[Inference]Support sentence transformers clip

benchmark/text-generation/llama2-13b.py

@@ -1,4 +1,3 @@
-import os
 from tempfile import TemporaryDirectory
 
 from transformers import AutoTokenizer

docs/source/tutorials/sentence_transformers.mdx

@@ -15,14 +15,14 @@ limitations under the License.
 -->
 # Sentence Transformers on AWS Inferentia with Optimum Neuron
 
+## Text Models
+
 _There is a notebook version of that tutorial [here](https://github.com/huggingface/optimum-neuron/blob/main/notebooks/sentence-transformers/getting-started.ipynb)._
 
 This guide explains how to compile, load, and use [Sentence Transformers (SBERT)](https://www.sbert.net/) models on AWS Inferentia2 with Optimum Neuron, enabling efficient calculation of embeddings. Sentence Transformers are powerful models for generating sentence embeddings. You can use this Sentence Transformers to compute sentence / text embeddings for more than 100 languages. These embeddings can then be compared e.g. with cosine-similarity to find sentences with a similar meaning. This can be useful for semantic textual similarity, semantic search, or paraphrase mining.
 
-_Note: Currently only text models are supported, we are working on vision support for CLIP._
-
 
-## Convert Sentence Transformers model to AWS Inferentia2  
+### Convert Sentence Transformers model to AWS Inferentia2  
 
 First, you need to convert your Sentence Transformers model to a format compatible with AWS Inferentia2. You can compile Sentence Transformers models with Optimum Neuron using the `optimum-cli` or `NeuronModelForSentenceTransformers` class. Below you will find an example for both approaches. We have to make sure `sentence-transformers` is installed. Thats only needed for exporting the model. 
 
@@ -52,7 +52,7 @@ Here we will use the `optimum-cli` to convert the model. Similar to the `NeuronM
 optimum-cli export neuron -m BAAI/bge-small-en-v1.5 --library-name sentence_transformers --sequence_length 384 --batch_size 1 --task feature-extraction bge_emb_inf2/
 ```
 
-## Load compiled Sentence Transformers model and run inference
+### Load compiled Sentence Transformers model and run inference
 
 Once we have a compiled Sentence Transformers model, which we either exported ourselves or is available on the Hugging Face Hub, we can load it and run inference. For loading the model we can use the `NeuronModelForSentenceTransformers` class, which is an abstraction layer for the `SentenceTransformer` class. The `NeuronModelForSentenceTransformers` class will automatically pad the input to the specified `sequence_length` and run inference on AWS Inferentia2.
 
@@ -79,6 +79,83 @@ print(f"token embeddings: {token_embeddings.shape}") # torch.Size([1, 7, 384])
 print(f"sentence_embedding: {sentence_embedding.shape}") # torch.Size([1, 384])
 ```
 
