39 lines
832 B
Bash
39 lines
832 B
Bash
|
|
source ~/environments/clearml/bin/activate
|
|
|
|
recipe_template=$(cat <<'EOF'
|
|
quant_stage:
|
|
quant_modifiers:
|
|
QuantizationModifier:
|
|
ignore: ["lm_head"]
|
|
scheme: FP8
|
|
targets: ["Linear"]
|
|
observer: "mse"
|
|
EOF
|
|
)
|
|
|
|
for size in 0.5B 1.5B 3B 7B 32B 72B
|
|
do
|
|
for version in base instruct
|
|
do
|
|
|
|
|
|
if [ $version = "base" ]; then
|
|
model="Qwen2.5-${size}"
|
|
else
|
|
model="Qwen2.5-${size}-Instruct"
|
|
fi
|
|
|
|
prefix="${model//./_}""__llm_compressor__calibration__mse__512__8196__damp01"
|
|
|
|
python /cache/git/research/automation/pipelines/pipeline_llmcompressor_oneshot.py \
|
|
--model-id "Qwen/"$model \
|
|
--project-name "LLM quantization - FP8/llmcompressor/Qwen2.5" \
|
|
--task-prefix $prefix \
|
|
--recipe "${recipe}" \
|
|
--num-samples 512 \
|
|
--max-seq-len 8196 \
|
|
--tags "Qwen2.5" "W4A16" "calibration" $size "MSE" $version
|
|
|
|
done
|
|
done |