update test paths

Signed-off-by: Alexandros Koumparoulis <[email protected]>

examples/llm/finetune/automodel.py

@@ -26,7 +26,7 @@
 
 # Run this example with torchrun, for example:
 # torchrun --nproc-per-node=8 \
-#   examples/llm/peft/automodel.py \
+#   examples/llm/finetune/automodel.py
 #   --strategy fsdp2 \
 #   --devices 8 \
 #   --model meta-llama/Llama-3.2-1B \
@@ -269,6 +269,11 @@ def main():
         'are currently supported only with position_ids and not attention_mask. Hence packed sequences needs to be'
         'run with --attn-implementation=flash_attention_2',
     )
+    parser.add_argument(
+        '--lora',
+        action='store_true',
+        help='Enables LoRA finetuning (PEFT); Default:  Supervised fine-tuning (SFT).'
+    )
 
     args = parser.parse_args()
 
@@ -341,7 +346,7 @@ def main():
         model,
         args.devices,
         args.num_nodes,
-        False,
+        args.lora,
         args.enable_cpu_offload,
         dp_size=args.dp_size,
         tp_size=args.tp_size,
@@ -411,6 +416,10 @@ def main():
         optim=optimizer,
         log=logger(args.ckpt_folder, args.max_steps // 2),
         resume=resume,
+        peft=llm.peft.LoRA(
+            target_modules=['*_proj'],
+            dim=8,
+        ) if args.lora else None,
     )
 
 

examples/llm/finetune/automodel_vllm.py

@@ -31,11 +31,19 @@
 
     parser = argparse.ArgumentParser()
     parser.add_argument('--model', required=True, type=str, help="Local path or model name on Hugging Face")
+    parser.add_argument('--lora', required=True, type=str, default=None, help="Local path of the lora model")
     parser.add_argument('--triton-model-name', required=True, type=str, help="Name for the service")
     args = parser.parse_args()
 
     exporter = vLLMHFExporter()
     exporter.export(model=args.model)
+    if args.lora is not None:
+        exporter.add_lora_models(lora_model_name=lora_model_name, lora_model=args.lora_model)
+        print(
+            "------------- Output: ", exporter.forward(input_texts=["How are you doing?"], lora_model_name=lora_model_name)
+        )
+        quit()
+
 
     nm = DeployPyTriton(
         model=exporter,
@@ -59,4 +67,4 @@
     )
 
     print("------------- Output: ", output_deployed)
-    nm.stop()
+    nm.stop()

tests/functional_tests/L2_HF_Transformer_PEFT_2gpu_FSDP2.sh

@@ -25,8 +25,8 @@ TRANSFORMERS_OFFLINE=1 coverage run -a --data-file=/workspace/.coverage --source
   --strategy fsdp2 \
   --ckpt-folder /tmp/hf_peft_ckpt_fsdp2 --auto-resume
 
-TRANSFORMERS_OFFLINE=1 HF_HOME=/home/TestData/automodel/hf_home coverage run -a --data-file=/workspace/.coverage --source=/workspace/nemo examples/llm/peft/automodel.py \
+TRANSFORMERS_OFFLINE=1 HF_HOME=/home/TestData/automodel/hf_home coverage run -a --data-file=/workspace/.coverage --source=/workspace/nemo examples/llm/finetune/automodel.py \
   --model /home/TestData/akoumparouli/hf_mixtral_2l/ \
   --max-steps 3 \
   --devices 2 \
-  --strategy fsdp2
+  --strategy fsdp2 --lora

tests/functional_tests/L2_HF_Transformer_PEFT_2gpu_FSDP2_fp8.sh

@@ -1,7 +1,7 @@
 export TRANSFORMERS_OFFLINE=1
 export HF_HOME=/home/TestData/automodel/hf_home
-coverage run -a --data-file=/workspace/.coverage --source=/workspace/nemo examples/llm/peft/automodel.py \
+coverage run -a --data-file=/workspace/.coverage --source=/workspace/nemo examples/llm/finetune/automodel.py \
     --model /home/TestData/akoumparouli/hf_mixtral_2l/ \
     --max-steps 3 \
     --devices 2 \
-    --strategy fsdp2 --fp8
+    --strategy fsdp2 --fp8 --lora

tests/functional_tests/L2_HF_Transformer_PEFT_2gpu_FSDP2_liger.sh

@@ -1,8 +1,8 @@
 export TRANSFORMERS_OFFLINE=1
 export HF_HOME=/home/TestData/automodel/hf_home
 
-coverage run -a --data-file=/workspace/.coverage --source=/workspace/nemo examples/llm/peft/automodel.py \
+coverage run -a --data-file=/workspace/.coverage --source=/workspace/nemo examples/llm/finetune/automodel.py \
     --model /home/TestData/akoumparouli/hf_mixtral_2l/ \
     --max-steps 3 \
     --devices 2 \
-    --strategy fsdp2 --liger
+    --strategy fsdp2 --liger --lora