Add files using upload-large-folder tool

lm-quant-toolkit/scripts/experiment-llama-sensi-7b-cfg-wise.sh

@@ -0,0 +1,119 @@
+#!/bin/bash
+
+# BUDGETS="2.13 2.25 2.51 3.13 3.25 3.51 4.13 4.25 4.51"
+BUDGETS="4.13 4.25 4.51"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+ATTEMPT="sensi-cfg-wise-7b"
+EXP_BASE_NAME=$ATTEMPT
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+
+weight_algo=sensi-milp
+
+log_file="logs/bench-$(date +%Y%m%d%H%M%S).log"
+
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+MODELS="0"
+EXP_NAME="${ATTEMPT}_7B"
+echo "=========Run perplexity evaluation on Llama-2-7b========="
+python ../src/cli.py llm \
+  --task eval_ppl \
+  --model $MODELS \
+  --algo mxq \
+  --weight-algo $weight_algo \
+  --config ${BUDGETS} \
+  --experiment-name "${EXP_NAME}_ppl" \
+  --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+  --result-dir=$RESULT_DIR \
+  2>&1 \
+  | tee -a $log_file
+EXIT_CODE=$?
+if [ $EXIT_CODE -ne 0 ]; then
+  echo "Perplexity evaluation failed!"
+  exit $EXIT_CODE
+fi
+echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_ppl \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+python ../src/cli.py dump \
+  --type quant_config \
+  --model $MODELS \
+  --budget ${BUDGETS} \
+  --attempt $ATTEMPT \
+  --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+  --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+  2>&1 \
+  | tee -a $log_file
+
+echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+algo=mxq
+model_ids=$MODELS
+for m in $model_ids; do
+  for cfg in ${BUDGETS}; do
+      python ../src/cli.py llm \
+          --model $m \
+          --algo ${algo} \
+          --config ${cfg} \
+          --task eval_model_storage \
+          --experiment-name "${EXP_NAME}_stor" \
+          --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+          --result-dir=$RESULT_DIR \
+          2>&1 \
+          | tee -a $log_file
+  done
+done
+echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_stor \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+# echo "=========Delete quantized models of batch ${batch_name}========="
+# find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs/allot ]; then
+    mkdir -p pdfs/allot
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+$OLD_DIR/../data-vis/plot-mxq-paired.R data/combined.csv
+$OLD_DIR/../data-vis/plot-mem-consumption.R data/combined.csv
+$OLD_DIR/../data-vis/plot-quant-speed.R data/combined.csv
+$OLD_DIR/../data-vis/gen-table-mxq-llm.R data/combined.csv
+pdflatex table.tex
+
+# plot configuration allocations for 3 * 12 MXQ combinations
+MODELS="Llama-2-7b-hf"
+BGS="4.13 4.25 4.51"
+for model in $MODELS; do
+    for budget in $BGS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+          -m $model \
+          -b $budget \
+          --attempt1 $ATTEMPT \
+          --attempt2 mxq1 \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/scripts/experiment-llama-sensi-7b-fnorm-avg.sh

@@ -0,0 +1,119 @@
+#!/bin/bash
+
+# BUDGETS="2.13 2.25 2.51 3.13 3.25 3.51 4.13 4.25 4.51"
+BUDGETS="4.13 4.25 4.51"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+ATTEMPT="fnorm-sensi-avg-7b"
+EXP_BASE_NAME=$ATTEMPT
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+
+weight_algo=sensi-milp
+
+log_file="logs/bench-$(date +%Y%m%d%H%M%S).log"
+
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+MODELS="0"
+EXP_NAME="${ATTEMPT}_7B"
+echo "=========Run perplexity evaluation on Llama-2-7b========="
+python ../src/cli.py llm \
+  --task eval_ppl \
+  --model $MODELS \
+  --algo mxq \
+  --weight-algo $weight_algo \
+  --config ${BUDGETS} \
+  --experiment-name "${EXP_NAME}_ppl" \
+  --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+  --result-dir=$RESULT_DIR \
+  2>&1 \
+  | tee -a $log_file
+EXIT_CODE=$?
+if [ $EXIT_CODE -ne 0 ]; then
+  echo "Perplexity evaluation failed!"
+  exit $EXIT_CODE
+fi
+echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_ppl \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+python ../src/cli.py dump \
+  --type quant_config \
+  --model $MODELS \
+  --budget ${BUDGETS} \
+  --attempt $ATTEMPT \
+  --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+  --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+  2>&1 \
+  | tee -a $log_file
+
+echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+algo=mxq
+model_ids=$MODELS
+for m in $model_ids; do
+  for cfg in ${BUDGETS}; do
+      python ../src/cli.py llm \
+          --model $m \
+          --algo ${algo} \
+          --config ${cfg} \
+          --task eval_model_storage \
+          --experiment-name "${EXP_NAME}_stor" \
+          --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+          --result-dir=$RESULT_DIR \
+          2>&1 \
+          | tee -a $log_file
+  done
+done
+echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_stor \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+# echo "=========Delete quantized models of batch ${batch_name}========="
+# find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs/allot ]; then
+    mkdir -p pdfs/allot
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+$OLD_DIR/../data-vis/plot-mxq-paired.R data/combined.csv
+$OLD_DIR/../data-vis/plot-mem-consumption.R data/combined.csv
+$OLD_DIR/../data-vis/plot-quant-speed.R data/combined.csv
+$OLD_DIR/../data-vis/gen-table-mxq-llm.R data/combined.csv
+pdflatex table.tex
+
+# plot configuration allocations for 3 * 12 MXQ combinations
+MODELS="Llama-2-7b-hf"
+BGS="4.13 4.25 4.51"
+for model in $MODELS; do
+    for budget in $BGS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+          -m $model \
+          -b $budget \
+          --attempt1 $ATTEMPT \
+          --attempt2 mxq1 \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/scripts/experiment-llama-sensi-7b-protect-mlp.sh

@@ -0,0 +1,119 @@
+#!/bin/bash
+
+# BUDGETS="2.13 2.25 2.51 3.13 3.25 3.51 4.13 4.25 4.51"
+BUDGETS="4.13 4.25 4.51"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+ATTEMPT="sensi-protect-mlp-7b"
+EXP_BASE_NAME=$ATTEMPT
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+
+weight_algo=sensi-milp
+
+log_file="logs/bench-$(date +%Y%m%d%H%M%S).log"
+
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+MODELS="0"
+EXP_NAME="${ATTEMPT}"
+echo "=========Run perplexity evaluation on Llama-2-7b========="
+python ../src/cli.py llm \
+  --task eval_ppl \
+  --model $MODELS \
+  --algo mxq \
+  --weight-algo $weight_algo \
+  --config ${BUDGETS} \
+  --experiment-name "${EXP_NAME}_ppl" \
+  --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+  --result-dir=$RESULT_DIR \
+  2>&1 \
+  | tee -a $log_file
+EXIT_CODE=$?
+if [ $EXIT_CODE -ne 0 ]; then
+  echo "Perplexity evaluation failed!"
+  exit $EXIT_CODE
+fi
+echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_ppl \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+python ../src/cli.py dump \
+  --type quant_config \
+  --model $MODELS \
+  --budget ${BUDGETS} \
+  --attempt $ATTEMPT \
+  --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+  --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+  2>&1 \
+  | tee -a $log_file
+
+echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+algo=mxq
+model_ids=$MODELS
+for m in $model_ids; do
+  for cfg in ${BUDGETS}; do
+      python ../src/cli.py llm \
+          --model $m \
+          --algo ${algo} \
+          --config ${cfg} \
+          --task eval_model_storage \
+          --experiment-name "${EXP_NAME}_stor" \
+          --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+          --result-dir=$RESULT_DIR \
+          2>&1 \
+          | tee -a $log_file
+  done
+done
+echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_stor \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+# echo "=========Delete quantized models of batch ${batch_name}========="
+# find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs/allot ]; then
+    mkdir -p pdfs/allot
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+$OLD_DIR/../data-vis/plot-mxq-paired.R data/combined.csv
+$OLD_DIR/../data-vis/plot-mem-consumption.R data/combined.csv
+$OLD_DIR/../data-vis/plot-quant-speed.R data/combined.csv
+$OLD_DIR/../data-vis/gen-table-mxq-llm.R data/combined.csv
+pdflatex table.tex
+
+# plot configuration allocations for 3 * 12 MXQ combinations
+MODELS="Llama-2-7b-hf"
+BGS="4.13 4.25 4.51"
+for model in $MODELS; do
+    for budget in $BGS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+          -m $model \
+          -b $budget \
+          --attempt1 $ATTEMPT \
+          --attempt2 mxq1 \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/scripts/experiment-llama-sensi-boost.sh

@@ -0,0 +1,133 @@
+#!/bin/bash
+
+BUDGETS="3.13 3.25 3.51 4.13 4.25 4.51"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+weight_algo=sensi-boost
+MODELS="0 1 2"
+MODEL_NAMES="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+
+
+# BOOST_STOPS="2 3"
+BOOST_STOPS="3"
+BOOST_TOP_MS="0"
+
+for BOOST_STOP in $BOOST_STOPS; do
+    for BOOST_TOP_M in $BOOST_TOP_MS; do
+        if [[ $BOOST_STOP -eq 2 && $BOOST_TOP_M -eq 1 ]]; then
+            continue
+        fi
+        ATTEMPT="sensi-boost-${BOOST_STOP}-${BOOST_TOP_M}"
+        EXP_BASE_NAME=$ATTEMPT
+        mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+        log_file="logs/bench-${ATTEMPT}-$(date +%Y%m%d%H%M%S).log"
+
+        mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+        mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+        mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+        # MODELS="0"
+        EXP_NAME="${ATTEMPT}"
+        echo "=========Run perplexity evaluation========="
+        python ../src/cli.py llm \
+          --task eval_ppl \
+          --model $MODELS \
+          --algo mxq \
+          --weight-algo $weight_algo \
+          --boost-stop $BOOST_STOP \
+          --top-m-layer $BOOST_TOP_M \
+          --config ${BUDGETS} \
+          --experiment-name "${EXP_NAME}_ppl" \
+          --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+          --result-dir=$RESULT_DIR \
+          2>&1 \
+          | tee -a $log_file
+        EXIT_CODE=$?
+        if [ $EXIT_CODE -ne 0 ]; then
+          echo "Perplexity evaluation failed!"
+          exit $EXIT_CODE
+        fi
+        echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+        find $RESULT_DIR/${EXP_NAME}_ppl \
+          -name "result-*.csv" \
+          -printf '%T@ %p\n' \
+          | sort -n \
+          | tail -1 \
+          | cut -d' '  -f2 \
+          | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+        echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+        mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+        python ../src/cli.py dump \
+          --type quant_config \
+          --model $MODELS \
+          --budget ${BUDGETS} \
+          --attempt $ATTEMPT \
+          --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+          --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+          2>&1 \
+          | tee -a $log_file
+
+        echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+        algo=mxq
+        model_ids=$MODELS
+        for m in $model_ids; do
+          for cfg in ${BUDGETS}; do
+              python ../src/cli.py llm \
+                  --model $m \
+                  --algo ${algo} \
+                  --config ${cfg} \
+                  --task eval_model_storage \
+                  --experiment-name "${EXP_NAME}_stor" \
+                  --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+                  --result-dir=$RESULT_DIR \
+                  2>&1 \
+                  | tee -a $log_file
+          done
+        done
+        echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+        find $RESULT_DIR/${EXP_NAME}_stor \
+          -name "result-*.csv" \
+          -printf '%T@ %p\n' \
+          | sort -n \
+          | tail -1 \
+          | cut -d' '  -f2 \
+          | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+        # echo "=========Delete quantized models of batch ${batch_name}========="
+        # find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+
+        OLD_DIR=$(pwd)
+        cd $RESULT_DIR/$EXP_BASE_NAME
+        if [ ! -d pdfs/allot ]; then
+            mkdir -p pdfs/allot
+        fi
+        $OLD_DIR/../data-vis/combine.R \
+            --baseline_data_dir $OLD_DIR/../data-vis/data \
+            --mxq_data_dir data
+        $OLD_DIR/../data-vis/plot-ppl-mem.R -d data/combined.csv
+        $OLD_DIR/../data-vis/plot-mem-consumption.R data/combined.csv
+        $OLD_DIR/../data-vis/plot-quant-speed.R data/combined.csv
+        $OLD_DIR/../data-vis/gen-table-mxq-llm.R --csv_file data/combined.csv --attempt $ATTEMPT
+        pdflatex table.tex
+
+        # plot configuration allocations for 3 * 12 MXQ combinations
+        for model in $MODEL_NAMES; do
+            for budget in $BUDGETS; do
+                $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+                  -m $model \
+                  -b $budget \
+                  --attempt1 $ATTEMPT \
+                  --attempt2 mxq1 \
+                  --fnorm_data_dir $OLD_DIR/../src/data \
+                  --quant_cfg_allot_file data/quant-cfg-allocation.csv
+            done
+        done
+        cd $OLD_DIR
+    done
+done
+

