Add Magpie and MagpieGenerator tasks (#778)

* Move `CudaDevicePlacementMixin` to new module

* Initial work for implementing Magpie

* Simplify magpie implementation

* Remove `use_open_ai` and add `MagpieChatTemplateMixin` to
`InferenceEndpointsLLM`

* Add `MagpieChatTemplateMixin` to `vLLM`

* Add `MagpieGenerator` task

* Fix unit tests

* Fix docstrings

* Mock `HF_TOKEN` environment variable

* Fix list index out of range

* Fix `MagpieGenerator` last batch

* Add `only_instruction` attribute

* Update categories

* testing

* Worth trying

* Add examples

* Add magpie unit tests

* Fix docstring

* Update docstrings

* Apply suggestions from code review

Co-authored-by: Agus <[email protected]>

* Update to `huggingface_hub >= 0.22.0`

* Add generation with `chat_completion`

* Update `agenerate` arguments

* Update unit tests

* Fix `tools` were not being used

* Update unit tests

* Fix list of tuples instead of list of list

* Add missing docstring

* Add `chat_completion` unit tests

* Fix `GroqLLM.generate` unit test after updating `_agenerate`

---------

Co-authored-by: Agus <[email protected]>

Author: gabrielmbmb

pyproject.toml

@@ -74,7 +74,7 @@ anthropic = ["anthropic >= 0.20.0"]
 argilla = ["argilla >= 1.29.0"]
 cohere = ["cohere >= 5.2.0"]
 groq = ["groq >= 0.4.1"]
-hf-inference-endpoints = ["huggingface_hub >= 0.19.0"]
+hf-inference-endpoints = ["huggingface_hub >= 0.22.0"]
 hf-transformers = ["transformers >= 4.34.1", "torch >= 2.0.0"]
 instructor = ["instructor >= 1.2.3"]
 litellm = ["litellm >= 1.30.0"]

src/distilabel/llms/__init__.py

@@ -22,7 +22,7 @@
 from distilabel.llms.litellm import LiteLLM
 from distilabel.llms.llamacpp import LlamaCppLLM
 from distilabel.llms.mistral import MistralLLM
-from distilabel.llms.mixins import CudaDevicePlacementMixin
+from distilabel.llms.mixins.cuda_device_placement import CudaDevicePlacementMixin
 from distilabel.llms.moa import MixtureOfAgentsLLM
 from distilabel.llms.ollama import OllamaLLM
 from distilabel.llms.openai import OpenAILLM

src/distilabel/llms/azure.py

@@ -45,7 +45,7 @@ class AzureOpenAILLM(OpenAILLM):
             `None` if not set.
 
     Icon:
-        `:simple-microsoftazure:`
+        `:material-microsoft-azure:`
 
     Examples:
 

src/distilabel/llms/base.py

@@ -329,7 +329,10 @@ async def _agenerate(
             for _ in range(num_generations)
         ]
         outputs = [outputs[0] for outputs in await asyncio.gather(*tasks)]
-        return list(grouper(outputs, n=num_generations, incomplete="ignore"))
+        return [
+            list(group)
+            for group in grouper(outputs, n=num_generations, incomplete="ignore")
+        ]
 
     def generate(
         self,

src/distilabel/llms/huggingface/inference_endpoints.py

@@ -19,7 +19,8 @@
 
 from distilabel.llms.base import LLM
 from distilabel.llms.chat_templates import CHATML_TEMPLATE
-from distilabel.llms.mixins import CudaDevicePlacementMixin
+from distilabel.llms.mixins.cuda_device_placement import CudaDevicePlacementMixin
+from distilabel.llms.mixins.magpie import MagpieChatTemplateMixin
 from distilabel.llms.typing import GenerateOutput
 from distilabel.mixins.runtime_parameters import RuntimeParameter
 from distilabel.steps.tasks.typing import OutlinesStructuredOutputType, StandardInput
@@ -32,7 +33,7 @@
     from distilabel.llms.typing import HiddenState
 
 
-class TransformersLLM(LLM, CudaDevicePlacementMixin):
+class TransformersLLM(LLM, MagpieChatTemplateMixin, CudaDevicePlacementMixin):
     """Hugging Face `transformers` library LLM implementation using the text generation
     pipeline.
 
@@ -64,6 +65,12 @@ class TransformersLLM(LLM, CudaDevicePlacementMixin):
             local configuration will be used. Defaults to `None`.
         structured_output: a dictionary containing the structured output configuration or if more
             fine-grained control is needed, an instance of `OutlinesStructuredOutput`. Defaults to None.
+        use_magpie_template: a flag used to enable/disable applying the Magpie pre-query
+            template. Defaults to `False`.
+        magpie_pre_query_template: the pre-query template to be applied to the prompt or
+            sent to the LLM to generate an instruction or a follow up user message. Valid
+            values are "llama3", "qwen2" or another pre-query template provided. Defaults
+            to `None`.
 
