kaiko-eva 0.3.2__py3-none-any.whl → 0.4.0__py3-none-any.whl

eva/language/models/networks/alibaba.py

@@ -0,0 +1,26 @@
+"""Models from Alibaba."""
+
+import torch
+
+from eva.language.models import wrappers
+from eva.language.models.networks.registry import model_registry
+
+
+@model_registry.register("alibaba/qwen2-0-5b-instruct")
+class Qwen205BInstruct(wrappers.HuggingFaceModel):
+    """Qwen2 0.5B Instruct model."""
+
+    def __init__(self, system_prompt: str | None = None, cache_dir: str | None = None):
+        """Initialize the model."""
+        super().__init__(
+            model_name_or_path="Qwen/Qwen2-0.5B-Instruct",
+            model_kwargs={
+                "torch_dtype": torch.bfloat16,
+                "cache_dir": cache_dir,
+            },
+            generation_kwargs={
+                "max_new_tokens": 512,
+            },
+            system_prompt=system_prompt,
+            chat_mode=True,
+        )

eva/language/models/networks/api/__init__.py

@@ -0,0 +1,11 @@
+"""Multimodal API networks."""
+
+from eva.language.models.networks.api.anthropic import (
+    Claude35Sonnet20240620,
+    Claude37Sonnet20250219,
+)
+
+__all__ = [
+    "Claude35Sonnet20240620",
+    "Claude37Sonnet20250219",
+]

eva/language/models/networks/api/anthropic.py

@@ -0,0 +1,34 @@
+"""Models from Anthropic."""
+
+import os
+
+from eva.language.models import wrappers
+from eva.language.models.networks.registry import model_registry
+
+
+class _Claude(wrappers.LiteLLMModel):
+    """Base class for Claude models."""
+
+    def __init__(self, model_name: str, system_prompt: str | None = None):
+        if not os.getenv("ANTHROPIC_API_KEY"):
+            raise ValueError("ANTHROPIC_API_KEY env variable must be set.")
+
+        super().__init__(model_name=model_name, system_prompt=system_prompt)
+
+
+@model_registry.register("anthropic/claude-3-5-sonnet-20240620")
+class Claude35Sonnet20240620(_Claude):
+    """Claude 3.5 Sonnet (June 2024) model."""
+
+    def __init__(self, system_prompt: str | None = None):
+        """Initialize the model."""
+        super().__init__(model_name="claude-3-5-sonnet-20240620", system_prompt=system_prompt)
+
+
+@model_registry.register("anthropic/claude-3-7-sonnet-20250219")
+class Claude37Sonnet20250219(_Claude):
+    """Claude 3.7 Sonnet (February 2025) model."""
+
+    def __init__(self, system_prompt: str | None = None):
+        """Initialize the model."""
+        super().__init__(model_name="claude-3-7-sonnet-20250219", system_prompt=system_prompt)

eva/language/models/networks/registry.py

@@ -0,0 +1,5 @@
+"""Language Model Registry."""
+
+from eva.core.utils.registry import Registry
+
+model_registry = Registry()

eva/language/models/typings.py

@@ -0,0 +1,39 @@
+"""Type definitions for language models."""
+
+from typing import Any, Dict, Generic, List, TypeVar
+
+from typing_extensions import NamedTuple
+
+from eva.language.data.messages import MessageSeries
+
+TargetType = TypeVar("TargetType")
+"""The target data type."""
+
+
+class TextBatch(NamedTuple, Generic[TargetType]):
+    """Text sample with target and metadata."""
+
+    text: List[MessageSeries]
+    """Text content."""
+
+    target: TargetType | None
+    """Target data."""
+
+    metadata: Dict[str, Any] | None
+    """Additional metadata."""
+
+
+class PredictionBatch(NamedTuple, Generic[TargetType]):
+    """Text sample with target and metadata."""
+
+    prediction: TargetType
+    """Prediction data."""
+
+    target: TargetType
+    """Target data."""
+
+    text: List[MessageSeries] | None
+    """Conversation messages that were used as input."""
+
+    metadata: Dict[str, Any] | None
+    """Additional metadata."""

eva/language/models/wrappers/__init__.py

@@ -1,11 +1,19 @@
 """Language Model Wrappers API."""
 
