feat: add model architecture back to wandb tracer (#6806)

# Description

This PR adds model architecture to the `WandbTracer` from the Serialized
Run kwargs. This allows visualization of the calling parameters of an
Agent, LLM and Tool in Weights & Biases.
    1. Safely serialize the run objects to WBTraceTree model_dict
    2. Refactors the run processing logic to be more organized.

- Twitter handle: @parambharat

---------

Co-authored-by: Bharat Ramanathan <ramanathan.parameshwaran@gohuddl.com>
Co-authored-by: Bagatur <baskaryan@gmail.com>
pull/6785/head^2
Bharat Ramanathan 1 year ago committed by GitHub
parent adc96d60b6
commit be29a6287d
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

@ -1,6 +1,7 @@
"""A Tracer Implementation that records activity to Weights & Biases."""
from __future__ import annotations
import json
from typing import (
TYPE_CHECKING,
Any,
@ -8,6 +9,7 @@ from typing import (
List,
Optional,
Sequence,
Tuple,
TypedDict,
Union,
)
@ -17,7 +19,7 @@ from langchain.callbacks.tracers.schemas import Run, RunTypeEnum
if TYPE_CHECKING:
from wandb import Settings as WBSettings
from wandb.sdk.data_types import trace_tree
from wandb.sdk.data_types.trace_tree import Span
from wandb.sdk.lib.paths import StrPath
from wandb.wandb_run import Run as WBRun
@ -25,115 +27,350 @@ if TYPE_CHECKING:
PRINT_WARNINGS = True
def _convert_lc_run_to_wb_span(trace_tree: Any, run: Run) -> trace_tree.Span:
if run.run_type == RunTypeEnum.llm:
return _convert_llm_run_to_wb_span(trace_tree, run)
elif run.run_type == RunTypeEnum.chain:
return _convert_chain_run_to_wb_span(trace_tree, run)
elif run.run_type == RunTypeEnum.tool:
return _convert_tool_run_to_wb_span(trace_tree, run)
def _serialize_inputs(run_inputs: dict) -> dict:
if "input_documents" in run_inputs:
docs = run_inputs["input_documents"]
return {f"input_document_{i}": doc.json() for i, doc in enumerate(docs)}
else:
return _convert_run_to_wb_span(trace_tree, run)
return run_inputs
def _convert_llm_run_to_wb_span(trace_tree: Any, run: Run) -> trace_tree.Span:
base_span = _convert_run_to_wb_span(trace_tree, run)
class RunProcessor:
"""Handles the conversion of a LangChain Runs into a WBTraceTree."""
base_span.results = [
trace_tree.Result(
inputs={"prompt": prompt},
outputs={
f"gen_{g_i}": gen["text"]
for g_i, gen in enumerate(run.outputs["generations"][ndx])
}
if (
run.outputs is not None
and len(run.outputs["generations"]) > ndx
and len(run.outputs["generations"][ndx]) > 0
def __init__(self, wandb_module: Any, trace_module: Any):
self.wandb = wandb_module
self.trace_tree = trace_module
def process_span(self, run: Run) -> Optional["Span"]:
"""Converts a LangChain Run into a W&B Trace Span.
:param run: The LangChain Run to convert.
:return: The converted W&B Trace Span.
"""
try:
span = self._convert_lc_run_to_wb_span(run)
return span
except Exception as e:
if PRINT_WARNINGS:
self.wandb.termwarn(
f"Skipping trace saving - unable to safely convert LangChain Run "
f"into W&B Trace due to: {e}"
)
return None
def _convert_run_to_wb_span(self, run: Run) -> "Span":
"""Base utility to create a span from a run.
:param run: The run to convert.
:return: The converted Span.
"""
attributes = {**run.extra} if run.extra else {}
attributes["execution_order"] = run.execution_order
return self.trace_tree.Span(
span_id=str(run.id) if run.id is not None else None,
name=run.name,
start_time_ms=int(run.start_time.timestamp() * 1000),
end_time_ms=int(run.end_time.timestamp() * 1000),
status_code=self.trace_tree.StatusCode.SUCCESS
if run.error is None
else self.trace_tree.StatusCode.ERROR,
status_message=run.error,
attributes=attributes,
)
def _convert_llm_run_to_wb_span(self, run: Run) -> "Span":
"""Converts a LangChain LLM Run into a W&B Trace Span.
:param run: The LangChain LLM Run to convert.
:return: The converted W&B Trace Span.
"""
base_span = self._convert_run_to_wb_span(run)
if base_span.attributes is None:
base_span.attributes = {}
base_span.attributes["llm_output"] = run.outputs.get("llm_output", {})
base_span.results = [
