Add the device_dtype init parameter to the transformers model

Author: RobinPicard

docs/features/models/transformers.md

@@ -12,10 +12,11 @@ title: Transformers
 
 ## Model Initialization
 
-To load the model, you can use the `from_transformers` function. It takes 2 arguments:
+To load the model, you can use the `from_transformers` function. It takes 3 arguments:
 
 - `model`: a `transformers` model (created with `AutoModelForCausalLM` for instance)
 - `tokenizer_or_processor`: a `transformers` tokenizer (created with `AutoTokenizer` for instance, it must be an instance of either `PreTrainedTokenizer` or `PreTrainedTokenizerFast`)
+- `device_dtype` (optional): the tensor dtype to use for inference. If not provided, the model will use the default dtype.
 
 For instance:
 

docs/features/models/transformers_multimodal.md

@@ -12,6 +12,7 @@ To load the model, you can use the `from_transformers` function. It takes 2 argu
 
 - `model`: a `transformers` model (created with `AutoModelForImageTextToText` for instance)
 - `tokenizer_or_processor`: a `transformers` processor (created with `AutoProcessor` for instance, it must be an instance of `ProcessorMixin`)
+- `device_dtype` (optional): the tensor dtype to use for inference. If not provided, the model will use the default dtype.
 
 For instance:
 

outlines/models/transformers.py

@@ -209,6 +209,8 @@ def __init__(
         self,
         model: "PreTrainedModel",
         tokenizer: "PreTrainedTokenizer",
+        *,
+        device_dtype: Optional["torch.dtype"] = None,
     ):
         """
         Parameters:
@@ -219,6 +221,9 @@ def __init__(
         tokenizer
             A `PreTrainedTokenizer`, or any tokenizer that is compatible with
             the `transformers` API for tokenizers.
+        device_dtype
+            The dtype to use for the model. If not provided, the model will use
+            the default dtype.
 
         """
         # We need to handle the cases in which jax/flax or tensorflow
@@ -237,6 +242,7 @@ def __init__(
         self.model = model
         self.hf_tokenizer = tokenizer
         self.tokenizer = TransformerTokenizer(tokenizer)
+        self.device_dtype = device_dtype
         self.type_adapter = TransformersTypeAdapter(tokenizer=tokenizer)
 
         if (
@@ -287,7 +293,11 @@ def _prepare_model_inputs(
         input_ids, attention_mask = self.tokenizer.encode(prompts)
         inputs = {
             "input_ids": input_ids.to(self.model.device),
-            "attention_mask": attention_mask.to(self.model.device),
+            "attention_mask": (
+                attention_mask.to(self.model.device, dtype=self.device_dtype)
+                if self.device_dtype is not None
+                else attention_mask.to(self.model.device)
+            ),
         }
 
         return prompts, inputs
@@ -600,7 +610,13 @@ class TransformersMultiModal(Transformers):
 
     """
 
-    def __init__(self, model: "PreTrainedModel", processor):
+    def __init__(
+        self,
+        model: "PreTrainedModel",
+        processor,
+        *,
+        device_dtype: Optional["torch.dtype"] = None,
+    ):
         """Create a TransformersMultiModal model instance
 
         We rely on the `__init__` method of the `Transformers` class to handle
@@ -614,6 +630,9 @@ def __init__(self, model: "PreTrainedModel", processor):
             `transformers` API for models.
         processor
             A `ProcessorMixin` instance.
+        device_dtype
+            The dtype to use for the model. If not provided, the model will use
+            the default dtype.
 
         """
         self.processor = processor
@@ -622,7 +641,7 @@ def __init__(self, model: "PreTrainedModel", processor):
 
         tokenizer: "PreTrainedTokenizer" = self.processor.tokenizer
 
-        super().__init__(model, tokenizer)
+        super().__init__(model, tokenizer, device_dtype=device_dtype)
 
         self.type_adapter = TransformersMultiModalTypeAdapter(
             tokenizer=tokenizer
@@ -655,14 +674,20 @@ def _prepare_model_inputs(
 
         inputs = self.processor(
             **merged_prompts, padding=True, return_tensors="pt"
-        ).to(self.model.device)
+        )
+        if self.device_dtype is not None:
+            inputs = inputs.to(self.model.device, dtype=self.device_dtype)
+        else:
+            inputs = inputs.to(self.model.device)
 