-from eva.language.models.wrappers.huggingface import HuggingFaceTextModel
-from eva.language.models.wrappers.litellm import LiteLLMTextModel
+from eva.language.models.wrappers.base import LanguageModel
+from eva.language.models.wrappers.from_registry import ModelFromRegistry
+from eva.language.models.wrappers.huggingface import HuggingFaceModel
+from eva.language.models.wrappers.litellm import LiteLLMModel
 
 try:
-    from eva.language.models.wrappers.vllm import VLLMTextModel
+    from eva.language.models.wrappers.vllm import VllmModel
 
-    __all__ = ["HuggingFaceTextModel", "LiteLLMTextModel", "VLLMTextModel"]
+    __all__ = [
+        "LanguageModel",
+        "HuggingFaceModel",
+        "LiteLLMModel",
+        "VllmModel",
+        "ModelFromRegistry",
+    ]
 except ImportError:
-    __all__ = ["HuggingFaceTextModel", "LiteLLMTextModel"]
+    __all__ = ["LanguageModel", "HuggingFaceModel", "LiteLLMModel", "ModelFromRegistry"]

eva/language/models/wrappers/base.py

@@ -0,0 +1,47 @@
+"""Base class for language model wrappers."""
+
+import abc
+from typing import Any, Callable, List
+
+from typing_extensions import override
+
+from eva.core.models.wrappers import base
+from eva.language.data.messages import ModelSystemMessage
+from eva.language.models.typings import TextBatch
+
+
+class LanguageModel(base.BaseModel[TextBatch, List[str]]):
+    """Base class for language models.
+
+    Classes that inherit from this should implement the following methods:
+    - `load_model`: Loads & instantiates the model.
+    - `model_forward`: Implements the forward pass of the model. For API models,
+        this can be an API call.
+    - `format_inputs`: Preprocesses and converts the input batch into the format
+        expected by the `model_forward` method.
+    """
+
+    def __init__(
+        self, system_prompt: str | None, output_transforms: Callable | None = None
+    ) -> None:
+        """Creates a new model instance.
+
+        Args:
+            system_prompt: The system prompt to use for the model (optional).
+            output_transforms: Optional transforms to apply to the output of
+                the model's forward pass.
+        """
+        super().__init__(transforms=output_transforms)
+
+        self.system_message = ModelSystemMessage(content=system_prompt) if system_prompt else None
+
+    @override
+    def forward(self, batch: TextBatch) -> List[str]:
+        """Forward pass of the model."""
+        inputs = self.format_inputs(batch)
+        return super().forward(inputs)
+
+    @abc.abstractmethod
+    def format_inputs(self, batch: TextBatch) -> Any:
+        """Converts the inputs into the format expected by the model."""
+        raise NotImplementedError

