter(torch.ones(1, num_heads, 1, 1))
+
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        if self.qkv_same_dim:
+            # Empirically observed the convergence to be much better with
+            # the scaled initialization
+            nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2))
+            nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2))
+            nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2))
+        else:
+            nn.init.xavier_uniform_(self.k_proj.weight)
+            nn.init.xavier_uniform_(self.v_proj.weight)
+            nn.init.xavier_uniform_(self.q_proj.weight)
+
+        nn.init.xavier_uniform_(self.out_proj.weight)
+        if self.out_proj.bias is not None:
+            nn.init.constant_(self.out_proj.bias, 0.0)
+        if self.bias_k is not None:
+            nn.init.xavier_normal_(self.bias_k)
+        if self.bias_v is not None:
+            nn.init.xavier_normal_(self.bias_v)
+        if self.has_relative_attention_bias:
+            nn.init.xavier_normal_(self.relative_attention_bias.weight)
+
+    def _relative_positions_bucket(self, relative_positions, bidirectional=True):
+        num_buckets = self.num_buckets
+        max_distance = self.max_distance
+        relative_buckets = 0
+
+        if bidirectional:
+            num_buckets = num_buckets // 2
+            relative_buckets += (relative_positions > 0).to(torch.long) * num_buckets
+            relative_positions = torch.abs(relative_positions)
+        else:
+            relative_positions = -torch.min(relative_positions, torch.zeros_like(relative_positions))
+
+        max_exact = num_buckets // 2
+        is_small = relative_positions < max_exact
+
+        relative_postion_if_large = max_exact + (
+                torch.log(relative_positions.float() / max_exact)
+                / math.log(max_distance / max_exact)
+                * (num_buckets - max_exact)
+        ).to(torch.long)
+        relative_postion_if_large = torch.min(
+            relative_postion_if_large, torch.full_like(relative_postion_if_large, num_buckets - 1)
+        )
+
+        relative_buckets += torch.where(is_small, relative_positions, relative_postion_if_large)
+        return relative_buckets
+
+    def compute_bias(self, query_length, key_length):
+        context_position = torch.arange(query_length, dtype=torch.long)[:, None]
+        memory_position = torch.arange(key_length, dtype=torch.long)[None, :]
+        relative_position = memory_position - context_position
+        relative_position_bucket = self._relative_positions_bucket(
+            relative_position,
+            bidirectional=True
+        )
+        relative_position_bucket = relative_position_bucket.to(self.relative_attention_bias.weight.device)
+        values = self.relative_attention_bias(relative_position_bucket)
+        values = values.permute([2, 0, 1])
+        return values
+
+    def forward(
+            self,
+            query,
+            key: Optional[Tensor],
+            value: Optional[Tensor],
+            key_padding_mask: Optional[Tensor] = None,
+            incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,
+            need_weights: bool = True,
+            static_kv: bool = False,
+            attn_mask: Optional[Tensor] = None,
+            before_softmax: bool = False,
+            need_head_weights: bool = False,
+            position_bias: Optional[Tensor] = None
+    ) -> Tuple[Tensor, Optional[Tensor], Optional[Tensor]]:
+        """Input shape: Time x Batch x Channel
+
+        Args:
+            key_padding_mask (ByteTensor, optional): mask to exclude
+                keys that are pads, of shape `(batch, src_len)`, where
+                padding elements are indicated by 1s.
+            need_weights (bool, optional): return the attention weights,
+                averaged over heads (default: False).
+            attn_mask (ByteTensor, optional): typically used to
+                implement causal attention, where the mask prevents the
+                attention from looking forward in time (default: None).
+            before_softmax (bool, optional): return the raw attention
+                weights and values before the attention softmax.
+            need_head_weights (bool, optional): return the attention
+                weights for each head. Implies *need_weights*. Default:
+                return the average attention weights over all heads.
+        """
+        if need_head_weights:
+            need_weights = True
+
+        is_tpu = query.device.type == "xla"
+
+        tgt_len, bsz, embed_dim = query.size()
+        src_len = tgt_len
+        assert embed_dim == self.embed_dim
+        assert list(query.size()) == [tgt_len, bsz, embed_dim]
+        if key is not None:
+            src_len, key_bsz, _ = key.size()
+            if not torch.jit.is_scripting():
+                assert key_bsz == bsz
+                assert value is not None
+                assert src_len, bsz == value.shape[:2]
+
+        if self.has_relative_attention_bias and position_bias is None:
+            position_bias = self.compute_bias(tgt_len, src_len)
+            position_bias = position_bias.unsqueeze(0).repeat(bsz, 1, 1, 1).view(bsz * self.num_heads, tgt_len, src_len)
+
+        if incremental_state is not None:
+            saved_state = self._get_input_buffer(incremental_state)
+            if saved_state is not None and "prev_key" in saved_state:
+                # previous time steps are cached - no need to recompute
+                # key and value if they are static
+                if static_kv:
+                    assert self.encoder_decoder_attention and not self.self_attention
+                    key = value = None
+        else:
+            saved_state = None
+
+        if self.self_attention:
+            q = self.q_proj(query)
+            k = self.k_proj(query)
+            v = self.v_proj(query)
+        elif self.encoder_decoder_attention:
+            # encoder-decoder attention
+            q = self.q_proj(query)
+            if key is None:
+                assert value is None
+                k = v = None
+            else:
+                k = self.k_proj(key)
+                v = self.v_proj(key)
+
+        else:
+            assert key is not None and value is not None
+            q = self.q_proj(query)
+            k = self.k_proj(key)
+            v = self.v_proj(value)
+        q *= self.scaling
+        alpha = 32
+        q *= 1 / alpha
+
+        if self.bias_k is not None:
+            assert self.bias_v is not None
+            k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)])
+            v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)])
+            if attn_mask is not None:
+                attn_mask = torch.cat(
+                    [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1
+                )
+            if key_padding_mask is not None:
+                key_padding_mask = torch.cat(
+                    [
+                        key_padding_mask,
+                        key_padding_mask.new_zeros(key_padding_mask.size(0), 1),
+                    ],
+                    dim=1,
+                )
+
+        q = (
+            q.contiguous()
+                .view(tgt_len, bsz * self.num_heads, self.q_head_dim)
+                .transpose(0, 1)
+        )
+        if k is not None:
+            k = (
+                k.contiguous()
+                    .view(-1, bsz * self.num_heads, self.k_head_dim)
+                    .transpose(0, 1)
+            )
+        if v is not None:
+            v = (
+                v.contiguous()
+                    .view(-1, bsz * self.num_heads, self.head_dim)
+                    .transpose(0, 1)
+            )
+
+        if saved_state is not None:
+            # saved states are stored with shape (bsz, num_heads, seq_len, head_dim)
+            if "prev_key" in saved_state:
+                _prev_key = saved_state["prev_key"]
+                assert _prev_key is not None
+                prev_key = _prev_key.view(bsz * self.num_heads, -1, self.head_dim)
+                if static_kv:
+                    k = prev_key
+                else:
+                    assert k is not None
+                    k = torch.cat([prev_key, k], dim=1)
+                src_len = k.size(1)
+            if "prev_value" in saved_state:
+                _prev_value = saved_state["prev_value"]
+                assert _prev_value is not None
+                prev_value = _prev_value.view(bsz * self.num_heads, -1, self.head_dim)
+                if static_kv:
+                    v = prev_value
+                else:
+                    assert v is not None
+                    v = torch.cat([prev_value, v], dim=1)
+            prev_key_padding_mask: Optional[Tensor] = None
+            if "prev_key_padding_mask" in saved_state:
+                prev_key_padding_mask = saved_state["prev_key_padding_mask"]
+            assert k is not None and v is not None
+            key_padding_mask = MultiheadAttention._append_prev_key_padding_mask(
+                key_padding_mask=key_padding_mask,
+                prev_key_padding_mask=prev_key_padding_mask,
+                batch_size=bsz,
+                src_len=k.size(1),
+                static_kv=static_kv,
+            )
+
+            saved_state["prev_key"] = k.view(bsz, self.num_heads, -1, self.head_dim)
+            saved_state["prev_value"] = v.view(bsz, self.num_heads, -1, self.head_dim)
+            saved_state["prev_key_padding_mask"] = key_padding_mask
+            # In this branch incremental_state is never None
+            assert incremental_state is not None
+            incremental_state = self._set_input_buffer(incremental_state, saved_state)
+        assert k is not None
+        assert k.size(1) == src_len
+
+        # This is part of a workaround to get around fork/join parallelism
+        # not supporting Optional types.
+        if key_padding_mask is not None and key_padding_mask.dim() == 0:
+            key_padding_mask = None
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz
+            assert key_padding_mask.size(1) == src_len
+
+        if self.add_zero_attn:
+            assert v is not None
+            src_len += 1
+            k = torch.cat([k, k.new_zeros((k.size(0), 1) + k.size()[2:])], dim=1)
+            v = torch.cat([v, v.new_zeros((v.size(0), 1) + v.size()[2:])], dim=1)
+            if attn_mask is not None:
+                attn_mask = torch.cat(
+                    [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1
+                )
+            if key_padding_mask is not None:
+                key_padding_mask = torch.cat(
+                    [
+                        key_padding_mask,
+                        torch.zeros(key_padding_mask.size(0), 1).type_as(
+                            key_padding_mask
+                        ),
+                    ],
+                    dim=1,
+                )
+
+        attn_weights = torch.bmm(q, k.transpose(1, 2))
+        attn_weights = (attn_weights - attn_weights.max(dim=-1, keepdim=True)[0]) * alpha
+        attn_weights = self.apply_sparse_mask(attn_weights, tgt_len, src_len, bsz)
+
+        assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len]
+
+        if attn_mask is not None:
+            attn_mask = attn_mask.unsqueeze(0)
+            attn_weights += attn_mask
+
+        if key_padding_mask is not None:
+            # don't attend to padding symbols
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            if not is_tpu:
+                attn_weights = attn_weights.masked_fill(
+                    key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool),
+                    float("-inf"),
+                )
+            else:
+                attn_weights = attn_weights.transpose(0, 2)
+                attn_weights = attn_weights.masked_fill(key_padding_mask, float("-inf"))
+                attn_weights = attn_weights.transpose(0, 2)
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        if before_softmax:
+            return attn_weights, v, position_bias
+
+        if position_bias is not None:
+            attn_mask_rel_pos = position_bias
+            if self.gru_rel_pos == 1:
+                query_layer = q.view(bsz, self.num_heads, tgt_len, self.q_head_dim) * alpha / self.scaling
+                _B, _H, _L, __ = query_layer.size()
+                gate_a, gate_b = torch.sigmoid(self.grep_linear(query_layer).view(
+                    _B, _H, _L, 2, 4).sum(-1, keepdim=False)).chunk(2, dim=-1)
+                gate_a_1 = gate_a * (gate_b * self.grep_a - 1.0) + 2.0
+                attn_mask_rel_pos = gate_a_1.view(bsz * self.num_heads, tgt_len, 1) * position_bias
+
+            attn_mask_rel_pos = attn_mask_rel_pos.view(attn_weights.size())
+
+            attn_weights = attn_weights + attn_mask_rel_pos
+
+        attn_weights_float = F.softmax(
+            attn_weights, dim=-1
+        )
+        attn_weights = attn_weights_float.type_as(attn_weights)
+        attn_probs = self.dropout_module(attn_weights)
+
+        assert v is not None
+        attn = torch.bmm(attn_probs, v)
+        assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim]
+        attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        attn = self.out_proj(attn)
+        attn_weights: Optional[Tensor] = None
+        if need_weights:
+            attn_weights = attn_weights_float.view(
+                bsz, self.num_heads, tgt_len, src_len
+            ).transpose(1, 0)
+            if not need_head_weights:
+                # average attention weights over heads
+                attn_weights = attn_weights.mean(dim=0)
+
+        return attn, attn_weights, position_bias
+
+    @staticmethod
+    def _append_prev_key_padding_mask(
+            key_padding_mask: Optional[Tensor],
+            prev_key_padding_mask: Optional[Tensor],
+            batch_size: int,
+            src_len: int,
+            static_kv: bool,
+    ) -> Optional[Tensor]:
+        # saved key padding masks have shape (bsz, seq_len)
+        if prev_key_padding_mask is not None and static_kv:
+            new_key_padding_mask = prev_key_padding_mask
+        elif prev_key_padding_mask is not None and key_padding_mask is not None:
+            new_key_padding_mask = torch.cat(
+                [prev_key_padding_mask.float(), key_padding_mask.float()], dim=1
+            )
+        # During incremental decoding, as the padding token enters and
+        # leaves the frame, there will be a time when prev or current
+        # is None
+        elif prev_key_padding_mask is not None:
+            if src_len > prev_key_padding_mask.size(1):
+                filler = torch.zeros(
+                    (batch_size, src_len - prev_key_padding_mask.size(1)),
+                    device=prev_key_padding_mask.device,
+                )
+                new_key_padding_mask = torch.cat(
+                    [prev_key_padding_mask.float(), filler.float()], dim=1
+                )
+            else:
+                new_key_padding_mask = prev_key_padding_mask.float()
+        elif key_padding_mask is not None:
+            if src_len > key_padding_mask.size(1):
+                filler = torch.zeros(
+                    (batch_size, src_len - key_padding_mask.size(1)),
+                    device=key_padding_mask.device,
+                )
+                new_key_padding_mask = torch.cat(
+                    [filler.float(), key_padding_mask.float()], dim=1
+                )
+            else:
+                new_key_padding_mask = key_padding_mask.float()
+        else:
+            new_key_padding_mask = prev_key_padding_mask
+        return new_key_padding_mask
+
+    def _get_input_buffer(
+            self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]]
+    ) -> Dict[str, Optional[Tensor]]:
+        result = self.get_incremental_state(incremental_state, "attn_state")
+        if result is not None:
+            return result
+        else:
+            empty_result: Dict[str, Optional[Tensor]] = {}
+            return empty_result
+
+    def _set_input_buffer(
+            self,
+            incremental_state: Dict[str, Dict[str, Optional[Tensor]]],
+            buffer: Dict[str, Optional[Tensor]],
+    ):
+        return self.set_incremental_state(incremental_state, "attn_state", buffer)
+
+    def apply_sparse_mask(self, attn_weights, tgt_len: int, src_len: int, bsz: int):
+        return attn_weights
+
+
+def init_bert_params(module):
+    """
+    Initialize the weights specific to the BERT Model.
+    This overrides the default initializations depending on the specified arguments.
+        1. If normal_init_linear_weights is set then weights of linear
+           layer will be initialized using the normal distribution and
+           bais will be set to the specified value.