-## Production Usage
+### Production Usage
 
 For deploying these models in a production environment, refer to the [Amazon SageMaker Blog](https://www.philschmid.de/inferentia2-embeddings).
+
+
+## CLIP
+
+
+### Compile CLIP for AWS Inferentia2 
+
+You can compile CLIP models with Optimum Neuron either by using the `optimum-cli` or `NeuronModelForSentenceTransformers` class. Adopt one approach that you prefer:
+
+* With the Optimum CLI
+
+```bash
+optimum-cli export neuron -m sentence-transformers/clip-ViT-B-32 --sequence_length 64 --text_batch_size 3 --image_batch_size 1 --num_channels 3 --height 224 --width 224 --task feature-extraction --library-name sentence_transformers --subfolder 0_CLIPModel clip_emb/
+```
+
+* With the `NeuronModelForSentenceTransformers` class
+
+```python
+from optimum.neuron import NeuronModelForSentenceTransformers
+
+model_id = "sentence-transformers/clip-ViT-B-32"
+
+# configs for compiling model
+input_shapes = {
+    "num_channels": 3,
+    "height": 224,
+    "width": 224,
+    "text_batch_size": 3,
+    "image_batch_size": 1,
+    "sequence_length": 64,
+}
+
+emb_model = NeuronModelForSentenceTransformers.from_pretrained(
+    model_id, subfolder="0_CLIPModel", export=True, library_name="sentence_transformers", dynamic_batch_size=False, **input_shapes
+)
+
+# Save locally or upload to the HuggingFace Hub
+save_directory = "clip_emb/"
+emb_model.save_pretrained(save_directory)
+```
+
+### Load compiled Sentence Transformers model and run inference
+
+```python
+from PIL import Image
+from sentence_transformers import util
+from transformers import CLIPProcessor
+
+from optimum.neuron import NeuronModelForSentenceTransformers
+
+save_directory = "clip_emb"
+emb_model = NeuronModelForSentenceTransformers.from_pretrained(save_directory)
+
+processor = CLIPProcessor.from_pretrained(save_directory)
+inputs = processor(
+    text=["Two dogs in the snow", 'A cat on a table', 'A picture of London at night'], images=Image.open("two_dogs_in_snow.jpg"), return_tensors="pt", padding=True
+)  
+
+outputs = emb_model(**inputs)
+
+
+# Compute cosine similarities
+cos_scores = util.cos_sim(outputs.image_embeds, outputs.text_embeds)
+print(cos_scores)
+
+# tensor([[0.3072, 0.1016, 0.1095]])
+```
+
+<Tip>
+
+**Caveat**
+
+Since compiled models with dynamic batching enabled only accept input tensors with the same batch size, we cannot set `dynamic_batch_size=True` if the input texts and images have different batch sizes. And as `NeuronModelForSentenceTransformers` class pads the inputs to the batch sizes (`text_batch_size` and `image_batch_size`) used during the compilation, you could use relatively larger batch sizes during the compilation for flexibility with the trade-off of compute.
+
+eg. if you want to encode 3 or 4 or 5 texts and 1 image, you could set `text_batch_size = 5 = max(3, 4, 5)` and `image_batch_size = 1` during the compilation.
+
+</Tip>

optimum/commands/export/neuronx.py

@@ -145,6 +145,16 @@ def parse_args_neuronx(parser: "ArgumentParser"):
         type=int,
         help=f"Batch size {doc_input}",
     )
+    input_group.add_argument(
+        "--text_batch_size",
+        type=int,
+        help=f"Batch size of the text inputs {doc_input} (Only applied for multi-modal models)",
+    )
+    input_group.add_argument(
+        "--image_batch_size",
+        type=int,
+        help=f"Batch size of the vision inputs {doc_input} (Only applied for multi-modal models)",
+    )
     input_group.add_argument(
         "--sequence_length",
         type=int,

optimum/exporters/neuron/__main__.py

@@ -37,9 +37,9 @@
     is_neuron_available,
     is_neuronx_available,
 )
+from ...neuron.utils.misc import maybe_save_preprocessors
 from ...neuron.utils.version_utils import check_compiler_compatibility_for_stable_diffusion
 from ...utils import is_diffusers_available, logging
-from ...utils.save_utils import maybe_save_preprocessors
 from ..error_utils import AtolError, OutputMatchError, ShapeError
 from ..tasks import TasksManager
 from .base import NeuronDecoderConfig
@@ -129,9 +129,11 @@ def get_input_shapes_and_config_class(task: str, args: argparse.Namespace) -> Di
 
 def normalize_sentence_transformers_input_shapes(args: argparse.Namespace) -> Dict[str, int]:
     args = vars(args) if isinstance(args, argparse.Namespace) else args
-    mandatory_axes = {"batch_size", "sequence_length"}
     if "clip" in args.get("model", "").lower():
-        mandatory_axes.update(["num_channels", "width", "height"])
+        mandatory_axes = {"text_batch_size", "image_batch_size", "sequence_length", "num_channels", "width", "height"}
+    else:
+        mandatory_axes = {"batch_size", "sequence_length"}
+
     if not mandatory_axes.issubset(set(args.keys())):
         raise AttributeError(
             f"Shape of {mandatory_axes} are mandatory for neuron compilation, while {mandatory_axes.difference(args.keys())} are not given."
@@ -237,6 +239,7 @@ def _get_submodels_and_neuron_configs(
     task: str,
     output: Path,
     library_name: Optional[str] = None,
+    subfolder: str = "",
     dynamic_batch_size: bool = False,
     model_name_or_path: Optional[Union[str, Path]] = None,
     submodels: Optional[Dict[str, Union[Path, str]]] = None,
@@ -284,7 +287,7 @@ def _get_submodels_and_neuron_configs(
         model_name = model_name.split("/")[-1] if model_name else model.config.model_type
         output_model_names = {model_name: "model.neuron"}
         models_and_neuron_configs = {model_name: (model, neuron_config)}
-        maybe_save_preprocessors(model_name_or_path, output)
+        maybe_save_preprocessors(model_name_or_path, output, src_subfolder=subfolder)
     return models_and_neuron_configs, output_model_names
 