eva/language/models/wrappers/from_registry.py

@@ -0,0 +1,54 @@
+"""Vision backbone helper class."""
+
+from typing import Any, Callable, Dict, List
+
+from torch import nn
+from typing_extensions import override
+
+from eva.core.models.wrappers import base
+from eva.core.utils import factory
+from eva.language.models.networks.registry import model_registry
+from eva.language.models.typings import TextBatch
+
+
+class ModelFromRegistry(base.BaseModel[TextBatch, List[str]]):
+    """Wrapper class for vision backbone models.
+
+    This class can be used by load backbones available in eva's
+    model registry by name. New backbones can be registered by using
+    the `@backbone_registry.register(model_name)` decorator.
+    """
+
+    def __init__(
+        self,
+        model_name: str,
+        model_kwargs: Dict[str, Any] | None = None,
+        model_extra_kwargs: Dict[str, Any] | None = None,
+        transforms: Callable | None = None,
+    ) -> None:
+        """Initializes the model.
+
+        Args:
+            model_name: The name of the model to load.
+            model_kwargs: The arguments used for instantiating the model.
+            model_extra_kwargs: Extra arguments used for instantiating the model.
+            transforms: The transforms to apply to the output tensor
+                produced by the model.
+        """
+        super().__init__(transforms=transforms)
+
+        self._model_name = model_name
+        self._model_kwargs = model_kwargs or {}
+        self._model_extra_kwargs = model_extra_kwargs or {}
+
+        self.model = self.load_model()
+
+    @override
+    def load_model(self) -> nn.Module:
+        ModelFromRegistry.__name__ = self._model_name
+
+        return factory.ModuleFactory(
+            registry=model_registry,
+            name=self._model_name,
+            init_args=self._model_kwargs | self._model_extra_kwargs,
+        )

eva/language/models/wrappers/huggingface.py

@@ -1,14 +1,16 @@
 """LLM wrapper for HuggingFace `transformers` models."""
 
-from typing import Any, Dict, List, Literal
+from typing import Any, Callable, Dict, List, Literal
 
 from transformers.pipelines import pipeline
 from typing_extensions import override
 
-from eva.core.models.wrappers import base
+from eva.language.models.typings import TextBatch
+from eva.language.models.wrappers import base
+from eva.language.utils.text import messages as message_utils
 
 
-class HuggingFaceTextModel(base.BaseModel[List[str], List[str]]):
+class HuggingFaceModel(base.LanguageModel):
     """Wrapper class for loading HuggingFace `transformers` models using pipelines."""
 
     def __init__(
@@ -16,7 +18,9 @@ class HuggingFaceTextModel(base.BaseModel[List[str], List[str]]):
         model_name_or_path: str,
         task: Literal["text-generation"] = "text-generation",
         model_kwargs: Dict[str, Any] | None = None,
+        system_prompt: str | None = None,
         generation_kwargs: Dict[str, Any] | None = None,
+        chat_mode: bool = True,
     ) -> None:
         """Initializes the model.
 
@@ -26,27 +30,59 @@ class HuggingFaceTextModel(base.BaseModel[List[str], List[str]]):
                 model hub.
             task: The pipeline task. Defaults to "text-generation".
             model_kwargs: Additional arguments for configuring the pipeline.
+            system_prompt: System prompt to use.
             generation_kwargs: Additional generation parameters (temperature, max_length, etc.).
+            chat_mode: Whether the specified model expects chat style messages. If set to False
+                the model is assumed to be a standard text completion model and will expect
+                plain text string inputs.
         """
-        super().__init__()
+        super().__init__(system_prompt=system_prompt)
 
         self._model_name_or_path = model_name_or_path
         self._task = task
         self._model_kwargs = model_kwargs or {}
         self._generation_kwargs = generation_kwargs or {}
+        self._chat_mode = chat_mode
 
-        self.load_model()
+        self.model = self.load_model()
 
     @override
-    def load_model(self) -> None:
+    def load_model(self) -> Callable:
         """Loads the model as a Hugging Face pipeline."""
-        self._pipeline = pipeline(
+        return pipeline(
             task=self._task,
             model=self._model_name_or_path,
             trust_remote_code=True,
             **self._model_kwargs,
         )
 