+        2. If normal_init_embed_weights is set then weights of embedding
+           layer will be initialized using the normal distribution.
+        3. If normal_init_proj_weights is set then weights of
+           in_project_weight for MultiHeadAttention initialized using
+           the normal distribution (to be validated).
+    """
+
+    def normal_(data):
+        # with FSDP, module params will be on CUDA, so we cast them back to CPU
+        # so that the RNG is consistent with and without FSDP
+        data.copy_(
+            data.cpu().normal_(mean=0.0, std=0.02).to(data.device)
+        )
+
+    if isinstance(module, nn.Linear):
+        normal_(module.weight.data)
+        if module.bias is not None:
+            module.bias.data.zero_()
+    if isinstance(module, nn.Embedding):
+        normal_(module.weight.data)
+        if module.padding_idx is not None:
+            module.weight.data[module.padding_idx].zero_()
+    if isinstance(module, MultiheadAttention):
+        normal_(module.q_proj.weight.data)
+        normal_(module.k_proj.weight.data)
+        normal_(module.v_proj.weight.data)

models/beats/modules.py

@@ -0,0 +1,218 @@
+# --------------------------------------------------------
+# BEATs: Audio Pre-Training with Acoustic Tokenizers (https://arxiv.org/abs/2212.09058)
+# Github source: https://github.com/microsoft/unilm/tree/master/beats
+# Copyright (c) 2022 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Based on fairseq code bases
+# https://github.com/pytorch/fairseq
+# --------------------------------------------------------
+
+import math
+import warnings
+import torch
+from torch import Tensor, nn
+import torch.nn.functional as F
+
+
+class GradMultiply(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x, scale):
+        ctx.scale = scale
+        res = x.new(x)
+        return res
+
+    @staticmethod
+    def backward(ctx, grad):
+        return grad * ctx.scale, None
+
+
+class SamePad(nn.Module):
+    def __init__(self, kernel_size, causal=False):
+        super().__init__()
+        if causal:
+            self.remove = kernel_size - 1
+        else:
+            self.remove = 1 if kernel_size % 2 == 0 else 0
+
+    def forward(self, x):
+        if self.remove > 0:
+            x = x[:, :, : -self.remove]
+        return x
+
+
+class Swish(nn.Module):
+    def __init__(self):
+        super(Swish, self).__init__()
+        self.act = torch.nn.Sigmoid()
+
+    def forward(self, x):
+        return x * self.act(x)
+
+
+class GLU_Linear(nn.Module):
+    def __init__(self, input_dim, output_dim, glu_type="sigmoid", bias_in_glu=True):
+        super(GLU_Linear, self).__init__()
+
+        self.glu_type = glu_type
+        self.output_dim = output_dim
+
+        if glu_type == "sigmoid":
+            self.glu_act = torch.nn.Sigmoid()
+        elif glu_type == "swish":
+            self.glu_act = Swish()
+        elif glu_type == "relu":
+            self.glu_act = torch.nn.ReLU()
+        elif glu_type == "gelu":
+            self.glu_act = torch.nn.GELU()
+
+        if bias_in_glu:
+            self.linear = nn.Linear(input_dim, output_dim * 2, True)
+        else:
+            self.linear = nn.Linear(input_dim, output_dim * 2, False)
+
+    def forward(self, x):
+        # to be consistent with GLU_Linear, we assume the input always has the #channel (#dim) in the last dimension of the tensor, so need to switch the dimension first for 1D-Conv case
+        x = self.linear(x)
+
+        if self.glu_type == "bilinear":
+            x = (x[:, :, 0:self.output_dim] * x[:, :, self.output_dim:self.output_dim * 2])
+        else:
+            x = (x[:, :, 0:self.output_dim] * self.glu_act(x[:, :, self.output_dim:self.output_dim * 2]))
+
+        return x
+
+
+def gelu_accurate(x):
+    if not hasattr(gelu_accurate, "_a"):
+        gelu_accurate._a = math.sqrt(2 / math.pi)
+    return (
+        0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3))))
+    )
+
+
+def gelu(x: torch.Tensor) -> torch.Tensor:
+    return torch.nn.functional.gelu(x.float()).type_as(x)
+
+
+def get_activation_fn(activation: str):
+    """Returns the activation function corresponding to `activation`"""
+
+    if activation == "relu":
+        return F.relu
+    elif activation == "gelu":
+        return gelu
+    elif activation == "gelu_fast":
+        warnings.warn(
+            "--activation-fn=gelu_fast has been renamed to gelu_accurate"
+        )
+        return gelu_accurate
+    elif activation == "gelu_accurate":
+        return gelu_accurate
+    elif activation == "tanh":
+        return torch.tanh
+    elif activation == "linear":
+        return lambda x: x
+    elif activation == "glu":
+        return lambda x: x
+    else:
+        raise RuntimeError("--activation-fn {} not supported".format(activation))
+
+
+def quant_noise(module, p, block_size):
+    """
+    Wraps modules and applies quantization noise to the weights for
+    subsequent quantization with Iterative Product Quantization as
+    described in "Training with Quantization Noise for Extreme Model Compression"
+
+    Args:
+        - module: nn.Module
+        - p: amount of Quantization Noise
+        - block_size: size of the blocks for subsequent quantization with iPQ
+
+    Remarks:
+        - Module weights must have the right sizes wrt the block size
+        - Only Linear, Embedding and Conv2d modules are supported for the moment
+        - For more detail on how to quantize by blocks with convolutional weights,
+          see "And the Bit Goes Down: Revisiting the Quantization of Neural Networks"
+        - We implement the simplest form of noise here as stated in the paper
+          which consists in randomly dropping blocks
+    """
+
+    # if no quantization noise, don't register hook
+    if p <= 0:
+        return module
+
+    # supported modules
+    assert isinstance(module, (nn.Linear, nn.Embedding, nn.Conv2d))
+
+    # test whether module.weight has the right sizes wrt block_size
+    is_conv = module.weight.ndim == 4
+
+    # 2D matrix
+    if not is_conv:
+        assert (
+            module.weight.size(1) % block_size == 0
+        ), "Input features must be a multiple of block sizes"
+
+    # 4D matrix
+    else:
+        # 1x1 convolutions
+        if module.kernel_size == (1, 1):
+            assert (
+                module.in_channels % block_size == 0
+            ), "Input channels must be a multiple of block sizes"
+        # regular convolutions
+        else:
+            k = module.kernel_size[0] * module.kernel_size[1]
+            assert k % block_size == 0, "Kernel size must be a multiple of block size"
+
+    def _forward_pre_hook(mod, input):
+        # no noise for evaluation
+        if mod.training:
+            if not is_conv:
+                # gather weight and sizes
+                weight = mod.weight
+                in_features = weight.size(1)
+                out_features = weight.size(0)
+
+                # split weight matrix into blocks and randomly drop selected blocks
+                mask = torch.zeros(
+                    in_features // block_size * out_features, device=weight.device
+                )
+                mask.bernoulli_(p)
+                mask = mask.repeat_interleave(block_size, -1).view(-1, in_features)
+
+            else:
+                # gather weight and sizes
+                weight = mod.weight
+                in_channels = mod.in_channels
+                out_channels = mod.out_channels
+
+                # split weight matrix into blocks and randomly drop selected blocks
+                if mod.kernel_size == (1, 1):
+                    mask = torch.zeros(
+                        int(in_channels // block_size * out_channels),
+                        device=weight.device,
+                    )
+                    mask.bernoulli_(p)
+                    mask = mask.repeat_interleave(block_size, -1).view(-1, in_channels)
+                else:
+                    mask = torch.zeros(
+                        weight.size(0), weight.size(1), device=weight.device
+                    )
+                    mask.bernoulli_(p)
+                    mask = (
+                        mask.unsqueeze(2)
+                        .unsqueeze(3)
+                        .repeat(1, 1, mod.kernel_size[0], mod.kernel_size[1])
+                    )
+
+            # scale weights and apply mask
+            mask = mask.to(
+                torch.bool
+            )  # x.bool() is not currently supported in TorchScript
+            s = 1 / (1 - p)
+            mod.weight.data = s * weight.masked_fill(mask, 0)
+
+    module.register_forward_pre_hook(_forward_pre_hook)
+    return module

models/beats/quantizer.py

@@ -0,0 +1,215 @@
+# --------------------------------------------------------
+# BEATs: Audio Pre-Training with Acoustic Tokenizers (https://arxiv.org/abs/2212.09058)
+# Github source: https://github.com/microsoft/unilm/tree/master/beats
+# Copyright (c) 2022 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Based on VQGAN code bases
+# https://github.com/CompVis/taming-transformers
+# --------------------------------------------------------'
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.distributed as distributed
+
+try:
+    from einops import rearrange, repeat
+except ImportError:
+    pass
+
+
+def l2norm(t):
+    return F.normalize(t, p=2, dim=-1)
+
+
+def ema_inplace(moving_avg, new, decay):
+    moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))
+
+
+def sample_vectors(samples, num):
+    num_samples, device = samples.shape[0], samples.device
+
+    if num_samples >= num:
+        indices = torch.randperm(num_samples, device=device)[:num]
+    else:
+        indices = torch.randint(0, num_samples, (num,), device=device)
+
+    return samples[indices]
+
+
+def kmeans(samples, num_clusters, num_iters=10, use_cosine_sim=False):
+    dim, dtype, device = samples.shape[-1], samples.dtype, samples.device
+
+    means = sample_vectors(samples, num_clusters)
+
+    for _ in range(num_iters):
+        if use_cosine_sim:
+            dists = samples @ means.t()
+        else:
+            diffs = rearrange(samples, 'n d -> n () d') \
+                    - rearrange(means, 'c d -> () c d')
+            dists = -(diffs ** 2).sum(dim=-1)
+
+        buckets = dists.max(dim=-1).indices
+        bins = torch.bincount(buckets, minlength=num_clusters)
+        zero_mask = bins == 0
+        bins_min_clamped = bins.masked_fill(zero_mask, 1)
+
+        new_means = buckets.new_zeros(num_clusters, dim, dtype=dtype)
+        new_means.scatter_add_(0, repeat(buckets, 'n -> n d', d=dim), samples)
+        new_means = new_means / bins_min_clamped[..., None]
+
+        if use_cosine_sim:
+            new_means = l2norm(new_means)
+
+        means = torch.where(zero_mask[..., None], means, new_means)
+
+    return means, bins
+
+
+class EmbeddingEMA(nn.Module):
+    def __init__(self, num_tokens, codebook_dim, decay=0.99, eps=1e-5, kmeans_init=True, codebook_init_path=''):
+        super().__init__()
+        self.num_tokens = num_tokens
+        self.codebook_dim = codebook_dim
+        self.decay = decay
+        self.eps = eps
+        if codebook_init_path == '':
+            if not kmeans_init:
+                weight = torch.randn(num_tokens, codebook_dim)
+                weight = l2norm(weight)
+            else:
+                weight = torch.zeros(num_tokens, codebook_dim)
+            self.register_buffer('initted', torch.Tensor([not kmeans_init]))
+        else:
+            print(f"load init codebook weight from {codebook_init_path}")
+            codebook_ckpt_weight = torch.load(codebook_init_path, map_location='cpu')
+            weight = codebook_ckpt_weight.clone()
+            self.register_buffer('initted', torch.Tensor([True]))
+
+        self.weight = nn.Parameter(weight, requires_grad=False)
+        self.cluster_size = nn.Parameter(torch.zeros(num_tokens), requires_grad=False)
+        self.embed_avg = nn.Parameter(weight.clone(), requires_grad=False)
+        # self.register_buffer('initted', torch.Tensor([not kmeans_init]))
+        self.update = True
+
+    @torch.jit.ignore
+    def init_embed_(self, data):
+        if self.initted:
+            return
+        print("Performing Kemans init for codebook")
+        embed, cluster_size = kmeans(data, self.num_tokens, 10, use_cosine_sim=True)
+        self.weight.data.copy_(embed)
+        self.cluster_size.data.copy_(cluster_size)
+        self.initted.data.copy_(torch.Tensor([True]))
+
+    def forward(self, embed_id):
+        return F.embedding(embed_id, self.weight)
+
+    def cluster_size_ema_update(self, new_cluster_size):
+        self.cluster_size.data.mul_(self.decay).add_(new_cluster_size, alpha=1 - self.decay)
+
+    def embed_avg_ema_update(self, new_embed_avg):
+        self.embed_avg.data.mul_(self.decay).add_(new_embed_avg, alpha=1 - self.decay)
+
+    def weight_update(self, num_tokens):
+        n = self.cluster_size.sum()
+        smoothed_cluster_size = (
+                (self.cluster_size + self.eps) / (n + num_tokens * self.eps) * n
+        )
+        # normalize embedding average with smoothed cluster size
+        embed_normalized = self.embed_avg / smoothed_cluster_size.unsqueeze(1)
+        # embed_normalized = l2norm(self.embed_avg / smoothed_cluster_size.unsqueeze(1))
+        self.weight.data.copy_(embed_normalized)
+
+
+def norm_ema_inplace(moving_avg, new, decay):
+    moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))
+    moving_avg.data.copy_(l2norm(moving_avg.data))
+
+
+class NormEMAVectorQuantizer(nn.Module):
+    def __init__(self, n_embed, embedding_dim, beta, decay=0.99, eps=1e-5,
+                 statistic_code_usage=True, kmeans_init=False, codebook_init_path=''):
+        super().__init__()
+        self.codebook_dim = embedding_dim
+        self.num_tokens = n_embed
+        self.beta = beta
+        self.decay = decay
+
+        # learnable = True if orthogonal_reg_weight > 0 else False
+        self.embedding = EmbeddingEMA(self.num_tokens, self.codebook_dim, decay, eps, kmeans_init, codebook_init_path)
+
+        self.statistic_code_usage = statistic_code_usage
+        if statistic_code_usage:
+            self.register_buffer('cluster_size', torch.zeros(n_embed))
+        if distributed.is_available() and distributed.is_initialized():
+            print("ddp is enable, so use ddp_reduce to sync the statistic_code_usage for each gpu!")
