Spaces:
Paused
Paused
Automate Docker building
Browse files- README.md +137 -47
- README.qmd +4 -4
- docker/scripts/setup-whisperbot.sh +2 -7
- docker/scripts/setup.sh +6 -0
- requirements.txt +4 -1
README.md
CHANGED
@@ -1,29 +1,52 @@
|
|
1 |
# WhisperBot
|
2 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
|
4 |
## Features
|
5 |
-
- **Real-Time Speech-to-Text**: Utilizes OpenAI WhisperLive to convert spoken language into text in real-time.
|
6 |
|
7 |
-
- **
|
|
|
|
|
|
|
|
|
|
|
8 |
|
9 |
-
- **TensorRT Optimization**: Both Mistral and Whisper are optimized to
|
|
|
|
|
10 |
|
11 |
## Prerequisites
|
12 |
-
|
13 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
|
15 |
### Build Whisper TensorRT Engine
|
16 |
|
17 |
> [!NOTE]
|
18 |
>
|
19 |
-
> These steps are included in `
|
20 |
|
21 |
Change working dir to the [whisper example
|
22 |
dir](https://github.com/NVIDIA/TensorRT-LLM/tree/main/examples/whisper)
|
23 |
in TensorRT-LLM.
|
24 |
|
25 |
``` bash
|
26 |
-
cd TensorRT-LLM
|
27 |
```
|
28 |
|
29 |
Currently, by default TensorRT-LLM only supports `large-v2` and
|
@@ -33,7 +56,7 @@ Download the required assets
|
|
33 |
|
34 |
``` bash
|
35 |
# the sound filter definitions
|
36 |
-
wget --directory-prefix=assets
|
37 |
# the small.en model weights
|
38 |
wget --directory-prefix=assets https://openaipublic.azureedge.net/main/whisper/models/f953ad0fd29cacd07d5a9eda5624af0f6bcf2258be67c92b79389873d91e0872/small.en.pt
|
39 |
```
|
@@ -62,15 +85,23 @@ model:
|
|
62 |
``` bash
|
63 |
pip install -r requirements.txt
|
64 |
python3 build.py --output_dir whisper_small_en --use_gpt_attention_plugin --use_gemm_plugin --use_layernorm_plugin --use_bert_attention_plugin --model_name small.en
|
|
|
|
|
65 |
```
|
66 |
|
67 |
### Build Mistral TensorRT Engine
|
68 |
-
|
69 |
-
|
70 |
-
|
|
|
|
|
|
|
|
|
71 |
```
|
72 |
-
|
73 |
-
|
|
|
|
|
74 |
python build.py --model_dir teknium/OpenHermes-2.5-Mistral-7B \
|
75 |
--dtype float16 \
|
76 |
--remove_input_padding \
|
@@ -78,20 +109,30 @@ python build.py --model_dir teknium/OpenHermes-2.5-Mistral-7B \
|
|
78 |
--enable_context_fmha \
|
79 |
--use_gemm_plugin float16 \
|
80 |
--output_dir ./tmp/mistral/7B/trt_engines/fp16/1-gpu/ \
|
81 |
-
--max_input_len 5000
|
82 |
--max_batch_size 1
|
|
|
|
|
83 |
```
|
84 |
|
85 |
### Build Phi TensorRT Engine
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
90 |
```
|
91 |
-
|
92 |
-
|
|
|
|
|
93 |
git lfs install
|
94 |
-
|
95 |
python3 build.py --dtype=float16 \
|
96 |
--log_level=verbose \
|
97 |
--use_gpt_attention_plugin float16 \
|
@@ -99,46 +140,95 @@ python3 build.py --dtype=float16 \
|
|
99 |
--max_batch_size=16 \
|
100 |
--max_input_len=1024 \
|
101 |
--max_output_len=1024 \
|
102 |
-
--output_dir=
|
103 |
-
--model_dir=
|
|
|
|
|
|
|
|
|
|
|
104 |
```
|
105 |
|
106 |
-
##
|
107 |
-
|
108 |
-
|
109 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
110 |
cd WhisperBot
|
111 |
apt update
|
112 |
apt install ffmpeg portaudio19-dev -y
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
113 |
pip install -r requirements.txt
|
|
|
114 |
```
|
115 |
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
|
|
|
|
|
|
|
|
|
|
123 |
```
|
124 |
|
125 |
-
### Whisper
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
132 |
```
|
133 |
|
134 |
-
- On the client side clone the repo, install the requirements and
|
135 |
-
|
|
|
|
|
136 |
cd WhisperBot
|
137 |
pip install -r requirements.txt
|
138 |
python3 run_client.py
|
139 |
```
|
140 |
|
141 |
-
|
142 |
## Contact Us
|
143 |
-
|
144 |
-
|
|
|
|
|
|
1 |
# WhisperBot
|
2 |
+
|
3 |
+
|
4 |
+
Welcome to WhisperBot. WhisperBot builds upon the capabilities of the
|
5 |
+
[WhisperLive](https://github.com/collabora/WhisperLive) and
|
6 |
+
[WhisperSpeech](https://github.com/collabora/WhisperSpeech) by
|
7 |
+
integrating Mistral, a Large Language Model (LLM), on top of the
|
8 |
+
real-time speech-to-text pipeline. WhisperLive relies on OpenAI Whisper,
|
9 |
+
a powerful automatic speech recognition (ASR) system. Both Mistral and
|
10 |
+
Whisper are optimized to run efficiently as TensorRT engines, maximizing
|
11 |
+
performance and real-time processing capabilities.