         return merged_prompts["text"], inputs
 
 
 def from_transformers(
     model: "PreTrainedModel",
     tokenizer_or_processor: Union["PreTrainedTokenizer", "ProcessorMixin"],
+    *,
+    device_dtype: Optional["torch.dtype"] = None,
 ) -> Union[Transformers, TransformersMultiModal]:
     """Create an Outlines `Transformers` or `TransformersMultiModal` model
     instance from a `PreTrainedModel` instance and a `PreTrainedTokenizer` or
@@ -679,6 +704,9 @@ def from_transformers(
     tokenizer_or_processor
         A `transformers.PreTrainedTokenizer` or
         `transformers.ProcessorMixin` instance.
+    device_dtype
+        The dtype to use for the model. If not provided, the model will use
+        the default dtype.
 
     Returns
     -------
@@ -693,10 +721,10 @@ def from_transformers(
         tokenizer_or_processor, (PreTrainedTokenizer, PreTrainedTokenizerFast)
     ):
         tokenizer = tokenizer_or_processor
-        return Transformers(model, tokenizer)
+        return Transformers(model, tokenizer, device_dtype=device_dtype)
     elif isinstance(tokenizer_or_processor, ProcessorMixin):
         processor = tokenizer_or_processor
-        return TransformersMultiModal(model, processor)
+        return TransformersMultiModal(model, processor, device_dtype=device_dtype)
     else:
         raise ValueError(
             "We could determine whether the model passed to `from_transformers`"

tests/models/test_transformers.py

@@ -3,6 +3,7 @@
 
 from pydantic import BaseModel
 import pytest
+import torch
 import transformers
 
 import outlines
@@ -47,15 +48,17 @@ def test_transformers_instantiate_mamba():
     assert isinstance(model, Transformers)
 
 
-def test_transformers_instantiate_tokenizer_kwargs():
+def test_transformers_instantiate_tokenizer_kwargs_dtype():
     model = outlines.from_transformers(
         transformers.AutoModelForCausalLM.from_pretrained(TEST_MODEL),
         transformers.AutoTokenizer.from_pretrained(
             TEST_MODEL, additional_special_tokens=["<t1>", "<t2>"]
         ),
+        device_dtype=torch.bfloat16,
     )
     assert "<t1>" in model.tokenizer.special_tokens
     assert "<t2>" in model.tokenizer.special_tokens
+    assert model.device_dtype == torch.bfloat16
 
 
 @pytest.fixture
@@ -88,6 +91,10 @@ def test_transformers_call(model, model_bart):
     result = model("Respond with one word. Not more.")
     assert isinstance(result, str)
 
+    model.device_dtype = torch.bfloat16
+    result = model("Respond with one word. Not more.")
+    assert isinstance(result, str)
+
     result = model_bart("Respond with one word. Not more.")
     assert isinstance(result, str)
 

tests/models/test_transformers_multimodal.py

@@ -2,6 +2,7 @@
 
 import io
 import re
+import torch
 from enum import Enum
 
 import pytest
@@ -47,15 +48,17 @@ def model():
     return model
 
 
-def test_transformers_multimodal_instantiate_simple():
+def test_transformers_multimodal_instantiate():
     model = outlines.from_transformers(
         LlavaForConditionalGeneration.from_pretrained(TEST_MODEL),
         AutoProcessor.from_pretrained(TEST_MODEL),
+        device_dtype=torch.bfloat16,
     )
     assert isinstance(model, TransformersMultiModal)
     assert isinstance(model.tokenizer, TransformerTokenizer)
     assert isinstance(model.type_adapter, TransformersMultiModalTypeAdapter)
     assert model.tensor_library_name == "torch"
+    assert model.device_dtype == torch.bfloat16
 
 
 def test_transformers_multimodal_simple(model, image):
@@ -74,6 +77,13 @@ def test_transformers_multimodal_call(model, image):
     )
     assert isinstance(result, str)
 
+    model.device_dtype = torch.bfloat16
+    result = model(
+        ["<image>Describe this image in one sentence:", Image(image)],
+        max_new_tokens=2,
+    )
+    assert isinstance(result, str)
+
 
 def test_transformers_multimodal_wrong_number_image(model, image):
     with pytest.raises(ValueError):