+            self.all_reduce_fn = distributed.all_reduce
+        else:
+            self.all_reduce_fn = nn.Identity()
+
+    def reset_cluster_size(self, device):
+        if self.statistic_code_usage:
+            self.register_buffer('cluster_size', torch.zeros(self.num_tokens))
+            self.cluster_size = self.cluster_size.to(device)
+
+    def forward(self, z):
+        # reshape z -> (batch, height, width, channel) and flatten
+        # z, 'b c h w -> b h w c'
+        # z = rearrange(z, 'b c h w -> b h w c')
+        # z = z.transpose(1, 2)
+        z = l2norm(z)
+        z_flattened = z.reshape(-1, self.codebook_dim)
+
+        self.embedding.init_embed_(z_flattened)
+
+        d = z_flattened.pow(2).sum(dim=1, keepdim=True) + \
+            self.embedding.weight.pow(2).sum(dim=1) - 2 * \
+            torch.einsum('bd,nd->bn', z_flattened, self.embedding.weight)  # 'n d -> d n'
+
+        encoding_indices = torch.argmin(d, dim=1)
+
+        z_q = self.embedding(encoding_indices).view(z.shape)
+
+        encodings = F.one_hot(encoding_indices, self.num_tokens).type(z.dtype)
+
+        if not self.training:
+            with torch.no_grad():
+                cluster_size = encodings.sum(0)
+                self.all_reduce_fn(cluster_size)
+                ema_inplace(self.cluster_size, cluster_size, self.decay)
+
+        if self.training and self.embedding.update:
+            # EMA cluster size
+
+            bins = encodings.sum(0)
+            self.all_reduce_fn(bins)
+
+            # self.embedding.cluster_size_ema_update(bins)
+            ema_inplace(self.cluster_size, bins, self.decay)
+
+            zero_mask = (bins == 0)
+            bins = bins.masked_fill(zero_mask, 1.)
+
+            embed_sum = z_flattened.t() @ encodings
+            self.all_reduce_fn(embed_sum)
+
+            embed_normalized = (embed_sum / bins.unsqueeze(0)).t()
+            embed_normalized = l2norm(embed_normalized)
+
+            embed_normalized = torch.where(zero_mask[..., None], self.embedding.weight,
+                                           embed_normalized)
+            norm_ema_inplace(self.embedding.weight, embed_normalized, self.decay)
+
+        # compute loss for embedding
+        loss = self.beta * F.mse_loss(z_q.detach(), z)
+
+        # preserve gradients
+        z_q = z + (z_q - z).detach()
+
+        # reshape back to match original input shape
+        # z_q, 'b h w c -> b c h w'
+        # z_q = rearrange(z_q, 'b h w c -> b c h w')
+        # z_q = z_q.transpose(1, 2)
+        return z_q, loss, encoding_indices

models/modeling_llama.py

@@ -0,0 +1,754 @@
+# This script is based on https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py
+
+""" PyTorch LLaMA model."""
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from transformers.models.llama.configuration_llama import LlamaConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "LlamaConfig"
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(
+    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+class LlamaRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        LlamaRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+
+        # convert into half-precision if necessary
+        if self.weight.dtype in [torch.float16, torch.bfloat16]:
+            hidden_states = hidden_states.to(self.weight.dtype)
+
+        return self.weight * hidden_states
+
+
+class LlamaRotaryEmbedding(torch.nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
+        self.register_buffer("inv_freq", inv_freq)
+
+        # Build here to make `torch.jit.trace` work.
+        self.max_seq_len_cached = max_position_embeddings
+        t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
+        freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+        # Different from paper, but it uses a different permutation in order to obtain the same calculation
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
+        self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
+
+    def forward(self, x, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        # This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
+        if seq_len > self.max_seq_len_cached:
+            self.max_seq_len_cached = seq_len
+            t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype)
+            freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+            # Different from paper, but it uses a different permutation in order to obtain the same calculation
+            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
+            self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
+            self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
+        return (
+            self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
+            self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
+        )
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
+    gather_indices = position_ids[:, None, :, None]  # [bs, 1, seq_len, 1]
+    gather_indices = gather_indices.repeat(1, cos.shape[1], 1, cos.shape[3])
+    cos = torch.gather(cos.repeat(gather_indices.shape[0], 1, 1, 1), 2, gather_indices)
+    sin = torch.gather(sin.repeat(gather_indices.shape[0], 1, 1, 1), 2, gather_indices)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class LlamaMLP(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+    ):
+        super().__init__()
+        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
+        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.act_fn = ACT2FN[hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class LlamaAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: LlamaConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.max_position_embeddings = config.max_position_embeddings
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+        self.rotary_emb = LlamaRotaryEmbedding(self.head_dim, max_position_embeddings=self.max_position_embeddings)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+        # [bsz, nh, t, hd]
+
+        if past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+            attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class LlamaDecoderLayer(nn.Module):
+    def __init__(self, config: LlamaConfig):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = LlamaAttention(config=config)
+        self.mlp = LlamaMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+        )
+        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        """
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+LLAMA_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`LlamaConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
+    LLAMA_START_DOCSTRING,
+)
+class LlamaPreTrainedModel(PreTrainedModel):
+    config_class = LlamaConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["LlamaDecoderLayer"]
+    _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, LlamaModel):
+            module.gradient_checkpointing = value
+
+
+LLAMA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
+    LLAMA_START_DOCSTRING,
+)
+class LlamaModel(LlamaPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]
+
+    Args:
+        config: LlamaConfig
+    """
+
+    def __init__(self, config: LlamaConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([LlamaDecoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape,
+                inputs_embeds.dtype,
+                device=inputs_embeds.device,
+                past_key_values_length=past_key_values_length,
+            )
+
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
+                inputs_embeds.device
+            )
+            combined_attention_mask = (
+                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
+            )
+
+        return combined_attention_mask
+
+    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        query_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+        if query_embeds is not None:
+            inputs_embeds = torch.cat([query_embeds, inputs_embeds], dim=1)
+            batch_size, seq_length, _ = inputs_embeds.shape
+
+        seq_length_with_past = seq_length
+        past_key_values_length = 0
+
+        if past_key_values is not None:
+            past_key_values_length = past_key_values[0][0].shape[2]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(
+                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+        else:
+            position_ids = position_ids.view(-1, seq_length).long()
+
+        # embed positions
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
+            )
+        attention_mask = self._prepare_decoder_attention_mask(
+            attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
+        )
+
+        hidden_states = inputs_embeds
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+
+        for idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, None)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    None,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class LlamaForCausalLM(LlamaPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = LlamaModel(config)
+
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        query_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, LlamaForCausalLM
+
+        >>> model = LlamaForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you consciours? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you."
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            query_embeds=query_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self, input_ids, query_embeds=None, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        if past_key_values:
+            input_ids = input_ids[:, -1:]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+                query_embeds = None
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "query_embeds": query_embeds,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
+        return reordered_past

models/modeling_whisper.py

@@ -0,0 +1,1770 @@
+# This script is based on https://github.com/huggingface/transformers/blob/v4.29.1/src/transformers/models/whisper/modeling_whisper.py
+
+""" PyTorch Whisper model."""
+
+import math
+import random
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from transformers.activations import ACT2FN
+from transformers.generation.logits_process import WhisperTimeStampLogitsProcessor
+from transformers.modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPastAndCrossAttentions,
+    Seq2SeqLMOutput,
+    Seq2SeqModelOutput,
+    SequenceClassifierOutput,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from transformers.models.whisper.configuration_whisper import WhisperConfig
+from transformers.models.whisper.tokenization_whisper import TASK_IDS, TO_LANGUAGE_CODE
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "WhisperConfig"
+_CHECKPOINT_FOR_DOC = "openai/whisper-tiny"
+
+
+WHISPER_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "openai/whisper-base",
+    # See all Whisper models at https://huggingface.co/models?filter=whisper
+]
+
+
+# Copied from transformers.models.bart.modeling_bart.shift_tokens_right
+def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
+    """
+    Shift input ids one token to the right.
+    """
+    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
+    shifted_input_ids[:, 0] = decoder_start_token_id
+
+    if pad_token_id is None:
+        raise ValueError("self.model.config.pad_token_id has to be defined.")
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
+
+    return shifted_input_ids
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(
+    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+# Copied from transformers.models.wav2vec2.modeling_wav2vec2._compute_mask_indices
+def _compute_mask_indices(
+    shape: Tuple[int, int],
+    mask_prob: float,
+    mask_length: int,
+    attention_mask: Optional[torch.LongTensor] = None,
+    min_masks: int = 0,
+) -> np.ndarray:
+    """
+    Computes random mask spans for a given shape. Used to implement [SpecAugment: A Simple Data Augmentation Method for
+    ASR](https://arxiv.org/abs/1904.08779). Note that this method is not optimized to run on TPU and should be run on
+    CPU as part of the preprocessing during training.
+
+    Args:
+        shape: The shape for which to compute masks. This should be of a tuple of size 2 where
+               the first element is the batch size and the second element is the length of the axis to span.
+        mask_prob:  The percentage of the whole axis (between 0 and 1) which will be masked. The number of
+                    independently generated mask spans of length `mask_length` is computed by
+                    `mask_prob*shape[1]/mask_length`. Note that due to overlaps, `mask_prob` is an upper bound and the
+                    actual percentage will be smaller.
+        mask_length: size of the mask
+        min_masks: minimum number of masked spans
+        attention_mask: A (right-padded) attention mask which independently shortens the feature axis of
+                        each batch dimension.
+    """
+    batch_size, sequence_length = shape
+
+    if mask_length < 1:
+        raise ValueError("`mask_length` has to be bigger than 0.")
+
+    if mask_length > sequence_length:
+        raise ValueError(
+            f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length}"
+            f" and `sequence_length`: {sequence_length}`"
+        )
+
+    # epsilon is used for probabilistic rounding
+    epsilon = np.random.rand(1).item()
+
+    def compute_num_masked_span(input_length):
+        """Given input length, compute how many spans should be masked"""
+        num_masked_span = int(mask_prob * input_length / mask_length + epsilon)
+        num_masked_span = max(num_masked_span, min_masks)
+
+        # make sure num masked span <= sequence_length
+        if num_masked_span * mask_length > sequence_length:
+            num_masked_span = sequence_length // mask_length
+
+        # make sure num_masked span is also <= input_length - (mask_length - 1)
+        if input_length - (mask_length - 1) < num_masked_span:
+            num_masked_span = max(input_length - (mask_length - 1), 0)
+
+        return num_masked_span
+
+    # compute number of masked spans in batch
+    input_lengths = (
+        attention_mask.sum(-1).detach().tolist()
+        if attention_mask is not None
+        else [sequence_length for _ in range(batch_size)]
+    )
+
+    # SpecAugment mask to fill
+    spec_aug_mask = np.zeros((batch_size, sequence_length), dtype=bool)
+    spec_aug_mask_idxs = []
+
+    max_num_masked_span = compute_num_masked_span(sequence_length)
+
+    if max_num_masked_span == 0:
+        return spec_aug_mask
+
+    for input_length in input_lengths:
+        # compute num of masked spans for this input
+        num_masked_span = compute_num_masked_span(input_length)
+
+        # get random indices to mask
+        spec_aug_mask_idx = np.random.choice(
+            np.arange(input_length - (mask_length - 1)), num_masked_span, replace=False
+        )
+
+        # pick first sampled index that will serve as a dummy index to pad vector
+        # to ensure same dimension for all batches due to probabilistic rounding
+        # Picking first sample just pads those vectors twice.
+        if len(spec_aug_mask_idx) == 0:
+            # this case can only happen if `input_length` is strictly smaller then
+            # `sequence_length` in which case the last token has to be a padding
+            # token which we can use as a dummy mask id
+            dummy_mask_idx = sequence_length - 1
+        else:
+            dummy_mask_idx = spec_aug_mask_idx[0]
+
+        spec_aug_mask_idx = np.concatenate(
+            [spec_aug_mask_idx, np.ones(max_num_masked_span - num_masked_span, dtype=np.int32) * dummy_mask_idx]
+        )
+        spec_aug_mask_idxs.append(spec_aug_mask_idx)
+
+    spec_aug_mask_idxs = np.array(spec_aug_mask_idxs)
+
+    # expand masked indices to masked spans
+    spec_aug_mask_idxs = np.broadcast_to(
+        spec_aug_mask_idxs[:, :, None], (batch_size, max_num_masked_span, mask_length)
+    )
+    spec_aug_mask_idxs = spec_aug_mask_idxs.reshape(batch_size, max_num_masked_span * mask_length)
+
+    # add offset to the starting indexes so that indexes now create a span
+    offsets = np.arange(mask_length)[None, None, :]
+    offsets = np.broadcast_to(offsets, (batch_size, max_num_masked_span, mask_length)).reshape(
+        batch_size, max_num_masked_span * mask_length
+    )
+    spec_aug_mask_idxs = spec_aug_mask_idxs + offsets
+
+    # ensure that we cannot have indices larger than sequence_length
+    if spec_aug_mask_idxs.max() > sequence_length - 1:
+        spec_aug_mask_idxs[spec_aug_mask_idxs > sequence_length - 1] = sequence_length - 1
+
+    # scatter indices to mask
+    np.put_along_axis(spec_aug_mask, spec_aug_mask_idxs, 1, -1)
+
+    return spec_aug_mask
+
+
+class WhisperPositionalEmbedding(nn.Embedding):
+    def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None):
+        super().__init__(num_positions, embedding_dim)
+
+    def forward(self, input_ids, past_key_values_length=0):
+        return self.weight[past_key_values_length : past_key_values_length + input_ids.shape[1]]
+
+
+class WhisperAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        is_decoder: bool = False,
+        bias: bool = True,
+    ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+
+        if (self.head_dim * num_heads) != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
+                f" and `num_heads`: {num_heads})."
+            )
+        self.scaling = self.head_dim**-0.5
+        self.is_decoder = is_decoder
+
+        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=False)
+        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+
+    # Copied from transformers.models.bart.modeling_bart.BartAttention._shape with BART->whisper
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    # Copied from transformers.models.bart.modeling_bart.BartAttention.forward with BART->whisper
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, tgt_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scaling
+        # get key, value proj
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+        key_states = key_states.reshape(*proj_shape)
+        value_states = value_states.reshape(*proj_shape)
+
+        src_len = key_states.size(1)
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+
+        if layer_head_mask is not None:
+            if layer_head_mask.size() != (self.num_heads,):
+                raise ValueError(
+                    f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
+                    f" {layer_head_mask.size()}"
+                )
+            attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        if output_attentions:
+            # this operation is a bit awkward, but it's required to
+            # make sure that attn_weights keeps its gradient.
+            # In order to do so, attn_weights have to be reshaped
+            # twice and have to be reused in the following
+            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+        else:
+            attn_weights_reshaped = None
+
+        attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+
+        attn_output = torch.bmm(attn_probs, value_states)
+
+        if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+        attn_output = attn_output.transpose(1, 2)
+
+        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
+        # partitioned across GPUs when using tensor-parallelism.
