CODEOWNERS

@@ -8,4 +8,4 @@
 #  https://help.github.com/en/articles/about-code-owners
 #
 
-*  @alex-jw-brooks @gkumbhat @evaline-ju @gabe-l-hart @tharapalanivel @Ssukriti
+*  @gkumbhat @evaline-ju @gabe-l-hart

caikit_nlp/modules/text_embedding/crossencoder.py

@@ -431,8 +431,10 @@ def __init__(
         num_labels: int = None,
         max_length: int = None,
         device: str = None,
-        tokenizer_args: Dict = None,
         automodel_args: Dict = None,
+        tokenizer_args: Dict = None,
+        config_args: Dict = None,
+        cache_dir: str = None,
         trust_remote_code: bool = False,
         revision: Optional[str] = None,
         local_files_only: bool = False,
@@ -444,8 +446,10 @@ def __init__(
             num_labels,
             max_length,
             device,
-            tokenizer_args,
             automodel_args,
+            tokenizer_args,
+            config_args,
+            cache_dir,
             trust_remote_code,
             revision,
             local_files_only,

caikit_nlp/modules/text_generation/text_generation_local.py

@@ -641,6 +641,10 @@ def _launch_training(
     ) -> None:
         """Utility function to wrap trainer and execute training"""
 
+        # gradient_accumulation_steps seems to be causing hang atleast on CPU training
+        # so we are removing it for now here
+        del training_args["gradient_accumulation_steps"]
+
         trainer = base_model.get_trainer(
             train_dataset=training_dataset, **training_args
         )

caikit_nlp/resources/pretrained_model/base.py

@@ -48,7 +48,8 @@
 
 
 class LoggingTrainer(Trainer):
-    def log(self, logs: Dict[str, float]) -> None:
+    # pylint: disable=unused-argument
+    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
         """
         Log `logs` on the various objects watching training.
 
@@ -58,6 +59,8 @@ def log(self, logs: Dict[str, float]) -> None:
             logs (`Dict[str, float]`):
                 The values to log.
         """
+        # start_time was added in default trainer log
+        # https://github.com/huggingface/transformers/pull/34507
         self.state = log_step(self.state, logs)
         self.control = self.callback_handler.on_log(
             self.args, self.state, self.control, logs

caikit_nlp/resources/pretrained_model/hf_auto_seq2seq_lm.py

@@ -16,7 +16,7 @@
 """
 # Standard
 from collections.abc import Mapping
-from typing import Dict, List, Union
+from typing import Dict, List, Optional, Union
 
 # Third Party
 from torch.utils.data import IterableDataset
@@ -46,7 +46,8 @@
 
 
 class LoggingTrainer(Seq2SeqTrainer):
-    def log(self, logs: Dict[str, float]) -> None:
+    # pylint: disable=unused-argument
+    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
         """
         Log `logs` on the various objects watching training.
 
@@ -56,6 +57,8 @@ def log(self, logs: Dict[str, float]) -> None:
             logs (`Dict[str, float]`):
                 The values to log.
         """
+        # start_time was added in default trainer log
+        # https://github.com/huggingface/transformers/pull/34507
         self.state = log_step(self.state, logs)
         self.control = self.callback_handler.on_log(
             self.args, self.state, self.control, logs

examples/utils.py

@@ -88,7 +88,16 @@ def get_distributed_model(model_path):
                 "initializers": {
                     "default": {
                         "config": {
-                            "backend_priority": [{"type": TGISBackend.backend_type}]
+                            "backend_priority": [
+                                {
+                                    "type": TGISBackend.backend_type,
+                                    "config": {
+                                        "connection": {
+                                            "hostname": "localhost:8033",
+                                        }
+                                    },
+                                }
+                            ]
                         }
                     }
                 }

pyproject.toml

@@ -20,19 +20,20 @@ dependencies = [
     "grpcio>=1.62.2", # explicitly pin grpc dependencies to a recent version to avoid pip backtracking
     "grpcio-reflection>=1.62.2",
     "grpcio-health-checking>=1.62.2",
-    "accelerate>=0.22.0",
-    "datasets>=2.4.0",
+    "accelerate>=1.3.0",
+    "datasets>=3.3.0",
     "huggingface-hub",
-    "numpy>=1.22.4",
-    "pandas>=1.5.0",
-    "scikit-learn>=1.1",
-    "scipy>=1.8.1",
-    "sentence-transformers>=3.0.0,<3.1.0",
-    "tokenizers>=0.13.3",
+    "numpy>=1.23.0",
+    "pandas>=2.2.3",
+    "scikit-learn>=1.6.1",
+    "scipy>=1.10.0",
+    "sentence-transformers>=3.4.0,<3.5.0",
+    "tf-keras>=2.18.0",
+    "tokenizers>=0.20.0",
     "torch>=2.3.1,<2.6.0",
-    "tqdm>=4.65.0",
-    "transformers>=4.38.0,<4.44.0",
-    "peft==0.6.0",
+    "tqdm>=4.67.0",
+    "transformers>=4.48.3,<4.50.0",
+    "peft==0.14.0",
 ]
 
 [tool.setuptools.packages.find]

tests/modules/text_embedding/test_crossencoder.py

@@ -507,7 +507,6 @@ def test_truncation(truncate_input_tokens, loaded_model):
         queries_scores = [x.score for x in r.scores]
         query_scores = [x.score for x in query_results[i].scores]
         assert np.array_equal(queries_scores, query_scores)
-
         # To compare scores based on the inputs, we need to use the index too
         indexed_query_scores = {s.index: s.score for s in query_results[i].scores}
 

tests/modules/text_generation/test_peft_prompt_tuning.py

@@ -89,7 +89,6 @@ def test_run_stream_out_model(causal_lm_dummy_model):
     pred_stream = causal_lm_dummy_model.run_stream_out("This text doesn't matter")
     assert isinstance(pred_stream, Iterable)
     for pred in pred_stream:
-        print(pred)
         assert isinstance(pred, GeneratedTextStreamResult)
 
 

tests/modules/text_generation/test_text_generation_local.py

@@ -160,7 +160,7 @@ def test_train_model_causallm(disable_wip, set_cpu_device):
 
 
 @pytest.mark.skipif(platform.processor() == "arm", reason="ARM training not supported")
-def test_train_model_causallm(disable_wip, set_cpu_device):
+def test_train_and_infer_model_causallm(disable_wip, set_cpu_device):
     """Ensure that we can finetune a causal-lm model on some toy data for 1+
     steps & run inference."""
     train_kwargs = {