Push logprob generation to LLMEngine (#3065)

Co-authored-by: Avnish Narayan <avnish@anyscale.com>
This commit is contained in:
Antoni Baum
2024-03-04 19:54:06 +00:00
committed by GitHub
co-authored by Avnish Narayan
parent 76e8a70476
commit 22de45235c
13 changed files with 551 additions and 331 deletions
+124 -104
View File
@@ -82,8 +82,12 @@ class OpenAIServingChat(OpenAIServing):
return self.chat_completion_stream_generator(
request, result_generator, request_id)
else:
return await self.chat_completion_full_generator(
request, raw_request, result_generator, request_id)
try:
return await self.chat_completion_full_generator(
request, raw_request, result_generator, request_id)
except ValueError as e:
# TODO: Use a vllm-specific Validation Error
return self.create_error_response(str(e))
def get_chat_request_role(self, request: ChatCompletionRequest) -> str:
if request.add_generation_prompt:
@@ -99,117 +103,133 @@ class OpenAIServingChat(OpenAIServing):
model_name = request.model
created_time = int(time.monotonic())
chunk_object_type = "chat.completion.chunk"
# Send first response for each request.n (index) with the role
role = self.get_chat_request_role(request)
for i in range(request.n):
choice_data = ChatCompletionResponseStreamChoice(
index=i,
delta=DeltaMessage(role=role),
logprobs=None,
finish_reason=None)
chunk = ChatCompletionStreamResponse(id=request_id,
object=chunk_object_type,
created=created_time,
choices=[choice_data],
model=model_name)
data = chunk.model_dump_json(exclude_unset=True)
yield f"data: {data}\n\n"
# Send response to echo the input portion of the last message
if request.echo:
last_msg_content = ""
if request.messages and isinstance(
request.messages, list) and request.messages[-1].get(
"content") and request.messages[-1].get(
"role") == role:
last_msg_content = request.messages[-1]["content"]
if last_msg_content:
for i in range(request.n):
choice_data = ChatCompletionResponseStreamChoice(
index=i,
delta=DeltaMessage(content=last_msg_content),
finish_reason=None)
chunk = ChatCompletionStreamResponse(
id=request_id,
object=chunk_object_type,
created=created_time,
choices=[choice_data],
logprobs=None,
model=model_name)
data = chunk.model_dump_json(exclude_unset=True)
yield f"data: {data}\n\n"
first_iteration = True
# Send response for each token for each request.n (index)
previous_texts = [""] * request.n
previous_num_tokens = [0] * request.n
finish_reason_sent = [False] * request.n
async for res in result_generator:
res: RequestOutput
for output in res.outputs:
i = output.index
try:
async for res in result_generator:
res: RequestOutput
# We need to do it here, because if there are exceptions in
# the result_generator, it needs to be sent as the FIRST
# response (by the try...catch).
if first_iteration:
# Send first response for each request.n (index) with the role
role = self.get_chat_request_role(request)
for i in range(request.n):
choice_data = ChatCompletionResponseStreamChoice(
index=i,
delta=DeltaMessage(role=role),
logprobs=None,
finish_reason=None)
chunk = ChatCompletionStreamResponse(
id=request_id,
object=chunk_object_type,
created=created_time,
choices=[choice_data],
model=model_name)
data = chunk.model_dump_json(exclude_unset=True)
yield f"data: {data}\n\n"
if finish_reason_sent[i]:
continue
# Send response to echo the input portion of the last message
if request.echo:
last_msg_content = ""
if request.messages and isinstance(
request.messages,
list) and request.messages[-1].get(
"content") and request.messages[-1].get(
"role") == role:
last_msg_content = request.messages[-1]["content"]
delta_token_ids = output.token_ids[previous_num_tokens[i]:]
top_logprobs = output.logprobs[
previous_num_tokens[i]:] if output.logprobs else None
if last_msg_content:
for i in range(request.n):
choice_data = ChatCompletionResponseStreamChoice(
index=i,
delta=DeltaMessage(
content=last_msg_content),
finish_reason=None)
chunk = ChatCompletionStreamResponse(
id=request_id,
object=chunk_object_type,
created=created_time,
choices=[choice_data],
logprobs=None,
model=model_name)
data = chunk.model_dump_json(
exclude_unset=True)
yield f"data: {data}\n\n"
first_iteration = False
if request.logprobs:
logprobs = self._create_logprobs(
token_ids=delta_token_ids,
top_logprobs=top_logprobs,
num_output_top_logprobs=request.logprobs,
initial_text_offset=len(previous_texts[i]),
)
else:
logprobs = None
for output in res.outputs:
i = output.index
delta_text = output.text[len(previous_texts[i]):]
previous_texts[i] = output.text
previous_num_tokens[i] = len(output.token_ids)
if output.finish_reason is None:
# Send token-by-token response for each request.n
choice_data = ChatCompletionResponseStreamChoice(
index=i,
delta=DeltaMessage(content=delta_text),
logprobs=logprobs,
finish_reason=None)
chunk = ChatCompletionStreamResponse(
id=request_id,
object=chunk_object_type,
created=created_time,
choices=[choice_data],
model=model_name)
data = chunk.model_dump_json(exclude_unset=True)
yield f"data: {data}\n\n"
else:
# Send the finish response for each request.n only once
prompt_tokens = len(res.prompt_token_ids)
final_usage = UsageInfo(
prompt_tokens=prompt_tokens,
completion_tokens=previous_num_tokens[i],
total_tokens=prompt_tokens + previous_num_tokens[i],
)
choice_data = ChatCompletionResponseStreamChoice(
index=i,
delta=DeltaMessage(content=delta_text),
logprobs=logprobs,
finish_reason=output.finish_reason)
chunk = ChatCompletionStreamResponse(
id=request_id,
object=chunk_object_type,
created=created_time,
choices=[choice_data],
model=model_name)
if final_usage is not None:
chunk.usage = final_usage
data = chunk.model_dump_json(exclude_unset=True,
exclude_none=True)
yield f"data: {data}\n\n"
finish_reason_sent[i] = True
if finish_reason_sent[i]:
continue
delta_token_ids = output.token_ids[previous_num_tokens[i]:]
top_logprobs = output.logprobs[
previous_num_tokens[i]:] if output.logprobs else None
if request.logprobs:
logprobs = self._create_logprobs(
token_ids=delta_token_ids,
top_logprobs=top_logprobs,
num_output_top_logprobs=request.logprobs,
initial_text_offset=len(previous_texts[i]),
)
else:
logprobs = None
delta_text = output.text[len(previous_texts[i]):]
previous_texts[i] = output.text
previous_num_tokens[i] = len(output.token_ids)
if output.finish_reason is None:
# Send token-by-token response for each request.n
choice_data = ChatCompletionResponseStreamChoice(
index=i,
delta=DeltaMessage(content=delta_text),
logprobs=logprobs,
finish_reason=None)
chunk = ChatCompletionStreamResponse(
id=request_id,
object=chunk_object_type,
created=created_time,
choices=[choice_data],
model=model_name)
data = chunk.model_dump_json(exclude_unset=True)
yield f"data: {data}\n\n"
else:
# Send the finish response for each request.n only once
prompt_tokens = len(res.prompt_token_ids)
final_usage = UsageInfo(
prompt_tokens=prompt_tokens,
completion_tokens=previous_num_tokens[i],
total_tokens=prompt_tokens +
previous_num_tokens[i],
)
choice_data = ChatCompletionResponseStreamChoice(
index=i,
delta=DeltaMessage(content=delta_text),
logprobs=logprobs,
finish_reason=output.finish_reason)
chunk = ChatCompletionStreamResponse(
id=request_id,
object=chunk_object_type,
created=created_time,
choices=[choice_data],
model=model_name)
if final_usage is not None:
chunk.usage = final_usage
data = chunk.model_dump_json(exclude_unset=True,
exclude_none=True)
yield f"data: {data}\n\n"
finish_reason_sent[i] = True
except ValueError as e:
# TODO: Use a vllm-specific Validation Error
data = self.create_streaming_error_response(str(e))
yield f"data: {data}\n\n"
# Send the final done message after all response.n are finished
yield "data: [DONE]\n\n"
+203 -188
View File
@@ -26,107 +26,6 @@ TypeCreateLogProbsFn = Callable[
[TypeTokenIDs, TypeTopLogProbs, Optional[int], int], LogProbs]
async def completion_stream_generator(
request: CompletionRequest,
raw_request: Request,
on_abort,
result_generator: AsyncIterator[Tuple[int, RequestOutput]],
create_logprobs_fn: TypeCreateLogProbsFn,
request_id: str,
created_time: int,
model_name: str,
num_prompts: int,
) -> AsyncGenerator[str, None]:
previous_texts = [""] * request.n * num_prompts
previous_num_tokens = [0] * request.n * num_prompts
has_echoed = [False] * request.n * num_prompts
async for prompt_idx, res in result_generator:
# Abort the request if the client disconnects.
if await raw_request.is_disconnected():
await on_abort(f"{request_id}-{prompt_idx}")
raise StopAsyncIteration()
for output in res.outputs:
i = output.index + prompt_idx * request.n
# TODO(simon): optimize the performance by avoiding full text O(n^2) sending.
if request.echo and request.max_tokens == 0:
# only return the prompt
delta_text = res.prompt
delta_token_ids = res.prompt_token_ids
top_logprobs = res.prompt_logprobs
has_echoed[i] = True
elif request.echo and request.max_tokens > 0 and not has_echoed[i]:
# echo the prompt and first token
delta_text = res.prompt + output.text
delta_token_ids = res.prompt_token_ids + output.token_ids
top_logprobs = res.prompt_logprobs + (output.logprobs or [])
has_echoed[i] = True
else:
# return just the delta
delta_text = output.text[len(previous_texts[i]):]
delta_token_ids = output.token_ids[previous_num_tokens[i]:]
top_logprobs = output.logprobs[
previous_num_tokens[i]:] if output.logprobs else None
if request.logprobs is not None:
assert top_logprobs is not None, "top_logprobs must be provided when logprobs is requested"
logprobs = create_logprobs_fn(
token_ids=delta_token_ids,
top_logprobs=top_logprobs,
num_output_top_logprobs=request.logprobs,
initial_text_offset=len(previous_texts[i]),
)
else:
logprobs = None
previous_texts[i] = output.text
previous_num_tokens[i] = len(output.token_ids)
finish_reason = output.finish_reason
response_json = CompletionStreamResponse(
id=request_id,
created=created_time,
model=model_name,
choices=[
CompletionResponseStreamChoice(
index=i,
text=delta_text,
logprobs=logprobs,
finish_reason=finish_reason,
)
]).model_dump_json()
yield f"data: {response_json}\n\n"
if output.finish_reason is not None: # return final usage
logprobs = LogProbs() if request.logprobs is not None else None
prompt_tokens = len(res.prompt_token_ids)
completion_tokens = len(output.token_ids)
final_usage = UsageInfo(
prompt_tokens=prompt_tokens,
completion_tokens=completion_tokens,
total_tokens=prompt_tokens + completion_tokens,
)
response_json = CompletionStreamResponse(
id=request_id,
created=created_time,
model=model_name,
choices=[
CompletionResponseStreamChoice(
index=i,
text="",
logprobs=logprobs,
finish_reason=output.finish_reason,
)
],
usage=final_usage,
).model_dump_json()
yield f"data: {response_json}\n\n"
yield "data: [DONE]\n\n"
def parse_prompt_format(prompt) -> Tuple[bool, list]:
# get the prompt, openai supports the following
# "a string, array of strings, array of tokens, or array of token arrays."
@@ -151,73 +50,6 @@ def parse_prompt_format(prompt) -> Tuple[bool, list]:
return prompt_is_tokens, prompts
def request_output_to_completion_response(
final_res_batch: List[RequestOutput],
request: CompletionRequest,
create_logprobs_fn: TypeCreateLogProbsFn,
request_id: str,
created_time: int,
model_name: str,
) -> CompletionResponse:
choices = []
num_prompt_tokens = 0
num_generated_tokens = 0
for final_res in final_res_batch:
assert final_res is not None
prompt_token_ids = final_res.prompt_token_ids
prompt_logprobs = final_res.prompt_logprobs
prompt_text = final_res.prompt
for output in final_res.outputs:
if request.echo and request.max_tokens == 0:
token_ids = prompt_token_ids
top_logprobs = prompt_logprobs
output_text = prompt_text
elif request.echo and request.max_tokens > 0:
token_ids = prompt_token_ids + output.token_ids
top_logprobs = prompt_logprobs + output.logprobs
output_text = prompt_text + output.text
else:
token_ids = output.token_ids
top_logprobs = output.logprobs
output_text = output.text
if request.logprobs is not None:
logprobs = create_logprobs_fn(
token_ids=token_ids,
top_logprobs=top_logprobs,
num_output_top_logprobs=request.logprobs,
)
else:
logprobs = None
choice_data = CompletionResponseChoice(
index=len(choices),
text=output_text,
logprobs=logprobs,
finish_reason=output.finish_reason,
)
choices.append(choice_data)
num_prompt_tokens += len(prompt_token_ids)
num_generated_tokens += sum(
len(output.token_ids) for output in final_res.outputs)
usage = UsageInfo(
prompt_tokens=num_prompt_tokens,
completion_tokens=num_generated_tokens,
total_tokens=num_prompt_tokens + num_generated_tokens,
)
return CompletionResponse(
id=request_id,
created=created_time,
model=model_name,
choices=choices,
usage=usage,
)
def merge_async_iterators(*iterators):
"""Merge multiple asynchronous iterators into a single iterator.
@@ -230,8 +62,11 @@ def merge_async_iterators(*iterators):
finished = [False] * len(iterators)
async def producer(i, iterator):
async for item in iterator:
await queue.put((i, item))
try:
async for item in iterator:
await queue.put((i, item))
except Exception as e:
await queue.put(e)
finished[i] = True
_tasks = [
@@ -242,6 +77,8 @@ def merge_async_iterators(*iterators):
async def consumer():
while not all(finished) or not queue.empty():
item = await queue.get()
if isinstance(item, Exception):
raise item
yield item
await asyncio.gather(*_tasks)
@@ -312,6 +149,7 @@ class OpenAIServingCompletion(OpenAIServing):
prompt_token_ids=input_ids,
lora_request=lora_request))
except ValueError as e:
# TODO: Use a vllm-specific Validation Error
return self.create_error_response(str(e))
result_generator: AsyncIterator[Tuple[
@@ -325,27 +163,28 @@ class OpenAIServingCompletion(OpenAIServing):
# Streaming response
if stream:
return completion_stream_generator(request,
raw_request,
self.engine.abort,
result_generator,
self._create_logprobs,
request_id,
created_time,
model_name,
num_prompts=len(prompts))
return self.completion_stream_generator(request,
raw_request,
result_generator,
request_id,
created_time,
model_name,
num_prompts=len(prompts))
# Non-streaming response
final_res_batch: RequestOutput = [None] * len(prompts)
async for i, res in result_generator:
if await raw_request.is_disconnected():
# Abort the request if the client disconnects.
await self.engine.abort(f"{request_id}-{i}")
return self.create_error_response("Client disconnected")
final_res_batch[i] = res
response = request_output_to_completion_response(
final_res_batch, request, self._create_logprobs, request_id,
created_time, model_name)
try:
async for i, res in result_generator:
if await raw_request.is_disconnected():
# Abort the request if the client disconnects.
await self.engine.abort(f"{request_id}-{i}")
return self.create_error_response("Client disconnected")
final_res_batch[i] = res
response = self.request_output_to_completion_response(
final_res_batch, request, request_id, created_time, model_name)
except ValueError as e:
# TODO: Use a vllm-specific Validation Error
return self.create_error_response(str(e))