lm-quant-toolkit/scripts/experiment-llama-sensi-milp-7b.sh

@@ -0,0 +1,119 @@
+#!/bin/bash
+
+# BUDGETS="2.13 2.25 2.51 3.13 3.25 3.51 4.13 4.25 4.51"
+BUDGETS="4.13 4.25 4.51"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+ATTEMPT="llama-sensi-milp-2x"
+EXP_BASE_NAME=$ATTEMPT
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+
+weight_algo=sensi-milp
+
+log_file="logs/bench-$(date +%Y%m%d%H%M%S).log"
+
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+MODELS="0 2"
+EXP_NAME="${ATTEMPT}"
+echo "=========Run perplexity evaluation on Llama-2-7b========="
+python ../src/cli.py llm \
+  --task eval_ppl \
+  --model $MODELS \
+  --algo mxq \
+  --weight-algo $weight_algo \
+  --config ${BUDGETS} \
+  --experiment-name "${EXP_NAME}_ppl" \
+  --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+  --result-dir=$RESULT_DIR \
+  2>&1 \
+  | tee -a $log_file
+EXIT_CODE=$?
+if [ $EXIT_CODE -ne 0 ]; then
+  echo "Perplexity evaluation failed!"
+  exit $EXIT_CODE
+fi
+echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_ppl \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+python ../src/cli.py dump \
+  --type quant_config \
+  --model $MODELS \
+  --budget ${BUDGETS} \
+  --attempt $ATTEMPT \
+  --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+  --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+  2>&1 \
+  | tee -a $log_file
+
+echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+algo=mxq
+model_ids=$MODELS
+for m in $model_ids; do
+  for cfg in ${BUDGETS}; do
+      python ../src/cli.py llm \
+          --model $m \
+          --algo ${algo} \
+          --config ${cfg} \
+          --task eval_model_storage \
+          --experiment-name "${EXP_NAME}_stor" \
+          --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+          --result-dir=$RESULT_DIR \
+          2>&1 \
+          | tee -a $log_file
+  done
+done
+echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_stor \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+# echo "=========Delete quantized models of batch ${batch_name}========="
+# find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs/allot ]; then
+    mkdir -p pdfs/allot
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+$OLD_DIR/../data-vis/plot-mxq-paired.R data/combined.csv
+$OLD_DIR/../data-vis/plot-mem-consumption.R data/combined.csv
+$OLD_DIR/../data-vis/plot-quant-speed.R data/combined.csv
+$OLD_DIR/../data-vis/gen-table-mxq-llm.R data/combined.csv
+pdflatex table.tex
+
+# plot configuration allocations for 3 * 12 MXQ combinations
+MODELS="Llama-2-7b-hf Meta-Llama-3-8B"
+BGS="4.13 4.25 4.51"
+for model in $MODELS; do
+    for budget in $BGS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+          -m $model \
+          -b $budget \
+          --attempt1 $ATTEMPT \
+          --attempt2 mxq1 \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/scripts/experiment-llama-sensi-milp-ablation.sh

@@ -0,0 +1,109 @@
+#!/bin/bash
+
+BUDGETS="6.89 5.72 5.02 4.51 4.25 4.21 4.17 4.13 4.11 4.07 3.95 3.87 3.83 3.65 3.51 3.25 3.19 3.15 3.13 3.11 3.07"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+weight_algo=sensi-milp
+MODELS="0 1 2"
+MODEL_NAMES="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+
+
+# for SensiMiLP ablation test, all topm values are equivalent
+
+ATTEMPT="sensi-milp-abl"
+EXP_BASE_NAME=$ATTEMPT
+EXP_NAME="${ATTEMPT}"
+
+log_file="logs/bench-${ATTEMPT}-$(date +%Y%m%d%H%M%S).log"
+
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+echo "=========Run perplexity evaluation========="
+python ../src/cli.py llm \
+  --task eval_ppl \
+  --model $MODELS \
+  --algo mxq \
+  --weight-algo $weight_algo \
+  --ablation \
+  --config ${BUDGETS} \
+  --experiment-name "${EXP_NAME}_ppl" \
+  --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+  --result-dir=$RESULT_DIR \
+  2>&1 \
+  | tee -a $log_file
+EXIT_CODE=$?
+if [ $EXIT_CODE -ne 0 ]; then
+  echo "Perplexity evaluation failed!"
+  exit $EXIT_CODE
+fi
+echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_ppl \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+python ../src/cli.py dump \
+  --type quant_config \
+  --model $MODELS \
+  --budget ${BUDGETS} \
+  --attempt $ATTEMPT \
+  --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+  --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+  2>&1 \
+  | tee -a $log_file
+
+echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+algo=mxq
+model_ids=$MODELS
+for m in $model_ids; do
+  for cfg in ${BUDGETS}; do
+      python ../src/cli.py llm \
+          --model $m \
+          --algo ${algo} \
+          --config ${cfg} \
+          --task eval_model_storage \
+          --experiment-name "${EXP_NAME}_stor" \
+          --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+          --result-dir=$RESULT_DIR \
+          2>&1 \
+          | tee -a $log_file
+  done
+done
+echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_stor \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+# echo "=========Delete quantized models of batch ${batch_name}========="
+# find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs ]; then
+    mkdir pdfs
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+$OLD_DIR/../data-vis/plot-ppl-mem.R -d data/combined.csv
+$OLD_DIR/../data-vis/gen-table-mxq-llm.R --csv_file data/combined.csv --attempt $ATTEMPT
+cd pdfs
+pdflatex table.tex
+cd ..
+
+cd $OLD_DIR

lm-quant-toolkit/scripts/experiment-llama-sensi-milp.sh

@@ -0,0 +1,129 @@
+#!/bin/bash
+
+BUDGETS="6.89 5.72 5.02 4.51 4.25 4.21 4.17 4.13 4.11 4.07 3.95 3.87 3.83 3.65 3.51 3.25 3.19 3.15 3.13 3.11 3.07"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+weight_algo=sensi-milp
+MODELS="0 1 2"
+MODEL_NAMES="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+
+
+BOOST_TOP_MS="3"
+
+for BOOST_TOP_M in $BOOST_TOP_MS; do
+    ATTEMPT="sensi-milp-${BOOST_TOP_M}"
+    EXP_BASE_NAME=$ATTEMPT
+    mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+    log_file="logs/bench-${ATTEMPT}-$(date +%Y%m%d%H%M%S).log"
+
+    mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+    mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+    mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+    # MODELS="0"
+    EXP_NAME="${ATTEMPT}"
+    echo "=========Run perplexity evaluation========="
+    python ../src/cli.py llm \
+      --task eval_ppl \
+      --model $MODELS \
+      --algo mxq \
+      --weight-algo $weight_algo \
+      --top-m-layer $BOOST_TOP_M \
+      --config ${BUDGETS} \
+      --experiment-name "${EXP_NAME}_ppl" \
+      --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+      --result-dir=$RESULT_DIR \
+      2>&1 \
+      | tee -a $log_file
+    EXIT_CODE=$?
+    if [ $EXIT_CODE -ne 0 ]; then
+      echo "Perplexity evaluation failed!"
+      exit $EXIT_CODE
+    fi
+    echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+    find $RESULT_DIR/${EXP_NAME}_ppl \
+      -name "result-*.csv" \
+      -printf '%T@ %p\n' \
+      | sort -n \
+      | tail -1 \
+      | cut -d' '  -f2 \
+      | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+    echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+    mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+    python ../src/cli.py dump \
+      --type quant_config \
+      --model $MODELS \
+      --budget ${BUDGETS} \
+      --attempt $ATTEMPT \
+      --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+      --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+      2>&1 \
+      | tee -a $log_file
+
+    echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+    algo=mxq
+    model_ids=$MODELS
+    for m in $model_ids; do
+      for cfg in ${BUDGETS}; do
+          python ../src/cli.py llm \
+              --model $m \
+              --algo ${algo} \
+              --config ${cfg} \
+              --task eval_model_storage \
+              --experiment-name "${EXP_NAME}_stor" \
+              --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+              --result-dir=$RESULT_DIR \
+              2>&1 \
+              | tee -a $log_file
+      done
+    done
+    echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+    find $RESULT_DIR/${EXP_NAME}_stor \
+      -name "result-*.csv" \
+      -printf '%T@ %p\n' \
+      | sort -n \
+      | tail -1 \
+      | cut -d' '  -f2 \
+      | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+    # echo "=========Delete quantized models of batch ${batch_name}========="
+    # find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+    OLD_DIR=$(pwd)
+    cd $RESULT_DIR/$EXP_BASE_NAME
+    if [ ! -d pdfs ]; then
+        mkdir pdfs
+    fi
+    $OLD_DIR/../data-vis/combine.R \
+        --baseline_data_dir $OLD_DIR/../data-vis/data \
+        --mxq_data_dir data
+    $OLD_DIR/../data-vis/plot-ppl-mem.R -d data/combined.csv
+    $OLD_DIR/../data-vis/plot-mem-consumption.R -d data/combined.csv
+    $OLD_DIR/../data-vis/plot-quant-speed.R -d data/combined.csv
+    $OLD_DIR/../data-vis/gen-table-mxq-llm.R --csv_file data/combined.csv --attempt $ATTEMPT
+    cd pdfs
+    pdflatex table.tex
+    cd ..
+
+    # plot configuration allocations for 3 * 12 MXQ combinations
+    MODELS="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+    BUDGETS=${MXQ_BATCH[@]:1}
+    for model in $MODELS; do
+        for budget in $BUDGETS; do
+            $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+              -m $model \
+              -b $budget \
+              --fnorm_data_dir $OLD_DIR/../src/data \
+              --attempt1 mxq1 \
+              --attempt2 $ATTEMPT \
+              --quant_cfg_allot_file data/quant-cfg-allocation.csv
+        done
+    done
+    cd $OLD_DIR
+
+done
+