+        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped, past_key_value
+
+
+# Copied from transformers.models.mbart.modeling_mbart.MBartEncoderLayer with MBart->Whisper
+class WhisperEncoderLayer(nn.Module):
+    def __init__(self, config: WhisperConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+        self.self_attn = WhisperAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.encoder_attention_heads,
+            dropout=config.attention_dropout,
+        )
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
+        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        layer_head_mask: torch.Tensor,
+        output_attentions: bool = False,
+    ) -> torch.Tensor:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
+                `(encoder_attention_heads,)`.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+        hidden_states, attn_weights, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            layer_head_mask=layer_head_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        if hidden_states.dtype == torch.float16 and (
+            torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any()
+        ):
+            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+# Copied from transformers.models.mbart.modeling_mbart.MBartDecoderLayer with MBart->Whisper
+class WhisperDecoderLayer(nn.Module):
+    def __init__(self, config: WhisperConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        self.self_attn = WhisperAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+        )
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.encoder_attn = WhisperAttention(
+            self.embed_dim,
+            config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+        )
+        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
+        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        cross_attn_layer_head_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = True,
+    ) -> torch.Tensor:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            encoder_hidden_states (`torch.FloatTensor`):
+                cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
+            encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
+                `(encoder_attention_heads,)`.
+            cross_attn_layer_head_mask (`torch.FloatTensor`): mask for cross-attention heads in a given layer of
+                size `(decoder_attention_heads,)`.
+            past_key_value (`Tuple(torch.FloatTensor)`): cached past key and value projection states
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Self Attention
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        # add present self-attn cache to positions 1,2 of present_key_value tuple
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            past_key_value=self_attn_past_key_value,
+            attention_mask=attention_mask,
+            layer_head_mask=layer_head_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        # Cross-Attention Block
+        cross_attn_present_key_value = None
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                layer_head_mask=cross_attn_layer_head_mask,
+                past_key_value=cross_attn_past_key_value,
+                output_attentions=output_attentions,
+            )
+            hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+            hidden_states = residual + hidden_states
+
+            # add cross-attn to positions 3,4 of present_key_value tuple
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights, cross_attn_weights)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+class WhisperPreTrainedModel(PreTrainedModel):
+    config_class = WhisperConfig
+    base_model_prefix = "model"
+    main_input_name = "input_features"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["WhisperEncoderLayer", "WhisperDecoderLayer"]
+
+    def _init_weights(self, module):
+        std = self.config.init_std
+        if isinstance(module, (nn.Linear, nn.Conv1d)):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (WhisperDecoder, WhisperEncoder)):
+            module.gradient_checkpointing = value
+
+    def _get_feat_extract_output_lengths(self, input_lengths: torch.LongTensor):
+        """
+        Computes the output length of the convolutional layers
+        """
+        input_lengths = (input_lengths - 1) // 2 + 1
+
+        return input_lengths
+
+
+WHISPER_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`WhisperConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+WHISPER_INPUTS_DOCSTRING = r"""
+    Args:
+        input_features (`torch.FloatTensor` of shape `(batch_size, feature_size, sequence_length)`):
+            Float values mel features extracted from the raw speech waveform. Raw speech waveform can be obtained by
+            loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
+            the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+            [`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
+            tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
+        attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing *SpecAugment* data augmentation on padding token indices. Mask values selected in
+            `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`WhisperTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are decoder input IDs?](../glossary#decoder-input-ids)
+
+            Whisper uses the `decoder_start_token_id` as the starting token for `decoder_input_ids` generation. If
+            `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+        decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default.
+
+            If you want to change padding behavior, you should read
+            [`modeling_whisper._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the BART
+            paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy.
+        head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        decoder_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
+            Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
+            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
+            hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
+            representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
+            input (see `past_key_values`). This is useful if you want more control over how to convert
+            `decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+WHISPER_ENCODER_INPUTS_DOCSTRING = r"""
+    Args:
+        input_features (`torch.FloatTensor` of shape `(batch_size, feature_size, sequence_length)`):
+            Float values mel features extracted from the raw speech waveform. Raw speech waveform can be obtained by
+            loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
+            the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+            [`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
+            tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
+        head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
+            Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
+            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
+            hidden-states at the output of the last layer of the encoder.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+class WhisperEncoder(WhisperPreTrainedModel):
+    """
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    [`WhisperEncoderLayer`].
+
+    Args:
+        config: WhisperConfig
+    """
+
+    def __init__(self, config: WhisperConfig):
+        super().__init__(config)
+        self.dropout = config.dropout
+        self.layerdrop = config.encoder_layerdrop
+
+        embed_dim = config.d_model
+        self.num_mel_bins = config.num_mel_bins
+        self.padding_idx = config.pad_token_id
+        self.max_source_positions = config.max_source_positions
+        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
+
+        self.conv1 = nn.Conv1d(self.num_mel_bins, embed_dim, kernel_size=3, padding=1)
+        self.conv2 = nn.Conv1d(embed_dim, embed_dim, kernel_size=3, stride=2, padding=1)
+
+        self.embed_positions = nn.Embedding(self.max_source_positions, embed_dim)
+
+        self.layers = nn.ModuleList([WhisperEncoderLayer(config) for _ in range(config.encoder_layers)])
+        self.layer_norm = nn.LayerNorm(config.d_model)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def _freeze_parameters(self):
+        for param in self.parameters():
+            param.requires_grad = False
+        self._requires_grad = False
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.conv1
+
+    def set_input_embeddings(self, value: nn.Module):
+        self.conv1 = value
+
+    def forward(
+        self,
+        input_features,
+        attention_mask=None,
+        head_mask=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            input_features (`torch.LongTensor` of shape `(batch_size, feature_size, sequence_length)`):
+                Float values of mel features extracted from the raw speech waveform. Raw speech waveform can be
+                obtained by loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a
+                `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
+                `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the mel features, padding
+                and conversion into a tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
+            attention_mask (`torch.Tensor`)`, *optional*):
+                Whisper does not support masking of the `input_features`, this argument is preserved for compatibility,
+                but it is not used. By default the silence in the input log mel spectrogram are ignored.
+            head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        inputs_embeds = nn.functional.gelu(self.conv1(input_features))
+        inputs_embeds = nn.functional.gelu(self.conv2(inputs_embeds))
+
+        inputs_embeds = inputs_embeds.permute(0, 2, 1)
+        embed_pos = self.embed_positions.weight
+
+        hidden_states = inputs_embeds + embed_pos
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        # check if head_mask has a correct number of layers specified if desired
+        if head_mask is not None:
+            assert head_mask.size()[0] == (
+                len(self.layers)
+            ), f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
+
+        for idx, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):  # skip the layer
+                layer_outputs = (None, None)
+            else:
+                if self.gradient_checkpointing and self.training:
+
+                    def create_custom_forward(module):
+                        def custom_forward(*inputs):
+                            return module(*inputs, output_attentions)
+
+                        return custom_forward
+
+                    layer_outputs = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(encoder_layer),
+                        hidden_states,
+                        None,
+                        (head_mask[idx] if head_mask is not None else None),
+                    )
+                else:
+                    layer_outputs = encoder_layer(
+                        hidden_states,
+                        None,
+                        layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                        output_attentions=output_attentions,
+                    )
+
+                hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        hidden_states = self.layer_norm(hidden_states)
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+class WhisperDecoder(WhisperPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`WhisperDecoderLayer`]
+
+    Args:
+        config: WhisperConfig
+    """
+
+    def __init__(self, config: WhisperConfig):
+        super().__init__(config)
+        self.dropout = config.dropout
+        self.layerdrop = config.decoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_target_positions
+        self.max_source_positions = config.max_source_positions
+        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
+        self.embed_positions = WhisperPositionalEmbedding(self.max_target_positions, config.d_model)
+
+        self.layers = nn.ModuleList([WhisperDecoderLayer(config) for _ in range(config.decoder_layers)])
+
+        self.layer_norm = nn.LayerNorm(config.d_model)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape,
+                inputs_embeds.dtype,
+                device=inputs_embeds.device,
+                past_key_values_length=past_key_values_length,
+            )
+
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
+            combined_attention_mask = (
+                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
+            )
+
+        return combined_attention_mask
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        encoder_hidden_states=None,
+        head_mask=None,
+        cross_attn_head_mask=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using [`WhisperTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+                [`PreTrainedTokenizer.__call__`] for details.
+
+                [What are input IDs?](../glossary#input-ids)
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules in encoder to avoid performing cross-attention
+                on hidden heads. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+                Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+                shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+                shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
+                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
+                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
+                shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
+                `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more
+                control over how to convert `input_ids` indices into associated vectors than the model's internal
+                embedding lookup matrix.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        attention_mask = self._prepare_decoder_attention_mask(
+            attention_mask, input_shape, inputs_embeds, past_key_values_length
+        )
+
+        # embed positions
+        if input_ids is not None:
+            positions = self.embed_positions(input_ids, past_key_values_length=past_key_values_length)
+        else:
+            positions = self.embed_positions(inputs_embeds, past_key_values_length=past_key_values_length)
+
+        hidden_states = inputs_embeds + positions
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache = True` is incompatible with gradient checkpointing. Setting `use_cache = False`..."
+                )
+                use_cache = False
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
+        next_decoder_cache = () if use_cache else None
+
+        # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
+        for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]):
+            if attn_mask is not None:
+                assert attn_mask.size()[0] == (len(self.layers)), (
+                    f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
+                    f" {head_mask.size()[0]}."
+                )
+        for idx, decoder_layer in enumerate(self.layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):
+                continue
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, use_cache)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    None,  # encoder attention mask
+                    head_mask[idx] if head_mask is not None else None,
+                    cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None,
+                    None,  # past_key_value
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                    cross_attn_layer_head_mask=(
+                        cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None
+                    ),
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
+        hidden_states = self.layer_norm(hidden_states)
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    "The bare Whisper Model outputting raw hidden-states without any specific head on top.",
+    WHISPER_START_DOCSTRING,
+)
+class WhisperModel(WhisperPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"proj_out.weight"]
+
+    def __init__(self, config: WhisperConfig):
+        super().__init__(config)
+
+        self.encoder = WhisperEncoder(config)
+        self.decoder = WhisperDecoder(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.decoder.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.decoder.embed_tokens = value
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    def freeze_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the Whisper encoder so that its parameters will
+        not be updated during training.
+        """
+        self.encoder._freeze_parameters()
+
+    def _mask_input_features(
+        self,
+        input_features: torch.FloatTensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+    ):
+        """
+        Masks extracted features along time axis and/or along feature axis according to
+        [SpecAugment](https://arxiv.org/abs/1904.08779).
+        """
+
+        # `config.apply_spec_augment` can set masking to False
+        if not getattr(self.config, "apply_spec_augment", True):
+            return input_features
+
+        # generate indices & apply SpecAugment along time axis
+        batch_size, hidden_size, sequence_length = input_features.size()
+
+        if self.config.mask_time_prob > 0 and self.training:
+            # generate indices & apply SpecAugment along time axis
+            mask_time_indices = _compute_mask_indices(
+                (batch_size, sequence_length),
+                mask_prob=self.config.mask_time_prob,
+                mask_length=self.config.mask_time_length,
+                attention_mask=attention_mask,
+                min_masks=self.config.mask_time_min_masks,
+            )
+            mask_time_indices = torch.tensor(mask_time_indices, device=input_features.device, dtype=torch.bool)
+            mask_time_indices = mask_time_indices[:, None].expand(-1, hidden_size, -1)
+            input_features[mask_time_indices] = 0
+
+        if self.config.mask_feature_prob > 0 and self.training:
+            # generate indices & apply SpecAugment along feature axis
+            mask_feature_indices = _compute_mask_indices(
+                (batch_size, hidden_size),
+                mask_prob=self.config.mask_feature_prob,
+                mask_length=self.config.mask_feature_length,
+                min_masks=self.config.mask_feature_min_masks,
+            )
+            mask_feature_indices = torch.tensor(mask_feature_indices, device=input_features.device, dtype=torch.bool)
+            input_features[mask_feature_indices] = 0
+
+        return input_features
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_features: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        decoder_attention_mask: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        decoder_head_mask: Optional[torch.Tensor] = None,
+        cross_attn_head_mask: Optional[torch.Tensor] = None,
+        encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        decoder_inputs_embeds: Optional[Tuple[torch.FloatTensor]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], Seq2SeqModelOutput]:
+        r"""
+        Returns:
+
+        Example:
+         ```python
+         >>> import torch
+         >>> from transformers import AutoFeatureExtractor, WhisperModel
+         >>> from datasets import load_dataset
+
+         >>> model = WhisperModel.from_pretrained("openai/whisper-base")
+         >>> feature_extractor = AutoFeatureExtractor.from_pretrained("openai/whisper-base")
+         >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+         >>> inputs = feature_extractor(ds[0]["audio"]["array"], return_tensors="pt")
+         >>> input_features = inputs.input_features
+         >>> decoder_input_ids = torch.tensor([[1, 1]]) * model.config.decoder_start_token_id
+         >>> last_hidden_state = model(input_features, decoder_input_ids=decoder_input_ids).last_hidden_state
+         >>> list(last_hidden_state.shape)
+         [1, 2, 512]
+         ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if encoder_outputs is None:
+            input_features = self._mask_input_features(input_features, attention_mask=attention_mask)
+
+            encoder_outputs = self.encoder(
+                input_features,
+                head_mask=head_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_outputs[0],
+            head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return Seq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    "The Whisper Model with a language modeling head. Can be used for automatic speech recognition.",
+    WHISPER_START_DOCSTRING,
+)
+class WhisperForConditionalGeneration(WhisperPreTrainedModel):
+    base_model_prefix = "model"
+    _keys_to_ignore_on_load_missing = [
+        r"encoder.version",
+        r"decoder.version",
+        r"proj_out.weight",
+    ]
+    _keys_to_ignore_on_save = [
+        r"proj_out.weight",
+    ]
+
+    def __init__(self, config: WhisperConfig):
+        super().__init__(config)
+        self.model = WhisperModel(config)
+        self.proj_out = nn.Linear(config.d_model, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_encoder(self):
+        return self.model.get_encoder()
+
+    def get_decoder(self):
+        return self.model.get_decoder()
+
+    def resize_token_embeddings(self, new_num_tokens: int) -> nn.Embedding:
+        new_embeddings = super().resize_token_embeddings(new_num_tokens)
+        return new_embeddings
+
+    def get_output_embeddings(self):
+        return self.proj_out
+
+    def set_output_embeddings(self, new_embeddings):
+        self.proj_out = new_embeddings
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.model.get_input_embeddings()
+
+    def freeze_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the Whisper encoder so that its parameters will
+        not be updated during training.
