hf_text-generation-inference/server/text_generation_server/server.py

158 lines
5.3 KiB
Python
Raw Normal View History

2022-10-08 04:30:12 -06:00
import asyncio
2022-10-17 06:59:00 -06:00
import os
import torch
2022-10-17 06:59:00 -06:00
2022-10-08 04:30:12 -06:00
from grpc import aio
from loguru import logger
2022-10-08 04:30:12 -06:00
from grpc_reflection.v1alpha import reflection
from pathlib import Path
2023-01-31 10:53:56 -07:00
from typing import List, Optional
2022-10-08 04:30:12 -06:00
2023-03-07 10:52:22 -07:00
from text_generation_server.cache import Cache
from text_generation_server.interceptor import ExceptionInterceptor
from text_generation_server.models import Model, get_model
from text_generation_server.pb import generate_pb2_grpc, generate_pb2
from text_generation_server.tracing import UDSOpenTelemetryAioServerInterceptor
2022-10-08 04:30:12 -06:00
class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer):
def __init__(self, model: Model, cache: Cache, server_urls: List[str]):
2022-10-08 04:30:12 -06:00
self.cache = cache
self.model = model
self.server_urls = server_urls
# For some reason, inference_mode does not work well with GLOO which we use on CPU
if model.device.type == "cuda":
# Force inference mode for the lifetime of TextGenerationService
self._inference_mode_raii_guard = torch._C._InferenceMode(True)
2022-10-08 04:30:12 -06:00
async def Info(self, request, context):
return self.model.info
async def Health(self, request, context):
if self.model.device.type == "cuda":
torch.zeros((2, 2)).cuda()
return generate_pb2.HealthResponse()
2022-10-08 04:30:12 -06:00
async def ServiceDiscovery(self, request, context):
return generate_pb2.ServiceDiscoveryResponse(urls=self.server_urls)
async def ClearCache(self, request, context):
if request.HasField("id"):
self.cache.delete(request.id)
else:
self.cache.clear()
if torch.cuda.is_available():
torch.cuda.empty_cache()
return generate_pb2.ClearCacheResponse()
2022-10-08 04:30:12 -06:00
async def FilterBatch(self, request, context):
batch = self.cache.pop(request.batch_id)
if batch is None:
raise ValueError(f"Batch ID {request.batch_id} not found in cache.")
filtered_batch = batch.filter(request.keep_requests)
self.cache.set(filtered_batch)
return generate_pb2.FilterBatchResponse(batch=filtered_batch.to_pb())
async def Prefill(self, request, context):
batch = self.model.batch_type.from_pb(
request.batch, self.model.tokenizer, self.model.device
)
generations, next_batch = self.model.generate_token(batch)
self.cache.set(next_batch)
return generate_pb2.PrefillResponse(
generations=[generation.to_pb() for generation in generations],
batch=next_batch.to_pb() if next_batch else None,
2022-10-08 04:30:12 -06:00
)
async def Decode(self, request, context):
if len(request.batches) == 0:
raise ValueError("Must provide at least one batch")
batches = []
for batch_pb in request.batches:
batch = self.cache.pop(batch_pb.id)
if batch is None:
raise ValueError(f"Batch ID {batch_pb.id} not found in cache.")
batches.append(batch)
if len(batches) == 0:
raise ValueError("All batches are empty")
if len(batches) > 1:
batch = self.model.batch_type.concatenate(batches)
else:
batch = batches[0]
generations, next_batch = self.model.generate_token(batch)
self.cache.set(next_batch)
return generate_pb2.DecodeResponse(
generations=[generation.to_pb() for generation in generations],
batch=next_batch.to_pb() if next_batch else None,
)
2022-10-08 04:30:12 -06:00
2022-10-18 07:19:03 -06:00
def serve(
model_id: str,
2023-01-31 10:53:56 -07:00
revision: Optional[str],
2022-10-18 07:19:03 -06:00
sharded: bool,
feat(server): GPTQ quantization (step1) (#277) Changes only the type from `bool` to `Option<Enum>` pretty much everywhere. - Use `Optional[str]` in Python (easier to manage than importing type everywhere). Except for the cli to get proper validation - Updated all models to handle gracefully new values. (Error out if unknown value, or gptq since not implemented). <!-- Congratulations! You've made it this far! You're not quite done yet though. Once merged, your PR is going to appear in the release notes with the title you set, so make sure it's a great title that fully reflects the extent of your awesome contribution. Then, please replace this with a description of the change and which issue is fixed (if applicable). Please also include relevant motivation and context. List any dependencies (if any) that are required for this change. Once you're done, someone will review your PR shortly (see the section "Who can review?" below to tag some potential reviewers). They may suggest changes to make the code even better. If no one reviewed your PR after a week has passed, don't hesitate to post a new comment @-mentioning the same persons---sometimes notifications get lost. --> <!