|
12 |
|
13 |
## Features
|
|
|
14 |
|
15 |
+
- **Real-Time Speech-to-Text**: Utilizes OpenAI WhisperLive to convert
|
16 |
+
spoken language into text in real-time.
|
17 |
+
|
18 |
+
- **Large Language Model Integration**: Adds Mistral, a Large Language
|
19 |
+
Model, to enhance the understanding and context of the transcribed
|
20 |
+
text.
|
21 |
|
22 |
+
- **TensorRT Optimization**: Both Mistral and Whisper are optimized to
|
23 |
+
run as TensorRT engines, ensuring high-performance and low-latency
|
24 |
+
processing.
|
25 |
|
26 |
## Prerequisites
|
27 |
+
|
28 |
+
Install
|
29 |
+
[TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM/blob/main/docs/source/installation.md)
|
30 |
+
to build Whisper and Mistral TensorRT engines. The README builds a
|
31 |
+
docker image for TensorRT-LLM. Instead of building a docker image, we
|
32 |
+
can also refer to the README and the
|
33 |
+
[Dockerfile.multi](https://github.com/NVIDIA/TensorRT-LLM/blob/main/docker/Dockerfile.multi)
|
34 |
+
to install the required packages in the base pytroch docker image. Just
|
35 |
+
make sure to use the correct base image as mentioned in the dockerfile
|
36 |
+
and everything should go nice and smooth.
|
37 |
|
38 |
### Build Whisper TensorRT Engine
|
39 |
|
40 |
> [!NOTE]
|
41 |
>
|
42 |
+
> These steps are included in `docker/scripts/build-whisper.sh`
|
43 |
|
44 |
Change working dir to the [whisper example
|
45 |
dir](https://github.com/NVIDIA/TensorRT-LLM/tree/main/examples/whisper)
|
46 |
in TensorRT-LLM.
|
47 |
|
48 |
``` bash
|
49 |
+
cd /root/TensorRT-LLM-examples/whisper
|
50 |
```
|
51 |
|
52 |
Currently, by default TensorRT-LLM only supports `large-v2` and
|
|
|
56 |
|
57 |
``` bash
|
58 |
# the sound filter definitions
|
59 |
+
wget --directory-prefix=assets https://raw.githubusercontent.com/openai/whisper/main/whisper/assets/mel_filters.npz
|
60 |
# the small.en model weights
|
61 |
wget --directory-prefix=assets https://openaipublic.azureedge.net/main/whisper/models/f953ad0fd29cacd07d5a9eda5624af0f6bcf2258be67c92b79389873d91e0872/small.en.pt
|
62 |
```
|
|
|
85 |
``` bash
|
86 |
pip install -r requirements.txt
|
87 |
python3 build.py --output_dir whisper_small_en --use_gpt_attention_plugin --use_gemm_plugin --use_layernorm_plugin --use_bert_attention_plugin --model_name small.en
|
88 |
+
mkdir -p /root/scratch-space/models
|
89 |
+
cp -r whisper_small_en /root/scratch-space/models
|
90 |
```
|
91 |
|
92 |
### Build Mistral TensorRT Engine
|
93 |
+
|
94 |
+
> [!NOTE]
|
95 |
+
>
|
96 |
+
> These steps are included in `docker/scripts/build-mistral.sh`
|
97 |
+
|
98 |
+
``` bash
|
99 |
+
cd /root/TensorRT-LLM-examples/llama
|
100 |
```
|
101 |
+
|
102 |
+
Build TensorRT for Mistral with `fp16`
|
103 |
+
|
104 |
+
``` bash
|
105 |
python build.py --model_dir teknium/OpenHermes-2.5-Mistral-7B \
|
106 |
--dtype float16 \
|
107 |
--remove_input_padding \
|
|
|
109 |
--enable_context_fmha \
|
110 |
--use_gemm_plugin float16 \
|
111 |
--output_dir ./tmp/mistral/7B/trt_engines/fp16/1-gpu/ \
|
112 |
+
--max_input_len 5000 \
|
113 |
--max_batch_size 1
|
114 |
+
mkdir -p /root/scratch-space/models
|
115 |
+
cp -r tmp/mistral/7B/trt_engines/fp16/1-gpu /root/scratch-space/models/mistral
|
116 |
```
|
117 |
|
118 |