# When user requests streaming but we don't stream, we still need to
# return a streaming response with a single event.
@@ -359,3 +198,179 @@ class OpenAIServingCompletion(OpenAIServing):
return fake_stream_generator()
return response
async def completion_stream_generator(
self,
request: CompletionRequest,
raw_request: Request,
result_generator: AsyncIterator[Tuple[int, RequestOutput]],
request_id: str,
created_time: int,
model_name: str,
num_prompts: int,
) -> AsyncGenerator[str, None]:
previous_texts = [""] * request.n * num_prompts
previous_num_tokens = [0] * request.n * num_prompts
has_echoed = [False] * request.n * num_prompts
try:
async for prompt_idx, res in result_generator:
# Abort the request if the client disconnects.
if await raw_request.is_disconnected():
await self.engine.abort(f"{request_id}-{prompt_idx}")
raise StopAsyncIteration()
for output in res.outputs:
i = output.index + prompt_idx * request.n
# TODO(simon): optimize the performance by avoiding full text O(n^2) sending.
if request.echo and request.max_tokens == 0:
# only return the prompt
delta_text = res.prompt
delta_token_ids = res.prompt_token_ids
top_logprobs = res.prompt_logprobs
has_echoed[i] = True
elif request.echo and request.max_tokens > 0 and not has_echoed[
i]:
# echo the prompt and first token
delta_text = res.prompt + output.text
delta_token_ids = res.prompt_token_ids + output.token_ids
top_logprobs = res.prompt_logprobs + (output.logprobs
or [])
has_echoed[i] = True
else:
# return just the delta
delta_text = output.text[len(previous_texts[i]):]
delta_token_ids = output.token_ids[
previous_num_tokens[i]:]
top_logprobs = output.logprobs[previous_num_tokens[
i]:] if output.logprobs else None
if request.logprobs is not None:
assert top_logprobs is not None, "top_logprobs must be provided when logprobs is requested"
logprobs = self._create_logprobs(
token_ids=delta_token_ids,
top_logprobs=top_logprobs,
num_output_top_logprobs=request.logprobs,
initial_text_offset=len(previous_texts[i]),
)
else:
logprobs = None
previous_texts[i] = output.text
previous_num_tokens[i] = len(output.token_ids)
finish_reason = output.finish_reason
response_json = CompletionStreamResponse(
id=request_id,
created=created_time,
model=model_name,
choices=[
CompletionResponseStreamChoice(
index=i,
text=delta_text,
logprobs=logprobs,
finish_reason=finish_reason,
)
]).model_dump_json()
yield f"data: {response_json}\n\n"
if output.finish_reason is not None: # return final usage
logprobs = LogProbs(
) if request.logprobs is not None else None
prompt_tokens = len(res.prompt_token_ids)
completion_tokens = len(output.token_ids)
final_usage = UsageInfo(
prompt_tokens=prompt_tokens,
completion_tokens=completion_tokens,
total_tokens=prompt_tokens + completion_tokens,
)
response_json = CompletionStreamResponse(
id=request_id,
created=created_time,
model=model_name,
choices=[
CompletionResponseStreamChoice(
index=i,
text="",
logprobs=logprobs,
finish_reason=output.finish_reason,
)
],
usage=final_usage,
).model_dump_json()
yield f"data: {response_json}\n\n"
except ValueError as e:
# TODO: Use a vllm-specific Validation Error
data = self.create_streaming_error_response(str(e))
print("yield", f"data: {data}\n\n")
yield f"data: {data}\n\n"
print("yield", "data: [DONE]\n\n")
yield "data: [DONE]\n\n"
def request_output_to_completion_response(
self,
final_res_batch: List[RequestOutput],
request: CompletionRequest,
request_id: str,