     Icon:
         `:hugging:`
@@ -157,14 +164,25 @@ def model_name(self) -> str:
         return self.model
 
     def prepare_input(self, input: "StandardInput") -> str:
-        """Prepares the input by applying the chat template to the input, which is formatted
-        as an OpenAI conversation, and adding the generation prompt.
+        """Prepares the input (applying the chat template and tokenization) for the provided
+        input.
+
+        Args:
+            input: the input list containing chat items.
+
+        Returns:
+            The prompt to send to the LLM.
         """
-        return self._pipeline.tokenizer.apply_chat_template(  # type: ignore
-            input,  # type: ignore
-            tokenize=False,
-            add_generation_prompt=True,
+        prompt: str = (
+            self._pipeline.tokenizer.apply_chat_template(  # type: ignore
+                input,  # type: ignore
+                tokenize=False,
+                add_generation_prompt=True,
+            )
+            if input
+            else ""
         )
+        return super().apply_magpie_pre_query_template(prompt, input)
 
     @validate_call
     def generate(  # type: ignore
@@ -209,6 +227,7 @@ def generate(  # type: ignore
             do_sample=do_sample,
             num_return_sequences=num_generations,
             prefix_allowed_tokens_fn=self._prefix_allowed_tokens_fn,
+            pad_token_id=self._pipeline.tokenizer.eos_token_id,  # type: ignore
         )
         return [
             [generation["generated_text"] for generation in output]

src/distilabel/llms/huggingface/transformers.py

@@ -0,0 +1,14 @@
+# Copyright 2023-present, Argilla, Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+

src/distilabel/llms/mixins/__init__.py

@@ -0,0 +1,89 @@
+# Copyright 2023-present, Argilla, Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING, Dict, Literal, Union
+
+from pydantic import BaseModel, field_validator, model_validator
+from typing_extensions import Self
+
+if TYPE_CHECKING:
+    from distilabel.steps.tasks.typing import StandardInput
+
+MagpieAvailablePreQueryTemplates = Literal["llama3", "qwen2"]
+"""The available predefined pre-query templates."""
+
+MAGPIE_PRE_QUERY_TEMPLATES: Dict[MagpieAvailablePreQueryTemplates, str] = {
+    "llama3": "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n",
+    "qwen2": "<|im_start|>user\n",
+}
+
+
+class MagpieChatTemplateMixin(BaseModel, validate_assignment=True):
+    """A simple mixin that adds the required logic to apply the pre-query template that
+    allows to an instruct fine-tuned LLM to generate user instructions as described in
+    the paper 'Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing'.
+
+    This mixin is meant to be used in combination with the [Magpie][distilabel.steps.tasks.magpie.base.Magpie]
+    task.
+
+    Attributes:
+        use_magpie_template: a flag used to enable/disable applying the Magpie pre-query
+            template. Defaults to `False`.
+        magpie_pre_query_template: the pre-query template to be applied to the prompt or
+            sent to the LLM to generate an instruction or a follow up user message. Valid
+            values are "llama3", "qwen2" or another pre-query template provided. Defaults
+            to `None`.
+
+    References:
+        - [Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing](https://arxiv.org/abs/2406.08464)
+    """
+
+    use_magpie_template: bool = False
+    magpie_pre_query_template: Union[MagpieAvailablePreQueryTemplates, str, None] = None
+
+    @field_validator("magpie_pre_query_template")
+    @classmethod
+    def magpie_pre_query_template_validator(cls, value: str) -> str:
+        """Resolves the pre-query template alias if it exists, otherwise, returns the
+        value with no modification."""
+        if value in MAGPIE_PRE_QUERY_TEMPLATES:
+            return MAGPIE_PRE_QUERY_TEMPLATES[value]
+        return value
+
+    @model_validator(mode="after")
+    def use_magpie_template_validation(self) -> Self:
+        """Checks that there is a pre-query template set if Magpie is going to be used."""
+        if self.use_magpie_template and self.magpie_pre_query_template is None:
+            raise ValueError(
+                f"Cannot set `use_magpie_template=True` if `magpie_pre_query_template` is"
+                f" `None`. To use Magpie with `{self.__class__.__name__}` you need to set"
+                f" the `magpie_pre_query_template` attribute."
+            )
+        return self
+
+    def apply_magpie_pre_query_template(
+        self, prompt: str, input: "StandardInput"
+    ) -> str:
+        """Applies the pre-query template to the prompt if Magpie is going to be used.
+
+        Args:
+            prompt: the prompt to which the pre-query template has to be applied.
+            input: the list with the chat items that were used to generate the prompt.
+
+        Returns:
+            The prompt with the pre-query template applied if needed.
+        """
+        if not self.use_magpie_template or (input and input[-1]["role"] == "user"):
+            return prompt
+        return prompt + self.magpie_pre_query_template  # type: ignore

src/distilabel/llms/mixins.py renamed to src/distilabel/llms/mixins/cuda_device_placement.py