lm-quant-toolkit/scripts/experiment-llama-sensi.sh

@@ -0,0 +1,190 @@
+#!/bin/bash
+
+# BUDGETS="2.13 2.25 2.51 3.13 3.25 3.51 4.13 4.25 4.51"
+BUDGETS="4.13 4.25 4.51"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+ATTEMPT="llama-sensi"
+EXP_BASE_NAME=$ATTEMPT
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+
+weight_algo=sensi-directive
+
+log_file="logs/bench-$(date +%Y%m%d%H%M%S).log"
+
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+MODELS="1"
+boost_layers="3 39"
+EXP_NAME="${ATTEMPT}_13b"
+echo "=========Run perplexity evaluation on Llama-2-13b========="
+python ../src/cli.py llm \
+  --task eval_ppl \
+  --model $MODELS \
+  --algo mxq \
+  --weight-algo $weight_algo \
+  --boost-layer $boost_layers \
+  --boost-stop 2 \
+  --config ${BUDGETS} \
+  --experiment-name "${EXP_NAME}_ppl" \
+  --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+  --result-dir=$RESULT_DIR \
+  2>&1 \
+  | tee -a $log_file
+EXIT_CODE=$?
+if [ $EXIT_CODE -ne 0 ]; then
+  echo "Perplexity evaluation failed!"
+  exit $EXIT_CODE
+fi
+echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_ppl \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+python ../src/cli.py dump \
+  --type quant_config \
+  --model $MODELS \
+  --budget ${BUDGETS} \
+  --attempt $ATTEMPT \
+  --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+  --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+  2>&1 \
+  | tee -a $log_file
+
+echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+algo=mxq
+model_ids=$MODELS
+for m in $model_ids; do
+  for cfg in ${BUDGETS}; do
+      python ../src/cli.py llm \
+          --model $m \
+          --algo ${algo} \
+          --config ${cfg} \
+          --task eval_model_storage \
+          --experiment-name "${EXP_NAME}_stor" \
+          --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+          --result-dir=$RESULT_DIR \
+          2>&1 \
+          | tee -a $log_file
+  done
+done
+echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_stor \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+MODELS="0 2"
+boost_layers="1 31"
+EXP_NAME="${ATTEMPT}_78"
+echo "=========Run perplexity evaluation on Llama-2-7b & Llama-3-8B========="
+python ../src/cli.py llm \
+  --task eval_ppl \
+  --model $MODELS \
+  --algo mxq \
+  --weight-algo $weight_algo \
+  --boost-layer $boost_layers \
+  --boost-stop 1 \
+  --config ${BUDGETS} \
+  --experiment-name "${EXP_NAME}_ppl" \
+  --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+  --result-dir=$RESULT_DIR \
+  2>&1 \
+  | tee -a $log_file
+EXIT_CODE=$?
+if [ $EXIT_CODE -ne 0 ]; then
+  echo "Perplexity evaluation failed!"
+  exit $EXIT_CODE
+fi
+echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_ppl \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+python ../src/cli.py dump \
+  --type quant_config \
+  --model $MODELS \
+  --budget ${BUDGETS} \
+  --attempt $ATTEMPT \
+  --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+  --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+  2>&1 \
+  | tee -a $log_file
+
+echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+algo=mxq
+model_ids=$MODELS
+for m in $model_ids; do
+  for cfg in ${BUDGETS}; do
+      python ../src/cli.py llm \
+          --model $m \
+          --algo ${algo} \
+          --config ${cfg} \
+          --task eval_model_storage \
+          --experiment-name "${EXP_NAME}_stor" \
+          --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+          --result-dir=$RESULT_DIR \
+          2>&1 \
+          | tee -a $log_file
+  done
+done
+echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_stor \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+# echo "=========Delete quantized models of batch ${batch_name}========="
+# find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs/allot ]; then
+    mkdir -p pdfs/allot
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+$OLD_DIR/../data-vis/plot-mxq-paired.R data/combined.csv
+$OLD_DIR/../data-vis/plot-mem-consumption.R data/combined.csv
+$OLD_DIR/../data-vis/plot-quant-speed.R data/combined.csv
+$OLD_DIR/../data-vis/gen-table-mxq-llm.R data/combined.csv
+pdflatex table.tex
+
+# plot configuration allocations for 3 * 12 MXQ combinations
+MODELS="Llama-2-7b-hf Llama-2-13b-hf"
+BGS="4.13 4.25 4.51"
+for model in $MODELS; do
+    for budget in $BGS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+          -m $model \
+          -b $budget \
+          --attempt1 $ATTEMPT \
+          --attempt2 mxq1 \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/scripts/experiment-llama2-13b-balanced.sh

@@ -0,0 +1,122 @@
+#!/bin/bash
+
+# BUDGETS="2.13 2.25 2.51 3.13 3.25 3.51 4.13 4.25 4.51"
+BUDGETS="4.13 4.25 4.51"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+MODELS="1"
+ATTEMPT="Llama2-13B-balanced"
+EXP_BASE_NAME=$ATTEMPT
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+
+weight_algo=sensi-directive
+boost_layers="3 39"
+decline_layers="0"
+EXP_NAME="${ATTEMPT}"
+log_file="logs/bench-$(date +%Y%m%d%H%M%S).log"
+
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+echo "=========Run perplexity evaluation on batch ${EXP_NAME}========="
+python ../src/cli.py llm \
+  --task eval_ppl \
+  --model $MODELS \
+  --algo mxq \
+  --weight-algo $weight_algo \
+  --boost-layer $boost_layers \
+  --decline-layer $decline_layers \
+  --config ${BUDGETS} \
+  --experiment-name "${EXP_NAME}_ppl" \
+  --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+  --result-dir=$RESULT_DIR \
+  2>&1 \
+  | tee -a $log_file
+EXIT_CODE=$?
+if [ $EXIT_CODE -ne 0 ]; then
+  echo "Perplexity evaluation failed!"
+  exit $EXIT_CODE
+fi
+echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_ppl \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+python ../src/cli.py dump \
+  --type quant_config \
+  --model $MODELS \
+  --budget ${BUDGETS} \
+  --attempt $ATTEMPT \
+  --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+  --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+  2>&1 \
+  | tee -a $log_file
+
+echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+algo=mxq
+model_ids=$MODELS
+for m in $model_ids; do
+  for cfg in ${BUDGETS}; do
+      python ../src/cli.py llm \
+          --model $m \
+          --algo ${algo} \
+          --config ${cfg} \
+          --task eval_model_storage \
+          --experiment-name "${EXP_NAME}_stor" \
+          --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+          --result-dir=$RESULT_DIR \
+          2>&1 \
+          | tee -a $log_file
+  done
+done
+echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_stor \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+# echo "=========Delete quantized models of batch ${batch_name}========="
+# find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs/allot ]; then
+    mkdir -p pdfs/allot
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+$OLD_DIR/../data-vis/plot-mxq-paired.R data/combined.csv
+$OLD_DIR/../data-vis/plot-mem-consumption.R data/combined.csv
+$OLD_DIR/../data-vis/plot-quant-speed.R data/combined.csv
+$OLD_DIR/../data-vis/gen-table-mxq-llm.R data/combined.csv
+pdflatex table.tex
+
+# plot configuration allocations for 3 * 12 MXQ combinations
+MODELS="Llama-2-13b-hf"
+BGS="4.13 4.25 4.51"
+for model in $MODELS; do
+    for budget in $BGS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+          -m $model \
+          -b $budget \
+          --attempt1 $ATTEMPT \
+          --attempt2 mxq1 \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/scripts/experiment-llama2-13b-boost.sh

@@ -0,0 +1,121 @@
+#!/bin/bash
+
+# BUDGETS="2.13 2.25 2.51 3.13 3.25 3.51 4.13 4.25 4.51"
+BUDGETS="4.13 4.25 4.51"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+MODELS="1"
+ATTEMPT="Llama2-13B-boost"
+EXP_BASE_NAME=$ATTEMPT
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+
+weight_algo=sensi-directive
+boost_layers="3 39"
+EXP_NAME="${ATTEMPT}"
+log_file="logs/bench-$(date +%Y%m%d%H%M%S).log"
+
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+echo "=========Run perplexity evaluation on batch ${EXP_NAME}========="
+python ../src/cli.py llm \
+  --task eval_ppl \
+  --model $MODELS \
+  --algo mxq \
+  --weight-algo $weight_algo \
+  --boost-layer $boost_layers \
+  --boost-stop 2 \
+  --config ${BUDGETS} \
+  --experiment-name "${EXP_NAME}_ppl" \
+  --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+  --result-dir=$RESULT_DIR \
+  2>&1 \
+  | tee -a $log_file
+EXIT_CODE=$?
+if [ $EXIT_CODE -ne 0 ]; then
+  echo "Perplexity evaluation failed!"
+  exit $EXIT_CODE
+fi
+echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_ppl \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+python ../src/cli.py dump \
+  --type quant_config \
+  --model $MODELS \
+  --budget ${BUDGETS} \
+  --attempt $ATTEMPT \
+  --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+  --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+  2>&1 \
+  | tee -a $log_file
+
+echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+algo=mxq
+model_ids=$MODELS
+for m in $model_ids; do
+  for cfg in ${BUDGETS}; do
+      python ../src/cli.py llm \
+          --model $m \
+          --algo ${algo} \
+          --config ${cfg} \
+          --task eval_model_storage \
+          --experiment-name "${EXP_NAME}_stor" \
+          --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+          --result-dir=$RESULT_DIR \
+          2>&1 \
+          | tee -a $log_file
+  done
+done
+echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_stor \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+# echo "=========Delete quantized models of batch ${batch_name}========="
+# find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs/allot ]; then
+    mkdir -p pdfs/allot
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+$OLD_DIR/../data-vis/plot-mxq-paired.R data/combined.csv
+$OLD_DIR/../data-vis/plot-mem-consumption.R data/combined.csv
+$OLD_DIR/../data-vis/plot-quant-speed.R data/combined.csv
+$OLD_DIR/../data-vis/gen-table-mxq-llm.R data/combined.csv
+pdflatex table.tex
+
+# plot configuration allocations for 3 * 12 MXQ combinations
+MODELS="Llama-2-13b-hf"
+BGS="4.13 4.25 4.51"
+for model in $MODELS; do
+    for budget in $BGS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+          -m $model \
+          -b $budget \
+          --attempt1 $ATTEMPT \
+          --attempt2 mxq1 \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/scripts/experiment-llama3-8b-aggressive-reduction.sh

@@ -0,0 +1,122 @@
+#!/bin/bash
+
+# BUDGETS="2.13 2.25 2.51 3.13 3.25 3.51 4.13 4.25 4.51"
+BUDGETS="4.13 4.25 4.51"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+MODELS="2"
+ATTEMPT="Llama3-8B-aggressive"
+EXP_BASE_NAME=$ATTEMPT
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+
+weight_algo=sensi-directive
+boost_layers="31 30"
+decline_layers="0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29"
+EXP_NAME="${ATTEMPT}"
+log_file="logs/bench-$(date +%Y%m%d%H%M%S).log"
+
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+echo "=========Run perplexity evaluation on batch ${EXP_NAME}========="
+python ../src/cli.py llm \
+  --task eval_ppl \
+  --model $MODELS \
+  --algo mxq \
+  --weight-algo $weight_algo \
+  --boost-layer $boost_layers \
+  --decline-layer $decline_layers \
+  --config ${BUDGETS} \
+  --experiment-name "${EXP_NAME}_ppl" \
+  --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+  --result-dir=$RESULT_DIR \
+  2>&1 \
+  | tee -a $log_file
+EXIT_CODE=$?
+if [ $EXIT_CODE -ne 0 ]; then
+  echo "Perplexity evaluation failed!"
+  exit $EXIT_CODE
+fi
+echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_ppl \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+python ../src/cli.py dump \
+  --type quant_config \
+  --model $MODELS \
+  --budget ${BUDGETS} \
+  --attempt $ATTEMPT \
+  --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+  --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+  2>&1 \
+  | tee -a $log_file
+
+echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+algo=mxq
+model_ids=$MODELS
+for m in $model_ids; do
+  for cfg in ${BUDGETS}; do
+      python ../src/cli.py llm \
+          --model $m \
+          --algo ${algo} \
+          --config ${cfg} \
+          --task eval_model_storage \
+          --experiment-name "${EXP_NAME}_stor" \
+          --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+          --result-dir=$RESULT_DIR \
+          2>&1 \
+          | tee -a $log_file
+  done
+done
+echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_stor \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+# echo "=========Delete quantized models of batch ${batch_name}========="
+# find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs/allot ]; then
+    mkdir -p pdfs/allot
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+$OLD_DIR/../data-vis/plot-mxq-paired.R data/combined.csv
+$OLD_DIR/../data-vis/plot-mem-consumption.R data/combined.csv
+$OLD_DIR/../data-vis/plot-quant-speed.R data/combined.csv
+$OLD_DIR/../data-vis/gen-table-mxq-llm.R data/combined.csv
+pdflatex table.tex
+
+# plot configuration allocations for 3 * 12 MXQ combinations
+MODELS="Meta-Llama-3-8B"
+BGS="4.13 4.25 4.51"
+for model in $MODELS; do
+    for budget in $BGS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+          -m $model \
+          -b $budget \
+          --attempt1 $ATTEMPT \
+          --attempt2 mxq1 \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/scripts/experiment-sensi-milp-mini-batch.sh

