DSPy: The framework for programming—not prompting—language models
Python
Fix chat adapter finetune (#8044)
Dilara Soylu53045805e686
diff --git a/docs/docs/tutorials/classification_finetuning/index.ipynb b/docs/docs/tutorials/classification_finetuning/index.ipynbindex 17120ed87..d7c7f4300 100644--- a/docs/docs/tutorials/classification_finetuning/index.ipynb+++ b/docs/docs/tutorials/classification_finetuning/index.ipynb@@ -12,16 +12,25 @@"\n","### Install dependencies and download data\n","\n",- "Install the latest DSPy via `pip install -U --pre dspy` and follow along. This tutorial depends on DSPy 2.6.0 (pre-release).\n",+ "Install the latest DSPy via `pip install -U dspy>=2.6.0` and follow along (or `uv pip`, if you prefer). This tutorial depends on DSPy >= 2.6.0.\n","\n","This tutorial requires a local GPU at the moment for inference, though we plan to support ollama serving for finetuned models as well.\n","\n","You will also need the following dependencies:\n",+ "1. Inference: We use SGLang to run local inference servers. You can install the latest version by following the instructions here: https://docs.sglang.ai/start/install.html\n",+ "Shared below is the most recent install command as of 04/02/2025, but we recommend that you follow the instructions in the most up to date version by navigating to the installation link.\n",+ "This ensures that the fine-tuning packages and the `sglang` package are in sync.\n",+ " ```shell\n",+ " > pip install --upgrade pip\n",+ " > pip install uv\n",+ " > uv pip install \"sglang[all]>=0.4.4.post3\" --find-links https://flashinfer.ai/whl/cu124/torch2.5/flashinfer-python\n",+ " ```\n",+ "1. Fine-tuning: We use the following packages. Note that we specify the version for the transformers package as a temporary fix to a recent issue: https://github.com/huggingface/trl/issues/2338\n",+ " ```shell\n",+ " > uv pip install -U torch transformers==4.48.3 accelerate trl peft\n",+ " ```\n","\n",- "```shell\n",- "> pip install \"sglang[all]\"; pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.4/\n",- "> pip install -U torch transformers accelerate trl peft\n",- "```"+ "We recommend using `uv` package manager to speed up the installation."]},{@@ -241,15 +250,39 @@"\n","On every question in the (unlabeled) training set, this will invoke the teacher program, which will produce reasoning and select a class. This will be traced and then constitute a training set for all modules (in this case, just the one CoT module) in the student program.\n","\n",+ "When the `compile` method is called, the `BootstrapFinetune` optimizer will use the passed teacher program (or programs, you can pass a list!) to create a training dataset.\n",+ "It will then use this training dataset to create a fine-tuned version of the LM set for the `student` program, replacing it with the trained LM.\n",+ "Note that the trained LM will be a new LM instance (the `student_lm` object we instantiated here will be untouched!)\n",+ "\n","Note: If you have labels, you can pass `metric` to the constructor of `BootstrapFinetune`. If you want to apply this in practice, you can pass `train_kwargs` to the constructor to control local LM training settings: `device`, `use_peft`, `num_train_epochs`, `per_device_train_batch_size`, `gradient_accumulation_steps`, `learning_rate`, `max_seq_length`, `packing`, `bf16`, and `output_dir`."]},+ {+ "cell_type": "code",+ "execution_count": null,+ "metadata": {},+ "outputs": [],+ "source": [+ "# Optional:\n",+ "# [1] You can set `DSPY_FINETUNEDIR` environment variable to control where the directory that will be used to store the\n",+ "# checkpoints and fine-tuning data. If this is not set, `DSPY_CACHEDIR` is used by default.\n",+ "# [2] You can set the `CUDA_VISIBLE_DEVICES` environment variable to control the GPU that will be used for fine-tuning\n",+ "# and inference. If this is not set and the default GPU that's used by HuggingFace's `transformers` library is\n",+ "# occupied, an OutOfMemoryError might be raised.\n",+ "#\n",+ "# import os\n",+ "# os.environ[\"CUDA_VISIBLE_DEVICES\"] = \"0\"\n",+ "# os.environ[\"DSPY_FINETUNEDIR\"] = \"/path/to/dir\""+ ]+ },{"cell_type": "code","execution_count": 7,"metadata": {},"outputs": [],"source": [+ "dspy.settings.experimental = True # fine-tuning is an experimental feature, so we set a flag to enable it\n",+ "\n","optimizer = dspy.BootstrapFinetune(num_threads=16) # if you *do* have labels, pass metric=your_metric here!\n","classify_ft = optimizer.compile(student_classify, teacher=teacher_classify, trainset=unlabeled_trainset)"]@@ -1088,7 +1121,7 @@"name": "python","nbconvert_exporter": "python","pygments_lexer": "ipython3",- "version": "3.11.9"+ "version": "3.11.11"}},"nbformat": 4,diff --git a/dspy/adapters/chat_adapter.py b/dspy/adapters/chat_adapter.pyindex abb381f96..98ed24aa2 100644--- a/dspy/adapters/chat_adapter.py+++ b/dspy/adapters/chat_adapter.py@@ -193,18 +193,26 @@ class ChatAdapter(Adapter):return "\n\n".join(output).strip()- # TODO(PR): Looks ok?def format_finetune_data(- self, signature: Type[Signature], demos: list[dict[str, Any]], inputs: dict[str, Any], outputs: dict[str, Any]+ self,+ signature: Type[Signature],+ demos: list[dict[str, Any]],+ inputs: dict[str, Any],+ outputs: dict[str, Any],) -> dict[str, list[Any]]:- # Get system + user messages- messages = self.format(signature, demos, inputs)-- # Add the assistant message- role = "assistant"- incomplete = False- assistant_message = self.format_turn(signature, outputs, role, incomplete)- messages.append(assistant_message)+ """+ Format the call data into finetuning data according to the OpenAI API specifications.- # Wrap the messages in a dictionary with a "messages" key+ For the chat adapter, this means formatting the data as a list of messages, where each message is a dictionary+ with a "role" and "content" key. The role can be "system", "user", or "assistant". Then, the messages are+ wrapped in a dictionary with a "messages" key.+ """+ system_user_messages = self.format( # returns a list of dicts with the keys "role" and "content"+ signature=signature, demos=demos, inputs=inputs+ )+ assistant_message_content = self.format_assistant_message_content( # returns a string, without the role+ signature=signature, outputs=outputs+ )+ assistant_message = dict(role="assistant", content=assistant_message_content)+ messages = system_user_messages + [assistant_message]return dict(messages=messages)diff --git a/dspy/clients/lm_local.py b/dspy/clients/lm_local.pyindex d74d079a6..8329db978 100644--- a/dspy/clients/lm_local.py+++ b/dspy/clients/lm_local.py@@ -31,8 +31,8 @@ class LocalProvider(Provider):import sglang # noqa: F401except ImportError:raise ImportError(- "For local model launching, please install sglang by running "- '`pip install "sglang[all]"; pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.4/`'+ "For local model launching, please install sglang."+ "Navigate to https://docs.sglang.ai/start/install.html for the latest installation instructions!")if hasattr(lm, "process"):@@ -153,6 +153,7 @@ class LocalProvider(Provider):raise ValueError("Only chat models are supported for local finetuning.")data_path = save_data(train_data)+ logger.info(f"Train data saved to {data_path}")output_dir = create_output_dir(model, data_path)default_train_kwargs = {