3084ef9fa1
kwargs is pretty ugly when figuring out which arguments to use. The base requests falls back to defaults anyways, so pass in the params object as is. However, since Python's typing isn't like TypeScript where types can be transformed, the type hinting has a possiblity of None showing up despite there always being a value for some params. Signed-off-by: kingbri <8082010+kingbri1@users.noreply.github.com>
504 lines
17 KiB
Python
504 lines
17 KiB
Python
"""Chat completion utilities for OAI server."""
|
|
|
|
import asyncio
|
|
import json
|
|
import pathlib
|
|
from asyncio import CancelledError
|
|
from typing import List, Optional
|
|
from fastapi import HTTPException, Request
|
|
from jinja2 import TemplateError
|
|
from loguru import logger
|
|
|
|
from common import model
|
|
from common.multimodal import MultimodalEmbeddingWrapper
|
|
from common.networking import (
|
|
get_generator_error,
|
|
handle_request_disconnect,
|
|
handle_request_error,
|
|
request_disconnect_loop,
|
|
)
|
|
from common.utils import unwrap
|
|
from endpoints.OAI.types.chat_completion import (
|
|
ChatCompletionLogprobs,
|
|
ChatCompletionLogprob,
|
|
ChatCompletionMessage,
|
|
ChatCompletionRequest,
|
|
ChatCompletionRespChoice,
|
|
ChatCompletionStreamChunk,
|
|
ChatCompletionResponse,
|
|
ChatCompletionStreamChoice,
|
|
)
|
|
from endpoints.OAI.types.common import UsageStats
|
|
from endpoints.OAI.utils.completion import _stream_collector
|
|
from endpoints.OAI.utils.tools import ToolCallProcessor
|
|
|
|
|
|
def _create_response(
|
|
request_id: str, generations: List[dict], model_name: Optional[str]
|
|
):
|
|
"""Create a chat completion response from the provided text."""
|
|
|
|
prompt_tokens = unwrap(generations[-1].get("prompt_tokens"), 0)
|
|
completion_tokens = unwrap(generations[-1].get("generated_tokens"), 0)
|
|
|
|
choices = []
|
|
for index, generation in enumerate(generations):
|
|
message = ChatCompletionMessage(
|
|
role="assistant", content=unwrap(generation.get("text"), "")
|
|
)
|
|
|
|
tool_calls = generation["tool_calls"]
|
|
if tool_calls:
|
|
message.tool_calls = ToolCallProcessor.from_json(tool_calls)
|
|
|
|
logprob_response = None
|
|
|
|
token_probs = unwrap(generation.get("token_probs"), {})
|
|
if token_probs:
|
|
logprobs = unwrap(generation.get("logprobs"), [])
|
|
|
|
collected_token_probs = []
|
|
for index, token in enumerate(token_probs.keys()):
|
|
top_logprobs = [
|
|
ChatCompletionLogprob(token=token, logprob=logprob)
|
|
for token, logprob in logprobs[index].items()
|
|
]
|
|
|
|
collected_token_probs.append(
|
|
ChatCompletionLogprob(
|
|
token=token,
|
|
logprob=token_probs[token],
|
|
top_logprobs=top_logprobs,
|
|
)
|
|
)
|
|
|
|
logprob_response = ChatCompletionLogprobs(content=collected_token_probs)
|
|
|
|
# Initialize finish_reason with a default value or from generation data
|
|
finish_reason = generation.get("finish_reason", "stop")
|
|
|
|
# If a tool call is present, mark the finish reason as such
|
|
if message.tool_calls:
|
|
finish_reason = "tool_calls"
|
|
|
|
choice = ChatCompletionRespChoice(
|
|
index=index,
|
|
finish_reason=finish_reason,
|
|
stop_str=generation.get("stop_str"),
|
|
message=message,
|
|
logprobs=logprob_response,
|
|
)
|
|
|
|
choices.append(choice)
|
|
|
|
response = ChatCompletionResponse(
|
|
id=f"chatcmpl-{request_id}",
|
|
choices=choices,
|
|
model=unwrap(model_name, ""),
|
|
usage=UsageStats(
|
|
prompt_tokens=prompt_tokens,
|
|
completion_tokens=completion_tokens,
|
|
total_tokens=prompt_tokens + completion_tokens,
|
|
),
|
|
)
|
|
|
|
return response
|
|
|
|
|
|
def _create_stream_chunk(
|
|
request_id: str,
|
|
generation: Optional[dict] = None,
|
|
model_name: Optional[str] = None,
|
|
is_usage_chunk: bool = False,
|
|
):
|
|
"""Create a chat completion stream chunk from the provided text."""