@@ -0,0 +1,128 @@
+#!/bin/bash
+
+
+# 100%, 99%, 98%, 97%, 96%, 95% of [3.51, 4.25]
+# MXQ2=(minibatch 3.51 3.47 3.44 3.40 3.37 3.33 4.25 4.21 4.17 4.12 4.08 4.04)
+MXQ2=(minibatch 6.89 5.72 5.02 4.51 4.25 4.21 4.17 4.13 4.11 4.07 3.95 3.87 3.83 3.65 3.51 3.25 3.19 3.15 3.13 3.11 3.07)
+MXQ_BATCHES=(MXQ2)
+declare -n MXQ_BATCH
+
+ATTEMPT="sensi-milp-mini2"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+EXP_BASE_NAME="${ATTEMPT}"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+# Setup data files directories for reporting
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+
+weight_algo=sensi-milp
+
+for MXQ_BATCH in "${MXQ_BATCHES[@]}"; do
+  batch_name=${MXQ_BATCH[@]:0:1}
+  EXP_NAME="${EXP_BASE_NAME}-${batch_name}"
+  mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+  mkdir -p $RESULT_DIR/${EXP_NAME}_stor
+  log_file="logs/bench-${EXP_NAME}-$(date +%Y%m%d%H%M%S).log"
+
+  echo "=========Run perplexity evaluation on batch ${batch_name}========="
+  python ../src/cli.py llm \
+      --task eval_ppl \
+      --model 0 1 2 \
+      --algo mxq \
+      --weight-algo ${weight_algo} \
+      --config ${MXQ_BATCH[@]:1} \
+      --experiment-name "${EXP_NAME}_ppl" \
+      --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+      --result-dir=$RESULT_DIR \
+      2>&1 \
+      | tee -a $log_file
+  EXIT_CODE=$?
+  if [ $EXIT_CODE -ne 0 ]; then
+      echo "Perplexity evaluation failed!"
+      exit $EXIT_CODE
+  fi
+  echo "=========Collect perplexity evaluation result on batch ${batch_name}========="
+  find $RESULT_DIR/${EXP_NAME}_ppl \
+      -name "result-*.csv" \
+      -printf '%T@ %p\n' \
+      | sort -n \
+      | tail -1 \
+      | cut -d' '  -f2 \
+      | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+  echo "=========Dump quantization configs on batch ${batch_name}========="
+  mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+  python ../src/cli.py dump \
+      --type quant_config \
+      --model 0 1 2 \
+      --budget ${MXQ_BATCH[@]:1} \
+      --attempt $ATTEMPT \
+      --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+      --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+      2>&1 \
+      | tee -a $log_file
+
+  echo "=========Run memory evaluation on batch ${batch_name}========="
+  algo=mxq
+  model_ids="0 1 2"
+  for m in $model_ids; do
+      for cfg in "${MXQ_BATCH[@]:1}"; do
+          python ../src/cli.py llm \
+              --model $m \
+              --algo ${algo} \
+              --config ${cfg} \
+              --task eval_model_storage \
+              --experiment-name "${EXP_NAME}_stor" \
+              --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+              --result-dir=$RESULT_DIR \
+              2>&1 \
+              | tee -a $log_file
+      done
+  done
+  echo "=========Collect memory evaluation result on batch ${batch_name}========="
+  find $RESULT_DIR/${EXP_NAME}_stor \
+      -name "result-*.csv" \
+      -printf '%T@ %p\n' \
+      | sort -n \
+      | tail -1 \
+      | cut -d' '  -f2 \
+      | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+
+  # echo "=========Delete quantized models of batch ${batch_name}========="
+  # find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+
+  OLD_DIR=$(pwd)
+  cd $RESULT_DIR/$EXP_BASE_NAME
+  if [ ! -d pdfs ]; then
+      mkdir pdfs
+  fi
+  $OLD_DIR/../data-vis/combine.R \
+      --baseline_data_dir $OLD_DIR/../data-vis/data \
+      --mxq_data_dir data
+  $OLD_DIR/../data-vis/plot-ppl-mem.R -d data/combined.csv
+  $OLD_DIR/../data-vis/plot-mem-consumption.R -d data/combined.csv
+  $OLD_DIR/../data-vis/plot-quant-speed.R -d data/combined.csv
+  $OLD_DIR/../data-vis/gen-table-mxq-llm.R --csv_file data/combined.csv --attempt $ATTEMPT
+  cd pdfs
+  pdflatex table.tex
+  cd ..
+
+  # plot configuration allocations for 3 * 12 MXQ combinations
+  MODELS="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+  BUDGETS=${MXQ_BATCH[@]:1}
+  for model in $MODELS; do
+      for budget in $BUDGETS; do
+          $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+            -m $model \
+            -b $budget \
+            --fnorm_data_dir $OLD_DIR/../src/data \
+            --attempt1 mxq1 \
+            --attempt2 $ATTEMPT \
+            --quant_cfg_allot_file data/quant-cfg-allocation.csv
+      done
+  done
+  cd $OLD_DIR
+done

lm-quant-toolkit/scripts/experiment-tail-2-4.sh

@@ -0,0 +1,135 @@
+#!/bin/bash
+
+HEADS=(head-prioritized 1.05 1.10 1.15 1.20)
+# TAILS=(tail-prioritized 1.05 1.10 1.15 1.20)
+TAILS=(tail-prioritized 2.00 4.00)
+# MXQ_BATCHES=(HEADS TAILS)
+MXQ_BATCHES=(TAILS)
+declare -n MXQ_BATCH
+
+# BUDGETS="2.13 2.25 2.51 3.13 3.25 3.51 4.13 4.25 4.51"
+BUDGETS="4.13 4.25 4.51"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+# --config ${MXQ_BATCH[@]:1}
+
+EXP_BASE_NAME="tail-big-factor"
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+
+for MXQ_BATCH in "${MXQ_BATCHES[@]}"; do
+  weight_algo=${MXQ_BATCH[@]:0:1}
+  for factor in ${MXQ_BATCH[@]:1}; do
+      ATTEMPT="${weight_algo}_${factor/./_}"
+      EXP_NAME="${ATTEMPT}"
+      log_file="logs/bench-$(date +%Y%m%d%H%M%S).log"
+
+      mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+      mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+      mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+      # echo "=========Run perplexity evaluation on batch ${EXP_NAME}========="
+      # python ../src/cli.py llm \
+      #     --task eval_ppl \
+      #     --model 0 1 2 \
+      #     --algo mxq \
+      #     --weight-algo $weight_algo \
+      #     --factor $factor \
+      #     --config ${BUDGETS} \
+      #     --experiment-name "${EXP_NAME}_ppl" \
+      #     --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+      #     --result-dir=$RESULT_DIR \
+      #     2>&1 \
+      #     | tee -a $log_file
+      # EXIT_CODE=$?
+      # if [ $EXIT_CODE -ne 0 ]; then
+      #     echo "Perplexity evaluation failed!"
+      #     exit $EXIT_CODE
+      # fi
+      # echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+      # find $RESULT_DIR/${EXP_NAME}_ppl \
+      #     -name "result-*.csv" \
+      #     -printf '%T@ %p\n' \
+      #     | sort -n \
+      #     | tail -1 \
+      #     | cut -d' '  -f2 \
+      #     | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+      echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+      mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+      python ../src/cli.py dump \
+          --type quant_config \
+          --model 0 1 2 \
+          --budget ${BUDGETS} \
+          --attempt $ATTEMPT \
+          --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+          --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+          2>&1 \
+          | tee -a $log_file
+
+      # echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+      # algo=mxq
+      # #model_ids="0 1 2"
+      # model_ids="0"
+      # for m in $model_ids; do
+      #     for cfg in ${BUDGETS}; do
+      #         python ../src/cli.py llm \
+      #             --model $m \
+      #             --algo ${algo} \
+      #             --config ${cfg} \
+      #             --task eval_model_storage \
+      #             --experiment-name "${EXP_NAME}_stor" \
+      #             --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+      #             --result-dir=$RESULT_DIR \
+      #             2>&1 \
+      #             | tee -a $log_file
+      #     done
+      # done
+      # echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+      # find $RESULT_DIR/${EXP_NAME}_stor \
+      #     -name "result-*.csv" \
+      #     -printf '%T@ %p\n' \
+      #     | sort -n \
+      #     | tail -1 \
+      #     | cut -d' '  -f2 \
+      #     | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+  done
+
+
+  # echo "=========Delete quantized models of batch ${batch_name}========="
+  # find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+done
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs/allot ]; then
+    mkdir -p pdfs/allot
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+$OLD_DIR/../data-vis/plot-mxq-paired.R data/combined.csv
+$OLD_DIR/../data-vis/plot-mem-consumption.R data/combined.csv
+$OLD_DIR/../data-vis/plot-quant-speed.R data/combined.csv
+$OLD_DIR/../data-vis/gen-table-mxq-llm.R data/combined.csv
+pdflatex table.tex
+
+# plot configuration allocations for 3 * 12 MXQ combinations
+# MODELS="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+MODELS="Llama-2-7b-hf"
+BGS="4.13 4.25 4.51"
+for model in $MODELS; do
+    for budget in $BGS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+          -m $model \
+          -b $budget \
+          --attempt1 tail-prioritized_2_00 \
+          --attempt2 tail-prioritized_4_00 \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/scripts/experiment-tail-boost.sh

@@ -0,0 +1,117 @@
+#!/bin/bash
+
+# BUDGETS="2.13 2.25 2.51 3.13 3.25 3.51 4.13 4.25 4.51"
+BUDGETS="4.13 4.25 4.51"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+EXP_BASE_NAME="tail-boost"
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+
+weight_algo=tail_boost
+ATTEMPT="${weight_algo}"
+EXP_NAME="${ATTEMPT}"
+log_file="logs/bench-$(date +%Y%m%d%H%M%S).log"
+
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+echo "=========Run perplexity evaluation on batch ${EXP_NAME}========="
+python ../src/cli.py llm \
+  --task eval_ppl \
+  --model 0 1 2 \
+  --algo mxq \
+  --weight-algo $weight_algo \
+  --config ${BUDGETS} \
+  --experiment-name "${EXP_NAME}_ppl" \
+  --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+  --result-dir=$RESULT_DIR \
+  2>&1 \
+  | tee -a $log_file
+EXIT_CODE=$?
+if [ $EXIT_CODE -ne 0 ]; then
+  echo "Perplexity evaluation failed!"
+  exit $EXIT_CODE
+fi
+echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_ppl \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+python ../src/cli.py dump \
+  --type quant_config \
+  --model 0 1 2 \
+  --budget ${BUDGETS} \
+  --attempt $ATTEMPT \
+  --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+  --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+  2>&1 \
+  | tee -a $log_file
+
+echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+algo=mxq
+model_ids="0 1 2"
+for m in $model_ids; do
+  for cfg in ${BUDGETS}; do
+      python ../src/cli.py llm \
+          --model $m \
+          --algo ${algo} \
+          --config ${cfg} \
+          --task eval_model_storage \
+          --experiment-name "${EXP_NAME}_stor" \
+          --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+          --result-dir=$RESULT_DIR \
+          2>&1 \
+          | tee -a $log_file
+  done
+done
+echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_stor \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+# echo "=========Delete quantized models of batch ${batch_name}========="
+# find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs/allot ]; then
+    mkdir -p pdfs/allot
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+$OLD_DIR/../data-vis/plot-mxq-paired.R data/combined.csv
+$OLD_DIR/../data-vis/plot-mem-consumption.R data/combined.csv
+$OLD_DIR/../data-vis/plot-quant-speed.R data/combined.csv
+$OLD_DIR/../data-vis/gen-table-mxq-llm.R data/combined.csv
+pdflatex table.tex
+
+# plot configuration allocations for 3 * 12 MXQ combinations
+MODELS="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+BGS="4.13 4.25 4.51"
+for model in $MODELS; do
+    for budget in $BGS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+          -m $model \
+          -b $budget \
+          --attempt1 tail_boost \
+          --attempt2 mxq1 \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/scripts/experiment-tail-reduction.sh

@@ -0,0 +1,118 @@
+#!/bin/bash
+
+# BUDGETS="2.13 2.25 2.51 3.13 3.25 3.51 4.13 4.25 4.51"
+BUDGETS="4.13 4.25 4.51"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+EXP_BASE_NAME="tail-reduction"
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+
+weight_algo=tail_boost
+ATTEMPT="${EXP_BASE_NAME}"
+EXP_NAME="${ATTEMPT}"
+log_file="logs/bench-$(date +%Y%m%d%H%M%S).log"
+
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+echo "=========Run perplexity evaluation on batch ${EXP_NAME}========="
+python ../src/cli.py llm \
+  --task eval_ppl \
+  --model 0 1 2 \
+  --algo mxq \
+  --weight-algo $weight_algo \
+  --factor -1 \
+  --config ${BUDGETS} \
+  --experiment-name "${EXP_NAME}_ppl" \
+  --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+  --result-dir=$RESULT_DIR \
+  2>&1 \
+  | tee -a $log_file
+EXIT_CODE=$?
+if [ $EXIT_CODE -ne 0 ]; then
+  echo "Perplexity evaluation failed!"
+  exit $EXIT_CODE
+fi
+echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+# find $RESULT_DIR/${EXP_NAME}_ppl \
+#   -name "result-*.csv" \
+#   -printf '%T@ %p\n' \
+#   | sort -n \
+#   | tail -1 \
+#   | cut -d' '  -f2 \
+#   | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+#
+# echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+# mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+# python ../src/cli.py dump \
+#   --type quant_config \
+#   --model 0 1 2 \
+#   --budget ${BUDGETS} \
+#   --attempt $ATTEMPT \
+#   --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+#   --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+#   2>&1 \
+#   | tee -a $log_file
+#
+# echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+# algo=mxq
+# model_ids="0 1 2"
+# for m in $model_ids; do
+#   for cfg in ${BUDGETS}; do
+#       python ../src/cli.py llm \
+#           --model $m \
+#           --algo ${algo} \
+#           --config ${cfg} \
+#           --task eval_model_storage \
+#           --experiment-name "${EXP_NAME}_stor" \
+#           --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+#           --result-dir=$RESULT_DIR \
+#           2>&1 \
+#           | tee -a $log_file
+#   done
+# done
+# echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+# find $RESULT_DIR/${EXP_NAME}_stor \
+#   -name "result-*.csv" \
+#   -printf '%T@ %p\n' \
+#   | sort -n \
+#   | tail -1 \
+#   | cut -d' '  -f2 \
+#   | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+# echo "=========Delete quantized models of batch ${batch_name}========="
+# find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs/allot ]; then
+    mkdir -p pdfs/allot
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+$OLD_DIR/../data-vis/plot-mxq-paired.R data/combined.csv
+$OLD_DIR/../data-vis/plot-mem-consumption.R data/combined.csv
+$OLD_DIR/../data-vis/plot-quant-speed.R data/combined.csv
+$OLD_DIR/../data-vis/gen-table-mxq-llm.R data/combined.csv
+pdflatex table.tex
+
+# plot configuration allocations for 3 * 12 MXQ combinations
+MODELS="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+BUDGETS="4.13 4.25 4.51"
+for model in $MODELS; do
+    for budget in $BUDGETS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+          -m $model \
+          -b $budget \
+          --attempt1 tail-reduction \
+          --attempt2 mxq1 \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/scripts/experiment-vit-zs-mxq-kurt-boost.sh