+    @override
+    def format_inputs(self, batch: TextBatch) -> List[List[Dict[str, Any]]] | List[str]:
+        """Formats inputs for HuggingFace models.
+
+        Note: If multiple system messages are present, they will be combined
+        into a single message, given that many models only support a single
+        system prompt.
+
+        Args:
+            batch: A batch of text and image inputs.
+
+        Returns:
+            When in chat mode, returns a batch of message series following
+            OpenAI's API format {"role": "user", "content": "..."}, for non-chat
+            models returns a list of plain text strings.
+        """
+        message_batch, _, _ = TextBatch(*batch)
+        message_batch = message_utils.batch_insert_system_message(
+            message_batch, self.system_message
+        )
+        message_batch = list(map(message_utils.combine_system_messages, message_batch))
+
+        if self._chat_mode:
+            return list(map(message_utils.format_chat_message, message_batch))
+        else:
+            return list(map(message_utils.merge_message_contents, message_batch))
+
     @override
     def model_forward(self, prompts: List[str]) -> List[str]:
         """Generates text using the pipeline.
@@ -57,7 +93,7 @@ class HuggingFaceTextModel(base.BaseModel[List[str], List[str]]):
         Returns:
             The generated text as a string.
         """
-        outputs = self._pipeline(prompts, return_full_text=False, **self._generation_kwargs)
+        outputs = self.model(prompts, return_full_text=False, **self._generation_kwargs)
         if outputs is None:
             raise ValueError("Outputs from the model are None.")
         results = []

eva/language/models/wrappers/litellm.py

@@ -1,77 +1,112 @@
-"""LLM wrapper for litellm models."""
+"""LiteLLM language model wrapper."""
 
+import logging
 from typing import Any, Dict, List
 
-from litellm import batch_completion  # type: ignore
+import backoff
+import litellm
+from litellm import batch_completion
+from litellm.exceptions import (
+    APIConnectionError,
+    InternalServerError,
+    RateLimitError,
+    ServiceUnavailableError,
+    Timeout,
+)
 from loguru import logger
 from typing_extensions import override
 
-from eva.core.models.wrappers import base
+from eva.language.models.typings import TextBatch
+from eva.language.models.wrappers import base
+from eva.language.utils.text import messages as message_utils
 
+RETRYABLE_ERRORS = (
+    RateLimitError,
+    Timeout,
+    InternalServerError,
+    APIConnectionError,
+    ServiceUnavailableError,
+)
 
-class LiteLLMTextModel(base.BaseModel[List[str], List[str]]):
-    """Wrapper class for using litellm for chat-based text generation.
 
-    This wrapper uses litellm's `completion` function which accepts a list of
-    message dicts. The `forward` method converts a string prompt into a chat
-    message with a default "user" role, optionally prepends a system message,
-    and includes an API key if provided.
-    """
+class LiteLLMModel(base.LanguageModel):
+    """Wrapper class for LiteLLM language models."""
 
     def __init__(
         self,
-        model_name_or_path: str,
+        model_name: str,
         model_kwargs: Dict[str, Any] | None = None,
-    ) -> None:
-        """Initializes the litellm chat model wrapper.
+        system_prompt: str | None = None,
+        log_level: int | None = logging.INFO,
+    ):
+        """Initialize the LiteLLM Wrapper.
 
         Args:
-            model_name_or_path: The model identifier (or name) for litellm
-                (e.g.,"openai/gpt-4o" or "anthropic/claude-3-sonnet-20240229").
+            model_name: The name of the model to use.
             model_kwargs: Additional keyword arguments to pass during
                 generation (e.g., `temperature`, `max_tokens`).
+            system_prompt: The system prompt to use (optional).
+            log_level: Optional logging level for LiteLLM. Defaults to WARNING.
         """
-        super().__init__()
-        self._model_name_or_path = model_name_or_path
-        self._model_kwargs = model_kwargs or {}
-        self.load_model()
+        super().__init__(system_prompt=system_prompt)
 
-    @override
-    def load_model(self) -> None:
-        """Prepares the litellm model.
+        self.model_name = model_name
+        self.model_kwargs = model_kwargs or {}
 
-        Note:
-            litellm doesn't require an explicit loading step; models are called
-            directly during generation. This method exists for API consistency.
-        """
-        pass
+        litellm.suppress_debug_info = True
+
+        if log_level is not None:
+            logging.getLogger("LiteLLM").setLevel(log_level)
 
     @override
-    def model_forward(self, prompts: List[str]) -> List[str]:
-        """Generates text using litellm.
+    def format_inputs(self, batch: TextBatch) -> List[List[Dict[str, Any]]]:
+        """Formats inputs for LiteLLM.
 