### Build Phi TensorRT Engine
|
119 |
+
|
120 |
+
> [!NOTE]
|
121 |
+
>
|
122 |
+
> These steps are included in `docker/scripts/build-phi-2.sh`
|
123 |
+
|
124 |
+
Note: Phi is only available in main branch and hasnt been released yet.
|
125 |
+
So, make sure to build TensorRT-LLM from main branch.
|
126 |
+
|
127 |
+
``` bash
|
128 |
+
cd /root/TensorRT-LLM-examples/phi
|
129 |
```
|
130 |
+
|
131 |
+
Build TensorRT for Phi-2 with `fp16`
|
132 |
+
|
133 |
+
``` bash
|
134 |
git lfs install
|
135 |
+
phi_path=$(huggingface-cli download --repo-type model --revision 834565c23f9b28b96ccbeabe614dd906b6db551a microsoft/phi-2)
|
136 |
python3 build.py --dtype=float16 \
|
137 |
--log_level=verbose \
|
138 |
--use_gpt_attention_plugin float16 \
|
|
|
140 |
--max_batch_size=16 \
|
141 |
--max_input_len=1024 \
|
142 |
--max_output_len=1024 \
|
143 |
+
--output_dir=phi-2 \
|
144 |
+
--model_dir="$phi_path" >&1 | tee build.log
|
145 |
+
dest=/root/scratch-space/models
|
146 |
+
mkdir -p "$dest/phi-2/tokenizer"
|
147 |
+
cp -r phi-2 "$dest"
|
148 |
+
(cd "$phi_path" && cp config.json tokenizer_config.json vocab.json merges.txt "$dest/phi-2/tokenizer")
|
149 |
+
cp -r "$phi_path" "$dest/phi-orig-model"
|
150 |
```
|
151 |
|
152 |
+
## Build WhisperBot
|
153 |
+
|
154 |
+
> [!NOTE]
|
155 |
+
>
|
156 |
+
> These steps are included in `docker/scripts/setup-whisperbot.sh`
|
157 |
+
|
158 |
+
Clone this repo and install requirements
|
159 |
+
|
160 |
+
``` bash
|
161 |
+
[ -d "WhisperBot" ] || git clone https://github.com/collabora/WhisperBot.git
|
162 |
cd WhisperBot
|
163 |
apt update
|
164 |
apt install ffmpeg portaudio19-dev -y
|
165 |
+
```
|
166 |
+
|
167 |
+
Install torchaudio matching the PyTorch from the base image
|
168 |
+
|
169 |
+
``` bash
|
170 |
+
pip install --extra-index-url https://download.pytorch.org/whl/cu121 torchaudio
|
171 |
+
```
|
172 |
+
|
173 |
+
Install all the other dependencies normally
|
174 |
+
|
175 |
+
``` bash
|
176 |
pip install -r requirements.txt
|
177 |
+
pip install openai-whisper whisperspeech soundfile
|
178 |
```
|
179 |
|
180 |
+
force update huggingface_hub (tokenizers 0.14.1 spuriously require and
|
181 |
+
ancient \<=0.18 version)
|
182 |
+
|
183 |
+
``` bash
|
184 |
+
pip install -U huggingface_hub
|
185 |
+
huggingface-cli download collabora/whisperspeech t2s-small-en+pl.model s2a-q4-tiny-en+pl.model
|
186 |
+
huggingface-cli download charactr/vocos-encodec-24khz
|
187 |
+
mkdir -p /root/.cache/torch/hub/checkpoints/
|
188 |
+
curl -L -o /root/.cache/torch/hub/checkpoints/encodec_24khz-d7cc33bc.th https://dl.fbaipublicfiles.com/encodec/v0/encodec_24khz-d7cc33bc.th
|
189 |
+
mkdir -p /root/.cache/whisper-live/
|
190 |
+
curl -L -o /root/.cache/whisper-live/silero_vad.onnx https://github.com/snakers4/silero-vad/raw/master/files/silero_vad.onnx
|
191 |
+
python -c 'from transformers.utils.hub import move_cache; move_cache()'
|
192 |
```
|
193 |
|
194 |
+
### Run WhisperBot with Whisper and Mistral/Phi-2
|
195 |
+
|
196 |
+
Take the folder path for Whisper TensorRT model, folder_path and
|
197 |
+
tokenizer_path for Mistral/Phi-2 TensorRT from the build phase. If a
|
198 |
+
huggingface model is used to build mistral/phi-2 then just use the
|
199 |
+
huggingface repo name as the tokenizer path.