@@ -30,7 +30,8 @@
 
 from distilabel.llms.base import LLM
 from distilabel.llms.chat_templates import CHATML_TEMPLATE
-from distilabel.llms.mixins import CudaDevicePlacementMixin
+from distilabel.llms.mixins.cuda_device_placement import CudaDevicePlacementMixin
+from distilabel.llms.mixins.magpie import MagpieChatTemplateMixin
 from distilabel.llms.typing import GenerateOutput
 from distilabel.mixins.runtime_parameters import RuntimeParameter
 from distilabel.steps.tasks.typing import FormattedInput, OutlinesStructuredOutputType
@@ -39,11 +40,13 @@
     from transformers import PreTrainedTokenizer
     from vllm import LLM as _vLLM
 
+    from distilabel.steps.tasks.typing import StandardInput
+
 
 SamplingParams = None
 
 
-class vLLM(LLM, CudaDevicePlacementMixin):
+class vLLM(LLM, MagpieChatTemplateMixin, CudaDevicePlacementMixin):
     """`vLLM` library LLM implementation.
 
     Attributes:
@@ -75,6 +78,12 @@ class vLLM(LLM, CudaDevicePlacementMixin):
         _tokenizer: the tokenizer instance used to format the prompt before passing it to
             the `LLM`. This attribute is meant to be used internally and should not be
             accessed directly. It will be set in the `load` method.
+        use_magpie_template: a flag used to enable/disable applying the Magpie pre-query
+            template. Defaults to `False`.
+        magpie_pre_query_template: the pre-query template to be applied to the prompt or
+            sent to the LLM to generate an instruction or a follow up user message. Valid
+            values are "llama3", "qwen2" or another pre-query template provided. Defaults
+            to `None`.
 
     References:
         - https://github.com/vllm-project/vllm/blob/main/vllm/entrypoints/llm.py
@@ -213,15 +222,26 @@ def model_name(self) -> str:
         """Returns the model name used for the LLM."""
         return self.model
 
-    def prepare_input(self, input: "FormattedInput") -> str:
-        """Prepares the input by applying the chat template to the input, which is formatted
-        as an OpenAI conversation, and adding the generation prompt.
+    def prepare_input(self, input: "StandardInput") -> str:
+        """Prepares the input (applying the chat template and tokenization) for the provided
+        input.
+
+        Args:
+            input: the input list containing chat items.
+
+        Returns:
+            The prompt to send to the LLM.
         """
-        return self._tokenizer.apply_chat_template(  # type: ignore
-            input,  # type: ignore
-            tokenize=False,
-            add_generation_prompt=True,  # type: ignore
+        prompt: str = (
+            self._tokenizer.apply_chat_template(  # type: ignore
+                input,  # type: ignore
+                tokenize=False,
+                add_generation_prompt=True,  # type: ignore
+            )
+            if input
+            else ""
         )
+        return super().apply_magpie_pre_query_template(prompt, input)
 
     def _prepare_batches(
         self, inputs: List[FormattedInput]
@@ -304,14 +324,13 @@ def generate(  # type: ignore
         if extra_sampling_params is None:
             extra_sampling_params = {}
         structured_output = None
-        needs_sorting = False
 
         if isinstance(inputs[0], tuple):
             prepared_batches, sorted_indices = self._prepare_batches(inputs)
-            needs_sorting = True
         else:
             # Simulate a batch without the structured output content
             prepared_batches = [([self.prepare_input(input) for input in inputs], None)]
+            sorted_indices = None
 
         # In case we have a single structured output for the dataset, we can
         logits_processors = None
@@ -348,7 +367,7 @@ def generate(  # type: ignore
 
         # If logits_processor is set, we need to sort the outputs back to the original order
         # (would be needed only if we have multiple structured outputs in the dataset)
-        if needs_sorting:
+        if sorted_indices is not None:
             batched_outputs = _sort_batches(
                 batched_outputs, sorted_indices, num_generations=num_generations
             )