@@ -0,0 +1,34 @@
+#!/bin/bash
+
+if [ ! -d logs ]; then
+    mkdir logs
+fi
+
+# python ../src/cli.py vit \
+#     --task eval_zeroshot_cls \
+#     --model 0 1 \
+#     --config 4.51 4.25 4.13 3.51 3.25 3.13\
+#     --algo mxq \
+#     --weight-algo kurt-boost \
+#     --boost-stop 2 \
+#     --top-m-layer 1 \
+#     --experiment-name eval_zs_BH_mxq_kurt_boost \
+#     --quant-snapshot-dir="/fdata/llm/mxq/snapshots" \
+#     --result-dir="/fdata/llm/mxq/results" \
+#     2>&1 \
+#     | tee logs/bench-vit-$(date +%Y%m%d%H%M%S).log
+
+python ../src/cli.py vit \
+    --task eval_zeroshot_cls \
+    --model 0 1 \
+    --config 4.51 4.25 4.13 3.51 3.25 3.13\
+    --algo mxq \
+    --weight-algo kurt-boost \
+    --boost-stop 2 \
+    --top-m-layer 2 \
+    --experiment-name eval_zs_BH_mxq_kurt_boost_22 \
+    --quant-snapshot-dir="/fdata/llm/mxq/snapshots" \
+    --result-dir="/fdata/llm/mxq/results" \
+    2>&1 \
+    | tee logs/bench-vit-$(date +%Y%m%d%H%M%S).log
+

lm-quant-toolkit/scripts/fix-3.25-llama-7b-sensi-boost.sh

@@ -0,0 +1,116 @@
+#!/bin/bash
+
+BUDGETS="3.25"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+weight_algo=sensi-boost
+MODELS="0"
+MODEL_NAMES="Llama-2-7b-hf"
+
+
+# BOOST_STOPS="2 3"
+BOOST_STOPS="3"
+BOOST_TOP_MS="3"
+
+for BOOST_STOP in $BOOST_STOPS; do
+    for BOOST_TOP_M in $BOOST_TOP_MS; do
+        if [[ $BOOST_STOP -eq 2 && $BOOST_TOP_M -eq 1 ]]; then
+            continue
+        fi
+        ATTEMPT="sb-3257_2-fix-${BOOST_STOP}-${BOOST_TOP_M}"
+        EXP_BASE_NAME=$ATTEMPT
+        mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+        log_file="logs/bench-${ATTEMPT}-$(date +%Y%m%d%H%M%S).log"
+
+        mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+        mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+        mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+        # MODELS="0"
+        EXP_NAME="${ATTEMPT}"
+        echo "=========Run perplexity evaluation========="
+        python ../src/cli.py llm \
+          --task eval_ppl \
+          --model $MODELS \
+          --algo mxq \
+          --weight-algo $weight_algo \
+          --boost-stop $BOOST_STOP \
+          --top-m-layer $BOOST_TOP_M \
+          --config ${BUDGETS} \
+          --experiment-name "${EXP_NAME}_ppl" \
+          --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+          --result-dir=$RESULT_DIR \
+          2>&1 \
+          | tee -a $log_file
+        EXIT_CODE=$?
+        if [ $EXIT_CODE -ne 0 ]; then
+          echo "Perplexity evaluation failed!"
+          exit $EXIT_CODE
+        fi
+        echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+        find $RESULT_DIR/${EXP_NAME}_ppl \
+          -name "result-*.csv" \
+          -printf '%T@ %p\n' \
+          | sort -n \
+          | tail -1 \
+          | cut -d' '  -f2 \
+          | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+        echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+        mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+        python ../src/cli.py dump \
+          --type quant_config \
+          --model $MODELS \
+          --budget ${BUDGETS} \
+          --attempt $ATTEMPT \
+          --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+          --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+          2>&1 \
+          | tee -a $log_file
+
+        echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+        algo=mxq
+        model_ids=$MODELS
+        for m in $model_ids; do
+          for cfg in ${BUDGETS}; do
+              python ../src/cli.py llm \
+                  --model $m \
+                  --algo ${algo} \
+                  --config ${cfg} \
+                  --task eval_model_storage \
+                  --experiment-name "${EXP_NAME}_stor" \
+                  --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+                  --result-dir=$RESULT_DIR \
+                  2>&1 \
+                  | tee -a $log_file
+          done
+        done
+        echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+        find $RESULT_DIR/${EXP_NAME}_stor \
+          -name "result-*.csv" \
+          -printf '%T@ %p\n' \
+          | sort -n \
+          | tail -1 \
+          | cut -d' '  -f2 \
+          | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+        # echo "=========Delete quantized models of batch ${batch_name}========="
+        # find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+
+        OLD_DIR=$(pwd)
+        cd $RESULT_DIR/$EXP_BASE_NAME
+        if [ ! -d pdfs/allot ]; then
+            mkdir -p pdfs/allot
+        fi
+        $OLD_DIR/../data-vis/combine.R \
+            --baseline_data_dir $OLD_DIR/../data-vis/data \
+            --mxq_data_dir data
+
+        cd $OLD_DIR
+    done
+done
+

lm-quant-toolkit/scripts/fix-experiment-llama-kurt-milp-ablation.sh

@@ -0,0 +1,82 @@
+#!/bin/bash
+
+BUDGETS="6.89 5.72 5.02 4.51 4.25 4.21 4.17 4.13 4.11 4.07 3.95 3.87 3.83 3.65 3.51 3.25 3.19 3.15 3.13 3.11 3.07"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+weight_algo=kurt-milp
+MODELS="0 1 2"
+MODEL_NAMES="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+
+
+# for SensiMiLP ablation test, all topm values are equivalent
+
+ATTEMPT="kurt-milp-abl"
+EXP_BASE_NAME=$ATTEMPT
+EXP_NAME="${ATTEMPT}"
+
+log_file="logs/bench-${ATTEMPT}-$(date +%Y%m%d%H%M%S).log"
+
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+python ../src/cli.py dump \
+  --type quant_config \
+  --model $MODELS \
+  --budget ${BUDGETS} \
+  --attempt $ATTEMPT \
+  --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+  --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+  2>&1 \
+  | tee -a $log_file
+
+echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+algo=mxq
+model_ids=$MODELS
+for m in $model_ids; do
+  for cfg in ${BUDGETS}; do
+      python ../src/cli.py llm \
+          --model $m \
+          --algo ${algo} \
+          --config ${cfg} \
+          --task eval_model_storage \
+          --experiment-name "${EXP_NAME}_stor" \
+          --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+          --result-dir=$RESULT_DIR \
+          2>&1 \
+          | tee -a $log_file
+  done
+done
+echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_stor \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+# echo "=========Delete quantized models of batch ${batch_name}========="
+# find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs ]; then
+    mkdir pdfs
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+$OLD_DIR/../data-vis/plot-ppl-mem.R -d data/combined.csv
+$OLD_DIR/../data-vis/gen-table-mxq-llm.R --csv_file data/combined.csv --attempt $ATTEMPT
+cd pdfs
+pdflatex table.tex
+cd ..
+
+cd $OLD_DIR

lm-quant-toolkit/scripts/fix-experiment-llama-sensi-milp-ablation.sh

@@ -0,0 +1,86 @@
+#!/bin/bash
+
+BUDGETS="6.89 5.72 5.02 4.51 4.25 4.21 4.17 4.13 4.11 4.07 3.95 3.87 3.83 3.65 3.51 3.25 3.19 3.15 3.13 3.11 3.07"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+weight_algo=sensi-milp
+MODELS="0 1 2"
+MODEL_NAMES="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+
+
+BOOST_TOP_MS="3"
+
+for BOOST_TOP_M in $BOOST_TOP_MS; do
+    ATTEMPT="sensi-milp-abl-${BOOST_TOP_M}"
+    EXP_BASE_NAME=$ATTEMPT
+    mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+    log_file="logs/bench-${ATTEMPT}-$(date +%Y%m%d%H%M%S).log"
+
+    mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+    mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+    mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+    EXP_NAME="${ATTEMPT}"
+
+    echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+    mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+    python ../src/cli.py dump \
+      --type quant_config \
+      --model $MODELS \
+      --budget ${BUDGETS} \
+      --attempt $ATTEMPT \
+      --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+      --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+      2>&1 \
+      | tee -a $log_file
+
+    echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+    algo=mxq
+    model_ids=$MODELS
+    for m in $model_ids; do
+      for cfg in ${BUDGETS}; do
+          python ../src/cli.py llm \
+              --model $m \
+              --algo ${algo} \
+              --config ${cfg} \
+              --task eval_model_storage \
+              --experiment-name "${EXP_NAME}_stor" \
+              --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+              --result-dir=$RESULT_DIR \
+              2>&1 \
+              | tee -a $log_file
+      done
+    done
+    echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+    find $RESULT_DIR/${EXP_NAME}_stor \
+      -name "result-*.csv" \
+      -printf '%T@ %p\n' \
+      | sort -n \
+      | tail -1 \
+      | cut -d' '  -f2 \
+      | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+    # echo "=========Delete quantized models of batch ${batch_name}========="
+    # find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+    OLD_DIR=$(pwd)
+    cd $RESULT_DIR/$EXP_BASE_NAME
+    if [ ! -d pdfs ]; then
+        mkdir pdfs
+    fi
+    $OLD_DIR/../data-vis/combine.R \
+        --baseline_data_dir $OLD_DIR/../data-vis/data \
+        --mxq_data_dir data
+    $OLD_DIR/../data-vis/plot-ppl-mem.R -d data/combined.csv
+    $OLD_DIR/../data-vis/gen-table-mxq-llm.R --csv_file data/combined.csv --attempt $ATTEMPT
+    cd pdfs
+    pdflatex table.tex
+    cd ..
+
+    cd $OLD_DIR
+
+done
+