|
200 |
+
|
201 |
+
> [!NOTE]
|
202 |
+
>
|
203 |
+
> These steps are included in `docker/scripts/run-whisperbot.sh`
|
204 |
+
|
205 |
+
``` bash
|
206 |
+
test -f /etc/shinit_v2 && source /etc/shinit_v2
|
207 |
+
cd WhisperBot
|
208 |
+
if [ "$1" != "mistral" ]; then
|
209 |
+
exec python3 main.py --phi \
|
210 |
+
--whisper_tensorrt_path /root/whisper_small_en \
|
211 |
+
--phi_tensorrt_path /root/phi-2 \
|
212 |
+
--phi_tokenizer_path /root/phi-2
|
213 |
+
else
|
214 |
+
exec python3 main.py --mistral \
|
215 |
+
--whisper_tensorrt_path /root/models/whisper_small_en \
|
216 |
+
--mistral_tensorrt_path /root/models/mistral \
|
217 |
+
--mistral_tokenizer_path teknium/OpenHermes-2.5-Mistral-7B
|
218 |
+
fi
|
219 |
```
|
220 |
|
221 |
+
- On the client side clone the repo, install the requirements and
|
222 |
+
execute `run_client.py`
|
223 |
+
|
224 |
+
``` bash
|
225 |
cd WhisperBot
|
226 |
pip install -r requirements.txt
|
227 |
python3 run_client.py
|
228 |
```
|
229 |
|
|
|
230 |
## Contact Us
|
231 |
+
|
232 |
+
For questions or issues, please open an issue. Contact us at:
|
233 |
+
marcus.edel@collabora.com, jpc@collabora.com,
|
234 |
+
vineet.suryan@collabora.com
|
README.qmd
CHANGED
@@ -29,7 +29,7 @@ These steps are included in `{fname}`
|
|
29 |
|
30 |
# WhisperBot
|
31 |
|
32 |
-
Welcome to WhisperBot. WhisperBot builds upon the capabilities of the [WhisperLive]() by integrating Mistral, a Large Language Model (LLM), on top of the real-time speech-to-text pipeline. WhisperLive relies on OpenAI Whisper, a powerful automatic speech recognition (ASR) system. Both Mistral and Whisper are optimized to run efficiently as TensorRT engines, maximizing performance and real-time processing capabilities.
|
33 |
|
34 |
## Features
|
35 |
- **Real-Time Speech-to-Text**: Utilizes OpenAI WhisperLive to convert spoken language into text in real-time.
|
@@ -45,19 +45,19 @@ Instead of building a docker image, we can also refer to the README and the [Doc
|
|
45 |
### Build Whisper TensorRT Engine
|
46 |
|
47 |
```{python}
|
48 |
-
include_file('docker/scripts/
|
49 |
```
|
50 |
|
51 |
### Build Mistral TensorRT Engine
|
52 |
|
53 |
```{python}
|
54 |
-
include_file('docker/scripts/
|
55 |
```
|
56 |
|
57 |
### Build Phi TensorRT Engine
|
58 |
|
59 |
```{python}
|
60 |
-
include_file('docker/scripts/
|
61 |
```
|
62 |
|
63 |
## Build WhisperBot
|
|
|
29 |
|
30 |
# WhisperBot
|
31 |
|
32 |
+
Welcome to WhisperBot. WhisperBot builds upon the capabilities of the [WhisperLive](https://github.com/collabora/WhisperLive) and [WhisperSpeech](https://github.com/collabora/WhisperSpeech) by integrating Mistral, a Large Language Model (LLM), on top of the real-time speech-to-text pipeline. WhisperLive relies on OpenAI Whisper, a powerful automatic speech recognition (ASR) system. Both Mistral and Whisper are optimized to run efficiently as TensorRT engines, maximizing performance and real-time processing capabilities.