+        """
+        self.model.encoder._freeze_parameters()
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_features: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        decoder_attention_mask: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        decoder_head_mask: Optional[torch.Tensor] = None,
+        cross_attn_head_mask: Optional[torch.Tensor] = None,
+        encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        decoder_inputs_embeds: Optional[Tuple[torch.FloatTensor]] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], Seq2SeqLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the language modeling loss. Indices should either be in `[0, ..., config.vocab_size]`
+            or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is
+            only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> import torch
+        >>> from transformers import AutoProcessor, WhisperForConditionalGeneration
+        >>> from datasets import load_dataset
+
+        >>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
+
+        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+
+        >>> inputs = processor(ds[0]["audio"]["array"], return_tensors="pt")
+        >>> input_features = inputs.input_features
+
+        >>> generated_ids = model.generate(inputs=input_features)
+
+        >>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+        >>> transcription
+        ' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if labels is not None:
+            if decoder_input_ids is None and decoder_inputs_embeds is None:
+                decoder_input_ids = shift_tokens_right(
+                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
+                )
+
+        outputs = self.model(
+            input_features,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            encoder_outputs=encoder_outputs,
+            decoder_attention_mask=decoder_attention_mask,
+            head_mask=head_mask,
+            decoder_head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            past_key_values=past_key_values,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        lm_logits = self.proj_out(outputs[0])
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # move labels to correct device to enable PP
+            labels = labels.to(lm_logits.device)
+            loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.reshape(-1))
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return Seq2SeqLMOutput(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=outputs.past_key_values,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        generation_config=None,
+        logits_processor=None,
+        stopping_criteria=None,
+        prefix_allowed_tokens_fn=None,
+        synced_gpus=False,
+        return_timestamps=None,
+        task=None,
+        language=None,
+        is_multilingual=None,
+        **kwargs,
+    ):
+        """
+
+        Generates sequences of token ids for models with a language modeling head.
+
+        <Tip warning={true}>
+
+        Most generation-controlling parameters are set in `generation_config` which, if not passed, will be set to the
+        model's default generation configuration. You can override any `generation_config` by passing the corresponding
+        parameters to generate(), e.g. `.generate(inputs, num_beams=4, do_sample=True)`.
+
+        For an overview of generation strategies and code examples, check out the [following
+        guide](./generation_strategies).
+
+        </Tip>
+
+        Parameters:
+            inputs (`torch.Tensor` of varying shape depending on the modality, *optional*):
+                The sequence used as a prompt for the generation or as model inputs to the encoder. If `None` the
+                method initializes it with `bos_token_id` and a batch size of 1. For decoder-only models `inputs`
+                should of in the format of `input_ids`. For encoder-decoder models *inputs* can represent any of
+                `input_ids`, `input_values`, `input_features`, or `pixel_values`.
+            generation_config (`~generation.GenerationConfig`, *optional*):
+                The generation configuration to be used as base parametrization for the generation call. `**kwargs`
+                passed to generate matching the attributes of `generation_config` will override them. If
+                `generation_config` is not provided, the default will be used, which had the following loading
+                priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
+                configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
+                default values, whose documentation should be checked to parameterize generation.
+            logits_processor (`LogitsProcessorList`, *optional*):
+                Custom logits processors that complement the default logits processors built from arguments and
+                generation config. If a logit processor is passed that is already created with the arguments or a
+                generation config an error is thrown. This feature is intended for advanced users.
+            stopping_criteria (`StoppingCriteriaList`, *optional*):
+                Custom stopping criteria that complement the default stopping criteria built from arguments and a
+                generation config. If a stopping criteria is passed that is already created with the arguments or a
+                generation config an error is thrown. This feature is intended for advanced users.
+            prefix_allowed_tokens_fn (`Callable[[int, torch.Tensor], List[int]]`, *optional*):
+                If provided, this function constraints the beam search to allowed tokens only at each step. If not
+                provided no constraint is applied. This function takes 2 arguments: the batch ID `batch_id` and
+                `input_ids`. It has to return a list with the allowed tokens for the next generation step conditioned
+                on the batch ID `batch_id` and the previously generated tokens `inputs_ids`. This argument is useful
+                for constrained generation conditioned on the prefix, as described in [Autoregressive Entity
+                Retrieval](https://arxiv.org/abs/2010.00904).
+            synced_gpus (`bool`, *optional*, defaults to `False`):
+                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
+            return_timestamps (`bool`, *optional*):
+                Whether to return the timestamps with the text. This enables the `WhisperTimestampsLogitsProcessor`.
+            task (`bool`, *optional*):
+                Task to use for generation, either "translate" or "transcribe". The `model.config.forced_decoder_ids`
+                will be updated accordingly.
+            language (`bool`, *optional*):
+                Language token to use for generation, can be either in the form of `<|en|>`, `en` or `english`. You can
+                find all the possible language tokens in the `model.generation_config.lang_to_id` dictionary.
+            is_multilingual (`bool`, *optional*):
+                Whether or not the model is multilingual.
+            kwargs:
+                Ad hoc parametrization of `generate_config` and/or additional model-specific kwargs that will be
+                forwarded to the `forward` function of the model. If the model is an encoder-decoder model, encoder
+                specific kwargs should not be prefixed and decoder specific kwargs should be prefixed with *decoder_*.
+
+        Return:
+            [`~utils.ModelOutput`] or `torch.LongTensor`: A [`~utils.ModelOutput`] (if `return_dict_in_generate=True`
+            or when `config.return_dict_in_generate=True`) or a `torch.FloatTensor`.
+
+                If the model is *not* an encoder-decoder model (`model.config.is_encoder_decoder=False`), the possible
+                [`~utils.ModelOutput`] types are:
+
+                    - [`~generation.GreedySearchDecoderOnlyOutput`],
+                    - [`~generation.SampleDecoderOnlyOutput`],
+                    - [`~generation.BeamSearchDecoderOnlyOutput`],
+                    - [`~generation.BeamSampleDecoderOnlyOutput`]
+
+                If the model is an encoder-decoder model (`model.config.is_encoder_decoder=True`), the possible
+                [`~utils.ModelOutput`] types are:
+
+                    - [`~generation.GreedySearchEncoderDecoderOutput`],
+                    - [`~generation.SampleEncoderDecoderOutput`],
+                    - [`~generation.BeamSearchEncoderDecoderOutput`],
+                    - [`~generation.BeamSampleEncoderDecoderOutput`]
+        """
+        if generation_config is None:
+            generation_config = self.generation_config
+
+        if return_timestamps is not None:
+            if not hasattr(generation_config, "no_timestamps_token_id"):
+                raise ValueError(
+                    "You are trying to return timestamps, but the generation config is not properly set."
+                    "Make sure to initialize the generation config with the correct attributes that are needed such as `no_timestamps_token_id`."
+                    "For more details on how to generate the approtiate config, refer to https://github.com/huggingface/transformers/issues/21878#issuecomment-1451902363"
+                )
+
+            generation_config.return_timestamps = return_timestamps
+        else:
+            generation_config.return_timestamps = False
+
+        if language is not None:
+            language = language.lower()
+            generation_config.language = language
+        if task is not None:
+            generation_config.task = task
+
+        forced_decoder_ids = []
+        if task is not None or language is not None:
+            if hasattr(generation_config, "language"):
+                if generation_config.language in generation_config.lang_to_id.keys():
+                    language_token = generation_config.language
+                elif generation_config.language in TO_LANGUAGE_CODE.keys():
+                    language_token = f"<|{TO_LANGUAGE_CODE[generation_config.language]}|>"
+                elif generation_config.language in TO_LANGUAGE_CODE.values():
+                    language_token = f"<|{generation_config.language}|>"
+                else:
+                    is_language_code = len(generation_config.language) == 2
+                    raise ValueError(
+                        f"Unsupported language: {generation_config.language}. Language should be one of:"
+                        f" {list(TO_LANGUAGE_CODE.values()) if is_language_code else list(TO_LANGUAGE_CODE.keys())}."
+                    )
+                forced_decoder_ids.append((1, generation_config.lang_to_id[language_token]))
+            else:
+                forced_decoder_ids.append((1, None))  # automatically detect the language
+
+            if hasattr(generation_config, "task"):
+                if generation_config.task in TASK_IDS:
+                    forced_decoder_ids.append((2, generation_config.task_to_id[generation_config.task]))
+                else:
+                    raise ValueError(
+                        f"The `{generation_config.task}`task is not supported. The task should be one of `{TASK_IDS}`"
+                    )
+            else:
+                forced_decoder_ids.append((2, generation_config.task_to_id["transcribe"]))  # defaults to transcribe
+            if hasattr(generation_config, "no_timestamps_token_id") and not generation_config.return_timestamps:
+                idx = forced_decoder_ids[-1][0] + 1 if forced_decoder_ids else 1
+                forced_decoder_ids.append((idx, generation_config.no_timestamps_token_id))
+
+        # Legacy code for backward compatibility
+        elif hasattr(self.config, "forced_decoder_ids") and self.config.forced_decoder_ids is not None:
+            forced_decoder_ids = self.config.forced_decoder_ids
+        elif (
+            hasattr(self.generation_config, "forced_decoder_ids")
+            and self.generation_config.forced_decoder_ids is not None
+        ):
+            forced_decoder_ids = self.generation_config.forced_decoder_ids
+
+        if generation_config.return_timestamps:
+            logits_processor = [WhisperTimeStampLogitsProcessor(generation_config)]
+
+        if len(forced_decoder_ids) > 0:
+            generation_config.forced_decoder_ids = forced_decoder_ids
+
+        return super().generate(
+            inputs,
+            generation_config,
+            logits_processor,
+            stopping_criteria,
+            prefix_allowed_tokens_fn,
+            synced_gpus,
+            **kwargs,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        decoder_input_ids,
+        past_key_values=None,
+        use_cache=None,
+        encoder_outputs=None,
+        attention_mask=None,
+        **kwargs,
+    ):
+        # cut decoder_input_ids if past is used
+        if past_key_values is not None:
+            decoder_input_ids = decoder_input_ids[:, -1:]
+
+        return {
+            "encoder_outputs": encoder_outputs,
+            "past_key_values": past_key_values,
+            "decoder_input_ids": decoder_input_ids,
+            "use_cache": use_cache,
+            "decoder_attention_mask": None,
+        }
+
+    #
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
+        return reordered_past
+
+
+@add_start_docstrings(
+    """
+    Whisper Encoder Model with a sequence classification head on top (a linear layer over the pooled output) for tasks
+    like SUPERB Keyword Spotting.
+    """,
+    WHISPER_ENCODER_INPUTS_DOCSTRING,
+)
+class WhisperForAudioClassification(WhisperPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.encoder = WhisperEncoder(config)
+        num_layers = config.num_hidden_layers + 1  # transformer layers + input embeddings
+        if config.use_weighted_layer_sum:
+            self.layer_weights = nn.Parameter(torch.ones(num_layers) / num_layers)
+        self.projector = nn.Linear(config.hidden_size, config.classifier_proj_size)
+        self.classifier = nn.Linear(config.classifier_proj_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def freeze_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the Whisper encoder so that its parameters will
+        not be updated during training. Only the projection layers and classification head will be updated.
+        """
+        self.encoder._freeze_parameters()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.encoder.get_input_embeddings()
+
+    def set_input_embeddings(self, value: nn.Module):
+        self.encoder.set_input_embeddings(value)
+
+    @add_start_docstrings_to_model_forward(WHISPER_ENCODER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=SequenceClassifierOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_features: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> import torch
+        >>> from transformers import AutoFeatureExtractor, WhisperForAudioClassification
+        >>> from datasets import load_dataset
+
+        >>> feature_extractor = AutoFeatureExtractor.from_pretrained("sanchit-gandhi/whisper-medium-fleurs-lang-id")
+        >>> model = WhisperForAudioClassification.from_pretrained("sanchit-gandhi/whisper-medium-fleurs-lang-id")
+
+        >>> ds = load_dataset("google/fleurs", "all", split="validation", streaming=True)
+        >>> sample = next(iter(ds))
+
+        >>> inputs = feature_extractor(
+        ...     sample["audio"]["array"], sampling_rate=sample["audio"]["sampling_rate"], return_tensors="pt"
+        ... )
+        >>> input_features = inputs.input_features
+
+        >>> with torch.no_grad():
+        ...     logits = model(input_features).logits
+
+        >>> predicted_class_ids = torch.argmax(logits).item()
+        >>> predicted_label = model.config.id2label[predicted_class_ids]
+        >>> predicted_label
+        'af_za'
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_features,
+                head_mask=head_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+
+        if self.config.use_weighted_layer_sum:
+            hidden_states = torch.stack(encoder_outputs, dim=1)
+            norm_weights = nn.functional.softmax(self.layer_weights, dim=-1)
+            hidden_states = (hidden_states * norm_weights.view(-1, 1, 1)).sum(dim=1)
+        else:
+            hidden_states = encoder_outputs[0]
+
+        hidden_states = self.projector(hidden_states)
+        pooled_output = hidden_states.mean(dim=1)
+
+        logits = self.classifier(pooled_output)
+
+        loss = None
+
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # move labels to correct device to enable PP
+            labels = labels.to(logits.device)
+            loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + encoder_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )

models/salmonn.py

@@ -0,0 +1,506 @@
+# Copyright (2024) Tsinghua University, Bytedance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import json
+import contextlib
+import random
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import LlamaTokenizer, StoppingCriteriaList
+from peft import LoraConfig, TaskType, get_peft_model
+
+from .Qformer import BertConfig, BertLMHeadModel
+from .modeling_llama import LlamaForCausalLM
+from .modeling_whisper import WhisperModel
+from .beats.BEATs import BEATsConfig, BEATs
+from .utils import StoppingCriteriaSub
+
+
+class SALMONN(nn.Module):
+    @classmethod
+    def init_speech_Qformer(cls, num_query_token, speech_width, num_hidden_layers=2):
+        encoder_config = BertConfig.from_pretrained("bert-base-uncased")
+        encoder_config.num_hidden_layers = num_hidden_layers
+        encoder_config.encoder_width = speech_width
+        # insert cross-attention layer every other block
+        encoder_config.add_cross_attention = True
+        encoder_config.cross_attention_freq = 1
+        encoder_config.query_length = num_query_token
+        Qformer = BertLMHeadModel(config=encoder_config)
+        query_tokens = nn.Parameter(
+            torch.zeros(1, num_query_token, encoder_config.hidden_size)
+        )
+        query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range)
+        return Qformer, query_tokens
+
+    @property
+    def device(self):
+        return list(self.parameters())[0].device
+
+    def maybe_autocast(self, dtype=torch.float16):
+        # if on cpu, don't use autocast
+        # if on gpu, use autocast with dtype if provided, otherwise use torch.float16
+        enable_autocast = self.device != torch.device("cpu")
+
+        if enable_autocast:
+            return torch.cuda.amp.autocast(dtype=dtype)
+        else:
+            return contextlib.nullcontext()
+
+    def __init__(
+        self,
+        llama_path="",
+        whisper_path="",
+        freeze_whisper=True,
+        beats_path="",
+        freeze_beats=True,
+
+        use_speech_Qformer=True,
+        num_speech_query_token=1,
+        freeze_speech_QFormer=False,
+        window_level_Qformer=True,
+        second_per_window=0.333333,
+        second_stride=0.333333,
+        
+        speech_llama_proj_model="",
+        freeze_speech_llama_proj=False,
+
+        lora=True,
+        lora_rank=8,
+        lora_alpha=32,
+        lora_dropout=0.1,
+
+        multi_prompt=False,
+        prompt_path="",
+        prompt_template="",
+        max_txt_len=128,
+        end_sym="</s>",
+        low_resource=False,  # use 8 bit
+        device_8bit=0,  # the device of 8bit model should be set when loading and cannot be changed anymore.
