training / flax /latency_scripts /run_speculative.sh
nullonesix's picture
Saving train state of step 1
a1be16b verified
raw
history blame
1.3 kB
#!/usr/bin/env bash
# batch_sizes=(1 4)
batch_sizes=(1)
names=("openai/whisper-large-v2" "openai/whisper-large-v2" "openai/whisper-medium.en" "openai/whisper-medium.en")
assistant_names=("patrickvonplaten/whisper-large-v2-32-2" "openai/whisper-small" "patrickvonplaten/whisper-medium-24-2" "openai/whisper-base.en")
# --assistant_model_name_or_path "patrickvonplaten/whisper-large-v2-32-2" \
# --use_pipeline \
# Double loop
for (( i=0; i<${#names[*]}; ++i)); do
name=${names[$i]}
assistant_name=${assistant_names[$i]}
for batch_size in "${batch_sizes[@]}"; do
CUDA_VISIBLE_DEVICES="0" python ./run_speed_pt.py \
--dataset_name "distil-whisper/chime4+distil-whisper/earnings22+google/fleurs+kensho/spgispeech" \
--wandb_name "FP16-RTX-4090-bsz${batch_size}-${name}-${assistant_name}" \
--model_name_or_path ${name} \
--wandb_project "distil-whisper-speed-bench-check-spec-dec-final" \
--dataset_config_name "1-channel+chunked+en_us+test" \
--dataset_split_name "test+test+test+test" \
--text_column_name "text+transcription+transcription+transcript" \
--attn_type "flash2" \
--assistant_model_name_or_path ${assistant_name} \
--samples_per_dataset "10" \
--batch_size ${batch_size}
done
done