File size: 8,164 Bytes
9c6594c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 |
import datetime
import io
import logging
from dataclasses import asdict, dataclass
from typing import Any, Dict, List, Optional, Sequence
import wandb
from wandb.sdk.data_types import trace_tree
from wandb.sdk.integration_utils.auto_logging import Response
logger = logging.getLogger(__name__)
@dataclass
class UsageMetrics:
elapsed_time: float = None
prompt_tokens: int = None
completion_tokens: int = None
total_tokens: int = None
@dataclass
class Metrics:
usage: UsageMetrics = None
stats: wandb.Table = None
trace: trace_tree.WBTraceTree = None
usage_metric_keys = {f"usage/{k}" for k in asdict(UsageMetrics())}
class OpenAIRequestResponseResolver:
def __init__(self):
self.define_metrics_called = False
def __call__(
self,
args: Sequence[Any],
kwargs: Dict[str, Any],
response: Response,
start_time: float, # pass to comply with the protocol, but use response["created"] instead
time_elapsed: float,
) -> Optional[Dict[str, Any]]:
request = kwargs
if not self.define_metrics_called:
# define metrics on first call
for key in usage_metric_keys:
wandb.define_metric(key, step_metric="_timestamp")
self.define_metrics_called = True
try:
if response.get("object") == "edit":
return self._resolve_edit(request, response, time_elapsed)
elif response.get("object") == "text_completion":
return self._resolve_completion(request, response, time_elapsed)
elif response.get("object") == "chat.completion":
return self._resolve_chat_completion(request, response, time_elapsed)
else:
# todo: properly treat failed requests
logger.info(
f"Unsupported OpenAI response object: {response.get('object')}"
)
except Exception as e:
logger.warning(f"Failed to resolve request/response: {e}")
return None
@staticmethod
def results_to_trace_tree(
request: Dict[str, Any],
response: Response,
results: List[trace_tree.Result],
time_elapsed: float,
) -> trace_tree.WBTraceTree:
"""Converts the request, response, and results into a trace tree.
params:
request: The request dictionary
response: The response object
results: A list of results object
time_elapsed: The time elapsed in seconds
returns:
A wandb trace tree object.
"""
start_time_ms = int(round(response["created"] * 1000))
end_time_ms = start_time_ms + int(round(time_elapsed * 1000))
span = trace_tree.Span(
name=f"{response.get('model', 'openai')}_{response['object']}_{response.get('created')}",
attributes=dict(response), # type: ignore
start_time_ms=start_time_ms,
end_time_ms=end_time_ms,
span_kind=trace_tree.SpanKind.LLM,
results=results,
)
model_obj = {"request": request, "response": response, "_kind": "openai"}
return trace_tree.WBTraceTree(root_span=span, model_dict=model_obj)
def _resolve_edit(
self,
request: Dict[str, Any],
response: Response,
time_elapsed: float,
) -> Dict[str, Any]:
"""Resolves the request and response objects for `openai.Edit`."""
request_str = (
f"\n\n**Instruction**: {request['instruction']}\n\n"
f"**Input**: {request['input']}\n"
)
choices = [
f"\n\n**Edited**: {choice['text']}\n" for choice in response["choices"]
]
return self._resolve_metrics(
request=request,
response=response,
request_str=request_str,
choices=choices,
time_elapsed=time_elapsed,
)
def _resolve_completion(
self,
request: Dict[str, Any],
response: Response,
time_elapsed: float,
) -> Dict[str, Any]:
"""Resolves the request and response objects for `openai.Completion`."""
request_str = f"\n\n**Prompt**: {request['prompt']}\n"
choices = [
f"\n\n**Completion**: {choice['text']}\n" for choice in response["choices"]
]
return self._resolve_metrics(
request=request,
response=response,
request_str=request_str,
choices=choices,
time_elapsed=time_elapsed,
)
def _resolve_chat_completion(
self,
request: Dict[str, Any],
response: Response,
time_elapsed: float,
) -> Dict[str, Any]:
"""Resolves the request and response objects for `openai.Completion`."""
prompt = io.StringIO()
for message in request["messages"]:
prompt.write(f"\n\n**{message['role']}**: {message['content']}\n")
request_str = prompt.getvalue()
choices = [
f"\n\n**{choice['message']['role']}**: {choice['message']['content']}\n"
for choice in response["choices"]
]
return self._resolve_metrics(
request=request,
response=response,
request_str=request_str,
choices=choices,
time_elapsed=time_elapsed,
)
def _resolve_metrics(
self,
request: Dict[str, Any],
response: Response,
request_str: str,
choices: List[str],
time_elapsed: float,
) -> Dict[str, Any]:
"""Resolves the request and response objects for `openai.Completion`."""
results = [
trace_tree.Result(
inputs={"request": request_str},
outputs={"response": choice},
)
for choice in choices
]
metrics = self._get_metrics_to_log(request, response, results, time_elapsed)
return self._convert_metrics_to_dict(metrics)
@staticmethod
def _get_usage_metrics(response: Response, time_elapsed: float) -> UsageMetrics:
"""Gets the usage stats from the response object."""
if response.get("usage"):
usage_stats = UsageMetrics(**response["usage"])
else:
usage_stats = UsageMetrics()
usage_stats.elapsed_time = time_elapsed
return usage_stats
def _get_metrics_to_log(
self,
request: Dict[str, Any],
response: Response,
results: List[Any],
time_elapsed: float,
) -> Metrics:
model = response.get("model") or request.get("model")
usage_metrics = self._get_usage_metrics(response, time_elapsed)
usage = []
for result in results:
row = {
"request": result.inputs["request"],
"response": result.outputs["response"],
"model": model,
"start_time": datetime.datetime.fromtimestamp(response["created"]),
"end_time": datetime.datetime.fromtimestamp(
response["created"] + time_elapsed
),
"request_id": response.get("id", None),
"api_type": response.get("api_type", "openai"),
"session_id": wandb.run.id,
}
row.update(asdict(usage_metrics))
usage.append(row)
usage_table = wandb.Table(
columns=list(usage[0].keys()),
data=[(item.values()) for item in usage],
)
trace = self.results_to_trace_tree(request, response, results, time_elapsed)
metrics = Metrics(stats=usage_table, trace=trace, usage=usage_metrics)
return metrics
@staticmethod
def _convert_metrics_to_dict(metrics: Metrics) -> Dict[str, Any]:
"""Converts metrics to a dict."""
metrics_dict = {
"stats": metrics.stats,
"trace": metrics.trace,
}
usage_stats = {f"usage/{k}": v for k, v in asdict(metrics.usage).items()}
metrics_dict.update(usage_stats)
return metrics_dict
|