 
@@ -425,6 +428,7 @@ def main_export(
         task=task,
         library_name=library_name,
         output=output,
+        subfolder=subfolder,
         dynamic_batch_size=dynamic_batch_size,
         model_name_or_path=model_name_or_path,
         submodels=submodels,

optimum/exporters/neuron/base.py

@@ -144,6 +144,8 @@ def __init__(
         compiler_type: Optional[str] = None,
         compiler_version: Optional[str] = None,
         batch_size: Optional[int] = None,
+        text_batch_size: Optional[int] = None,
+        image_batch_size: Optional[int] = None,
         dynamic_batch_size: bool = False,
         sequence_length: Optional[int] = None,
         num_choices: Optional[int] = None,
@@ -176,6 +178,8 @@ def __init__(
         # To avoid using **kwargs.
         axes_values = {
             "batch_size": batch_size,
+            "text_batch_size": text_batch_size,
+            "image_batch_size": image_batch_size,
             "sequence_length": sequence_length,
             "num_choices": num_choices,
             "width": width,

optimum/exporters/neuron/convert.py

@@ -167,7 +167,7 @@ def validate_model_outputs(
     input_shapes = {}
     for axis in config.mandatory_axes:
         input_shapes[axis] = getattr(config, axis)
-    if config.dynamic_batch_size is True:
+    if config.dynamic_batch_size is True and "batch_size" in input_shapes:
         input_shapes["batch_size"] *= 2
 
     # Reference outputs

optimum/exporters/neuron/model_configs.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 """Model specific Neuron configurations."""
 
-
+import copy
 from typing import TYPE_CHECKING, Dict, List
 
 import torch
@@ -23,6 +23,7 @@
 from ...utils import (
     DummyInputGenerator,
     DummySeq2SeqDecoderTextInputGenerator,
+    DummyTextInputGenerator,
     DummyTimestepInputGenerator,
     DummyVisionInputGenerator,
     NormalizedConfig,
@@ -276,7 +277,7 @@ def outputs(self) -> List[str]:
 class SentenceTransformersCLIPNeuronConfig(CLIPNeuronConfig):
     CUSTOM_MODEL_WRAPPER = SentenceTransformersCLIPNeuronWrapper
     ATOL_FOR_VALIDATION = 1e-3
-    INPUT_ARGS = ("batch_size", "sequence_length", "num_channels", "width", "height")
+    INPUT_ARGS = ("text_batch_size", "image_batch_size", "sequence_length", "num_channels", "width", "height")
 
     @property
     def outputs(self) -> List[str]:
@@ -285,6 +286,21 @@ def outputs(self) -> List[str]:
     def patch_model_for_export(self, model, dummy_inputs):
         return self.CUSTOM_MODEL_WRAPPER(model, list(dummy_inputs.keys()))
 
+    def _create_dummy_input_generator_classes(self, **kwargs) -> List["DummyInputGenerator"]:
+        for name, axis_dim in self._axes.items():
+            self._axes[name] = kwargs.pop(name, axis_dim)
+
+        self._validate_mandatory_axes()
+
+        other_axes = copy.deepcopy(self._axes)
+        text_batch_size = other_axes.pop("text_batch_size")
+        images_batch_size = other_axes.pop("image_batch_size")
+
+        return [
+            DummyTextInputGenerator(self.task, self._normalized_config, batch_size=text_batch_size, **other_axes),
+            DummyVisionInputGenerator(self.task, self._normalized_config, batch_size=images_batch_size, **other_axes),
+        ]
+
 
