Skip to content

armgit / dspy

Public Imported from GitHub
0

DSPy: The framework for programming—not prompting—language models

Python

Merge remote-tracking branch 'private/vllm-offline' into origin/dev_finetune

Isaac Miller731f5a3db6ee
diff --cc dspy/clients/lm.py
index 5175fe2be,45a14f706..dbfaabc28
--- a/dspy/clients/lm.py
+++ b/dspy/clients/lm.py
@@@ -94,59 -60,13 +94,67 @@@ class LM
def inspect_history(self, n: int = 1):
_inspect_history(self, n)
+ def launch(self):
+ """Send a request to the provider to launch the model, if needed."""
+ if is_self_hosted_model(self.model):
+ self_hosted_model_launch(self.model, self.launch_kwargs)
+ elif is_anyscale_model(self.model):
+ anyscale_model_launch(self.model, self.launch_kwargs)
+ logger.debug(f"`LM.launch()` is called for the auto-launched model {self.model} -- no action is taken.")
+
+ def kill(self):
+ """Send a request to the provider to kill the model, if needed."""
+ if is_self_hosted_model(self.model):
+ self_hosted_model_kill(self.model, self.launch_kwargs)
+ elif is_anyscale_model(self.model):
+ anyscale_model_kill(self.model, self.launch_kwargs)
+ logger.debug(f"`LM.kill()` is called for the auto-launched model {self.model} -- no action is taken.")
+
+ def finetune(self,
+ message_completion_pairs: List[Dict[str, str]],
+ train_kwargs: Optional[Dict[str, Any]]=None,
+ launch_kwargs: Optional[Dict[str, Any]]=None,
+ cache_finetune: bool = True,
+ ) -> FinetuneJob:
+ """Start model fine-tuning, if supported."""
+ # Fine-tuning is experimental and requires the experimental flag
+ from dspy import settings as settings
+ err = "Fine-tuning is an experimental feature and requires `dspy.settings.experimental = True`."
+ assert settings.experimental, err
+
+ # Get the fine-tuning provider, if it is supported
+ try:
+ provider = _get_supported_finetune_provider(self.model)
+ except ValueError as err:
+ raise err
+
+ # Initialize the finetune job
+ FinetuneJobClass = get_provider_finetune_job_class(provider=provider)
+ finetune_job = FinetuneJobClass(
+ model=self.model,
+ message_completion_pairs=message_completion_pairs,
+ train_kwargs=train_kwargs,
+ )
+
+ # Start asyncronous training
+ executor = ThreadPoolExecutor(max_workers=1)
+ executor.submit(
+ execute_finetune_job,
+ finetune_job,
+ launch_kwargs=launch_kwargs,
+ cache_finetune=cache_finetune
+ )
+ executor.shutdown(wait=False)
+
+ return finetune_job
++
+ def copy(self, **kwargs):
+ """Returns a copy of the language model with the same parameters."""
+ kwargs = {**self.kwargs, **kwargs}
+ # model = kwargs.pop("model") or self.model
+ init_kwargs = dict(model=self.model, model_type=self.model_type, cache=self.cache, temperature=self.kwargs["temperature"], max_tokens=self.kwargs["max_tokens"])
+ init_kwargs = {**init_kwargs, **kwargs}
+ return self.__class__(**init_kwargs)
@functools.lru_cache(maxsize=None)
diff --cc dspy/teleprompt/finetune_teleprompter.py
index 421053073,bcb97e636..5f750cc6a
--- a/dspy/teleprompt/finetune_teleprompter.py
+++ b/dspy/teleprompt/finetune_teleprompter.py
@@@ -8,7 -8,7 +8,6 @@@ from dspy.primitives.program import Pro
from dspy.primitives.prediction import Prediction
from dspy.signatures.signature import signature_to_template
-
-import concurrent.futures
#-------------------------------------------------------------------------------
# Templates for the user-facing strings used by this module
#-------------------------------------------------------------------------------
@@@ -387,7 -357,7 +356,8 @@@ def bootstrap_data_for_round
# Ensure that the LM consistency is satisfied, which ensures that either (1)
# the global LM is set or (2) all the predictors have an LM set.
- program._assert_lm_consistency()
++ # TODO(isaac): Uncomment this line after the LM consistency property is
+ # program._assert_lm_consistency()
# Deepcopy the program and copy the dataset to avoid modifying the original
program = program.deepcopy()