+    ):
+        super().__init__()
+
+        self.beats_path = beats_path
+        self.use_speech_Qformer = use_speech_Qformer
+        self.window_level_Qformer = window_level_Qformer
+        self.second_per_window = second_per_window
+        self.second_stride = second_stride
+        self.lora = lora
+        self.multi_prompt = multi_prompt
+        self.max_txt_len = max_txt_len
+        self.end_sym = end_sym
+        self.low_resource = low_resource
+
+        logging.info('Loading LLaMA Tokenizer')
+        self.llama_tokenizer = LlamaTokenizer.from_pretrained(llama_path, use_fast=False)
+        self.llama_tokenizer.add_special_tokens({'pad_token': '[PAD]'})
+        self.llama_tokenizer.padding_side = "right"
+
+        logging.info('Loading LLaMA Model')
+        if self.low_resource:
+            self.llama_model = LlamaForCausalLM.from_pretrained(
+                llama_path,
+                torch_dtype=torch.float16,
+                load_in_8bit=True,
+                device_map={"": device_8bit},
+            )
+        else:
+            self.llama_model = LlamaForCausalLM.from_pretrained(
+                llama_path,
+                torch_dtype=torch.float16,
+            )
+
+        self.llama_model.resize_token_embeddings(len(self.llama_tokenizer))
+        for name, param in self.llama_model.named_parameters():
+            param.requires_grad = False
+        logging.info('Loading LLaMA Done')
+
+        if self.lora:
+            self.peft_config = LoraConfig(
+                task_type=TaskType.CAUSAL_LM, 
+                inference_mode=False, 
+                r=lora_rank, 
+                lora_alpha=lora_alpha, 
+                lora_dropout=lora_dropout,
+            )
+            self.llama_model = get_peft_model(self.llama_model, self.peft_config)
+            self.llama_model.print_trainable_parameters()
+            logging.info('LoRA Training')
+
+        assert whisper_path
+        logging.info('Loading Whisper Model')
+        self.speech_encoder = WhisperModel.from_pretrained(whisper_path).encoder
+        self.ln_speech = nn.LayerNorm(self.speech_encoder.config.d_model)
+        if freeze_whisper:
+            for name, param in self.speech_encoder.named_parameters():
+                param.requires_grad = False
+            self.speech_encoder.eval()
+            logging.info("freeze Whisper")
+        
+        if self.beats_path:
+            logging.info("Loading BEATs Model")
+            beats_ckpt = torch.load(self.beats_path, map_location='cpu')
+            beats_cfg = BEATsConfig(beats_ckpt['cfg'])
+            self.beats = BEATs(beats_cfg)
+            self.beats.load_state_dict(beats_ckpt['model'])
+            self.ln_audio = nn.LayerNorm(self.beats.cfg.encoder_embed_dim)
+            if freeze_beats:
+                for name, param in self.beats.named_parameters():
+                    param.requires_grad = False
+                self.beats.eval()
+                logging.info("freeze BEATs")
+
+        if self.use_speech_Qformer:
+            if self.beats_path:
+                self.speech_Qformer, self.speech_query_tokens = self.init_speech_Qformer(
+                    num_query_token=num_speech_query_token, speech_width=self.speech_encoder.config.d_model + self.beats.cfg.encoder_embed_dim
+                )
+            else:
+                self.speech_Qformer, self.speech_query_tokens = self.init_speech_Qformer(
+                    num_query_token=num_speech_query_token, speech_width=self.speech_encoder.config.d_model
+                )
+            self.speech_Qformer.bert.embeddings.word_embeddings = None
+            self.speech_Qformer.bert.embeddings.position_embeddings = None
+            for layer in self.speech_Qformer.bert.encoder.layer:
+                layer.output = None
+                layer.intermediate = None
+            self.speech_Qformer.cls = None
+            if freeze_speech_QFormer:
+                for name, param in self.speech_Qformer.named_parameters():
+                    param.requires_grad = False
+                self.speech_Qformer.eval()
+                self.speech_query_tokens.requires_grad = False
+                logging.info("freeze Speech QFormer")
+
+            logging.info('Loading speech LLAMA proj')
+            self.speech_llama_proj = nn.Linear(
+                self.speech_Qformer.config.hidden_size, self.llama_model.config.hidden_size
+            )
+            if speech_llama_proj_model:
+                logging.info("Loading speech LLAMA proj from {}".format(speech_llama_proj_model))
+                speech_llama_proj_weight = torch.load(speech_llama_proj_model, map_location="cpu")
+                self.load_state_dict(speech_llama_proj_weight['model'], strict=False)
+            if freeze_speech_llama_proj:
+                for name, param in self.speech_llama_proj.named_parameters():
+                    param.requires_grad = False
+                self.speech_llama_proj.eval()
+                logging.info("freeze speech LLAMA proj")
+        else:
+            # feel free to add other aligners here
+            raise NotImplementedError
+
+        # prepare prompts
+        self.prompt_dict = {}
+        if prompt_path:
+            try:
+                raw_prompts = json.load(open(prompt_path, "r"))
+            except:
+                print("Failed to load prompt! Try to use utf-8 encoding.")
+                raw_prompts = json.load(open(prompt_path, "r", encoding='utf-8'))
+            for task in raw_prompts.keys():
+                filted_prompts = [raw_prompt for raw_prompt in raw_prompts[task] if "<SpeechHere>" in raw_prompt]
+                self.prompt_dict[task] = [prompt_template.format(p) for p in filted_prompts]
+            print("Loading training prompts done!")
+
+    def _encode_auditory_feature(self, speech_embeds, audio_embeds=None):
+        with self.maybe_autocast():
+            if self.use_speech_Qformer:
+                speech_embeds = self.ln_speech(speech_embeds)
+                if audio_embeds is not None:
+                    audio_embeds = self.ln_audio(audio_embeds)
+                    if audio_embeds.size(1) < speech_embeds.size(1):
+                        audio_embeds = F.pad(audio_embeds, (0, 0, 0, speech_embeds.size(1) - audio_embeds.size(1)))
+                    elif audio_embeds.size(1) > speech_embeds.size(1):
+                        speech_embeds = F.pad(speech_embeds, (0, 0, 0, audio_embeds.size(1) - speech_embeds.size(1)))
+                    speech_embeds = torch.cat((speech_embeds, audio_embeds), dim=-1)
+                speech_atts = torch.ones(speech_embeds.size()[:-1], dtype=torch.long).to(speech_embeds.device)
+
+                if self.window_level_Qformer:
+                    B, T, C = speech_embeds.shape
+                    kernel = round(1500 * self.second_per_window / 30.0)
+                    stride = round(1500 * self.second_stride / 30.0)
+                    kernel = (1, kernel)
+                    stride = (1, stride)
+                    speech_embeds_tr = speech_embeds.transpose(1, 2).unsqueeze(2)
+                    speech_embeds_overlap = F.unfold(speech_embeds_tr, kernel_size=kernel, dilation=1, padding=0, stride=stride)
+                    _, _, L = speech_embeds_overlap.shape
+                    speech_embeds_overlap = speech_embeds_overlap.view(B, -1, kernel[1], L)
+                    speech_embeds_overlap = torch.permute(speech_embeds_overlap, [0, 3, 2, 1])
+                    speech_embeds = speech_embeds_overlap.reshape(-1, kernel[1], C)
+                    speech_atts = torch.ones(speech_embeds.size()[:-1], dtype=torch.long, device=speech_embeds.device)
+
+                query_tokens = self.speech_query_tokens.expand(speech_embeds.shape[0], -1, -1)
+                query_output = self.speech_Qformer.bert(
+                    query_embeds=query_tokens,
+                    encoder_hidden_states=speech_embeds,
+                    encoder_attention_mask=speech_atts,
+                    return_dict=True,
+                )
+                speech_embeds = self.speech_llama_proj(query_output.last_hidden_state)
+
+                if self.window_level_Qformer:
+                    speech_embeds = speech_embeds.view(B, -1, speech_embeds.size(2)).contiguous()
+
+                speech_atts = torch.ones(speech_embeds.size()[:-1], dtype=torch.long).to(speech_embeds.device)
+            else:
+                raise NotImplementedError
+
+        return speech_embeds, speech_atts
+
+    def encode_speech(self, spectrogram, raw_wav=None, audio_padding_mask=None):
+        with self.maybe_autocast():
+            speech_embeds = self.speech_encoder(spectrogram, return_dict=True).last_hidden_state
+
+            if self.beats_path and raw_wav is not None:
+                audio_embeds, _ = self.beats.extract_features(raw_wav, padding_mask=audio_padding_mask, feature_only=True)
+            else:
+                audio_embeds = None
+
+        return self._encode_auditory_feature(speech_embeds, audio_embeds=audio_embeds)
+
+    def prompt_wrap(self, embeds, atts, prompt, multi_prompt=False):
+        if prompt:
+            if multi_prompt:
+                p_before = []
+                p_after = []
+                for i, p in enumerate(prompt):
+                    b, a = p.split("<SpeechHere>")
+                    p_before.append(b)
+                    p_after.append(a)
+                
+                p_before_tokens = self.llama_tokenizer(
+                    p_before, return_tensors="pt", add_special_tokens=False
+                ).to(embeds.device)
+                p_before_embeds = self.llama_model.model.embed_tokens(p_before_tokens.input_ids) if not self.lora else self.llama_model.model.model.embed_tokens(p_before_tokens.input_ids)
+
+                # speech_embeds wrapped with prompts_embeds are padded to the same length here
+                p_after_tokens = self.llama_tokenizer(
+                    p_after, return_tensors="pt", padding="longest", add_special_tokens=False
+                ).to(embeds.device)
+                p_after_embeds = self.llama_model.model.embed_tokens(p_after_tokens.input_ids) if not self.lora else self.llama_model.model.model.embed_tokens(p_after_tokens.input_ids)
+
+                wrapped_embeds = torch.cat([p_before_embeds, embeds, p_after_embeds], dim=1)
+                wrapped_atts = torch.cat([p_before_tokens.attention_mask, atts, p_after_tokens.attention_mask], dim=1)
+            else:
+                batch_size = embeds.shape[0]
+                p_before, p_after = prompt.split("<SpeechHere>")
+
+                p_before_tokens = self.llama_tokenizer(
+                    p_before, return_tensors="pt", add_special_tokens=False
+                ).to(embeds.device)
+                p_after_tokens = self.llama_tokenizer(
+                    p_after, return_tensors="pt", add_special_tokens=False
+                ).to(embeds.device)
+                p_before_embeds = self.llama_model.model.embed_tokens(p_before_tokens.input_ids).expand(batch_size, -1, -1) if not self.lora else self.llama_model.model.model.embed_tokens(p_before_tokens.input_ids).expand(batch_size, -1, -1)
+                p_after_embeds = self.llama_model.model.embed_tokens(p_after_tokens.input_ids).expand(batch_size, -1, -1) if not self.lora else self.llama_model.model.model.embed_tokens(p_after_tokens.input_ids).expand(batch_size, -1, -1)
+
+                wrapped_embeds = torch.cat([p_before_embeds, embeds, p_after_embeds], dim=1)
+                wrapped_atts = torch.cat([p_before_tokens.attention_mask, atts, p_after_tokens.attention_mask], dim=1)
+            return wrapped_embeds, wrapped_atts
+        else:
+            return embeds, atts
+
+    def forward(self, samples, verbose=False):
+        # detect whether there are multi tasks in this batch
+        task = list(set(samples["task"]))
+        if len(task) > 1 or "QA" in task:
+            self.multi_prompt = True
+
+        # prepare prompts
+        if self.prompt_dict:
+            if self.multi_prompt:
+                prompt = [random.choice(self.prompt_dict[task]) for task in samples["task"]]
+                if "Q" in samples:
+                    prompt = [p.format(q) if '{}' in p else p for p, q in zip(prompt, samples["Q"]) ]
+            else:
+                prompt = random.choice(self.prompt_dict[samples["task"][0]])
+
+        # use speech/audio encoder to encode speech/audio
+        spectrogram = samples["spectrogram"]
+        raw_wav = samples.get("raw_wav", None)
+        audio_padding_mask = samples.get("padding_mask", None)
+
+        speech_embeds, speech_atts = self.encode_speech(spectrogram, raw_wav=raw_wav, audio_padding_mask=audio_padding_mask)
+
+        # wrap speech_embeds with prompts
+        if self.prompt_dict:
+            speech_embeds, speech_atts = self.prompt_wrap(speech_embeds, speech_atts, prompt, multi_prompt=self.multi_prompt)
+
+        # prepare inputs for LLM
+        text = [t + self.end_sym for t in samples["text"]]
+        to_regress_tokens = self.llama_tokenizer(
+            text,
+            return_tensors="pt",
+            padding="longest",
+            truncation=True,
+            max_length=self.max_txt_len,
+            add_special_tokens=False
+        ).to(spectrogram.device)
+        to_regress_embeds = self.llama_model.model.embed_tokens(to_regress_tokens.input_ids) if not self.lora else self.llama_model.model.model.embed_tokens(to_regress_tokens.input_ids)
+        targets = to_regress_tokens.input_ids.masked_fill(
+            to_regress_tokens.input_ids == self.llama_tokenizer.pad_token_id, -100
+        )
+        empty_targets = (
+            torch.ones(
+                [speech_atts.shape[0], speech_atts.shape[1] + 1],
+                dtype=torch.long
+            ).to(spectrogram.device).fill_(-100)
+        )
+        targets = torch.cat([empty_targets, targets], dim=1)
+
+        batch_size = speech_embeds.shape[0]
+        bos = torch.ones(
+            [batch_size, 1],
+            dtype=to_regress_tokens.input_ids.dtype,
+            device=to_regress_tokens.input_ids.device,
