Add extraction script: extract_lora_256.sh

extraction_scripts/extract_lora_256.sh

@@ -0,0 +1,36 @@
+#!/bin/bash
+#SBATCH --job-name=dequant_extract
+#SBATCH --output=process_%j.log
+#SBATCH --error=process_%j.err
+#SBATCH --time=04:00:00
+#SBATCH -p kisski-h100
+#SBATCH -G H100:4
+#SBATCH --nodes=1
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=64
+#SBATCH -C inet
+
+source activate /projects/extern/kisski/kisski-narges-llm-interactive/dir.project/surgery
+
+# PATHS
+BIOMNI_MODEL="/projects/extern/kisski/kisski-narges-llm-interactive/dir.project/Biomni-R0-32B-Preview"
+TEMP_BASE="./temp_qwen_dequantized_bf16" # The folder the python script creates
+OUTPUT_LORA="./lora_extraction_results/lora_fp8_corrected_rank_256"
+
+# 2. RUN MERGEKIT
+# Now we use the TEMP_BASE (which is now BF16) as the base model.
+# MergeKit will see two BF16 models and be happy.
+echo "------------------------------------------------"
+echo "STEP 2: extracting LoRA using MergeKit..."
+echo "------------------------------------------------"
+
+mergekit-extract-lora \
+    --model "$BIOMNI_MODEL" \
+    --base-model "$TEMP_BASE" \
+    --out-path "$OUTPUT_LORA" \
+    --max-rank 256 \
+    --device cuda \
+    --no-lazy-unpickle
+
+
+echo "✅ ALL DONE. Your mathematically corrected LoRA is in $OUTPUT_LORA"