|
33 |
|
34 |
## Features
|
35 |
- **Real-Time Speech-to-Text**: Utilizes OpenAI WhisperLive to convert spoken language into text in real-time.
|
|
|
45 |
### Build Whisper TensorRT Engine
|
46 |
|
47 |
```{python}
|
48 |
+
include_file('docker/scripts/build-whisper.sh')
|
49 |
```
|
50 |
|
51 |
### Build Mistral TensorRT Engine
|
52 |
|
53 |
```{python}
|
54 |
+
include_file('docker/scripts/build-mistral.sh')
|
55 |
```
|
56 |
|
57 |
### Build Phi TensorRT Engine
|
58 |
|
59 |
```{python}
|
60 |
+
include_file('docker/scripts/build-phi-2.sh')
|
61 |
```
|
62 |
|
63 |
## Build WhisperBot
|
docker/scripts/setup-whisperbot.sh
CHANGED
@@ -7,15 +7,11 @@ cd WhisperBot
|
|
7 |
apt update
|
8 |
apt install ffmpeg portaudio19-dev -y
|
9 |
|
10 |
-
##
|
11 |
-
|
12 |
-
#apt install -y cmake
|
13 |
-
#TORCH_CUDA_ARCH_LIST="8.9 9.0" pip install --no-build-isolation git+https://github.com/pytorch/audio.git
|
14 |
|
15 |
## Install all the other dependencies normally
|
16 |
-
pip install --extra-index-url https://download.pytorch.org/whl/cu121 torchaudio
|
17 |
pip install -r requirements.txt
|
18 |
-
pip install openai-whisper whisperspeech soundfile
|
19 |
|
20 |
## force update huggingface_hub (tokenizers 0.14.1 spuriously require and ancient <=0.18 version)
|
21 |
pip install -U huggingface_hub
|
@@ -29,4 +25,3 @@ mkdir -p /root/.cache/whisper-live/
|
|
29 |
curl -L -o /root/.cache/whisper-live/silero_vad.onnx https://github.com/snakers4/silero-vad/raw/master/files/silero_vad.onnx
|
30 |
|
31 |
python -c 'from transformers.utils.hub import move_cache; move_cache()'
|
32 |
-
|
|
|
7 |
apt update
|
8 |
apt install ffmpeg portaudio19-dev -y
|
9 |
|
10 |
+
## Install torchaudio matching the PyTorch from the base image
|
11 |
+
pip install --extra-index-url https://download.pytorch.org/whl/cu121 torchaudio
|
|
|
|
|
12 |
|
13 |
## Install all the other dependencies normally
|
|
|
14 |
pip install -r requirements.txt
|
|
|
15 |
|
16 |
## force update huggingface_hub (tokenizers 0.14.1 spuriously require and ancient <=0.18 version)
|
17 |
pip install -U huggingface_hub
|
|
|
25 |
curl -L -o /root/.cache/whisper-live/silero_vad.onnx https://github.com/snakers4/silero-vad/raw/master/files/silero_vad.onnx
|
26 |
|
27 |
python -c 'from transformers.utils.hub import move_cache; move_cache()'
|
|
docker/scripts/setup.sh
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/bin/bash -e
|
2 |
+
|
3 |
+
./setup-whisper.sh
|
4 |
+
#./setup-mistral.sh
|
5 |
+
./setup-phi-2.sh
|
6 |
+
./setup-whisperbot.sh
|
requirements.txt
CHANGED
@@ -6,4 +6,7 @@ scipy
|
|
6 |
websocket-client
|
7 |
tiktoken==0.3.3
|
8 |
kaldialign
|
9 |
-
braceexpand
|
|
|
|
|
|
|
|
6 |
websocket-client
|
7 |
tiktoken==0.3.3
|
8 |
kaldialign
|
9 |
+
braceexpand
|
10 |
+
openai-whisper
|
11 |
+
whisperspeech
|
12 |
+
soundfile
|