Merge pull request #165 from georgian-io/bumps-20240507

[Version Bumps] May 7, 2024

Author: benjaminye

llmtune/config.yml

@@ -38,14 +38,16 @@ lora:
   r: 32
   lora_alpha: 64
   lora_dropout: 0.1
-  target_modules:
-    - q_proj
-    - v_proj
-    - k_proj
-    - o_proj
-    - up_proj
-    - down_proj
-    - gate_proj
+  target_modules: "all-linear"
+  # to target specific modules
+  # target_modules:
+  #   - q_proj
+  #   - v_proj
+  #   - k_proj
+  #   - o_proj
+  #   - up_proj
+  #   - down_proj
+  #   - gate_proj
 
 # Training -------------------
 training:

llmtune/pydantic_models/config_model.py

@@ -125,7 +125,9 @@ class LoraConfig(BaseModel):
     lora_alpha: Optional[int] = Field(16, description="The alpha parameter for Lora scaling")
     bias: Optional[str] = Field("none", description="Bias type for Lora. Can be 'none', 'all' or 'lora_only'")
     lora_dropout: Optional[float] = Field(0.1, description="The dropout probability for Lora layers")
-    target_modules: Optional[List[str]] = Field(None, description="The names of the modules to apply Lora to")
+    target_modules: Optional[Union[List[str], Literal["all-linear"]]] = Field(
+        "all-linear", description="The names of the modules to apply Lora to"
+    )
     fan_in_fan_out: Optional[bool] = Field(
         False,
         description="Flag to indicate if the layer to replace stores weight like (fan_in, fan_out)",

poetry.lock

@@ -32,7 +32,7 @@ folders = [
 
 [tool.poetry.dependencies]
 python = ">=3.9, <=3.12"
-transformers = "~4.37.2"
+transformers = "~4.40.2"
 datasets = "^2.17.0"
 peft = "^0.8.2"
 pandas = "^2.2.0"
@@ -44,7 +44,7 @@ einops = "^0.7.0"
 bitsandbytes = "^0.42.0"
 nltk = "^3.8.1"
 accelerate = "^0.27.0"
-trl = "~0.7.10"
+trl = "~0.8.6"
 rouge-score = "^0.1.2"
 absl-py = "^2.1.0"
 py7zr = "^0.20.8"

pyproject.toml

@@ -1,5 +1,7 @@
+from unittest.mock import MagicMock, mock_open
+
 import pytest
-from unittest.mock import patch, MagicMock, mock_open
+from datasets import Dataset
 
 from llmtune.data.ingestor import (
     CsvIngestor,
@@ -9,8 +11,6 @@
     get_ingestor,
 )
 
-from datasets import Dataset
-
 
 def test_get_ingestor():
     assert isinstance(get_ingestor("json")(""), JsonIngestor)
@@ -31,9 +31,7 @@ def test_json_ingestor_to_dataset(mocker):
 
 
 def test_jsonl_ingestor_to_dataset(mocker):
-    mock_generator = mocker.patch(
-        "llmtune.data.ingestor.JsonlIngestor._jsonl_generator"
-    )
+    mock_generator = mocker.patch("llmtune.data.ingestor.JsonlIngestor._jsonl_generator")
     mock_dataset = mocker.patch("llmtune.data.ingestor.Dataset")
     JsonlIngestor("").to_dataset()
 
@@ -52,9 +50,7 @@ def test_huggingface_to_dataset(mocker):
     # Setup
     path = "some_path"
     ingestor = HuggingfaceIngestor(path)
-    mock_concatenate_datasets = mocker.patch(
-        "llmtune.data.ingestor.concatenate_datasets"
-    )
+    mock_concatenate_datasets = mocker.patch("llmtune.data.ingestor.concatenate_datasets")
     mock_load_dataset = mocker.patch("llmtune.data.ingestor.load_dataset")
     mock_dataset = mocker.patch("llmtune.data.ingestor.Dataset")
 