+        ) * self.llama_tokenizer.bos_token_id
+        bos_embeds = self.llama_model.model.embed_tokens(bos) if not self.lora else self.llama_model.model.model.embed_tokens(bos)
+        atts_bos = speech_atts[:, :1]
+
+        inputs_embeds = torch.cat([bos_embeds, speech_embeds, to_regress_embeds], dim=1)
+        attention_mask = torch.cat([atts_bos, speech_atts, to_regress_tokens.attention_mask], dim=1)
+
+        # calulate loss
+        with self.maybe_autocast():
+            outputs = self.llama_model(
+                inputs_embeds=inputs_embeds,
+                attention_mask=attention_mask,
+                return_dict=True,
+                labels=targets,
+            )
+            loss = outputs.loss
+
+        if verbose:
+            nvocab = self.llama_model.config.vocab_size
+            results = outputs.logits[:, empty_targets.size(1) - 1: -1, :].contiguous().view(-1, nvocab).argmax(dim=-1)
+            labels = targets[:, empty_targets.size(1):].contiguous().view(-1)
+            mask = (labels != -100)
+            correct = (results[mask] == labels[mask]).float().sum()
+            total = len(labels[mask])
+
+        if verbose:
+            return {"loss": loss, "correct": correct, "total": total}
+
+        return {"loss": loss}
+
+    def generate(self, samples, generate_cfg, prompts=None):
+        batch_size = samples["spectrogram"].shape[0]
+
+        spectrogram = samples["spectrogram"]
+        raw_wav = samples.get("raw_wav", None)
+        audio_padding_mask = samples.get("padding_mask", None)
+
+        speech_embeds, speech_atts = self.encode_speech(spectrogram, raw_wav=raw_wav, audio_padding_mask=audio_padding_mask)
+
+        if prompts is not None:
+            speech_embeds, speech_atts = self.prompt_wrap(speech_embeds, speech_atts, prompts, multi_prompt=True)
+
+        bos = torch.ones(
+            [batch_size, 1],
+            dtype=torch.int32,
+            device=speech_embeds.device,
+        ) * self.llama_tokenizer.bos_token_id
+        bos_embeds = self.llama_model.model.embed_tokens(bos) if not self.lora else self.llama_model.model.model.embed_tokens(bos)
+        atts_bos = speech_atts[:, :1]
+
+        embeds = torch.cat([bos_embeds, speech_embeds], dim=1)
+        attns = torch.cat([atts_bos, speech_atts], dim=1)
+
+        stop_words_ids = [torch.tensor([2]).cuda()]  
+        stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(stops=stop_words_ids)])
+        outputs = self.llama_model.generate(
+            inputs_embeds=embeds,
+            max_new_tokens=generate_cfg.get("max_new_tokens", 200),
+            stopping_criteria=stopping_criteria,
+            num_beams=generate_cfg.get("num_beams", 4),
+            do_sample=generate_cfg.get("do_sample", False),
+            min_length=generate_cfg.get("min_length", 1),
+            temperature=generate_cfg.get("temperature", 1.0),
+            top_p=generate_cfg.get("top_p", 0.9),
+            repetition_penalty=generate_cfg.get("repetition_penalty", 1.0),
+            length_penalty=generate_cfg.get("length_penalty", 1.0),
+            attention_mask=attns,
+        )
+        text = self.llama_tokenizer.batch_decode(outputs, add_special_tokens=False)
+
+        return text
+
+    @classmethod
+    def from_config(cls, config):
+        llama_path = config.get("llama_path")
+        whisper_path = config.get("whisper_path")
+        freeze_whisper = config.get("freeze_whisper", True)
+        beats_path = config.get("beats_path", "")
+        freeze_beats = config.get("freeze_beats", True)
+
+        use_speech_Qformer = config.get("use_speech_Qformer", True)
+        num_speech_query_token = config.get("num_speech_query_token", 1)
+        freeze_speech_QFormer = config.get("freeze_speech_QFormer", False)
+        window_level_Qformer = config.get("window_level_Qformer", True)
+        second_per_window = config.get("second_per_window", 0.333333)
+        second_stride = config.get("second_stride", 0.333333)
+
+        speech_llama_proj_model = config.get("speech_llama_proj_model", "")
+        freeze_speech_llama_proj = config.get("freeze_speech_llama_proj", False)
+
+        lora = config.get("lora", True)
+        lora_rank = config.get("lora_rank", 8)
+        lora_alpha = config.get("lora_alpha", 32)
+        lora_dropout = config.get("lora_dropout", 0.1)
+
+        multi_prompt = config.get("multi_prompt", False)
+        prompt_path = config.get("prompt_path", "")
+        prompt_template = config.get("prompt_template", "")
+        max_txt_len = config.get("max_txt_len", 128)
+        end_sym = config.get("end_sym", "</s>")
+        low_resource = config.get("low_resource", False)
+        device_8bit = config.get("device_8bit", 0)
+
+        model = cls(
+            llama_path=llama_path,
+            whisper_path=whisper_path,
+            freeze_whisper=freeze_whisper,
+            beats_path=beats_path,
+            freeze_beats=freeze_beats,
+            use_speech_Qformer=use_speech_Qformer,
+            num_speech_query_token=num_speech_query_token,
+            freeze_speech_QFormer=freeze_speech_QFormer,
+            window_level_Qformer=window_level_Qformer,
+            second_per_window=second_per_window,
+            second_stride=second_stride,
+            speech_llama_proj_model=speech_llama_proj_model,
+            freeze_speech_llama_proj=freeze_speech_llama_proj,
+            lora=lora,
+            lora_rank=lora_rank,
+            lora_alpha=lora_alpha,
+            lora_dropout=lora_dropout,
+            multi_prompt=multi_prompt,
+            prompt_path=prompt_path,
+            prompt_template=prompt_template,
+            max_txt_len=max_txt_len,
+            end_sym=end_sym,
+            low_resource=low_resource,
+            device_8bit=device_8bit,
+        )
+
+        ckpt_path = config.get("ckpt", "")
+        if ckpt_path:
+            logging.info("Load SALMONN ckpt from: {}".format(ckpt_path))
+            ckpt = torch.load(ckpt_path, map_location="cpu")
+            model.load_state_dict(ckpt['model'], strict=False)
+
+        return model

models/utils.py

@@ -0,0 +1,30 @@
+# Copyright (2024) Tsinghua University, Bytedance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+from transformers import StoppingCriteria
+
+
+class StoppingCriteriaSub(StoppingCriteria):
+
+    def __init__(self, stops=[], encounters=1):
+        super().__init__()
+        self.stops = stops
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor):
+        for stop in self.stops:
+            if torch.all((stop == input_ids[0][-len(stop):])).item():
+                return True
+
+        return False

prompts/test_prompt.json

@@ -0,0 +1,27 @@
+{
+    "asr": "<Speech><SpeechHere></Speech> Recognize the speech and give me the transcription.",
+    "asr_zh": "<Speech><SpeechHere></Speech> 请将语音中的内容写下来。",
+    "asr_de": "<Speech><SpeechHere></Speech> Hören Sie sich die Rede an und schreiben Sie ihren Inhalt auf.",
+    "translation_ec": "<Speech><SpeechHere></Speech> Listen to the speech and translate it into Chinese.",
+    "audiocaption": "<Speech><SpeechHere></Speech> Please describe the audio.",
+    "audiocaption_v2": "<Speech><SpeechHere></Speech> Please write down what your hear in the audio.",
+    "QA": "<Speech><SpeechHere></Speech> {}",
+    "gender_QA": "<Speech><SpeechHere></Speech> {}",
+    "phone_recognition": "<Speech><SpeechHere></Speech> Provide the phonetic transcription for the speech.",
+    "speech_query": "<Speech><SpeechHere></Speech> Please answer the question in detail.",
+    "emotion_recognition": "<Speech><SpeechHere></Speech> Describe the emotion of the speaker in one word.", 
+    "lyrics_recognition": "<Speech><SpeechHere></Speech> Listen to the song and write down its content.",
+    "audio_speech_description": "<Speech><SpeechHere></Speech> Describe the speech and the background audio",
+    "speaker_verification": "<Speech><SpeechHere></Speech> Do you only hear the same person talking? Answer yes or no.",
+    "fluent_speech_audio": "<Speech><SpeechHere></Speech> Describe the background audio and the speech in a fluent sentence.",
+    "speech_separation": "<Speech><SpeechHere></Speech> Please write down what you hear each person says.",
+    "audio_story_telling": "<Speech><SpeechHere></Speech> Based on the audio, write a story in detail. Your story should be highly related to the audio.",
+    "speech_audio_query": "<Speech><SpeechHere></Speech> Please answer the speaker's question in detail based on the background sound.",
+    "slot_filling": "<Speech><SpeechHere></Speech> According to the speech, what is the {}?",
+    "music_description": "<Speech><SpeechHere></Speech> Listen to this music clip and describe the music.",
+    "translation_en2ja": "<Speech><SpeechHere></Speech> Listen to the speech and translate it into Japanese.",
+    "translation_en2de": "<Speech><SpeechHere></Speech> Listen to the speech and translate it into German.",
+    "speech_audio_coreasoning": "<Speech><SpeechHere></Speech> Use your strong reasoning skills to answer the speaker's question in detail based on the background sound.",
+    "keywords": "<Speech><SpeechHere></Speech> Give me only three keywords of the text.",
+    "speaker_diarization_asr": "<Speech><SpeechHere></Speech> Please recognize each speaker and transcribe their speech content."
+}

prompts/train_prompt.json

@@ -0,0 +1,132 @@
+{
+    "asr": [
+        "<Speech><SpeechHere></Speech> Can you transcribe the speech into a written format?",
+        "<Speech><SpeechHere></Speech> Listen to the speech and write down its content.",
+        "<Speech><SpeechHere></Speech> What is the content of the speech you heard?",
+        "<Speech><SpeechHere></Speech> Please write down the transcription of the speech.",
+        "<Speech><SpeechHere></Speech> Please transcribe the speech into a written format.",
+        "<Speech><SpeechHere></Speech> Write down the content of the speech you heard.",
+        "<Speech><SpeechHere></Speech> Can you write down the transcription of the speech?",
+        "<Speech><SpeechHere></Speech> Put the speech into a written format.",
+        "<Speech><SpeechHere></Speech> Please help me to transcribe the speech into a written format.",
+        "<Speech><SpeechHere></Speech> Recognize the content of the speech you heard.",
+        "<Speech><SpeechHere></Speech> Can you recognize what you heard in the speech?",
+        "<Speech><SpeechHere></Speech> Recognize the speech and write it down in a written format.",
+        "<Speech><SpeechHere></Speech> Listen to the speech and recognize its content.",
+        "<Speech><SpeechHere></Speech> Give me the transcription of the speech you heard.",
+        "<Speech><SpeechHere></Speech> Recognize the speech and give me the transcription."
+    ],
+    "asr_zh": [
+        "<Speech><SpeechHere></Speech> 前面的语音说了什么？",
+        "<Speech><SpeechHere></Speech> 请将语音中的内容写下来。",
+        "<Speech><SpeechHere></Speech> 请识别这段中文语音。",
+        "<Speech><SpeechHere></Speech> 听前面的音频，写出对方说的内容。",
+        "<Speech><SpeechHere></Speech> 写下你听到的内容。",
+        "<Speech><SpeechHere></Speech> 请记下语音中人说的话。",
+        "<Speech><SpeechHere></Speech> 仔细听这段语音，记下语音中的话",
+        "<Speech><SpeechHere></Speech> 将你听到的话写下来",
+        "<Speech><SpeechHere></Speech> 这个人说了什么？请记下来。",
+        "<Speech><SpeechHere></Speech> 请将语音转换为文字",
+        "<Speech><SpeechHere></Speech> 请识别这个人说的内容"
+    ],
+    "asr_de": [
+        "<Speech><SpeechHere></Speech> Können Sie die Rede in ein schriftliches Format übertragen?",
+        "<Speech><SpeechHere></Speech> Hören Sie sich die Rede an und schreiben Sie ihren Inhalt auf.",
+        "<Speech><SpeechHere></Speech> Bitte notieren Sie die Transkription der Rede.",
+        "<Speech><SpeechHere></Speech> Geben Sie mir die Transkription der Rede, die Sie gehört haben.",
+        "<Speech><SpeechHere></Speech> Was hat dieser Mann gesagt? Bitte schreiben Sie es auf.",
+        "<Speech><SpeechHere></Speech> Können Sie die Transkription der Rede aufschreiben?",
+        "<Speech><SpeechHere></Speech> Hören Sie der Stimme aufmerksam zu und notieren Sie die Wörter in der Stimme",
+        "<Speech><SpeechHere></Speech> Schreiben Sie auf, was Sie hören.",
+        "<Speech><SpeechHere></Speech> Bitte Sprache in Text umwandeln.",
+        "<Speech><SpeechHere></Speech> Erkennen Sie den Inhalt der Rede, die Sie gehört haben."
+    ],
+    "translation_ec": [
+        "<Speech><SpeechHere></Speech> Can you translate the speech into Chinese?",
+        "<Speech><SpeechHere></Speech> Please translate the speech you heard into Chinese.",
+        "<Speech><SpeechHere></Speech> Listen to the speech and translate it into Chinese.",
+        "<Speech><SpeechHere></Speech> Give me the Chinese translation of this speech.",
+        "<Speech><SpeechHere></Speech> Could you please provide a Chinese translation for the speech?",
+        "<Speech><SpeechHere></Speech> Would you be willing to translate the speech into Chinese for me?",
+        "<Speech><SpeechHere></Speech> Would you be able to render the speech in Chinese?",
+        "<Speech><SpeechHere></Speech> Could you assist me in translating the speech into Chinese?",
+        "<Speech><SpeechHere></Speech> Can you help me convert the speech into Chinese text?",
+        "<Speech><SpeechHere></Speech> Please convert the speech into Chinese text.",
+        "<Speech><SpeechHere></Speech> 请将这段语音的内容翻译成中文。",
+        "<Speech><SpeechHere></Speech> 你能把这段语音用中文表达出来吗？",
+        "<Speech><SpeechHere></Speech> 请将你听到的语音用中文写出来。"
+    ],
+    "audiocaption": [
+        "<Speech><SpeechHere></Speech> Listen to this audio clip and provide its caption.",
+        "<Speech><SpeechHere></Speech> Describe the following audio in a caption.",
+        "<Speech><SpeechHere></Speech> Based on the sound you hear, create a caption for this audio.",
+        "<Speech><SpeechHere></Speech> Can you describe the scene or event depicted in this audio?",
+        "<Speech><SpeechHere></Speech> Could you summarise what's happening in this audio?",
+        "<Speech><SpeechHere></Speech> What does this audio describe?",
+        "<Speech><SpeechHere></Speech> Please describe the audio."
