Fix configuration errors

configs/fine_tuning.yaml

@@ -28,7 +28,7 @@ training_args:
   report_to: "wandb"
   seed: 46
   packing: True
-  torch_compile: True
+  torch_compile: False # setting to True might cause libcuda.so errors
 
 peft_args:
   r: 8

configs/model/llama_3.1_8b_instruct.yaml

@@ -1,11 +1,8 @@
 name: meta-llama/Llama-3.1-8B-Instruct
 tokenizer_name: ${.name}
-
 adapter_path: null
-
 max_seq_length: 32_000
 batch_size: 1
 padding: longest
 use_4bit: false
-
 use_unsloth: false

slurm/run_on_cluster.sh

@@ -2,11 +2,11 @@
 
 #SBATCH --job-name=juddges_sft
 #SBATCH --output=logs/%j-%x.log
-#SBATCH --time=72:00:00
+#SBATCH --time=48:00:00
 #SBATCH --nodes=1
-#SBATCH --gpus-per-node=1
-#SBATCH --cpus-per-gpu=8
-#SBATCH --mem=64G
+#SBATCH --gpus-per-node=4
+#SBATCH --cpus-per-gpu=4
+#SBATCH --mem=128G
 # NOTE: You can reconfigure the above parameters to your needs in the sbatch call.
 # NOTE: All env variables must be exported to be available after calling srun.
 # NOTE: You may need to specify some NCCL args in .env file depending on your cluster configuration