@@ -108,9 +104,7 @@ def test_jsonl_ingestor_generator(file_content, expected_output, mocker):
     mocker.patch("builtins.open", mock_open(read_data=file_content))
     mocker.patch(
         "ijson.items",
-        side_effect=lambda f, prefix, multiple_values: (
-            iter(expected_output) if multiple_values else iter([])
-        ),
+        side_effect=lambda f, prefix, multiple_values: (iter(expected_output) if multiple_values else iter([])),
     )
     ingestor = JsonlIngestor("dummy_path.jsonl")
 

tests/data/test_ingestor.py

@@ -18,15 +18,11 @@ def test_lora_finetune_initialization(mocker):
     )
 
     # Initialize LoRAFinetune with the sample configuration
-    lora_finetune = LoRAFinetune(
-        config=get_sample_config(), directory_helper=MagicMock()
-    )
+    lora_finetune = LoRAFinetune(config=get_sample_config(), directory_helper=MagicMock())
     # Assertions to ensure that LoRAFinetune is initialized as expected
     mock_lora_config.assert_called_once_with(**get_sample_config().lora.model_dump())
 
-    assert (
-        lora_finetune.config == get_sample_config()
-    ), "Configuration should match the input configuration"
+    assert lora_finetune.config == get_sample_config(), "Configuration should match the input configuration"
 
 
 def test_model_and_tokenizer_loading(mocker):
@@ -37,9 +33,7 @@ def test_model_and_tokenizer_loading(mocker):
         "llmtune.finetune.lora.AutoModelForCausalLM.from_pretrained",
         return_value=MagicMock(),
     )
-    mock_tokenizer = mocker.patch(
-        "llmtune.finetune.lora.AutoTokenizer.from_pretrained", return_value=MagicMock()
-    )
+    mock_tokenizer = mocker.patch("llmtune.finetune.lora.AutoTokenizer.from_pretrained", return_value=MagicMock())
     mock_inject_lora = mocker.patch(
         "llmtune.finetune.lora.LoRAFinetune._inject_lora",
         return_value=None,  # _inject_lora doesn't return a value
@@ -89,18 +83,14 @@ def test_model_finetune(mocker):
         "llmtune.finetune.lora.AutoModelForCausalLM.from_pretrained",
         return_value=MagicMock(),
     )
-    mocker.patch(
-        "llmtune.finetune.lora.AutoTokenizer.from_pretrained", return_value=MagicMock()
-    )
+    mocker.patch("llmtune.finetune.lora.AutoTokenizer.from_pretrained", return_value=MagicMock())
     mocker.patch(
         "llmtune.finetune.lora.LoRAFinetune._inject_lora",
         return_value=None,  # _inject_lora doesn't return a value
     )
 
     mock_trainer = mocker.MagicMock()
-    mock_sft_trainer = mocker.patch(
-        "llmtune.finetune.lora.SFTTrainer", return_value=mock_trainer
-    )
+    mock_sft_trainer = mocker.patch("llmtune.finetune.lora.SFTTrainer", return_value=mock_trainer)
 
     directory_helper = MagicMock()
 
@@ -144,9 +134,7 @@ def test_save_model(mocker):
     mocker.patch("llmtune.finetune.lora.AutoModelForCausalLM.from_pretrained")
 
     mock_tok = mocker.MagicMock()
-    mocker.patch(
-        "llmtune.finetune.lora.AutoTokenizer.from_pretrained", return_value=mock_tok
-    )
+    mocker.patch("llmtune.finetune.lora.AutoTokenizer.from_pretrained", return_value=mock_tok)
     mocker.patch(
         "llmtune.finetune.lora.LoRAFinetune._inject_lora",
         return_value=None,

tests/finetune/test_finetune_lora.py

@@ -1,37 +1,31 @@
-import pytest
 from unittest.mock import MagicMock
+
 from datasets import Dataset
+from transformers import BitsAndBytesConfig
 
 from llmtune.inference.lora import LoRAInference
-from llmtune.utils.save_utils import DirectoryHelper
 from test_utils.test_config import get_sample_config  # Adjust import path as needed
 