|
|
|
|
index = generation.get("index")
|
|
choices = []
|
|
usage_stats = None
|
|
|
|
if is_usage_chunk:
|
|
prompt_tokens = unwrap(generation.get("prompt_tokens"), 0)
|
|
completion_tokens = unwrap(generation.get("generated_tokens"), 0)
|
|
|
|
usage_stats = UsageStats(
|
|
prompt_tokens=prompt_tokens,
|
|
completion_tokens=completion_tokens,
|
|
total_tokens=prompt_tokens + completion_tokens,
|
|
)
|
|
elif "finish_reason" in generation:
|
|
# Get the finish reason from the generation
|
|
finish_reason = generation.get("finish_reason")
|
|
choice = ChatCompletionStreamChoice(index=index, finish_reason=finish_reason)
|
|
|
|
# lets check if we have tool calls since we are at the end of the generation
|
|
# Mark finish_reason as tool_calls since this is the last chunk
|
|
if "tool_calls" in generation:
|
|
tool_calls = generation["tool_calls"]
|
|
message = ChatCompletionMessage(
|
|
tool_calls=ToolCallProcessor.from_json(tool_calls)
|
|
)
|
|
choice.delta = message
|
|
choice.finish_reason = "tool_calls"
|
|
|
|
choices.append(choice)
|
|
|
|
else:
|
|
message = ChatCompletionMessage(
|
|
role="assistant", content=unwrap(generation.get("text"), "")
|
|
)
|
|
|
|
logprob_response = None
|
|
|
|
token_probs = unwrap(generation.get("token_probs"), {})
|
|
if token_probs:
|
|
logprobs = unwrap(generation.get("logprobs"), {})
|
|
top_logprobs = [
|
|
ChatCompletionLogprob(token=token, logprob=logprob)
|
|
for token, logprob in logprobs.items()
|
|
]
|
|
|
|
generated_token = next(iter(token_probs))
|
|
token_prob_response = ChatCompletionLogprob(
|
|
token=generated_token,
|
|
logprob=token_probs[generated_token],
|
|
top_logprobs=top_logprobs,
|
|
)
|
|
|
|
logprob_response = ChatCompletionLogprobs(content=[token_prob_response])
|
|
|
|
choice = ChatCompletionStreamChoice(
|
|
index=index,
|
|
delta=message,
|
|
logprobs=logprob_response,
|
|
)
|
|
|
|
choices.append(choice)
|
|
|
|
chunk = ChatCompletionStreamChunk(
|
|
id=f"chatcmpl-{request_id}",
|
|
choices=choices,
|
|
model=unwrap(model_name, ""),
|
|
usage=usage_stats,
|
|
)
|
|
|
|
return chunk
|
|
|
|
|
|
async def _append_template_metadata(data: ChatCompletionRequest, template_vars: dict):
|
|
"""Adding metadata is a one-time process."""
|
|
|
|
template_metadata = await model.container.prompt_template.extract_metadata(
|
|
template_vars
|
|
)
|
|
|
|
# Stop strings
|
|
if isinstance(data.stop, str):
|
|
data.stop = [data.stop] + template_metadata.stop_strings
|
|
else:
|
|
data.stop += template_metadata.stop_strings
|
|
|
|
# Tool call start strings
|
|
if template_metadata.tool_starts:
|
|
if data.tool_call_start is None:
|
|
data.tool_call_start = template_metadata.tool_starts
|
|
|
|
# Append to stop strings to halt for a tool call generation
|
|
data.stop.extend(template_metadata.tool_starts)
|
|
|
|
|
|
async def format_messages_with_template(
|
|
messages: List[ChatCompletionMessage],
|
|
existing_template_vars: Optional[dict] = None,
|
|
add_bos_token: bool = True,
|
|
ban_eos_token: bool = False,
|
|
):
|
|
"""Barebones function to format chat completion messages into a prompt."""
|
|
|
|
template_vars = unwrap(existing_template_vars, {})
|
|
mm_embeddings = MultimodalEmbeddingWrapper() if model.container.use_vision else None
|
|
|
|
for message in messages:
|
|
if isinstance(message.content, list):
|
|
concatenated_content = ""
|
|
for content in message.content:
|
|
if content.type == "text":
|
|
concatenated_content += content.text
|
|
elif content.type == "image_url" and mm_embeddings:
|
|
await mm_embeddings.add(content.image_url.url)
|
|
concatenated_content += mm_embeddings.text_alias[-1]
|
|
|
|
# Convert the message content into a concatenated string
|
|
message.content = concatenated_content
|
|
|
|
if message.tool_calls:
|
|
message.tool_calls_json = ToolCallProcessor.to_json(message.tool_calls)
|
|
|
|
# The tools variable is inspectable in the template, so
|
|
# store the list of dicts rather than the ToolCallProcessor object.
|
|
message.tool_calls = ToolCallProcessor.dump(message.tool_calls)
|
|
|
|
special_tokens_dict = model.container.get_special_tokens(
|
|
add_bos_token, ban_eos_token
|
|
)
|
|
|
|
template_vars.update({"messages": messages, **special_tokens_dict})
|
|
|
|
prompt = await model.container.prompt_template.render(template_vars)
|
|
return prompt, mm_embeddings, template_vars
|
|
|
|
|
|
async def apply_chat_template(
|
|
data: ChatCompletionRequest, tool_precursor: Optional[str] = None
|
|
):
|
|
"""
|
|
Compile the prompt and get any additional stop strings from the template.
