Skip to content

armgit / dspy

Public Imported from GitHub
0

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

Python

simplify dspy.LM (#7940)

Chen Qian6ecc92d4274c
diff --git a/dspy/clients/base_lm.py b/dspy/clients/base_lm.py
index 1cb97575e..fe1fc16ab 100644
--- a/dspy/clients/base_lm.py
+++ b/dspy/clients/base_lm.py
@@ -16,6 +16,22 @@ class BaseLM(ABC):
def __call__(self, prompt=None, messages=None, **kwargs):
pass
+ def copy(self, **kwargs):
+ """Returns a copy of the language model with possibly updated parameters."""
+
+ import copy
+
+ new_instance = copy.deepcopy(self)
+ new_instance.history = []
+
+ for key, value in kwargs.items():
+ if hasattr(self, key):
+ setattr(new_instance, key, value)
+ if (key in self.kwargs) or (not hasattr(self, key)):
+ new_instance.kwargs[key] = value
+
+ return new_instance
+
def inspect_history(self, n: int = 1):
_inspect_history(self.history, n)
diff --git a/dspy/clients/lm.py b/dspy/clients/lm.py
index c94bd5f9d..0a56d7da3 100644
--- a/dspy/clients/lm.py
+++ b/dspy/clients/lm.py
@@ -6,7 +6,7 @@ import threading
import uuid
from datetime import datetime
from hashlib import sha256
-from typing import Any, Dict, List, Literal, Optional, cast, TYPE_CHECKING
+from typing import Any, Dict, List, Literal, Optional, cast
import litellm
import pydantic
@@ -21,8 +21,6 @@ from dspy.clients.openai import OpenAIProvider
from dspy.clients.provider import Provider, TrainingJob
from dspy.clients.utils_finetune import TrainDataFormat
from dspy.utils.callback import BaseCallback, with_callbacks
-if TYPE_CHECKING:
- from dspy.adapters.base import Adapter
from .base_lm import BaseLM
@@ -142,17 +140,20 @@ class LM(BaseLM):
# Logging, with removed api key & where `cost` is None on cache hit.
kwargs = {k: v for k, v in kwargs.items() if not k.startswith("api_")}
- entry = dict(prompt=prompt, messages=messages, kwargs=kwargs, response=response)
- entry = dict(**entry, outputs=outputs, usage=dict(response["usage"]))
- entry = dict(**entry, cost=response.get("_hidden_params", {}).get("response_cost"))
- entry = dict(
- **entry,
- timestamp=datetime.now().isoformat(),
- uuid=str(uuid.uuid4()),
- model=self.model,
- response_model=response["model"],
- model_type=self.model_type,
- )
+ entry = {
+ "prompt": prompt,
+ "messages": messages,
+ "kwargs": kwargs,
+ "response": response,
+ "outputs": outputs,
+ "usage": dict(response["usage"]),
+ "cost": response.get("_hidden_params", {}).get("response_cost"),
+ "timestamp": datetime.now().isoformat(),
+ "uuid": str(uuid.uuid4()),
+ "model": self.model,
+ "response_model": response["model"],
+ "model_type": self.model_type,
+ }
self.history.append(entry)
self.update_global_history(entry)
@@ -216,38 +217,8 @@ class LM(BaseLM):
def infer_provider(self) -> Provider:
if OpenAIProvider.is_provider_model(self.model):
return OpenAIProvider()
- # TODO(PR): Keeping this function here will require us to import all
- # providers in this file. Is this okay?
return Provider()
- def infer_adapter(self) -> "Adapter":
- import dspy
-
- if dspy.settings.adapter:
- return dspy.settings.adapter
-
- model_type_to_adapter = {
- "chat": dspy.ChatAdapter(),
- }
- model_type = self.model_type
- return model_type_to_adapter[model_type]
-
- def copy(self, **kwargs):
- """Returns a copy of the language model with possibly updated parameters."""
-
- import copy
-
- new_instance = copy.deepcopy(self)
- new_instance.history = []
-
- for key, value in kwargs.items():
- if hasattr(self, key):
- setattr(new_instance, key, value)
- if (key in self.kwargs) or (not hasattr(self, key)):
- new_instance.kwargs[key] = value
-
- return new_instance
-
def dump_state(self):
state_keys = ["model", "model_type", "cache", "cache_in_memory", "num_retries", "finetuning_model", "launch_kwargs", "train_kwargs"]
return { key: getattr(self, key) for key in state_keys } | self.kwargs
diff --git a/dspy/teleprompt/bootstrap_finetune.py b/dspy/teleprompt/bootstrap_finetune.py
index 21167ac82..6e5bf569a 100644
--- a/dspy/teleprompt/bootstrap_finetune.py
+++ b/dspy/teleprompt/bootstrap_finetune.py
@@ -4,8 +4,10 @@ from typing import Any, Callable, Dict, List, Optional, Union
import dspy
from dspy.adapters.base import Adapter
+from dspy.adapters.chat_adapter import ChatAdapter
from dspy.clients.lm import LM
from dspy.clients.utils_finetune import infer_data_format
+from dspy.dsp.utils.settings import settings
from dspy.evaluate.evaluate import Evaluate
from dspy.predict.predict import Predict
from dspy.primitives.example import Example
@@ -160,7 +162,7 @@ class BootstrapFinetune(FinetuneTeleprompter):
logger.info(f"After filtering with the metric, {len(trace_data)} examples remain")
data = []
- adapter = self.adapter[lm] or lm.infer_adapter()
+ adapter = self.adapter[lm] or settings.adapter or ChatAdapter()
data_format = infer_data_format(adapter)
for item in trace_data:
for pred_ind, _ in enumerate(item["trace"]):
@@ -181,18 +183,12 @@ class BootstrapFinetune(FinetuneTeleprompter):
def build_call_data_from_trace(
trace: List[Dict],
pred_ind: int,
- adapter: Optional[Adapter] = None,
+ adapter: Adapter,
exclude_demos: bool = False,
) -> Dict[str, List[Dict[str, Any]]]:
# Find data that's relevant to the predictor
pred, inputs, outputs = trace[pred_ind] # assuming that the order is kept
- if not adapter:
- # TODO(feature): A trace is collected using a particular adapter. It
- # would be nice to get this adapter information from the trace (e.g.
- # pred.lm.adapter) as opposed to using the inference method below.
- adapter = pred.lm.infer_adapter()
-
demos = [] if exclude_demos else pred.demos
call_data = adapter.format_finetune_data(
signature=pred.signature,
@@ -209,8 +205,8 @@ def bootstrap_trace_data(
metric: Optional[Callable] = None,
num_threads=6,
) -> List[Dict[str, Any]]:
- # Return a list of dicts with the following keys:
- # example_ind, example, prediction, trace, and score (if metric != None)
+ # Return a list of dicts with the following keys: example_ind, example, prediction, trace, and score
+ # (if metric != None)
evaluator = Evaluate(
devset=dataset,
num_threads=num_threads,