-from transformers import BitsAndBytesConfig
-
 
 def test_lora_inference_initialization(mocker):
     # Mock dependencies
     mock_model = mocker.patch(
         "llmtune.inference.lora.AutoPeftModelForCausalLM.from_pretrained",
         return_value=MagicMock(),
     )
-    mock_tokenizer = mocker.patch(
-        "llmtune.inference.lora.AutoTokenizer.from_pretrained", return_value=MagicMock()
-    )
+    mock_tokenizer = mocker.patch("llmtune.inference.lora.AutoTokenizer.from_pretrained", return_value=MagicMock())
 
     # Mock configuration and directory helper
     config = get_sample_config()
-    dir_helper = MagicMock(
-        save_paths=MagicMock(results="results_dir", weights="weights_dir")
-    )
+    dir_helper = MagicMock(save_paths=MagicMock(results="results_dir", weights="weights_dir"))
     test_dataset = Dataset.from_dict(
         {
             "formatted_prompt": ["prompt1", "prompt2"],
             "label_column_name": ["label1", "label2"],
         }
     )
 
-    inference = LoRAInference(
+    _ = LoRAInference(
         test_dataset=test_dataset,
         label_column_name="label_column_name",
         config=config,
@@ -45,34 +39,24 @@ def test_lora_inference_initialization(mocker):
         device_map=config.model.device_map,
         attn_implementation=config.model.attn_implementation,
     )
-    mock_tokenizer.assert_called_once_with(
-        "weights_dir", device_map=config.model.device_map
-    )
+    mock_tokenizer.assert_called_once_with("weights_dir", device_map=config.model.device_map)
 
 
 def test_infer_all(mocker):
     mocker.patch(
         "llmtune.inference.lora.AutoPeftModelForCausalLM.from_pretrained",
         return_value=MagicMock(),
     )
-    mocker.patch(
-        "llmtune.inference.lora.AutoTokenizer.from_pretrained", return_value=MagicMock()
-    )
+    mocker.patch("llmtune.inference.lora.AutoTokenizer.from_pretrained", return_value=MagicMock())
     mocker.patch("os.makedirs")
     mock_open = mocker.patch("builtins.open", mocker.mock_open())
     mock_csv_writer = mocker.patch("csv.writer")
 
-    mock_infer_one = mocker.patch.object(
-        LoRAInference, "infer_one", return_value="predicted"
-    )
+    mock_infer_one = mocker.patch.object(LoRAInference, "infer_one", return_value="predicted")
 
     config = get_sample_config()
-    dir_helper = MagicMock(
-        save_paths=MagicMock(results="results_dir", weights="weights_dir")
-    )
-    test_dataset = Dataset.from_dict(
-        {"formatted_prompt": ["prompt1"], "label_column_name": ["label1"]}
-    )
+    dir_helper = MagicMock(save_paths=MagicMock(results="results_dir", weights="weights_dir"))
+    test_dataset = Dataset.from_dict({"formatted_prompt": ["prompt1"], "label_column_name": ["label1"]})
 
     inference = LoRAInference(
         test_dataset=test_dataset,

tests/inference/test_inference_lora.py

@@ -1,7 +1,7 @@
 import pytest
-from unittest.mock import Mock
 from pandas import DataFrame
-from llmtune.qa.generics import LLMQaTest, QaTestRegistry, LLMTestSuite
+
+from llmtune.qa.generics import LLMQaTest, LLMTestSuite
 
 
 @pytest.fixture

tests/qa/test_generics.py

@@ -1,14 +1,14 @@
 import pytest
+
 from llmtune.qa.qa_tests import (
-    QaTestRegistry,
-    LengthTest,
-    JaccardSimilarityTest,
+    AdjectivePercent,
     DotProductSimilarityTest,
+    JaccardSimilarityTest,
+    LengthTest,
+    NounPercent,
     RougeScoreTest,
-    WordOverlapTest,
     VerbPercent,
-    AdjectivePercent,
-    NounPercent,
+    WordOverlapTest,
 )
 
 

tests/qa/test_qa_tests.py

@@ -1,12 +1,12 @@
 import pytest
 from pydantic import BaseModel
+
 from llmtune.utils.ablation_utils import (
-    get_types_from_dict,
+    generate_permutations,
     get_annotation,
     get_model_field_type,
-    validate_and_get_ablations,
+    get_types_from_dict,
     patch_with_permutation,
-    generate_permutations,
 )