         Args:
-            prompts: A list of prompts to be converted into a "user" message.
+            batch: A batch of text inputs.
 
         Returns:
-            A list of generated text responses. Failed generations will contain
-            error messages instead of generated text.
+            A list of messages in the following format:
+            [
+                {
+                    "role": ...
+                    "content": ...
+                },
+                ...
+            ]
         """
-        messages = [[{"role": "user", "content": prompt}] for prompt in prompts]
+        message_batch, _, _ = TextBatch(*batch)
 
-        responses = batch_completion(
-            model=self._model_name_or_path,
-            messages=messages,
-            **self._model_kwargs,
+        message_batch = message_utils.batch_insert_system_message(
+            message_batch, self.system_message
         )
+        message_batch = list(map(message_utils.combine_system_messages, message_batch))
 
-        results = []
-        for i, response in enumerate(responses):
-            if isinstance(response, Exception):
-                error_msg = f"Error generating text for prompt {i}: {response}"
-                logger.error(error_msg)
-                raise RuntimeError(error_msg)
-            else:
-                results.append(response["choices"][0]["message"]["content"])
+        return list(map(message_utils.format_chat_message, message_batch))
 
-        return results
+    @override
+    @backoff.on_exception(
+        backoff.expo,
+        RETRYABLE_ERRORS,
+        max_tries=20,
+        jitter=backoff.full_jitter,
+        on_backoff=lambda details: logger.warning(
+            f"Retrying due to {details.get('exception') or 'Unknown error'}"
+        ),
+    )
+    def model_forward(self, batch: List[List[Dict[str, Any]]]) -> List[str]:
+        """Generates output text through API calls via LiteLLM's batch completion functionality."""
+        outputs = batch_completion(model=self.model_name, messages=batch, **self.model_kwargs)
+        self._raise_exceptions(outputs)
+
+        return [
+            output["choices"][0]["message"]["content"]
+            for output in outputs
+            if output["choices"][0]["message"]["role"] == "assistant"
+        ]
+
+    def _raise_exceptions(self, outputs: list):
+        for output in outputs:
+            if isinstance(output, Exception):
+                logger.error(f"Model {self.model_name} encountered an error: {output}")
+                raise output

eva/language/models/wrappers/vllm.py

@@ -1,6 +1,6 @@
 """LLM wrapper for vLLM models."""
 
-from typing import Any, Dict, List, Sequence
+from typing import Any, Dict, List
 
 from loguru import logger
 from typing_extensions import override
@@ -11,17 +11,20 @@ try:
     from vllm.transformers_utils.tokenizer import AnyTokenizer  # type: ignore
 except ImportError as e:
     raise ImportError(
-        "vLLM is required for VLLMTextModel but not installed. "
+        "vLLM is required for VllmModel but not installed. "
         "vLLM must be installed manually as it requires CUDA and is not included in dependencies. "
         "Install with: pip install vllm "
         "Note: vLLM requires Linux with CUDA support for optimal performance. "
-        "For alternatives, consider using HuggingFaceTextModel or LiteLLMTextModel."
+        "For alternatives, consider using HuggingFaceModel or LiteLLMModel."
     ) from e
 
-from eva.core.models.wrappers import base
+from eva.language.data.messages import MessageSeries
+from eva.language.models.typings import TextBatch
+from eva.language.models.wrappers import base
+from eva.language.utils.text import messages as message_utils
 
 
-class VLLMTextModel(base.BaseModel):
+class VllmModel(base.LanguageModel):
     """Wrapper class for using vLLM for text generation.
 
