quantize_qwen2.5_w8a8.sh

source ~/environments/clearml/bin/activate

size="7B"
model="Qwen2.5-${size}"
prefix="${model//./_}""__llm_compressor__calibration__mse__512__8196"

python /cache/git/research/automation/pipelines/pipeline_llmcompressor_oneshot.py \
  --model-id "Qwen/"$model \
  --project-name "LLM quantization - W8A8/llmcompressor/Qwen2.5" \
  --task-prefix $prefix \
  --recipe "/network/alexandre/quantization/recipe_w8a8_mse_damp01.yaml" \
  --num-samples 512 \
  --max-seq-len 8196 \
  --tags "Qwen2.5" "W8A8" "calibration" $size "MSE"

size="72B"
model="Qwen2.5-${size}"
prefix="${model//./_}""__llm_compressor__calibration__mse__512__8196"

python /cache/git/research/automation/pipelines/pipeline_llmcompressor_oneshot.py \
  --model-id "Qwen/"$model \
  --project-name "LLM quantization - W8A8/llmcompressor/Qwen2.5" \
  --oneshot-queue oneshot-a100x4 \
  --evaluation-queue oneshot-a100x4 \
  --task-prefix $prefix \
  --recipe "/network/alexandre/quantization/recipe_w8a8_mse_damp01.yaml" \
  --num-samples 512 \
  --max-seq-len 8196 \
  --tags "Qwen2.5" "W8A8" "calibration" $size "MSE"

model="Qwen2.5-${size}-Instruct"
prefix="${model//./_}""__llm_compressor__calibration__mse__512__8196"

python /cache/git/research/automation/pipelines/pipeline_llmcompressor_oneshot.py \
  --model-id "Qwen/"$model \
  --project-name "LLM quantization - W8A8/llmcompressor/Qwen2.5" \
  --oneshot-queue oneshot-a100x4 \
  --evaluation-queue oneshot-a100x4 \
  --task-prefix $prefix \
  --recipe "/network/alexandre/quantization/recipe_w8a8_mse_damp01.yaml" \
  --num-samples 512 \
  --max-seq-len 8196 \
  --tags "Qwen2.5" "W8A8" "calibration" $size "MSE" "Instruct"


<<END
for size in 0.5B 1.5B 3B 7B
do

model="Qwen2.5-${size}"
prefix="${model//./_}""__llm_compressor__calibration__mse__512__8196"

python /cache/git/research/automation/pipelines/pipeline_llmcompressor_oneshot.py \
  --model-id "Qwen/"$model \
  --project-name "LLM quantization - W8A8/llmcompressor/Qwen2.5" \
  --task-prefix $prefix \
  --recipe "/network/alexandre/quantization/recipe_w8a8_mse.yaml" \
  --num-samples 512 \
  --max-seq-len 8196 \
  --tags "Qwen2.5" "W8A8" "calibration" $size "MSE"

model="Qwen2.5-${size}-Instruct"
prefix="${model//./_}""__llm_compressor__calibration__mse__512__8196"

python /cache/git/research/automation/pipelines/pipeline_llmcompressor_oneshot.py \
  --model-id "Qwen/"$model \
  --project-name "LLM quantization - W8A8/llmcompressor/Qwen2.5" \
  --task-prefix $prefix \
  --recipe "/network/alexandre/quantization/recipe_w8a8_mse.yaml" \
  --num-samples 512 \
  --max-seq-len 8196 \
  --tags "Qwen2.5" "W8A8" "calibration" $size "MSE" "Instruct"

done

for size in 32B 72B
do

model="Qwen2.5-${size}"
prefix="${model//./_}""__llm_compressor__calibration__mse__512__8196"

python /cache/git/research/automation/pipelines/pipeline_llmcompressor_oneshot.py \
  --model-id "Qwen/"$model \
  --project-name "LLM quantization - W8A8/llmcompressor/Qwen2.5" \
  --task-prefix $prefix \
  --oneshot-queue oneshot-a100x2 \
  --evaluation-queue oneshot-a100x2 \
  --recipe "/network/alexandre/quantization/recipe_w8a8_mse.yaml" \
  --num-samples 512 \
  --max-seq-len 8196 \
  --tags "Qwen2.5" "W8A8" "calibration" $size "MSE"

model="Qwen2.5-${size}-Instruct"
prefix="${model//./_}""__llm_compressor__calibration__mse__512__8196"

python /cache/git/research/automation/pipelines/pipeline_llmcompressor_oneshot.py \
  --model-id "Qwen/"$model \
  --project-name "LLM quantization - W8A8/llmcompressor/Qwen2.5" \
  --task-prefix $prefix \
  --oneshot-queue oneshot-a100x2 \
  --evaluation-queue oneshot-a100x2 \
  --recipe "/network/alexandre/quantization/recipe_w8a8_mse.yaml" \
  --num-samples 512 \
  --max-seq-len 8196 \
  --tags "Qwen2.5" "W8A8" "calibration" $size "MSE" "Instruct"

done
END