Fix transcription worker to handle bot detection.

See JuanBindez/pytubefix#209

Also up the TFLOPS to avoid fp16 incapable GPUs and then
use the large-v3-turbo model for smaller size and faster
inference.

.github/workflows/build-all.yml

@@ -81,4 +81,4 @@ jobs:
 
       - name: Deploy Functions
         if: ${{ inputs.deploy-functions != '' }}
-        run: npm --prefix functions deploy -P ${{ inputs.deploy-functions }}
+        run: npm --prefix functions run deploy -P ${{ inputs.deploy-functions }}

functions-python/main.py

@@ -74,7 +74,7 @@ def start_transcribe(
                 storage=DISK_GB,
                 order="dph_total",
                 query=('cpu_cores_effective>=10 cpu_ram>=32 gpu_total_ram>=8 '
-                       'reliability>=0.95 dph<2 total_flops>=10'))
+                       'reliability>=0.95 dph<2 total_flops>=12'))
             cheapest = json.loads(offers_json)[0]
             print(f"cheapest ask id {cheapest['ask_contract_id']} cost "
                   f"{json.dumps(cheapest)}")

tools/process_new_vids/Dockerfile

@@ -7,13 +7,18 @@ COPY prime_model.mp4 /workspace/app/
 RUN apt-get -y update && \
   apt-get -y install git ffmpeg && \
   pip install git+https://github.com/m-bain/whisperx.git && \
-  whisperx --model=large-v3 --language=en --thread=4 --hf_token=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 --diarize --output_form=json --compute_type=int8 --output_dir=/tmp -- prime_model.mp4
+  whisperx --model=large-v3-turbo --language=en --thread=4 --hf_token=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 --diarize --output_form=json --compute_type=int8 --output_dir=/tmp -- prime_model.mp4
 
 RUN pip install vastai
 
 COPY requirements.txt /workspace/app/
-RUN  pip install -r requirements.txt
+RUN pip install -r requirements.txt
 
+RUN <<EOF
+apt-get -y install curl unzip
+curl -fsSL https://fnm.vercel.app/install | bash
+~/.local/share/fnm/fnm install --lts
+EOF
 
 COPY transcribe_worker.py lysine_protocol.sh onstart_hook.sh /workspace/app/
 RUN chmod 755 /workspace/app/transcribe_worker.py /workspace/app/lysine_protocol.sh /workspace/app/onstart_hook.sh

tools/process_new_vids/onstart_hook.sh

@@ -1,4 +1,9 @@
 #!/bin/bash
 
+export PATH=${HOME}/.local/share/fnm:$PATH
+
+eval "$(fnm env --use-on-cd --shell bash)"
+fnm use lts/latest
+
 # Default to giving 10 mins on failure.
 /workspace/app/lysine_protocol.sh "${3-:30}" & python /workspace/app/transcribe_worker.py -w /tmp/transcribe -t "${1:-4}" -x "$2" -m large-v3 -c -s; /workspace/app/lysine_protocol.sh "${4:-10}"

tools/process_new_vids/transcribe_worker.py

@@ -158,7 +158,7 @@ def main():
                         required=True)
     parser.add_argument('-m', '--model', dest='model', metavar="MODEL",
                         type=str, help='Downloads whisper MODEL',
-                        default="large-v3")
+                        default="large-v3-turbo")
     parser.add_argument('--compute_type', dest='compute_type',
                         metavar="COMPUTE_TYPE", type=str,
                         help='The compute type to use', default="float16")