|
|
Template stop strings can be overriden by sampler overrides if force is true.
|
|
"""
|
|
|
|
# Locally store tools dict
|
|
tools = data.model_dump()["tools"]
|
|
|
|
try:
|
|
data.template_vars.update(
|
|
{
|
|
"add_generation_prompt": data.add_generation_prompt,
|
|
"tools": tools,
|
|
"tools_json": json.dumps(tools, indent=2),
|
|
"functions": data.functions,
|
|
"functions_json": json.dumps(data.functions, indent=2),
|
|
"tool_precursor": tool_precursor,
|
|
}
|
|
)
|
|
|
|
prompt, mm_embeddings, template_vars = await format_messages_with_template(
|
|
data.messages, data.template_vars, data.add_bos_token, data.ban_eos_token
|
|
)
|
|
|
|
# Append response prefix if present
|
|
if data.response_prefix:
|
|
if data.add_generation_prompt:
|
|
prompt += data.response_prefix
|
|
else:
|
|
logger.warning(
|
|
"Could not add response prefix because "
|
|
"add_generation_prompt is False"
|
|
)
|
|
|
|
# Removes the starting BOS token if present
|
|
# This is to prevent add_bos_token from adding multiple bos tokens
|
|
bos_token = template_vars.get("bos_token")
|
|
if bos_token and prompt.startswith(bos_token):
|
|
prompt = prompt.removeprefix(bos_token)
|
|
|
|
# Add template metadata
|
|
await _append_template_metadata(data, template_vars)
|
|
|
|
return prompt, mm_embeddings
|
|
|
|
except KeyError as exc:
|
|
error_message = handle_request_error(
|
|
"Could not find a Conversation from prompt template "
|
|
f"'{model.container.prompt_template.name}'. "
|
|
"Check your spelling?",
|
|
).error.message
|
|
|
|
raise HTTPException(400, error_message) from exc
|
|
except TemplateError as exc:
|
|
error_message = handle_request_error(f"TemplateError: {str(exc)}").error.message
|
|
|
|
raise HTTPException(400, error_message) from exc
|
|
|
|
|
|
async def stream_generate_chat_completion(
|
|
prompt: str,
|
|
embeddings: MultimodalEmbeddingWrapper,
|
|
data: ChatCompletionRequest,
|
|
request: Request,
|
|
model_path: pathlib.Path,
|
|
):
|
|
"""Generator for the generation process."""
|
|
abort_event = asyncio.Event()
|
|
gen_queue = asyncio.Queue()
|
|
gen_tasks: List[asyncio.Task] = []
|
|
disconnect_task = asyncio.create_task(request_disconnect_loop(request))
|
|
|
|
try:
|
|
logger.info(f"Received chat completion streaming request {request.state.id}")
|
|
|
|
for n in range(0, data.n):
|
|
task_gen_params = data.model_copy(deep=True)
|
|
|
|
gen_task = asyncio.create_task(
|
|
_stream_collector(
|
|
n,
|
|
gen_queue,
|
|
request.state.id,
|
|
prompt,
|
|
task_gen_params,
|
|
abort_event,
|
|
mm_embeddings=embeddings,
|
|
)
|
|
)
|
|
|
|
gen_tasks.append(gen_task)
|
|
|
|
# We need to keep track of the text generated so we can resume the tool calls
|
|
current_generation_text = ""
|
|
|
|
# Consumer loop
|
|
while True:
|
|
if disconnect_task.done():
|
|
abort_event.set()
|
|
handle_request_disconnect(
|
|
f"Chat completion generation {request.state.id} cancelled by user."
