add generation script

Author: Dando18

generate/generate-codellama-70b-prompted.sbatch

@@ -0,0 +1,46 @@
+#!/bin/bash
+#SBATCH -n 1
+#SBATCH -c 16
+#SBATCH --ntasks-per-node=1
+#SBATCH --gpus-per-task=1
+#SBATCH --mem=164000
+#SBATCH -t 23:59:59
+#SBATCH -A m2404
+#SBATCH -C gpu&hbm80g
+#SBATCH -q regular
+#SBATCH -J generate-codellama-70b-prompted
+#SBATCH -o generate-codellama-70b-prompted-%A.out
+
+# settings
+MODEL="codellama/CodeLlama-70b-hf"
+TEMP=0.2
+TOPP=0.95
+MAX_NEW_TKNS=1024
+SAMPLES_PER_PROMPT=20
+BATCH_SIZE=2
+hash=$(md5sum ../prompts/generation-prompts.json | cut -d' ' -f1)
+OUTPUT="../outputs/output_${hash:0:8}_${MODEL//\//--}_prompted_temp${TEMP}.json"
+CACHE="../outputs/cache/cache_${hash:0:8}_${MODEL//\//--}_prompted_temp${TEMP}.jsonl"
+echo "Writing to $OUTPUT"
+echo "model=$MODEL   MAX_NEW_TKNS=$MAX_NEW_TKNS   SAMPLES_PER_PROMPT=$SAMPLES_PER_PROMPT   BATCH_SIZE=$BATCH_SIZE"
+
+# setup
+#ml cuda/11.8.0
+source .env/bin/activate
+export HF_HOME=/pscratch/sd/d/dnicho/.cache/huggingface
+export OMP_NUM_THREADS=16
+export SLURM_CPU_BIND="cores"
+
+# generate
+srun python generate.py \
+    --model $MODEL \
+    --prompts ../prompts/generation-prompts.json \
+    --cache $CACHE \
+    --output $OUTPUT \
+    --temperature $TEMP \
+    --top_p $TOPP \
+    --do_sample \
+    --max_new_tokens $MAX_NEW_TKNS \
+    --num_samples_per_prompt $SAMPLES_PER_PROMPT \
+    --batch_size $BATCH_SIZE \
+    --prompted