lm-quant-toolkit/scripts/fix-llama-sensi-milp.sh

@@ -0,0 +1,129 @@
+#!/bin/bash
+
+BUDGETS="6.89 5.72 5.02 4.51 4.25 4.21 4.17 4.13 4.11 4.07 3.95 3.87 3.83 3.65 3.51 3.25 3.19 3.15 3.13 3.11 3.07"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+weight_algo=sensi-milp
+MODELS="0 1 2"
+MODEL_NAMES="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+
+
+BOOST_TOP_MS="3"
+
+for BOOST_TOP_M in $BOOST_TOP_MS; do
+    ATTEMPT="sensi-milp-${BOOST_TOP_M}"
+    EXP_BASE_NAME=$ATTEMPT
+    mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+    log_file="logs/bench-${ATTEMPT}-$(date +%Y%m%d%H%M%S).log"
+
+    mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+    mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+    mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+    # MODELS="0"
+    EXP_NAME="${ATTEMPT}"
+    echo "=========Run perplexity evaluation========="
+    python ../src/cli.py llm \
+      --task eval_ppl \
+      --model $MODELS \
+      --algo mxq \
+      --weight-algo $weight_algo \
+      --top-m-layer $BOOST_TOP_M \
+      --config ${BUDGETS} \
+      --experiment-name "${EXP_NAME}_ppl" \
+      --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+      --result-dir=$RESULT_DIR \
+      2>&1 \
+      | tee -a $log_file
+    EXIT_CODE=$?
+    if [ $EXIT_CODE -ne 0 ]; then
+      echo "Perplexity evaluation failed!"
+      exit $EXIT_CODE
+    fi
+    echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+    find $RESULT_DIR/${EXP_NAME}_ppl \
+      -name "result-*.csv" \
+      -printf '%T@ %p\n' \
+      | sort -n \
+      | tail -1 \
+      | cut -d' '  -f2 \
+      | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+    echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+    mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+    python ../src/cli.py dump \
+      --type quant_config \
+      --model $MODELS \
+      --budget ${BUDGETS} \
+      --attempt $ATTEMPT \
+      --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+      --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+      2>&1 \
+      | tee -a $log_file
+
+    echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+    algo=mxq
+    model_ids=$MODELS
+    for m in $model_ids; do
+      for cfg in ${BUDGETS}; do
+          python ../src/cli.py llm \
+              --model $m \
+              --algo ${algo} \
+              --config ${cfg} \
+              --task eval_model_storage \
+              --experiment-name "${EXP_NAME}_stor" \
+              --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+              --result-dir=$RESULT_DIR \
+              2>&1 \
+              | tee -a $log_file
+      done
+    done
+    echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+    find $RESULT_DIR/${EXP_NAME}_stor \
+      -name "result-*.csv" \
+      -printf '%T@ %p\n' \
+      | sort -n \
+      | tail -1 \
+      | cut -d' '  -f2 \
+      | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+    # echo "=========Delete quantized models of batch ${batch_name}========="
+    # find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+    OLD_DIR=$(pwd)
+    cd $RESULT_DIR/$EXP_BASE_NAME
+    if [ ! -d pdfs ]; then
+        mkdir pdfs
+    fi
+    $OLD_DIR/../data-vis/combine.R \
+        --baseline_data_dir $OLD_DIR/../data-vis/data \
+        --mxq_data_dir data
+    $OLD_DIR/../data-vis/plot-ppl-mem.R -d data/combined.csv
+    $OLD_DIR/../data-vis/plot-mem-consumption.R -d data/combined.csv
+    $OLD_DIR/../data-vis/plot-quant-speed.R -d data/combined.csv
+    $OLD_DIR/../data-vis/gen-table-mxq-llm.R --csv_file data/combined.csv --attempt $ATTEMPT
+    cd pdfs
+    pdflatex table.tex
+    cd ..
+
+    # plot configuration allocations for 3 * 12 MXQ combinations
+    MODELS="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+    BUDGETS=${MXQ_BATCH[@]:1}
+    for model in $MODELS; do
+        for budget in $BUDGETS; do
+            $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+              -m $model \
+              -b $budget \
+              --fnorm_data_dir $OLD_DIR/../src/data \
+              --attempt1 mxq1 \
+              --attempt2 $ATTEMPT \
+              --quant_cfg_allot_file data/quant-cfg-allocation.csv
+        done
+    done
+    cd $OLD_DIR
+
+done
+

lm-quant-toolkit/scripts/fix_experiment-llama-kurt-milp.sh

@@ -0,0 +1,87 @@
+#!/bin/bash
+
+BUDGETS="6.89 5.72 5.02 4.51 4.25 4.21 4.17 4.13 4.11 4.07 3.95 3.87 3.83 3.65 3.51 3.25 3.19 3.15 3.13 3.11 3.07"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+weight_algo=kurt-milp
+MODELS="0 1 2"
+MODEL_NAMES="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+
+
+BOOST_TOP_MS="1"
+
+for BOOST_TOP_M in $BOOST_TOP_MS; do
+    ATTEMPT="kurt-milp-${BOOST_TOP_M}"
+    EXP_BASE_NAME=$ATTEMPT
+    mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+    log_file="logs/bench-${ATTEMPT}-$(date +%Y%m%d%H%M%S).log"
+
+    mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+    mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+    mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+    # MODELS="0"
+    EXP_NAME="${ATTEMPT}"
+
+    echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+    mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+    python ../src/cli.py dump \
+      --type quant_config \
+      --model $MODELS \
+      --budget ${BUDGETS} \
+      --attempt $ATTEMPT \
+      --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+      --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+      2>&1 \
+      | tee -a $log_file
+
+    echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+    algo=mxq
+    model_ids=$MODELS
+    for m in $model_ids; do
+      for cfg in ${BUDGETS}; do
+          python ../src/cli.py llm \
+              --model $m \
+              --algo ${algo} \
+              --config ${cfg} \
+              --task eval_model_storage \
+              --experiment-name "${EXP_NAME}_stor" \
+              --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+              --result-dir=$RESULT_DIR \
+              2>&1 \
+              | tee -a $log_file
+      done
+    done
+    echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+    find $RESULT_DIR/${EXP_NAME}_stor \
+      -name "result-*.csv" \
+      -printf '%T@ %p\n' \
+      | sort -n \
+      | tail -1 \
+      | cut -d' '  -f2 \
+      | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+    # echo "=========Delete quantized models of batch ${batch_name}========="
+    # find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+    OLD_DIR=$(pwd)
+    cd $RESULT_DIR/$EXP_BASE_NAME
+    if [ ! -d pdfs ]; then
+        mkdir pdfs
+    fi
+    $OLD_DIR/../data-vis/combine.R \
+        --baseline_data_dir $OLD_DIR/../data-vis/data \
+        --mxq_data_dir data
+    $OLD_DIR/../data-vis/plot-ppl-mem.R -d data/combined.csv
+    $OLD_DIR/../data-vis/gen-table-mxq-llm.R --csv_file data/combined.csv --attempt $ATTEMPT
+    cd pdfs
+    pdflatex table.tex
+    cd ..
+
+    cd $OLD_DIR
+
+done
+

lm-quant-toolkit/scripts/plot-allot-kurt-scaled-dense.sh

@@ -0,0 +1,46 @@
+#!/bin/bash
+
+# MXQ1=(2bit 2.13 2.15 2.17 2.19 2.21 2.23 2.25 2.27 2.29 2.31 2.45 2.47 2.49 2.51 2.53 2.55 2.57)
+# MXQ2=(3bit 3.07 3.09 3.11 3.13 3.15 3.17 3.19 3.21 3.23 3.25 3.27 3.29 3.31 3.33 3.35 3.37 3.39 3.41 3.42 3.43 3.45 3.47 3.49 3.51 3.53 3.55 3.57 3.59 3.61 3.63 3.65 3.67 3.69 3.71 3.73 3.75 3.77 3.79 3.81 3.83 3.85 3.87 3.89 3.91 3.93 3.95 3.97 3.99)
+# Removed 4.99 which is unsolvable with kurt-scaled scheme
+MXQ4=(4bit 4.01 4.03 4.05 4.07 4.09 4.11 4.13 4.15 4.17 4.19 4.21 4.23 4.25 4.27 4.29 4.31 4.33 4.35 4.37 4.39 4.41 4.43 4.45 4.47 4.49 4.51 4.53 4.55 4.57 4.59 4.61 4.63 4.65 4.67 4.69 4.71 4.73 4.75 4.77 4.79 4.81 4.83 4.85 4.87 4.89 4.91 4.93 4.95 4.97)
+# Remove 5.01 5.02, 5.04 5.10 5.32 which is unsolvable with kurt-scaled scheme
+#MXQ5=(5bit 5.00 5.03 5.06 5.08 5.09 5.12 5.14 5.16 5.18 5.20 5.22 5.24 5.26 5.28 5.30 5.33 5.34 5.36 5.38 5.40 5.42 5.44 5.46 5.48 5.50 5.52 5.54 5.56 5.58 5.60 5.62 5.64 5.66 5.68 5.70 5.72 5.74 5.76 5.78 5.80 5.82 5.84 5.86 5.88 5.90 5.92 5.94 5.96 5.98)
+#MXQ6=(6bit 6.00 6.02 6.03 6.04 6.05 6.06 6.07 6.09 6.11 6.13 6.15 6.17 6.19 6.21 6.23 6.25 6.27 6.29 6.31 6.33 6.35 6.37 6.39 6.41 6.43 6.45 6.47 6.49 6.51 6.53 6.55 6.57 6.59 6.61 6.63 6.65 6.68 6.69 6.71 6.72 6.75 6.77 6.78 6.81 6.83 6.86 6.87 6.89 6.92 6.95 6.96 6.98)
+# MXQ7=(78bit 7.01 7.02 7.04 7.06 7.08 7.10 7.13 7.14 7.16 7.19 7.20 7.22 7.24 7.26 7.29 7.30 7.32 7.34 7.36 7.38 7.41 7.42 7.44 7.47 7.48 7.50 7.53 7.54 7.56 7.57 7.60 7.62 7.63 7.66 7.68 7.70 7.72 7.74 7.76 8.13 8.25 8.51)
+MXQ_BATCHES=(MXQ4)
+declare -n MXQ_BATCH
+
+ATTEMPT="kurt-scaled"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+EXP_BASE_NAME="${ATTEMPT}-dense"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+# Setup data files directories for reporting
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs ]; then
+    mkdir pdfs
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+
+# plot configuration allocations for 3 * 12 MXQ combinations
+# MODELS="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+MODELS="Llama-2-7b-hf"
+BUDGETS="4.13 4.25 4.51"
+for model in $MODELS; do
+    for budget in $BUDGETS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R -m $model -b $budget \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --attempt1 mxq1 \
+          --attempt2 $ATTEMPT \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/scripts/plot-llama-sensi.sh

@@ -0,0 +1,43 @@
+#!/bin/bash
+
+# BUDGETS="2.13 2.25 2.51 3.13 3.25 3.51 4.13 4.25 4.51"
+BUDGETS="4.13 4.25 4.51"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+ATTEMPT="llama-sensi"
+EXP_BASE_NAME=$ATTEMPT
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+
+weight_algo=sensi-directive
+
+log_file="logs/bench-$(date +%Y%m%d%H%M%S).log"
+
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs/allot ]; then
+    mkdir -p pdfs/allot
+fi
+
+# plot configuration allocations for 3 * 12 MXQ combinations
+MODELS="Llama-2-7b-hf"
+BGS="4.13 4.25 4.51"
+for model in $MODELS; do
+    for budget in $BGS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+          -m $model \
+          -b $budget \
+          --fnorm FALSE \
+          --attempt1 $ATTEMPT \
+          --attempt2 mxq1 \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/scripts/plot-milp-low-bit.sh

@@ -0,0 +1,99 @@
+#!/bin/bash
+
+BUDGETS="4.25"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+
+FACTOR=2
+ATTEMPT="sensi-milp-debug-lowbits9"
+EXP_BASE_NAME=$ATTEMPT
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+
+# Use cached dataset to speedup wikitext, c4 ppl evaluation
+export HF_DATASETS_OFFLINE=1
+
+weight_algo=sensi-milp
+
+log_file="logs/bench-${ATTEMPT}-$(date +%Y%m%d%H%M%S).log"
+
+# MODELS="0 1 2"
+MODELS="0"
+EXP_NAME="${ATTEMPT}"
+
+echo "=========Collect perplexity evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_ppl \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/ppl/mxq/$ATTEMPT
+
+echo "=========Dump quantization configs on batch ${EXP_NAME}========="
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+python ../src/cli.py dump \
+  --type quant_config \
+  --model $MODELS \
+  --budget ${BUDGETS} \
+  --attempt $ATTEMPT \
+  --quant-snapshot-dir=$QUANT_SNAPSHOT_DIR \
+  --output-file "$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/quant-allot-${EXP_NAME}.csv" \
+  2>&1 \
+  | tee -a $log_file
+
+echo "=========Run memory evaluation on batch ${EXP_NAME}========="
+algo=mxq
+model_ids=$MODELS
+for m in $model_ids; do
+  for cfg in ${BUDGETS}; do
+      python ../src/cli.py llm \
+          --model $m \
+          --algo ${algo} \
+          --config ${cfg} \
+          --task eval_model_storage \
+          --experiment-name "${EXP_NAME}_stor" \
+          --quant-snapshot-dir="$QUANT_SNAPSHOT_DIR/$ATTEMPT" \
+          --result-dir=$RESULT_DIR \
+          2>&1 \
+          | tee -a $log_file
+  done
+done
+echo "=========Collect memory evaluation result on batch ${EXP_NAME}========="
+find $RESULT_DIR/${EXP_NAME}_stor \
+  -name "result-*.csv" \
+  -printf '%T@ %p\n' \
+  | sort -n \
+  | tail -1 \
+  | cut -d' '  -f2 \
+  | xargs -i cp {} $RESULT_DIR/$EXP_BASE_NAME/data/stor/mxq/$ATTEMPT
+
+# echo "=========Delete quantized models of batch ${batch_name}========="
+# find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs/allot ]; then
+    mkdir -p pdfs/allot
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+$OLD_DIR/../data-vis/plot-mxq-paired.R data/combined.csv
+$OLD_DIR/../data-vis/plot-mem-consumption.R data/combined.csv
+$OLD_DIR/../data-vis/plot-quant-speed.R data/combined.csv
+$OLD_DIR/../data-vis/gen-table-mxq-llm.R --csv_file data/combined.csv --attempt $ATTEMPT
+pdflatex table.tex
+
+# plot configuration allocations for 3 * 12 MXQ combinations
+MODELS="Llama-2-7b-hf"
+for model in $MODELS; do
+    for budget in $BUDGETS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+          -m $model \
+          -b $budget \
+          --attempt1 $ATTEMPT \
+          --attempt2 mxq1 \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/scripts/plot-qwen-sensi-metrics.sh