     This wrapper loads a vLLM model, sets up the tokenizer and sampling
@@ -34,6 +37,7 @@ class VLLMTextModel(base.BaseModel):
         self,
         model_name_or_path: str,
         model_kwargs: Dict[str, Any] | None = None,
+        system_prompt: str | None = None,
         generation_kwargs: Dict[str, Any] | None = None,
     ) -> None:
         """Initializes the vLLM model wrapper.
@@ -44,12 +48,13 @@ class VLLMTextModel(base.BaseModel):
             model_kwargs: Arguments required to initialize the vLLM model,
                 see [link](https://github.com/vllm-project/vllm/blob/main/vllm/entrypoints/llm.py)
                 for more information.
+            system_prompt: System prompt to use.
             generation_kwargs: Arguments required to generate the output,
                 need to align with the arguments of
                 [vllm.SamplingParams](https://github.com/vllm-project/vllm/blob/main/vllm/sampling_params.py).
 
         """
-        super().__init__()
+        super().__init__(system_prompt=system_prompt)
         self._model_name_or_path = model_name_or_path
         self._model_kwargs = model_kwargs or {}
         self._generation_kwargs = generation_kwargs or {}
@@ -71,11 +76,11 @@ class VLLMTextModel(base.BaseModel):
             raise RuntimeError("Model not initialized")
         self._llm_tokenizer = self._llm_model.get_tokenizer()
 
-    def _apply_chat_template(self, prompts: Sequence[str]) -> list[TokensPrompt]:
+    def _tokenize_messages(self, messages: List[MessageSeries]) -> List[TokensPrompt]:
         """Apply chat template to the messages.
 
         Args:
-            prompts: List of raw user strings.
+            messages: List of raw user strings.
 
         Returns:
             List of encoded messages.
@@ -90,7 +95,8 @@ class VLLMTextModel(base.BaseModel):
         if not hasattr(self._llm_tokenizer, "chat_template"):
             raise ValueError("Tokenizer does not have a chat template.")
 
-        chat_messages = [[{"role": "user", "content": p}] for p in prompts]
+        chat_messages = list(map(message_utils.format_chat_message, messages))
+
         encoded_messages = self._llm_tokenizer.apply_chat_template(
             chat_messages,  # type: ignore
             tokenize=True,
@@ -131,11 +137,30 @@ class VLLMTextModel(base.BaseModel):
 
         return result
 
-    def generate(self, prompts: List[str]) -> List[str]:
+    @override
+    def format_inputs(self, batch: TextBatch) -> List[TokensPrompt]:
+        """Formats inputs for vLLM models.
+
+        Args:
+            batch: A batch of text and image inputs.
+
+        Returns:
+            List of formatted prompts.
+        """
+        message_batch, _, _ = TextBatch(*batch)
+        message_batch = message_utils.batch_insert_system_message(
+            message_batch, self.system_message
+        )
+        message_batch = list(map(message_utils.combine_system_messages, message_batch))
+
+        return self._tokenize_messages(message_batch)
+
+    @override
+    def model_forward(self, batch: List[TokensPrompt]) -> List[str]:
         """Generates text for the given prompt using the vLLM model.
 
         Args:
-            prompts: A list of string prompts for generation.
+            batch: A list encoded / tokenized messages (TokensPrompt objects).
 
         Returns:
             The generated text response.
@@ -144,6 +169,5 @@ class VLLMTextModel(base.BaseModel):
         if self._llm_model is None:
             raise RuntimeError("Model not initialized")
 
-        prompt_tokens = self._apply_chat_template(prompts)
-        outputs = self._llm_model.generate(prompt_tokens, SamplingParams(**self._generation_kwargs))
+        outputs = self._llm_model.generate(batch, SamplingParams(**self._generation_kwargs))
         return [output.outputs[0].text for output in outputs]

eva/language/utils/__init__.py

@@ -1,5 +1,6 @@
 """Language utilities and helper functions."""
 
 from eva.language.utils.str_to_int_tensor import CastStrToIntTensor
+from eva.language.utils.text.messages import format_chat_message
 
-__all__ = ["CastStrToIntTensor"]
+__all__ = ["CastStrToIntTensor", "format_chat_message"]