+    ],
+    "audiocaption_v2": [
+        "<Speech><SpeechHere></Speech> Please write down what your hear in the audio."
+    ],
+    "QA": [
+        "<Speech><SpeechHere></Speech> {}"
+    ],
+    "inference_QA": [
+        "<Speech><SpeechHere></Speech> {}"
+    ],
+    "gender_QA": [
+        "<Speech><SpeechHere></Speech> {}"
+    ],
+    "gender_recognition": [
+        "<Speech><SpeechHere></Speech> What is the gender of the speaker?",
+        "<Speech><SpeechHere></Speech> Use one word to describe the speaker's gender.",
+        "<Speech><SpeechHere></Speech> Describe the speaker's gender.",
+        "<Speech><SpeechHere></Speech> Can you accurately identify the gender of the speaker?",
+        "<Speech><SpeechHere></Speech> Can you distinguish the gender of the speaker?",
+        "<Speech><SpeechHere></Speech> Describe the gender of the person speaking.",
+        "<Speech><SpeechHere></Speech> What is the speaker's gender based on the audio?",
+        "<Speech><SpeechHere></Speech> Tell me about the gender of the person you hear.",
+        "<Speech><SpeechHere></Speech> Is the speaker male or female?"
+    ],
+    "phone_recognition": [
+        "<Speech><SpeechHere></Speech> Please transcribe the audio clip into its corresponding phonetic representation.",
+        "<Speech><SpeechHere></Speech> Write the sequence of phonemes corresponding to this speech.",
+        "<Speech><SpeechHere></Speech> Provide the phonetic transcription for the speech.",
+        "<Speech><SpeechHere></Speech> Transcribe the phonemes for the speech please.",
+        "<Speech><SpeechHere></Speech> Can you recognize the phonetic representation in the speech?",
+        "<Speech><SpeechHere></Speech> Listen to the speech and recognize its phonetic representation",
+        "<Speech><SpeechHere></Speech> What is the phoneme transcription of the speech?"
+    ],
+    "speech_separation": [
+        "<Speech><SpeechHere></Speech> There are two people talking in the audio, please write what they say in order.",
+        "<Speech><SpeechHere></Speech> Please write down what you hear each person says.",
+        "<Speech><SpeechHere></Speech> Can you record what each person says?",
+        "<Speech><SpeechHere></Speech> Transcribe the words spoken by each person in the audio."
+    ],
+    "emotion_recognition": [
+        "<Speech><SpeechHere></Speech> Describe the emotion of the speaker in one word.",
+        "<Speech><SpeechHere></Speech> Use one word to describe the speaker's emotion."
+    ],
+    "music_description": [
+        "<Speech><SpeechHere></Speech> Listen to this music clip and describe the music.",
+        "<Speech><SpeechHere></Speech> Please describe the music.",
+        "<Speech><SpeechHere></Speech> Provide a description of the music.",
+        "<Speech><SpeechHere></Speech> Analyze the music in this clip and offer a description.",
+        "<Speech><SpeechHere></Speech> Give me a description of the music in this clip."
+    ],
+    "speaker_verification": [
+        "<Speech><SpeechHere></Speech> Are the two people speaking successively the same person? Answer yes or no.",
+        "<Speech><SpeechHere></Speech> Do you only hear the same person talking? Answer yes or no.",
+        "<Speech><SpeechHere></Speech> Is only one person speaking in the audio? Answer yes or no."
+    ],
+    "audio_story_telling": [
+        "<Speech><SpeechHere></Speech> Based on the audio, write a story in detail. Your story should be highly related to the audio.",
+        "<Speech><SpeechHere></Speech> Please write a story in detail based on the audio. Your story should contain all the elements in the audio.",
+        "<Speech><SpeechHere></Speech> Please generate a long story that is highly related to the audio."
+    ],
+    "speaker_diarization_asr": [
+        "<Speech><SpeechHere></Speech> Identify each speaker in turn and what is said.",
+        "<Speech><SpeechHere></Speech> Write down the content of each speaker and the corresponding speech in turn.",
+        "<Speech><SpeechHere></Speech> Please recognize each speaker and transcribe their speech content."
+    ]
+}

requirements.txt

@@ -0,0 +1,14 @@
+# SALMONN Inference Requirements
+torch>=2.0.0
+transformers>=4.35.0
+peft>=0.6.0
+accelerate>=0.24.0
+soundfile>=0.12.0
+librosa>=0.10.0
+numpy<2.0.0
+omegaconf>=2.3.0
+fastapi>=0.104.0
+uvicorn>=0.24.0
+python-multipart>=0.0.6
+pydantic>=2.0.0
+aiofiles>=23.0.0

server.py

@@ -0,0 +1,184 @@
+"""
+SALMONN FastAPI Server
+HTTP API for audio understanding and transcription.
+"""
+
+import os
+import tempfile
+import shutil
+from pathlib import Path
+from typing import Optional
+
+import yaml
+import uvicorn
+from fastapi import FastAPI, File, UploadFile, Form, HTTPException
+from fastapi.responses import JSONResponse
+from pydantic import BaseModel
+from omegaconf import OmegaConf
+
+from inference import SALMONNInference
+
+# Load config
+CONFIG_PATH = os.environ.get("SALMONN_CONFIG", "config.yaml")
+
+with open(CONFIG_PATH, "r") as f:
+    config = OmegaConf.create(yaml.safe_load(f))
+
+# Initialize FastAPI app
+app = FastAPI(
+    title="SALMONN API",
+    description="Audio Language Model for Speech, Audio Events, and Music Understanding",
+    version="1.0.0",
+)
+
+# Global model instance (loaded on startup)
+model: Optional[SALMONNInference] = None
+
+
+class TranscribeResponse(BaseModel):
+    text: str
+    status: str = "success"
+
+
+class ChatResponse(BaseModel):
+    question: str
+    answer: str
+    status: str = "success"
+
+
+class HealthResponse(BaseModel):
+    status: str
+    model_loaded: bool
+    device: str
+
+
+@app.on_event("startup")
+async def startup_event():
+    """Load model on startup."""
+    global model
+    print("Starting SALMONN server...")
+    model = SALMONNInference(CONFIG_PATH)
+    model.load()
+    print("Server ready!")
+
+
+@app.get("/", response_model=dict)
+async def root():
+    """Root endpoint with API info."""
+    return {
+        "name": "SALMONN API",
+        "version": "1.0.0",
+        "endpoints": {
+            "/health": "Health check",
+            "/transcribe": "Transcribe audio (POST)",
+            "/chat": "Ask questions about audio (POST)",
+        }
+    }
+
+
+@app.get("/health", response_model=HealthResponse)
+async def health():
+    """Health check endpoint."""
+    return HealthResponse(
+        status="healthy" if model and model._loaded else "loading",
+        model_loaded=model._loaded if model else False,
+        device=str(model.device) if model else "unknown",
+    )
+
+
+@app.post("/transcribe", response_model=TranscribeResponse)
+async def transcribe(
+    audio: UploadFile = File(..., description="Audio file (wav, mp3, etc.)"),
+):
+    """
+    Transcribe an audio file to text.
+    
+    - **audio**: Audio file to transcribe
+    
+    Returns transcribed text.
+    """
+    if not model or not model._loaded:
+        raise HTTPException(status_code=503, detail="Model not loaded yet")
+    
+    # Save uploaded file temporarily
+    suffix = Path(audio.filename).suffix if audio.filename else ".wav"
+    with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
+        shutil.copyfileobj(audio.file, tmp)
+        tmp_path = tmp.name
+    
+    try:
+        text = model.transcribe(tmp_path)
+        return TranscribeResponse(text=text)
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+    finally:
+        os.unlink(tmp_path)
+
+
+@app.post("/chat", response_model=ChatResponse)
+async def chat(
+    audio: UploadFile = File(..., description="Audio file (wav, mp3, etc.)"),
+    question: str = Form(..., description="Question about the audio"),
+):
+    """
+    Ask a question about an audio file.
+    
+    - **audio**: Audio file to analyze
+    - **question**: Question about the audio content
+    
+    Returns the model's answer.
+    """
+    if not model or not model._loaded:
+        raise HTTPException(status_code=503, detail="Model not loaded yet")
+    
+    # Save uploaded file temporarily
+    suffix = Path(audio.filename).suffix if audio.filename else ".wav"
+    with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
+        shutil.copyfileobj(audio.file, tmp)
+        tmp_path = tmp.name
+    
+    try:
+        answer = model.chat(tmp_path, question)
+        return ChatResponse(question=question, answer=answer)
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+    finally:
+        os.unlink(tmp_path)
+
+
+@app.post("/describe")
+async def describe(
+    audio: UploadFile = File(..., description="Audio file (wav, mp3, etc.)"),
+):
+    """
+    Get a detailed description of the audio content.
+    
+    - **audio**: Audio file to describe
+    
+    Returns description of the audio.
+    """
+    if not model or not model._loaded:
+        raise HTTPException(status_code=503, detail="Model not loaded yet")
+    
+    # Save uploaded file temporarily
+    suffix = Path(audio.filename).suffix if audio.filename else ".wav"
+    with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
+        shutil.copyfileobj(audio.file, tmp)
+        tmp_path = tmp.name
+    
+    try:
+        description = model.describe(tmp_path)
+        return {"description": description, "status": "success"}
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+    finally:
+        os.unlink(tmp_path)
+
+
+if __name__ == "__main__":
+    uvicorn.run(
+        "server:app",
+        host=config.server.host,
+        port=config.server.port,
+        reload=config.server.get("reload", False),
+    )

utils.py

@@ -0,0 +1,158 @@
+# Copyright (2024) Tsinghua University, Bytedance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import time
+
+import torch
+from torch.utils.data import DataLoader, DistributedSampler
+import soundfile as sf
+import numpy as np
+
+from dist_utils import is_main_process, get_world_size, get_rank
+
+
+def now():
+    from datetime import datetime
+
+    return datetime.now().strftime("%Y%m%d%H%M")
+
+
+def setup_logger():
+    logging.basicConfig(
+        level=logging.INFO if is_main_process() else logging.WARN,
+        format="%(asctime)s [%(levelname)s] %(message)s",
+        handlers=[logging.StreamHandler()],
+    )
+
+
+def get_dataloader(dataset, config, is_train=True, use_distributed=True):
+    if use_distributed:
+        sampler = DistributedSampler(
+            dataset,
+            shuffle=is_train,
+            num_replicas=get_world_size(),
+            rank=get_rank()
+        )
+    else:
+        sampler = None
+
+    loader = DataLoader(
+        dataset,
+        batch_size=config.batch_size_train if is_train else config.batch_size_eval,
+        num_workers=config.num_workers,
+        pin_memory=True,
+        sampler=sampler,
+        shuffle=sampler is None and is_train,
+        collate_fn=dataset.collater,
+        drop_last=is_train,
+    )
+
+    if is_train:
+        loader = IterLoader(loader, use_distributed=use_distributed)
+
+    return loader
+
+
+def apply_to_sample(f, sample):
+    if len(sample) == 0:
+        return {}
+
+    def _apply(x):
+        if torch.is_tensor(x):
+            return f(x)
+        elif isinstance(x, dict):
+            return {key: _apply(value) for key, value in x.items()}
+        elif isinstance(x, list):
+            return [_apply(x) for x in x]
+        else:
+            return x
+
+    return _apply(sample)
+
+
+def move_to_cuda(sample):
+    def _move_to_cuda(tensor):
+        return tensor.cuda()
+
+    return apply_to_sample(_move_to_cuda, sample)
+
+
+def prepare_sample(samples, cuda_enabled=True):
+    if cuda_enabled:
+        samples = move_to_cuda(samples)
+
+    # TODO fp16 support
+
+    return samples
+
+
+class IterLoader:
+    """
+    A wrapper to convert DataLoader as an infinite iterator.
+
+    Modified from:
+        https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/iter_based_runner.py
+    """
+
+    def __init__(self, dataloader: DataLoader, use_distributed: bool = False):
+        self._dataloader = dataloader
+        self.iter_loader = iter(self._dataloader)
+        self._use_distributed = use_distributed
+        self._epoch = 0
+
+    @property
+    def epoch(self) -> int:
+        return self._epoch
+
+    def __next__(self):
+        try:
+            data = next(self.iter_loader)
+        except StopIteration:
+            self._epoch += 1
+            if hasattr(self._dataloader.sampler, "set_epoch") and self._use_distributed:
+                self._dataloader.sampler.set_epoch(self._epoch)
+            time.sleep(2)  # Prevent possible deadlock during epoch transition
+            self.iter_loader = iter(self._dataloader)
+            data = next(self.iter_loader)
+
+        return data
+
+    def __iter__(self):
+        return self
+
+    def __len__(self):
+        return len(self._dataloader)
+
+
+def prepare_one_sample(wav_path, wav_processor, cuda_enabled=True):
+    audio, sr = sf.read(wav_path)
+    if len(audio.shape) == 2: # stereo to mono
+        audio = audio[:, 0]
+    if len(audio) < sr: # pad audio to at least 1s
+        sil = np.zeros(sr - len(audio), dtype=float)
+        audio = np.concatenate((audio, sil), axis=0)
+    audio = audio[: sr * 30] # truncate audio to at most 30s
+
+    spectrogram = wav_processor(audio, sampling_rate=sr, return_tensors="pt", padding="max_length")["input_features"]
+
+    samples = {
+        "spectrogram": spectrogram,
+        "raw_wav": torch.from_numpy(audio).unsqueeze(0),
+        "padding_mask": torch.zeros(len(audio), dtype=torch.bool).unsqueeze(0),
+    }
+    if cuda_enabled:
+        samples = move_to_cuda(samples)
+
+    return samples