Fix llm hp optimization error

sdk/python/kubeflow/storage_initializer/hugging_face.py

@@ -72,7 +72,9 @@ def download_model_and_tokenizer(self):
             trust_remote_code=True,
         )
         transformers.AutoTokenizer.from_pretrained(
-            self.model, cache_dir=VOLUME_PATH_MODEL
+            self.model,
+            token=self.config.access_token,
+            cache_dir=VOLUME_PATH_MODEL,
         )
 
 

sdk/python/kubeflow/storage_initializer/requirements.txt

@@ -1,4 +1,4 @@
-peft==0.3.0
-datasets==2.21.0
-transformers==4.38.0
-boto3==1.33.9
+peft==0.15.1
+datasets==3.5.0
+transformers==4.50.2
+boto3==1.37.22

sdk/python/kubeflow/trainer/hf_llm_training.py

@@ -13,6 +13,7 @@
     AutoModelForImageClassification,
     AutoTokenizer,
     DataCollatorForLanguageModeling,
+    DataCollatorWithPadding,
     Trainer,
     TrainingArguments,
 )
@@ -59,6 +60,9 @@ def setup_model_and_tokenizer(model_uri, transformer_type, model_dir, num_labels
     for param in model.parameters():
         param.requires_grad = False
 
+    if not tokenizer.pad_token:
+        tokenizer.pad_token = tokenizer.eos_token
+
     return model, tokenizer
 
 
@@ -148,6 +152,13 @@ def train_model(model, transformer_type, train_data, eval_data, tokenizer, train
             pad_to_multiple_of=8,
             mlm=False,
         )
+    else:
+        logger.info("Add general data collator with padding")
+        logger.info("-" * 40)
+        trainer.data_collator = DataCollatorWithPadding(
+            tokenizer,
+            pad_to_multiple_of=8,
+        )
 
     # Train the model.
     trainer.train()

sdk/python/kubeflow/trainer/requirements.txt

@@ -1,4 +1,5 @@
-peft==0.3.0
-datasets==2.21.0
-transformers==4.38.0
-accelerate==0.28.0
+peft==0.15.1
+datasets==3.5.0
+transformers==4.50.2
+accelerate==1.5.2
+tensorboard==2.19.0

sdk/python/kubeflow/training/api/training_client.py

@@ -294,7 +294,7 @@ def train(
                 VOLUME_PATH_DATASET,
                 "--lora_config",
                 json.dumps(
-                    trainer_parameters.lora_config.__dict__, cls=utils.SetEncoder
+                    trainer_parameters.lora_config.to_dict(), cls=utils.SetEncoder
                 ),
                 "--training_parameters",
                 json.dumps(trainer_parameters.training_parameters.to_dict()),

sdk/python/setup.py

@@ -64,6 +64,6 @@
     tests_require=TESTS_REQUIRES,
     extras_require={
         "test": TESTS_REQUIRES,
-        "huggingface": ["transformers==4.38.0", "peft==0.3.0"],
+        "huggingface": ["transformers==4.50.2", "peft==0.15.1"],
     },
 )

sdk/python/test/e2e-fine-tune-llm/test_e2e_pytorch_fine_tune_llm.py

@@ -55,7 +55,7 @@ def test_sdk_e2e_create_from_train_api(job_namespace="default"):
             training_parameters=transformers.TrainingArguments(
                 output_dir="test_trainer",
                 save_strategy="no",
-                evaluation_strategy="no",
+                eval_strategy="no",
                 do_eval=False,
                 disable_tqdm=True,
                 log_level="info",