|
|
)
|
|
|
|
generation = await gen_queue.get()
|
|
# lets only append the text if we need it for tool calls later
|
|
if data.tool_call_start and "text" in generation:
|
|
current_generation_text += generation["text"]
|
|
|
|
# check if we are running a tool model, and that we are at stop
|
|
if data.tool_call_start and "stop_str" in generation:
|
|
generations = await generate_tool_calls(
|
|
data,
|
|
[generation],
|
|
request,
|
|
current_generations=current_generation_text,
|
|
)
|
|
generation = generations[0] # We only have one generation in this case
|
|
|
|
# Stream collector will push an exception to the queue if it fails
|
|
if isinstance(generation, Exception):
|
|
raise generation
|
|
|
|
response = _create_stream_chunk(
|
|
request.state.id, generation, model_path.name
|
|
)
|
|
yield response.model_dump_json()
|
|
|
|
# Check if all tasks are completed
|
|
if all(task.done() for task in gen_tasks) and gen_queue.empty():
|
|
# Send a usage chunk
|
|
if data.stream_options and data.stream_options.include_usage:
|
|
usage_chunk = _create_stream_chunk(
|
|
request.state.id,
|
|
generation,
|
|
model_path.name,
|
|
is_usage_chunk=True,
|
|
)
|
|
yield usage_chunk.model_dump_json()
|
|
|
|
logger.info(
|
|
f"Finished chat completion streaming request {request.state.id}"
|
|
)
|
|
|
|
yield "[DONE]"
|
|
break
|
|
except CancelledError:
|
|
# Get out if the request gets disconnected
|
|
|
|
if not disconnect_task.done():
|
|
abort_event.set()
|
|
handle_request_disconnect("Chat completion generation cancelled by user.")
|
|
except Exception:
|
|
yield get_generator_error(
|
|
"Chat completion aborted. Please check the server console."
|
|
)
|
|
|
|
|
|
async def generate_chat_completion(
|
|
prompt: str,
|
|
embeddings: MultimodalEmbeddingWrapper,
|
|
data: ChatCompletionRequest,
|
|
request: Request,
|
|
model_path: pathlib.Path,
|
|
):
|
|
gen_tasks: List[asyncio.Task] = []
|
|
|
|
try:
|
|
for _ in range(0, data.n):
|
|
gen_tasks.append(
|
|
asyncio.create_task(
|
|
model.container.generate(
|
|
request.state.id,
|
|
prompt,
|
|
data,
|
|
mm_embeddings=embeddings,
|
|
)
|
|
)
|
|
)
|
|
|
|
generations = await asyncio.gather(*gen_tasks)
|
|
|
|
# Let's not waste our time if we arn't running a tool model
|
|
if data.tool_call_start:
|
|
generations = await generate_tool_calls(data, generations, request)
|
|
|
|
response = _create_response(request.state.id, generations, model_path.name)
|
|
|
|
logger.info(f"Finished chat completion request {request.state.id}")
|
|
|
|
return response
|
|
except Exception as exc:
|
|
error_message = handle_request_error(
|
|
f"Chat completion {request.state.id} aborted. "
|
|
"Maybe the model was unloaded? "
|
|
"Please check the server console."
|
|
).error.message
|
|
|
|
# Server error if there's a generation exception
|
|
raise HTTPException(503, error_message) from exc
|
|
|
|
|
|
async def generate_tool_calls(
|
|
data: ChatCompletionRequest,
|
|
generations: List[str],
|
|
request: Request,
|
|
current_generations: str = None,
|
|
):
|
|
gen_tasks: List[asyncio.Task] = []
|
|
tool_idx: List[int] = []
|
|
|
|
# Copy to make sure the parent JSON schema doesn't get modified
|
|
# FIXME: May not be necessary depending on how the codebase evolves
|
|
tool_data = data.model_copy(deep=True)
|
|
tool_data.json_schema = tool_data.tool_call_schema
|
|
|
|
for idx, gen in enumerate(generations):
|
|
if gen["stop_str"] in tool_data.tool_call_start:
|
|
logger.info(
|
|
f"Detected tool call in chat completion request {request.state.id}"
|
|
)
|
|
|
|
if "text" in gen:
|
|
# non streaming, all generations will have the text they generated
|
|
pre_tool_prompt, mm_embeddings = await apply_chat_template(
|
|
data, gen["text"]
|
|
)
|
|
elif current_generations is not None:
|
|
# streaming, we wont have text in the generation,
|
|
# we'll have to use the current_generations
|
|
pre_tool_prompt, mm_embeddings = await apply_chat_template(
|
|
data, current_generations
|
|
)
|
|
|
|
gen_tasks.append(
|
|
asyncio.create_task(
|
|
model.container.generate(
|
|
request.state.id,
|
|
pre_tool_prompt,
|
|
tool_data,
|
|
embeddings=mm_embeddings,
|
|
)
|
|
)
|
|
)
|
|
tool_idx.append(idx)
|
|
|
|
tool_calls = await asyncio.gather(*gen_tasks)
|
|
for outer_idx in range(0, len(tool_idx)):
|
|
gen_idx = tool_idx[outer_idx]
|
|
generations[gen_idx]["tool_calls"] = tool_calls[outer_idx]["text"]
|
|
|
|
return generations
|