⚗️ Style Transfer: add experiments for azure and jean-zay, bump trans…

…formers version to use llama3.2

lib/style-transfer/bash/experiment/rb_gen/az/llama3.2-3b.sh

@@ -0,0 +1,7 @@
+python style_transfer/run_rb_gen.py model.name=meta-llama/Llama-3.2-3B-Instruct \
+            model.peft_config.target_modules='["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]' \
+            dataset.name=bio-datasets/mimic_style_transfer \
+            max_steps=3 \
+            score.model.model_name_or_path=FremyCompany/BioLORD-2023-C \
+            dataset.sft_ratio=0.06 \
+            dataset.gen_ratio=0.7

lib/style-transfer/bash/experiment/rb_gen/az/test-azure.sh

@@ -0,0 +1 @@
+python style_transfer/run_rb_gen.py

lib/style-transfer/bash/experiment/rb_gen/jz-slurm/llama3b.sh

@@ -0,0 +1,6 @@
+sbatch bash/jz-slurm/submit-a100.sh \
+            model.name=meta-llama/Llama-3.2-3B-Instruct \
+            model.lora.target_modules='["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]' \
+            dataset.name=bio-datasets/mimic_style_transfer \
+            max_steps=3 \
+            score.model_name_or_path=FremyCompany/BioLORD-2023-C

lib/style-transfer/bash/jz-slurm/singularity-conv.sh → lib/style-transfer/bash/experiment/rb_gen/jz-slurm/singularity-conv.sh

@@ -4,7 +4,7 @@
 #SBATCH --cpus-per-task=10
 #SBATCH --output=log/conversion%j.out
 #SBATCH --partition=prepost
-#SBATCH --account=oha@cpu
+#SBATCH --account=oha@v100
 
 module purge
 module load singularity

lib/style-transfer/bash/experiment/rb_gen/jz-slurm/submit-a100.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+#SBATCH --job-name=gpu-job
+#SBATCH --nodes=1
+#SBATCH --gres=gpu:1
+#SBATCH --cpus-per-task=10
+#SBATCH --output=log/test.out
+#SBATCH --partition=gpu_p5
+#SBATCH --account=oha@a100
+#SBATCH -C a100
+#SBATCH --qos=qos_gpu_a100-t3
+
+module purge
+module load singularity
+
+SCRIPT_PATH="style_transfer/run_rb_gen.py"
+PYTHON="/opt/pysetup/.venv/bin/python"
+LIB_PATH="$WORK/open-nlp/lib/style-transfer/"
+singularity exec --bind $WORK/open-nlp,$HF_HOME,$WANDB_CACHE_DIR,$WANDB_DIR \
+                 --env PYTHONPATH=$PYTHONPATH:$LIB_PATH,HF_HUB_OFFLINE=True \
+                 --nv $SINGULARITY_ALLOWED_DIR/style-transfer.sif \
+                 bash -c "cd $LIB_PATH && $PYTHON $SCRIPT_PATH $1"

lib/style-transfer/bash/jz-slurm/test-pipeline.sh → lib/style-transfer/bash/experiment/rb_gen/jz-slurm/test-pipeline.sh

@@ -7,7 +7,7 @@
 #SBATCH --partition=gpu_p5
 #SBATCH --account=oha@a100
 #SBATCH -C a100
-#SBATCH --qos=qos_gpu-dev
+#SBATCH --qos=qos_gpu_a100-dev
 
 module purge
 module load singularity

lib/style-transfer/pyproject.toml

@@ -64,7 +64,7 @@ rich = "^13.5.2"
 sentence-transformers = "^2.3.1"
 torch = "^2.3.0"
 tqdm = "^4.66.1"
-transformers = "4.42.3"
+transformers = "4.45.1"
 trl = "^0.9.6"
 vllm = "^0.5.0.post1"
 wandb = "^0.16.3"

lib/style-transfer/style_transfer/rb_gen/steps/generate.py

@@ -56,8 +56,8 @@ def generate(
     tokenizer.save_pretrained("models/merged/")
     del model
     del tokenizer
-    gc.collect()
-    torch.cuda.empty_cache()
+    #gc.collect()
+    #torch.cuda.empty_cache()
     logging.info("🫧 Building VLLM Pipeline ...")
     llm = LLM(
         model="models/merged/",