@@ -0,0 +1,17 @@
+#!/bin/bash
+
+RESULT_BASE_DIR="/fdata/llm/mxq/results"
+CALIB_DATASETS="bos pileval wikitext c4"
+CONFIGS="b2g128 b2g64 b2g32 b3g128 b3g64 b3g32 b4g128 b4g64 b4g32 b8g128 b8g64 b8g32"
+MODELS="Qwen/Qwen2.5-7B Qwen/Qwen2.5-Coder-7B Qwen/Qwen2.5-Coder-7B-Instruct Qwen/Qwen2.5-Math-7B"
+
+EXP_NAME=sensi_qwen25
+RESULT_DIR=$RESULT_BASE_DIR/$EXP_NAME
+mkdir -p $RESULT_DIR/data
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR
+if [ ! -d pdfs ]; then
+    mkdir -p pdfs
+fi
+$OLD_DIR/../data-vis/plot-variant-sensi.R data/

lm-quant-toolkit/scripts/plot-sensi-milp-mini-batch.sh

@@ -0,0 +1,53 @@
+#!/bin/bash
+
+
+# 100%, 99%, 98%, 97%, 96%, 95% of [3.51, 4.25]
+# MXQ2=(minibatch 3.51 3.47 3.44 3.40 3.37 3.33 4.25 4.21 4.17 4.12 4.08 4.04)
+MXQ2=(minibatch 6.89 5.72 5.02 4.51 4.25 4.21 4.17 4.13 4.11 4.07 3.95 3.87 3.83 3.65 3.51 3.25 3.19 3.15 3.13 3.11 3.07)
+MXQ_BATCHES=(MXQ2)
+declare -n MXQ_BATCH
+
+ATTEMPT="sensi-milp-mini2"
+RESULT_DIR="/fdata/llm/mxq/results"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+EXP_BASE_NAME="${ATTEMPT}"
+QUANT_SNAPSHOT_DIR="/fdata/llm/mxq/snapshots"
+# Setup data files directories for reporting
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,qnt,stor}
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/{ppl,stor}/mxq/$ATTEMPT
+mkdir -p $QUANT_SNAPSHOT_DIR/$ATTEMPT
+
+weight_algo=sensi-milp
+
+for MXQ_BATCH in "${MXQ_BATCHES[@]}"; do
+  batch_name=${MXQ_BATCH[@]:0:1}
+  EXP_NAME="${EXP_BASE_NAME}-${batch_name}"
+  mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+  mkdir -p $RESULT_DIR/${EXP_NAME}_stor
+  log_file="logs/bench-${EXP_NAME}-$(date +%Y%m%d%H%M%S).log"
+
+  # echo "=========Delete quantized models of batch ${batch_name}========="
+  # find $QUANT_SNAPSHOT_DIR/$ATTEMPT -maxdepth 1 -type d | xargs rm -fr
+
+  OLD_DIR=$(pwd)
+  cd $RESULT_DIR/$EXP_BASE_NAME
+  if [ ! -d pdfs ]; then
+      mkdir pdfs
+  fi
+
+  # plot configuration allocations for 3 * 12 MXQ combinations
+  MODELS="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+  BUDGETS=${MXQ_BATCH[@]:1}
+  for model in $MODELS; do
+      for budget in $BUDGETS; do
+          $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+            -m $model \
+            -b $budget \
+            --fnorm_data_dir $OLD_DIR/../src/data \
+            --attempt1 mxq1 \
+            --attempt2 $ATTEMPT \
+            --quant_cfg_allot_file data/quant-cfg-allocation.csv
+      done
+  done
+  cd $OLD_DIR
+done

lm-quant-toolkit/scripts/quant-llm-13b-awq.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+if [ ! -d logs ]; then
+    mkdir logs
+fi
+
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+python ../src/cli.py llm \
+    --task quant \
+    --model 1 \
+    --algo awq \
+    --config b4g32 b4g64 b4g128 \
+    --experiment-name quant_llm_13B-awq2 \
+    --quant-snapshot-dir="/fdata/llm/mxq/snapshots" \
+    --result-dir="/fdata/llm/mxq/results" \
+    2>&1 \
+    | tee logs/bench-$(date +%Y%m%d%H%M%S).log

lm-quant-toolkit/scripts/quant-llm-2bit-dense1-mxq.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+
+# export HF_HOME=/data/hugginface
+# conda activate quant-eval
+
+if [ ! -d logs ]; then
+    mkdir logs
+fi
+
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+python ../src/cli.py llm \
+    --task quant \
+    --model 0 1 2 \
+    --algo mxq \
+    --config 2.57 2.55 2.53 2.51 2.49 2.47 2.45 2.31 2.29 2.27 2.25 2.23 2.21 2.19 2.17 2.15 2.13 \
+    --experiment-name quant_llm_2bit_dense1-mxq \
+    --quant-snapshot-dir="/fdata/llm/mxq/snapshots" \
+    --result-dir="/fdata/llm/mxq/results" \
+    2>&1 \
+    | tee logs/bench-$(date +%Y%m%d%H%M%S).log

lm-quant-toolkit/scripts/quant-llm-3bit-dense1-mxq.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+
+# export HF_HOME=/data/hugginface
+# conda activate quant-eval
+
+if [ ! -d logs ]; then
+    mkdir logs
+fi
+
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+python ../src/cli.py llm \
+    --task quant \
+    --model 0 1 2 \
+    --algo mxq \
+    --config 3.57 3.55 3.53 3.51 3.49 3.47 3.45 3.31 3.29 3.27 3.25 3.23 3.21 3.19 3.17 3.15 3.13 3.11 3.09 3.07 \
+    --experiment-name quant_llm_3bit_dense1-mxq \
+    --quant-snapshot-dir="/fdata/llm/mxq/snapshots" \
+    --result-dir="/fdata/llm/mxq/results" \
+    2>&1 \
+    | tee logs/bench-$(date +%Y%m%d%H%M%S).log

lm-quant-toolkit/scripts/quant-llm-4_51-mxq.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+
+# export HF_HOME=/data/hugginface
+# conda activate quant-eval
+
+if [ ! -d logs ]; then
+    mkdir logs
+fi
+
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+python ../src/cli.py llm \
+    --task quant \
+    --model 0 \
+    --algo mxq \
+    --config 4.51 \
+    --experiment-name quant_llm_4_51-mxq \
+    --quant-snapshot-dir="/fdata/llm/mxq/snapshots" \
+    --result-dir="/fdata/llm/mxq/results" \
+    2>&1 \
+    | tee logs/bench-$(date +%Y%m%d%H%M%S).log

lm-quant-toolkit/scripts/quant-llm-4bit-dense1-mxq.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+
+# export HF_HOME=/data/hugginface
+# conda activate quant-eval
+
+if [ ! -d logs ]; then
+    mkdir logs
+fi
+
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+python ../src/cli.py llm \
+    --task quant \
+    --model 0 1 2 \
+    --algo mxq \
+    --config 4.61 4.59 4.57 4.55 4.53 4.51 4.49 4.47 4.45 4.43 4.41 4.35 4.33 4.31 4.29 4.27 4.25 4.23 4.21 4.19 4.17 4.15 4.13 4.11 4.09 4.07 4.05 4.03 \
+    --experiment-name quant_llm_4bit_dense1-mxq \
+    --quant-snapshot-dir="/fdata/llm/mxq/snapshots" \
+    --result-dir="/fdata/llm/mxq/results" \
+    2>&1 \
+    | tee logs/bench-$(date +%Y%m%d%H%M%S).log

lm-quant-toolkit/scripts/quant-llm-8bit-gptq.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+if [ ! -d logs ]; then
+    mkdir logs
+fi
+
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+python ../src/cli.py llm \
+    --task quant \
+    --model 0 1 2 \
+    --algo gptq \
+    --config b8g32 b8g64 b8g128 \
+    --experiment-name quant_llm-8bit-gptq \
+    --quant-snapshot-dir="/fdata/llm/mxq/snapshots" \
+    --result-dir="/fdata/llm/mxq/results" \
+    2>&1 \
+    | tee logs/bench-$(date +%Y%m%d%H%M%S).log

lm-quant-toolkit/scripts/quant-llm-awq.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+if [ ! -d logs ]; then
+    mkdir logs
+fi
+
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+python ../src/cli.py llm \
+    --task quant \
+    --model 1 \
+    --algo awq \
+    --config b4g32 b4g64 b4g128 \
+    --experiment-name quant_llm_13B-awq \
+    --quant-snapshot-dir="/fdata/llm/mxq/snapshots" \
+    --result-dir="/fdata/llm/mxq/results" \
+    2>&1 \
+    | tee logs/bench-$(date +%Y%m%d%H%M%S).log

lm-quant-toolkit/scripts/quant-llm-b4-mxq.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+
+# export HF_HOME=/data/hugginface
+# conda activate quant-eval
+
+if [ ! -d logs ]; then
+    mkdir logs
+fi
+
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+python ../src/cli.py llm \
+    --task quant \
+    --model 0 1 2 \
+    --algo mxq \
+    --config 4.51 4.25 4.13 \
+    --experiment-name quant_llm_b4-mxq2 \
+    --quant-snapshot-dir="/fdata/llm/mxq/snapshots" \
+    --result-dir="/fdata/llm/mxq/results" \
+    2>&1 \
+    | tee logs/bench-$(date +%Y%m%d%H%M%S).log

lm-quant-toolkit/scripts/quant-llm-gptq.sh

@@ -0,0 +1,17 @@
+#!/bin/bash
+
+if [ ! -d logs ]; then
+    mkdir logs
+fi
+
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+python ../src/lm-quant-toolkit/src/cli.py llm \
+    --task quant \
+    --model 0 1 2 \
+    --algo gptq \
+    --experiment-name quant_llm-gptq \
+    --quant-snapshot-dir="/fdata/llm/mxq/snapshots" \
+    --result-dir="/fdata/llm/mxq/results" \
+    2>&1 \
+    | tee logs/bench-$(date +%Y%m%d%H%M%S).log

lm-quant-toolkit/scripts/quant-llm-hqq.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+
+# export HF_HOME=/data/hugginface
+# conda activate quant-eval
+
+if [ ! -d logs ]; then
+    mkdir logs
+fi
+
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+python ../src/cli.py llm \
+    --task quant \
+    --model 0 1 2 \
+    --algo hqq \
+    --experiment-name quant_llm-hqq \
+    --quant-snapshot-dir="/fdata/llm/mxq/snapshots" \
+    --result-dir="/fdata/llm/mxq/results" \
+    2>&1 \
+    | tee logs/bench-$(date +%Y%m%d%H%M%S).log

lm-quant-toolkit/scripts/quant-llm-kurt-13b-mxq.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+
+# export HF_HOME=/data/hugginface
+# conda activate quant-eval
+
+if [ ! -d logs ]; then
+    mkdir logs
+fi
+
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+python ../src/cli.py llm \
+    --task quant \
+    --model 1 \
+    --algo mxq \
+    --config 4.51 4.25 4.13 \
+    --experiment-name quant_llm-weighted-113b-mxq \
+    --quant-snapshot-dir="/fdata/llm/mxq/snapshots-kurt" \
+    --result-dir="/fdata/llm/mxq/results" \
+    2>&1 \
+    | tee logs/bench-$(date +%Y%m%d%H%M%S).log