created_time: int,
model_name: str,
) -> CompletionResponse:
choices = []
num_prompt_tokens = 0
num_generated_tokens = 0
for final_res in final_res_batch:
assert final_res is not None
prompt_token_ids = final_res.prompt_token_ids
prompt_logprobs = final_res.prompt_logprobs
prompt_text = final_res.prompt
for output in final_res.outputs:
if request.echo and request.max_tokens == 0:
token_ids = prompt_token_ids
top_logprobs = prompt_logprobs
output_text = prompt_text
elif request.echo and request.max_tokens > 0:
token_ids = prompt_token_ids + output.token_ids
top_logprobs = prompt_logprobs + output.logprobs
output_text = prompt_text + output.text
else:
token_ids = output.token_ids
top_logprobs = output.logprobs
output_text = output.text
if request.logprobs is not None:
logprobs = self._create_logprobs(
token_ids=token_ids,
top_logprobs=top_logprobs,
num_output_top_logprobs=request.logprobs,
)
else:
logprobs = None
choice_data = CompletionResponseChoice(
index=len(choices),
text=output_text,
logprobs=logprobs,
finish_reason=output.finish_reason,
)
choices.append(choice_data)
num_prompt_tokens += len(prompt_token_ids)
num_generated_tokens += sum(
len(output.token_ids) for output in final_res.outputs)
usage = UsageInfo(
prompt_tokens=num_prompt_tokens,
completion_tokens=num_generated_tokens,
total_tokens=num_prompt_tokens + num_generated_tokens,
)
return CompletionResponse(
id=request_id,
created=created_time,
model=model_name,
choices=choices,
usage=usage,
)
+19 -4
View File
@@ -1,4 +1,5 @@
import asyncio
import json
from dataclasses import dataclass
from http import HTTPStatus
from typing import Dict, List, Optional, Union
@@ -11,6 +12,7 @@ from vllm.entrypoints.openai.protocol import (CompletionRequest,
ModelCard, ModelList,
ModelPermission)
from vllm.lora.request import LoRARequest
from vllm.sequence import Logprob
logger = init_logger(__name__)
@@ -83,7 +85,7 @@ class OpenAIServing:
def _create_logprobs(
self,
token_ids: List[int],
top_logprobs: Optional[List[Optional[Dict[int, float]]]] = None,
top_logprobs: Optional[List[Optional[Dict[int, Logprob]]]] = None,
num_output_top_logprobs: Optional[int] = None,
initial_text_offset: int = 0,
) -> LogProbs:
@@ -95,10 +97,10 @@ class OpenAIServing:
for i, token_id in enumerate(token_ids):
step_top_logprobs = top_logprobs[i]
if step_top_logprobs is not None:
token_logprob = step_top_logprobs[token_id]
token_logprob = step_top_logprobs[token_id].logprob
else:
token_logprob = None
token = self.tokenizer.convert_ids_to_tokens(token_id)
token = step_top_logprobs[token_id].decoded_token
logprobs.tokens.append(token)
logprobs.token_logprobs.append(token_logprob)
if len(logprobs.text_offset) == 0:
@@ -110,7 +112,7 @@ class OpenAIServing:
if num_output_top_logprobs:
logprobs.top_logprobs.append({
self.tokenizer.convert_ids_to_tokens(i): p
p.decoded_token: p.logprob
for i, p in step_top_logprobs.items()
} if step_top_logprobs else None)
return logprobs
@@ -124,6 +126,19 @@ class OpenAIServing:
type=err_type,
code=status_code.value)
def create_streaming_error_response(
self,
message: str,
err_type: str = "BadRequestError",
status_code: HTTPStatus = HTTPStatus.BAD_REQUEST) -> str:
json_str = json.dumps({
"error":
self.create_error_response(message=message,
err_type=err_type,
status_code=status_code).model_dump()
})
return json_str
async def _check_model(self, request) -> Optional[ErrorResponse]:
if request.model == self.served_model:
return