Apple Silicon compatibility

.gitignore

@@ -7,4 +7,8 @@ __pycache__/
 /checkpoints
 /asset
 data/tos.py
-evaluation/livesports3kcc/livecc/*/*.json
+evaluation/livesports3kcc/livecc/*/*.json
+.gradio/certificate.pem
+# Custom ignored
+.DS_Store
+/decord

README.md

@@ -27,6 +27,81 @@ pip install livecc-utils==0.0.2
 
 We finished all things in ```torch==2.6.0```, ```transformers==4.50.0```, ```liger-kernel==4.50.0```. But other versions should also work. Our full environment is [requirements.txt](requirements.txt).
 
+---
+#### Installation on Apple Silicon
+Tested on M1 Max and M2 Max with 32Gb RAM. Not really real time but working ([Live demo](demo/sources/LiveCC_on_Apple_Silicon.mp4)).
+**/!\ _Do not use requirements.txt_**
+##### Requirements
+Install developper tools
+```sh
+xcode-select --install
+```
+Install [oh-my-zsh](https://ohmyz.sh) (optionnal)
+```sh
+sh -c "$(curl -fsSL https://raw.githubusercontent.com/ohmyzsh/ohmyzsh/master/tools/install.sh)"
+```
+Install [miniconda](https://www.anaconda.com/docs/getting-started/miniconda/install#mac-os-terminal-installer)
+```sh
+curl -O https://repo.anaconda.com/miniconda/Miniconda3-latest-MacOSX-arm64.sh
+bash ~/Miniconda3-latest-MacOSX-arm64.sh
+```
+
+Install [brew](https://brew.sh)
+```sh
+/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
+
+echo >> ~/.zprofile
+echo 'eval "$(/opt/homebrew/bin/brew shellenv)"' >> ~/.zprofile
+eval "$(/opt/homebrew/bin/brew shellenv)"
+```
+Install git-lfs
+```sh
+pip3 install git-lfs
+git lfs install
+```
+##### Install LiveCC for Apple Silicon
+Clone this other repo and switch to apple_silicon branch
+```sh
+git clone https://github.com/LeJeko/livecc.git
+cd livecc
+git checkout -f apple_silicon
+```
+Create conda environment
+```sh
+conda create -n livecc python=3.12
+conda activate livecc 
+```
+Install this specific version of ffmpeg
+```sh
+brew install ffmpeg@4
+echo 'export PATH="/opt/homebrew/opt/ffmpeg@4/bin:$PATH"' >> ~/.zshrc
+brew link ffmpeg@4
+```
+Compile decord for arm64
+```sh
+pip install cmake
+
+git clone --recursive https://github.com/dmlc/decord
+cd decord
+mkdir build && cd build
+cmake .. -DCMAKE_BUILD_TYPE=Release
+make
+cd ../python
+pip install . --user
+cd ../..
+```
+Install dependencies
+```sh
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
+pip install transformers accelerate peft opencv-python datasets tensorboard gradio pillow-heif gpustat timm sentencepiece openai av qwen_vl_utils
+pip install livecc-utils
+```
+##### Launch LiveCC
+```sh
+python -m demo.app
+```
+
+---
 #### Advanced
 
 If you want to delve into our data production pipeline:

demo/app.py

@@ -52,6 +52,7 @@ def __call__(self, message: str = None, history: list[str] = None, state: dict =
             gr_examples = gr.Examples(
                 examples=[
                     'demo/sources/howto_fix_laptop_mute_1080p.mp4',
+                    'demo/sources/LiveCC_on_Apple_Silicon.mp4',
                     'demo/sources/writing_mute_1080p.mp4',
                     'demo/sources/spacex_falcon9_mute_1080p.mp4',
                     'demo/sources/warriors_vs_rockets_2025wcr1_mute_1080p.mp4',

demo/infer.py

@@ -1,6 +1,4 @@
 import functools, torch
-from liger_kernel.transformers import apply_liger_kernel_to_qwen2_vl
-apply_liger_kernel_to_qwen2_vl()
 from transformers import Qwen2VLForConditionalGeneration, AutoProcessor, LogitsProcessor, logging
 from livecc_utils import prepare_multiturn_multimodal_inputs_for_generation, get_smart_resized_clip, get_smart_resized_video_reader, _read_video_decord_plus, _spatial_resize_video
 from qwen_vl_utils import process_vision_info
@@ -32,11 +30,18 @@ class LiveCCDemoInfer:
     streaming_time_interval = streaming_fps_frames / fps
     frame_time_interval = 1 / fps
 
-    def __init__(self, model_path: str = None, device: str = 'cuda'):
+    def __init__(self, model_path: str = None):
+        if torch.backends.mps.is_available():
+            device = 'mps'
+        elif torch.cuda.is_available():
+            device = 'cuda'
+        else:
+            device = 'cpu'
+        print(f"Using device: {device}")
         self.model = Qwen2VLForConditionalGeneration.from_pretrained(
-            model_path, torch_dtype="auto", 
-            device_map=device, 
-            attn_implementation='flash_attention_2'
+            model_path, torch_dtype="auto",
+            device_map=device,
+            attn_implementation='flash_attention_2' if device == 'cuda' else None
         )
         self.processor = AutoProcessor.from_pretrained(model_path, use_fast=False)
         self.streaming_eos_token_id = self.processor.tokenizer(' ...').input_ids[-1]

requirements.txt

@@ -312,4 +312,4 @@ xxhash==3.5.0
 yarl==1.18.3
 yt-dlp==2025.2.19
 zipp==3.21.0
-zstandard==0.19.0
+zstandard==0.19.0