lm-quant-toolkit/scripts/quant-llm-mxq.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+
+# export HF_HOME=/data/hugginface
+# conda activate quant-eval
+
+if [ ! -d logs ]; then
+    mkdir logs
+fi
+
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+python ../src/cli.py llm \
+    --task quant \
+    --model 0 1 2 \
+    --algo mxq \
+    --experiment-name quant_llm-mxq \
+    --quant-snapshot-dir="/fdata/llm/mxq/snapshots" \
+    --result-dir="/fdata/llm/mxq/results" \
+    2>&1 \
+    | tee logs/bench-$(date +%Y%m%d%H%M%S).log

lm-quant-toolkit/scripts/sim-quant-allot.sh

@@ -0,0 +1,60 @@
+#!/bin/bash
+
+# BUDGETS="2.13 2.25 2.51 3.13 3.25 3.51 4.13 4.25 4.51"
+BUDGETS="4.13 4.25 4.51"
+RESULT_DIR="/fdata/llm/mxq/results"
+
+TAILS=(tail-prioritized 2.00 4.00)
+# MXQ_BATCHES=(HEADS TAILS)
+MXQ_BATCHES=(TAILS)
+declare -n MXQ_BATCH
+
+EXP_BASE_NAME="tail-factor-search"
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME
+
+for MXQ_BATCH in "${MXQ_BATCHES[@]}"; do
+  weight_algo=${MXQ_BATCH[@]:0:1}
+  for factor in ${MXQ_BATCH[@]:1}; do
+      ATTEMPT="${weight_algo}_${factor/./_}"
+      EXP_NAME="${ATTEMPT}"
+      log_file="logs/bench-$(date +%Y%m%d%H%M%S).log"
+
+      mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+      mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+
+      # python -m pdb ../src/cli.py dump \
+      python ../src/cli.py dump \
+        --type quant_config_sim \
+        --model 0 1 2 \
+        --weight-algo $weight_algo \
+        --factor $factor \
+        --budget ${BUDGETS} \
+        --output-file="$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/${EXP_NAME}.csv"\
+        2>&1 \
+        | tee -a $log_file
+
+  done
+
+done
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs/allot ]; then
+    mkdir -p pdfs/allot
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+# plot configuration allocations for 3 * 12 MXQ combinations
+MODELS="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+for model in $MODELS; do
+    for budget in $BUDGETS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+          -m $model \
+          -b $budget \
+          --attempt1 tail-prioritized_2_00 \
+          --attempt2 tail-prioritized_4_00 \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/scripts/sim-quant-milp.sh

@@ -0,0 +1,4 @@
+#!/bin/bash
+
+# python ../src/lm_quant_toolkit/eval/common.py > 2-8-quant-bits.txt
+python ../src/lm_quant_toolkit/eval/common.py > 2-8-quant-bits-sensi.txt

lm-quant-toolkit/scripts/sim-tail-boost-quant-allot.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+
+# BUDGETS="2.13 2.25 2.51 3.13 3.25 3.51 4.13 4.25 4.51"
+BUDGETS="4.13 4.25 4.51"
+RESULT_DIR="/fdata/llm/mxq/results"
+weight_algo="tail_boost"
+EXP_BASE_NAME="$weight_algo"
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME
+
+ATTEMPT="$weight_algo"
+EXP_NAME="${ATTEMPT}"
+log_file="logs/bench-$(date +%Y%m%d%H%M%S).log"
+
+mkdir -p $RESULT_DIR/${EXP_NAME}_ppl
+mkdir -p $RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT
+
+# python -m pdb ../src/cli.py dump \
+python ../src/cli.py dump \
+    --type quant_config_sim \
+    --model 0 1 2 \
+    --weight-algo $weight_algo \
+    --budget ${BUDGETS} \
+    --output-file="$RESULT_DIR/$EXP_BASE_NAME/data/allot/mxq/$ATTEMPT/${EXP_NAME}.csv"\
+    2>&1 \
+    | tee -a $log_file
+
+
+OLD_DIR=$(pwd)
+cd $RESULT_DIR/$EXP_BASE_NAME
+if [ ! -d pdfs/allot ]; then
+    mkdir -p pdfs/allot
+fi
+$OLD_DIR/../data-vis/combine.R \
+    --baseline_data_dir $OLD_DIR/../data-vis/data \
+    --mxq_data_dir data
+# plot configuration allocations for 3 * 12 MXQ combinations
+MODELS="Llama-2-7b-hf Llama-2-13b-hf Meta-Llama-3-8B"
+for model in $MODELS; do
+    for budget in $BUDGETS; do
+        $OLD_DIR/../data-vis/plot-mxq-allocation.R \
+          -m $model \
+          -b $budget \
+          --attempt1 $weight_algo \
+          --attempt2 mxq1 \
+          --fnorm_data_dir $OLD_DIR/../src/data \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv
+    done
+done

lm-quant-toolkit/src/data/kurtosis-CLIP-ViT-L-14-laion2B-s32B-b82K.csv

@@ -0,0 +1,73 @@
+module,layer,kurtosis
+vision.mlp.c_fc,0,82.60432030937712
+vision.mlp.c_fc,1,12.176342485174775
+vision.mlp.c_fc,2,9.386186464371631
+vision.mlp.c_fc,3,6.0384986521099115
+vision.mlp.c_fc,4,5.121874436272312
+vision.mlp.c_fc,5,4.184969377709542
+vision.mlp.c_fc,6,4.256117717855523
+vision.mlp.c_fc,7,3.953278547908379
+vision.mlp.c_fc,8,3.923302040046366
+vision.mlp.c_fc,9,3.640527020435186
+vision.mlp.c_fc,10,3.8766116638521972
+vision.mlp.c_fc,11,3.364467358103123
+vision.mlp.c_fc,12,3.2501048983174172
+vision.mlp.c_fc,13,3.144904609090658
+vision.mlp.c_fc,14,3.1176022019209833
+vision.mlp.c_fc,15,3.170790033142059
+vision.mlp.c_fc,16,3.1901952176386157
+vision.mlp.c_fc,17,3.1728736833856566
+vision.mlp.c_fc,18,3.1092420753715064
+vision.mlp.c_fc,19,3.1289105323616995
+vision.mlp.c_fc,20,3.1633778054509616
+vision.mlp.c_fc,21,3.1948778928908625
+vision.mlp.c_fc,22,3.153364851937131
+vision.mlp.c_fc,23,3.1481862277653847
+vision.mlp.c_proj,0,48.91602226406029
+vision.mlp.c_proj,1,7.994586159177927
+vision.mlp.c_proj,2,9.259903305960076
+vision.mlp.c_proj,3,12.161504492865403
+vision.mlp.c_proj,4,11.137208824577082
+vision.mlp.c_proj,5,4.50118105140538
+vision.mlp.c_proj,6,3.9747365034192206
+vision.mlp.c_proj,7,4.846873977315018
+vision.mlp.c_proj,8,4.914742513787588
+vision.mlp.c_proj,9,5.0506453475209065
+vision.mlp.c_proj,10,13.838509154283804
+vision.mlp.c_proj,11,4.258906842824289
+vision.mlp.c_proj,12,3.4482411935554316
+vision.mlp.c_proj,13,3.4563538182296156
+vision.mlp.c_proj,14,3.1506796697623156
+vision.mlp.c_proj,15,3.140821870180245
+vision.mlp.c_proj,16,3.1870137578290736
+vision.mlp.c_proj,17,3.1622263243928037
+vision.mlp.c_proj,18,3.1727003318990294
+vision.mlp.c_proj,19,3.126420350320576
+vision.mlp.c_proj,20,3.214422564736524
+vision.mlp.c_proj,21,3.5410765909033515
+vision.mlp.c_proj,22,3.885907825953595
+vision.mlp.c_proj,23,3.241221353196137
+text.mlp.c_fc,0,3.111321629333239
+text.mlp.c_fc,1,3.1040094243226832
+text.mlp.c_fc,2,3.0476798189406455
+text.mlp.c_fc,3,3.0567396405641647
+text.mlp.c_fc,4,3.072398311749821
+text.mlp.c_fc,5,3.0581450022889753
+text.mlp.c_fc,6,3.053470998692992
+text.mlp.c_fc,7,3.0531601497538126
+text.mlp.c_fc,8,3.0696775698776992
+text.mlp.c_fc,9,3.065453985089178
+text.mlp.c_fc,10,3.067657384247557
+text.mlp.c_fc,11,3.086129858871849
+text.mlp.c_proj,0,16.411416535361464
+text.mlp.c_proj,1,10.636934397271798
+text.mlp.c_proj,2,5.467879831695406
+text.mlp.c_proj,3,3.117101945617834
+text.mlp.c_proj,4,3.162125318135728
+text.mlp.c_proj,5,3.114353715482094
+text.mlp.c_proj,6,3.1560610360682673
+text.mlp.c_proj,7,3.230989551532942
+text.mlp.c_proj,8,3.273957420333387
+text.mlp.c_proj,9,3.322224898837691
+text.mlp.c_proj,10,3.252371370612472
+text.mlp.c_proj,11,3.2207147468979427

lm-quant-toolkit/src/lm_quant_toolkit.egg-info/SOURCES.txt

@@ -0,0 +1,46 @@
+LICENSE
+MANIFEST.in
+README.md
+pyproject.toml
+docs/Makefile
+docs/source/CHANGELOG.rst
+docs/source/conf.py
+docs/source/index.rst
+docs/source/modules/index.rst
+src/cli.py
+src/dump.py
+src/lm_quant_toolkit/__init__.py
+src/lm_quant_toolkit.egg-info/PKG-INFO
+src/lm_quant_toolkit.egg-info/SOURCES.txt
+src/lm_quant_toolkit.egg-info/dependency_links.txt
+src/lm_quant_toolkit.egg-info/requires.txt
+src/lm_quant_toolkit.egg-info/top_level.txt
+src/lm_quant_toolkit/adapter/__init__.py
+src/lm_quant_toolkit/adapter/autoawq.py
+src/lm_quant_toolkit/adapter/autogptq.py
+src/lm_quant_toolkit/adapter/awq.py
+src/lm_quant_toolkit/adapter/bnb.py
+src/lm_quant_toolkit/adapter/common.py
+src/lm_quant_toolkit/adapter/fp16.py
+src/lm_quant_toolkit/adapter/hqq.py
+src/lm_quant_toolkit/adapter/mxq.py
+src/lm_quant_toolkit/adapter/vit/hqq.py
+src/lm_quant_toolkit/eval/__init__.py
+src/lm_quant_toolkit/eval/bench.py
+src/lm_quant_toolkit/eval/bench_vit.py
+src/lm_quant_toolkit/eval/clipbenchmark.py
+src/lm_quant_toolkit/eval/common.py
+src/lm_quant_toolkit/eval/leaderboard.py
+src/lm_quant_toolkit/eval/perplexity.py
+src/lm_quant_toolkit/misc/__init__.py
+src/lm_quant_toolkit/misc/lp_norm.py
+src/lm_quant_toolkit/misc/quant_sim.py
+src/lm_quant_toolkit/misc/qweight.py
+src/lm_quant_toolkit/prep/__init__.py
+src/lm_quant_toolkit/prep/fnorm.py
+src/lm_quant_toolkit/prep/sensitivity.py
+src/lm_quant_toolkit/prep/wdist.py
+src/lm_quant_toolkit/utils/__init__.py
+src/lm_quant_toolkit/utils/hub.py
+src/lm_quant_toolkit/utils/pickle.py
+src/lm_quant_toolkit/utils/safetensors.py

lm-quant-toolkit/src/lm_quant_toolkit.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

lm-quant-toolkit/src/lm_quant_toolkit.egg-info/requires.txt

@@ -0,0 +1,28 @@
+accelerate>=0.30.1
+bitsandbytes>0.37.0
+antlr4-python3-runtime==4.11.0
+datasets==2.20.0
+Jinja2==3.1.4
+langdetect==1.0.9
+nltk==3.9.1
+numpy==1.26.4
+optimum>=1.21.4
+pandas==2.2.2
+safetensors==0.4.3
+scikit-learn==1.4.2
+scipy==1.13.0
+sentencepiece==0.2.0
+tokenizers>=0.19.1
+torch>=2.1.0
+tqdm==4.66.4
+transformers>=4.41.2
+
+[dev]
+bumpver
+pip-tools
+
+[doc]
+sphinx
+
+[test]
+tox

lm-quant-toolkit/src/lm_quant_toolkit/__init__.py

@@ -0,0 +1,8 @@
+"""Packages for LLM Quantization Evaluation."""
+
+from importlib.metadata import PackageNotFoundError, version
+
+try:
+    __version__ = version("llm_quant_eval")
+except PackageNotFoundError:
+    __version__ = "unknown version"