 @register_in_tasks_manager("unet", *["semantic-segmentation"], library_name="diffusers")
 class UNetNeuronConfig(VisionNeuronConfig):

optimum/neuron/modeling.py

@@ -218,28 +218,41 @@ def forward(
         self,
         input_ids: torch.Tensor,
         attention_mask: torch.Tensor,
+        pixel_values: Optional[torch.Tensor] = None,
         token_type_ids: Optional[torch.Tensor] = None,
         **kwargs,
     ):
-        neuron_inputs = {
-            "input_ids": input_ids,
-            "attention_mask": attention_mask,
-        }
+        model_type = self.config.neuron["model_type"]
+        neuron_inputs = {"input_ids": input_ids}
+        if pixel_values is not None:
+            neuron_inputs["pixel_values"] = pixel_values
+        neuron_inputs["attention_mask"] = (
+            attention_mask  # The input order for clip is: input_ids, pixel_values, attention_mask.
+        )
 
         with self.neuron_padding_manager(neuron_inputs) as inputs:
             outputs = self.model(*inputs)
-            # token_embeddings -> (batch_size, sequencen_len, hidden_size)
-            token_embeddings = self.remove_padding(
-                [outputs[0]], dims=[0, 1], indices=[input_ids.shape[0], input_ids.shape[1]]
-            )[
-                0
-            ]  # Remove padding on batch_size(0), and sequence_length(1)
-            # sentence_embedding -> (batch_size, hidden_size)
-            sentence_embedding = self.remove_padding([outputs[1]], dims=[0], indices=[input_ids.shape[0]])[
-                0
-            ]  # Remove padding on batch_size(0)
+            if "clip" in model_type:
+                text_embeds = self.remove_padding([outputs[0]], dims=[0], indices=[input_ids.shape[0]])[
+                    0
+                ]  # Remove padding on batch_size(0)
+                image_embeds = self.remove_padding([outputs[1]], dims=[0], indices=[pixel_values.shape[0]])[
+                    0
+                ]  # Remove padding on batch_size(0)
+                return ModelOutput(text_embeds=text_embeds, image_embeds=image_embeds)
+            else:
+                # token_embeddings -> (batch_size, sequencen_len, hidden_size)
+                token_embeddings = self.remove_padding(
+                    [outputs[0]], dims=[0, 1], indices=[input_ids.shape[0], input_ids.shape[1]]
+                )[
+                    0
+                ]  # Remove padding on batch_size(0), and sequence_length(1)
+                # sentence_embedding -> (batch_size, hidden_size)
+                sentence_embedding = self.remove_padding([outputs[1]], dims=[0], indices=[input_ids.shape[0]])[
+                    0
+                ]  # Remove padding on batch_size(0)
 
-        return ModelOutput(token_embeddings=token_embeddings, sentence_embedding=sentence_embedding)
+                return ModelOutput(token_embeddings=token_embeddings, sentence_embedding=sentence_embedding)
 
 
 MASKED_LM_EXAMPLE = r"""

optimum/neuron/modeling_base.py

@@ -31,7 +31,6 @@
 from ..exporters.neuron.model_configs import *  # noqa: F403
 from ..exporters.tasks import TasksManager
 from ..modeling_base import OptimizedModel
-from ..utils.save_utils import maybe_load_preprocessors, maybe_save_preprocessors
 from .utils import (
     NEURON_FILE_NAME,
     check_if_weights_replacable,
@@ -40,6 +39,7 @@
     store_compilation_config,
 )
 from .utils.import_utils import is_neuronx_available
+from .utils.misc import maybe_load_preprocessors, maybe_save_preprocessors
 from .utils.version_utils import check_compiler_compatibility, get_neuroncc_version, get_neuronxcc_version
 
 
@@ -280,7 +280,7 @@ def _export(
 
         input_shapes = {}
         for name in neuron_config_constructor.func.get_mandatory_axes_for_task(task):
-            static_shape = kwargs_shapes.get(name, None) or config.neuron.get("static_" + name, None)
+            static_shape = kwargs_shapes.get(name, None)
             if static_shape is None:
                 raise AttributeError(
                     f"Cannot find the value of `{name}` from arguments nor the `config`. `{name}` is mandatory"