self.trace_tree.Result(
inputs={"prompt": prompt},
outputs={
f"gen_{g_i}": gen["text"]
for g_i, gen in enumerate(run.outputs["generations"][ndx])
}
if (
run.outputs is not None
and len(run.outputs["generations"]) > ndx
and len(run.outputs["generations"][ndx]) > 0
)
else None,
)
for ndx, prompt in enumerate(run.inputs["prompts"] or [])
]
base_span.span_kind = self.trace_tree.SpanKind.LLM
return base_span
def _convert_chain_run_to_wb_span(self, run: Run) -> "Span":
"""Converts a LangChain Chain Run into a W&B Trace Span.
:param run: The LangChain Chain Run to convert.
:return: The converted W&B Trace Span.
"""
base_span = self._convert_run_to_wb_span(run)
base_span.results = [
self.trace_tree.Result(
inputs=_serialize_inputs(run.inputs), outputs=run.outputs
)
else None,
]
base_span.child_spans = [
self._convert_lc_run_to_wb_span(child_run) for child_run in run.child_runs
]
base_span.span_kind = (
self.trace_tree.SpanKind.AGENT
if "agent" in run.name.lower()
else self.trace_tree.SpanKind.CHAIN
)
for ndx, prompt in enumerate(run.inputs["prompts"] or [])
]
base_span.span_kind = trace_tree.SpanKind.LLM
return base_span
return base_span
def _convert_tool_run_to_wb_span(self, run: Run) -> "Span":
"""Converts a LangChain Tool Run into a W&B Trace Span.
:param run: The LangChain Tool Run to convert.
:return: The converted W&B Trace Span.
"""
base_span = self._convert_run_to_wb_span(run)
base_span.results = [
self.trace_tree.Result(
inputs=_serialize_inputs(run.inputs), outputs=run.outputs
)
]
base_span.child_spans = [
self._convert_lc_run_to_wb_span(child_run) for child_run in run.child_runs
]
base_span.span_kind = self.trace_tree.SpanKind.TOOL
return base_span
def _convert_lc_run_to_wb_span(self, run: Run) -> "Span":
"""Utility to convert any generic LangChain Run into a W&B Trace Span.
:param run: The LangChain Run to convert.
:return: The converted W&B Trace Span.
"""
if run.run_type == RunTypeEnum.llm:
return self._convert_llm_run_to_wb_span(run)
elif run.run_type == RunTypeEnum.chain:
return self._convert_chain_run_to_wb_span(run)
elif run.run_type == RunTypeEnum.tool:
return self._convert_tool_run_to_wb_span(run)
else:
return self._convert_run_to_wb_span(run)
def process_model(self, run: Run) -> Optional[Dict[str, Any]]:
"""Utility to process a run for wandb model_dict serialization.
:param run: The run to process.
:return: The convert model_dict to pass to WBTraceTree.
"""
try:
data = json.loads(run.json())
processed = self.flatten_run(data)
keep_keys = (
"id",
"name",
"serialized",
"inputs",
"outputs",
"parent_run_id",
"execution_order",
)
processed = self.truncate_run_iterative(processed, keep_keys=keep_keys)
exact_keys, partial_keys = ("lc", "type"), ("api_key",)
processed = self.modify_serialized_iterative(
processed, exact_keys=exact_keys, partial_keys=partial_keys
)
output = self.build_tree(processed)
return output
except Exception as e:
if PRINT_WARNINGS:
self.wandb.termwarn(f"WARNING: Failed to serialize model: {e}")
return None
def flatten_run(self, run: Dict[str, Any]) -> List[Dict[str, Any]]:
"""Utility to flatten a nest run object into a list of runs.
:param run: The base run to flatten.
:return: The flattened list of runs.
"""
def flatten(child_runs: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
"""Utility to recursively flatten a list of child runs in a run.
:param child_runs: The list of child runs to flatten.
:return: The flattened list of runs.
"""
if child_runs is None:
return []
result = []
for item in child_runs:
child_runs = item.pop("child_runs", [])
result.append(item)
result.extend(flatten(child_runs))
return result
return flatten([run])
def truncate_run_iterative(
self, runs: List[Dict[str, Any]], keep_keys: Tuple[str, ...] = ()
) -> List[Dict[str, Any]]:
"""Utility to truncate a list of runs dictionaries to only keep the specified
keys in each run.
:param runs: The list of runs to truncate.
:param keep_keys: The keys to keep in each run.
:return: The truncated list of runs.
"""
def truncate_single(run: Dict[str, Any]) -> Dict[str, Any]:
"""Utility to truncate a single run dictionary to only keep the specified
keys.
:param run: The run dictionary to truncate.
:return: The truncated run dictionary
"""
new_dict = {}
for key in run:
if key in keep_keys:
new_dict[key] = run.get(key)
return new_dict
return list(map(truncate_single, runs))
def modify_serialized_iterative(
self,
runs: List[Dict[str, Any]],
exact_keys: Tuple[str, ...] = (),
partial_keys: Tuple[str, ...] = (),
) -> List[Dict[str, Any]]:
"""Utility to modify the serialized field of a list of runs dictionaries.
removes any keys that match the exact_keys and any keys that contain any of the
partial_keys.
recursively moves the dictionaries under the kwargs key to the top level.
changes the "id" field to a string "_kind" field that tells WBTraceTree how to
visualize the run. promotes the "serialized" field to the top level.
:param runs: The list of runs to modify.
:param exact_keys: A tuple of keys to remove from the serialized field.
:param partial_keys: A tuple of partial keys to remove from the serialized
field.
:return: The modified list of runs.
"""
def _serialize_inputs(run_inputs: dict) -> Union[dict, list]:
if "input_documents" in run_inputs:
docs = run_inputs["input_documents"]
return [doc.json() for doc in docs]
else:
return run_inputs
def remove_exact_and_partial_keys(obj: Dict[str, Any]) -> Dict[str, Any]:
"""Recursively removes exact and partial keys from a dictionary.
:param obj: The dictionary to remove keys from.
:return: The modified dictionary.
"""
if isinstance(obj, dict):
obj = {
k: v
for k, v in obj.items()
if k not in exact_keys
and not any(partial in k for partial in partial_keys)
}
for k, v in obj.items():
obj[k] = remove_exact_and_partial_keys(v)
elif isinstance(obj, list):
obj = [remove_exact_and_partial_keys(x) for x in obj]
return obj
def handle_id_and_kwargs(
obj: Dict[str, Any], root: bool = False
) -> Dict[str, Any]:
"""Recursively handles the id and kwargs fields of a dictionary.
changes the id field to a string "_kind" field that tells WBTraceTree how
to visualize the run. recursively moves the dictionaries under the kwargs
key to the top level.
:param obj: a run dictionary with id and kwargs fields.
:param root: whether this is the root dictionary or the serialized
dictionary.
:return: The modified dictionary.
"""
if isinstance(obj, dict):
if ("id" in obj or "name" in obj) and not root:
_kind = obj.get("id")
if not _kind:
_kind = [obj.get("name")]
obj["_kind"] = _kind[-1]
obj.pop("id", None)
obj.pop("name", None)
if "kwargs" in obj:
kwargs = obj.pop("kwargs")
for k, v in kwargs.items():
obj[k] = v
for k, v in obj.items():
obj[k] = handle_id_and_kwargs(v)
elif isinstance(obj, list):
obj = [handle_id_and_kwargs(x) for x in obj]
return obj
def transform_serialized(serialized: Dict[str, Any]) -> Dict[str, Any]:
"""Transforms the serialized field of a run dictionary to be compatible
with WBTraceTree.
:param serialized: The serialized field of a run dictionary.
:return: The transformed serialized field.
"""
serialized = handle_id_and_kwargs(serialized, root=True)
serialized = remove_exact_and_partial_keys(serialized)
return serialized
def transform_run(run: Dict[str, Any]) -> Dict[str, Any]:
"""Transforms a run dictionary to be compatible with WBTraceTree.
:param run: The run dictionary to transform.
:return: The transformed run dictionary.
"""
transformed_dict = transform_serialized(run)
serialized = transformed_dict.pop("serialized")
for k, v in serialized.items():
transformed_dict[k] = v
_kind = transformed_dict.get("_kind", None)
name = transformed_dict.pop("name", None)
exec_ord = transformed_dict.pop("execution_order", None)
if not name:
name = _kind
output_dict = {
f"{exec_ord}_{name}": transformed_dict,
}
return output_dict
return list(map(transform_run, runs))
def _convert_chain_run_to_wb_span(trace_tree: Any, run: Run) -> trace_tree.Span:
base_span = _convert_run_to_wb_span(trace_tree, run)
base_span.results = [
trace_tree.Result(inputs=_serialize_inputs(run.inputs), outputs=run.outputs)
]
base_span.child_spans = [