-- Remove if not applicable --> Fixes # (issue) ## Before submitting - [ ] This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). - [ ] Did you read the [contributor guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#start-contributing-pull-requests), Pull Request section? - [ ] Was this discussed/approved via a Github issue or the [forum](https://discuss.huggingface.co/)? Please add a link to it if that's the case. - [ ] Did you make sure to update the documentation with your changes? Here are the [documentation guidelines](https://github.com/huggingface/transformers/tree/main/docs), and [here are tips on formatting docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation). - [ ] Did you write any new necessary tests? ## Who can review? Anyone in the community is free to review the PR once the tests have passed. Feel free to tag members/contributors who may be interested in your PR. <!-- Your PR will be replied to more quickly if you can figure out the right person to tag with @ @OlivierDehaene OR @Narsil -->
2023-05-12 06:46:41 -06:00
quantize: Optional[str],
trust_remote_code: bool,
2022-10-18 07:19:03 -06:00
uds_path: Path,
):
2022-10-08 04:30:12 -06:00
async def serve_inner(
model_id: str,
2023-01-31 10:53:56 -07:00
revision: Optional[str],
2022-10-08 04:30:12 -06:00
sharded: bool = False,
feat(server): GPTQ quantization (step1) (#277) Changes only the type from `bool` to `Option<Enum>` pretty much everywhere. - Use `Optional[str]` in Python (easier to manage than importing type everywhere). Except for the cli to get proper validation - Updated all models to handle gracefully new values. (Error out if unknown value, or gptq since not implemented). <!-- Congratulations! You've made it this far! You're not quite done yet though. Once merged, your PR is going to appear in the release notes with the title you set, so make sure it's a great title that fully reflects the extent of your awesome contribution. Then, please replace this with a description of the change and which issue is fixed (if applicable). Please also include relevant motivation and context. List any dependencies (if any) that are required for this change. Once you're done, someone will review your PR shortly (see the section "Who can review?" below to tag some potential reviewers). They may suggest changes to make the code even better. If no one reviewed your PR after a week has passed, don't hesitate to post a new comment @-mentioning the same persons---sometimes notifications get lost. --> <!-- Remove if not applicable --> Fixes # (issue) ## Before submitting - [ ] This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). - [ ] Did you read the [contributor guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#start-contributing-pull-requests), Pull Request section? - [ ] Was this discussed/approved via a Github issue or the [forum](https://discuss.huggingface.co/)? Please add a link to it if that's the case. - [ ] Did you make sure to update the documentation with your changes? Here are the [documentation guidelines](https://github.com/huggingface/transformers/tree/main/docs), and [here are tips on formatting docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation). - [ ] Did you write any new necessary tests? ## Who can review? Anyone in the community is free to review the PR once the tests have passed. Feel free to tag members/contributors who may be interested in your PR. <!-- Your PR will be replied to more quickly if you can figure out the right person to tag with @ @OlivierDehaene OR @Narsil -->
2023-05-12 06:46:41 -06:00
quantize: Optional[str] = None,
trust_remote_code: bool = False,
2022-10-08 04:30:12 -06:00
):
2022-10-18 07:19:03 -06:00
unix_socket_template = "unix://{}-{}"
2022-10-08 04:30:12 -06:00
if sharded:
server_urls = [
2022-10-18 07:19:03 -06:00
unix_socket_template.format(uds_path, rank)
for rank in range(int(os.environ["WORLD_SIZE"]))
2022-10-08 04:30:12 -06:00
]
local_url = server_urls[int(os.environ["RANK"])]
2022-10-08 04:30:12 -06:00
else:
2022-10-18 07:19:03 -06:00
local_url = unix_socket_template.format(uds_path, 0)
2022-10-08 04:30:12 -06:00
server_urls = [local_url]
try:
model = get_model(model_id, revision, sharded, quantize, trust_remote_code)
except Exception:
logger.exception("Error when initializing model")
raise
2023-02-13 05:02:45 -07:00
server = aio.server(
interceptors=[
ExceptionInterceptor(),
UDSOpenTelemetryAioServerInterceptor(),
]
)
generate_pb2_grpc.add_TextGenerationServiceServicer_to_server(
TextGenerationService(model, Cache(), server_urls), server
2022-10-08 04:30:12 -06:00
)
SERVICE_NAMES = (
generate_pb2.DESCRIPTOR.services_by_name["TextGenerationService"].full_name,
2022-10-08 04:30:12 -06:00
reflection.SERVICE_NAME,
)
reflection.enable_server_reflection(SERVICE_NAMES, server)
server.add_insecure_port(local_url)
2022-10-08 04:30:12 -06:00
await server.start()
logger.info("Server started at {}".format(local_url))
2022-10-18 07:19:03 -06:00
try:
await server.wait_for_termination()
except KeyboardInterrupt:
logger.info("Signal received. Shutting down")
2022-10-18 07:19:03 -06:00
await server.stop(0)
2022-10-08 04:30:12 -06:00
asyncio.run(serve_inner(model_id, revision, sharded, quantize, trust_remote_code))