_convert_lc_run_to_wb_span(trace_tree, child_run)
for child_run in run.child_runs
]
base_span.span_kind = (
trace_tree.SpanKind.AGENT
if "agent" in run.serialized.get("name", "").lower()
else trace_tree.SpanKind.CHAIN
)
return base_span
def _convert_tool_run_to_wb_span(trace_tree: Any, run: Run) -> trace_tree.Span:
base_span = _convert_run_to_wb_span(trace_tree, run)
base_span.results = [
trace_tree.Result(inputs=_serialize_inputs(run.inputs), outputs=run.outputs)
]
base_span.child_spans = [
_convert_lc_run_to_wb_span(trace_tree, child_run)
for child_run in run.child_runs
]
base_span.span_kind = trace_tree.SpanKind.TOOL
return base_span
def _convert_run_to_wb_span(trace_tree: Any, run: Run) -> trace_tree.Span:
attributes = {**run.extra} if run.extra else {}
attributes["execution_order"] = run.execution_order
return trace_tree.Span(
span_id=str(run.id) if run.id is not None else None,
name=run.serialized.get("name"),
start_time_ms=int(run.start_time.timestamp() * 1000),
end_time_ms=int(run.end_time.timestamp() * 1000),
status_code=trace_tree.StatusCode.SUCCESS
if run.error is None
else trace_tree.StatusCode.ERROR,
status_message=run.error,
attributes=attributes,
)
def _replace_type_with_kind(data: Any) -> Any:
if isinstance(data, dict):
# W&B TraceTree expects "_kind" instead of "_type" since `_type` is special
# in W&B.
if "_type" in data:
_type = data.pop("_type")
data["_kind"] = _type
return {k: _replace_type_with_kind(v) for k, v in data.items()}
elif isinstance(data, list):
return [_replace_type_with_kind(v) for v in data]
elif isinstance(data, tuple):
return tuple(_replace_type_with_kind(v) for v in data)
elif isinstance(data, set):
return {_replace_type_with_kind(v) for v in data}
else:
return data
def build_tree(self, runs: List[Dict[str, Any]]) -> Dict[str, Any]:
"""Builds a nested dictionary from a list of runs.
:param runs: The list of runs to build the tree from.
:return: The nested dictionary representing the langchain Run in a tree
structure compatible with WBTraceTree.
"""
id_to_data = {}
child_to_parent = {}
for entity in runs:
for key, data in entity.items():
id_val = data.pop("id", None)
parent_run_id = data.pop("parent_run_id", None)
id_to_data[id_val] = {key: data}
if parent_run_id:
child_to_parent[id_val] = parent_run_id
for child_id, parent_id in child_to_parent.items():
parent_dict = id_to_data[parent_id]
parent_dict[next(iter(parent_dict))][
next(iter(id_to_data[child_id]))
] = id_to_data[child_id][next(iter(id_to_data[child_id]))]
root_dict = next(
data for id_val, data in id_to_data.items() if id_val not in child_to_parent
)
return root_dict
class WandbRunArgs(TypedDict):
@ -201,12 +438,13 @@ class WandbTracer(BaseTracer):
except ImportError as e:
raise ImportError(
"Could not import wandb python package."
"Please install it with `pip install wandb`."
"Please install it with `pip install -U wandb`."
) from e
self._wandb = wandb
self._trace_tree = trace_tree
self._run_args = run_args
self._ensure_run(should_print_url=(wandb.run is None))
self.run_processor = RunProcessor(self._wandb, self._trace_tree)
def finish(self) -> None:
"""Waits for all asynchronous processes to finish and data to upload.
@ -219,23 +457,11 @@ class WandbTracer(BaseTracer):
"""Logs a LangChain Run to W*B as a W&B Trace."""
self._ensure_run()
try:
root_span = _convert_lc_run_to_wb_span(self._trace_tree, run)
except Exception as e:
if PRINT_WARNINGS:
self._wandb.termwarn(
f"Skipping trace saving - unable to safely convert LangChain Run "
f"into W&B Trace due to: {e}"
)
return
model_dict = None
root_span = self.run_processor.process_span(run)
model_dict = self.run_processor.process_model(run)
# TODO: Add something like this once we have a way to get the clean serialized
# parent dict from a run:
# serialized_parent = safely_get_span_producing_model(run)
# if serialized_parent is not None:
# model_dict = safely_convert_model_to_dict(serialized_parent)
if root_span is None:
return
model_trace = self._trace_tree.WBTraceTree(
root_span=root_span,

Loading…
Cancel
Save