[Core][Bugfix][Perf] Refactor Server to Avoid `AsyncLLMEngine` #8092

robertgshaw2-neuralmagic · 2024-09-02T20:54:33Z

SUMMARY:

@njhill and I removed almost all the overhead from the OpenAI server, but still saw significant slowdown running in AsyncLLMEngine rather than LLMEngine on H100, including when we ran "headless" (e.g. with no uvicorn server).
This caused us to believe the asyncio event loop in AsyncLLMEngine was the root cause of the slowdown
This PR replaces AsyncLLMEngine with MPLLMEngine. MPLLMEngine works similarly to AsyncLLMEngine (i.e. it runs a background loop, accepts new requests, and streams requests back to the clients). We use zeromq as the message passing protocol rather than pulling from queues and pushing to generators
This PR also fixes the number of sockets in use by the RPCClient, avoiding all issues with Too Many Open Files

Summary Performance vs Offline:

pr	scenario	offline	serving	slowdown %
`main`	`multistep`	`42.9`	`33.3`	-22%
`pr`	`multistep`	`42.9`	`40.3`	-6%
`main`	`single-step`	`34.8`	`14.9`	-55%
`pr`	`single-step`	`34.8`	`31.6`	-9%

NOTE: the multistep performance on main is "less-bad" because we currently only stream 1/8 tokens. Once we enable incremental streaming, the performance will be closer to -50% on main.
NOTE: there is still some remaining performance to get by switching the inner loop to using protobufs

Multistep Performance

1xH100 PERFORMANCE BASELINE:

MODEL="meta-llama/Meta-Llama-3.1-8B-Instruct"
python3 benchmarks/benchmark_throughput.py --model $MODEL --dataset benchmarks/ShareGPT_V3_unfiltered_cleaned_split.json --num-scheduler-steps 8

1xH100 SERVING PERFORMANCE

Client:

MODEL="meta-llama/Meta-Llama-3.1-8B-Instruct"
python3 benchmarks/benchmark_serving.py --model $MODEL --dataset-path benchmarks/ShareGPT_V3_unfiltered_cleaned_split.json

Server:

MODEL="meta-llama/Meta-Llama-3.1-8B-Instruct"
vllm serve $MODEL --disable-log-requests --num-scheduler-steps 8 --max-model-len 8192

SUMMARY

branch	setup	throughput
`main`	offline	`42.9`
`main`	serving `mp`	`33.3`
`main`	serving `--disable-frontend-multiprocessing`	`28.9`
`pr`	serving `mp`	`40.3`
`pr`	serving `--disable-frontend-multiprocessing`	`27.9`

Single-Step Performance

1xH100 PERFORMANCE BASELINE:

MODEL="meta-llama/Meta-Llama-3.1-8B-Instruct"
python3 benchmarks/benchmark_throughput.py --model $MODEL --dataset benchmarks/ShareGPT_V3_unfiltered_cleaned_split.json

1xH100 SERVING PERFORMANCE

Client:

MODEL="meta-llama/Meta-Llama-3.1-8B-Instruct"
python3 benchmarks/benchmark_serving.py --model $MODEL --dataset-path benchmarks/ShareGPT_V3_unfiltered_cleaned_split.json

Server:

MODEL="meta-llama/Meta-Llama-3.1-8B-Instruct"
vllm serve $MODEL --disable-log-requests --max-model-len 8192

SUMMARY

branch	setup	throughput
`main`	offline	`34.8`
`main`	serving `mp`	`14.9`
`pr`	serving `mp`	`31.6`

TODOS:

battle test with high load (abort | HWM?)
improve robustness (startup / teardown)
feature set (ray, non-gpu, pipeline parallel, profiler)
hardening / unit tests for MPLLMEngine
tests passing

co-authored by @njhill

FIX #7920

BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE

PR Checklist (Click to Expand)

Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.

PR Title and Classification

Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:

[Bugfix] for bug fixes.
[CI/Build] for build or continuous integration improvements.
[Doc] for documentation fixes and improvements.
[Model] for adding a new model or improving an existing model. Model name should appear in the title.
[Frontend] For changes on the vLLM frontend (e.g., OpenAI API server, LLM class, etc.)
[Kernel] for changes affecting CUDA kernels or other compute kernels.
[Core] for changes in the core vLLM logic (e.g., LLMEngine, AsyncLLMEngine, Scheduler, etc.)
[Hardware][Vendor] for hardware-specific changes. Vendor name should appear in the prefix (e.g., [Hardware][AMD]).
[Misc] for PRs that do not fit the above categories. Please use this sparingly.

Note: If the PR spans more than one category, please include all relevant prefixes.

Code Quality

The PR need to meet the following code quality standards:

We adhere to Google Python style guide and Google C++ style guide.
Pass all linter checks. Please use format.sh to format your code.
The code need to be well-documented to ensure future contributors can easily understand the code.
Include sufficient tests to ensure the project to stay correct and robust. This includes both unit tests and integration tests.
Please add documentation to docs/source/ if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.

Notes for Large Changes

Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with rfc-required and might not go through the PR.

What to Expect for the Reviews

The goal of the vLLM team is to be a transparent reviewing machine. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process:

After the PR is submitted, the PR will be assigned to a reviewer. Every reviewer will pick up the PRs based on their expertise and availability.
After the PR is assigned, the reviewer will provide status update every 2-3 days. If the PR is not reviewed within 7 days, please feel free to ping the reviewer or the vLLM team.
After the review, the reviewer will put an action-required label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.
Please respond to all comments within a reasonable time frame. If a comment isn't clear or you disagree with a suggestion, feel free to ask for clarification or discuss the suggestion.

Thank You

Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone!

Like benchmark_throughput but using AsyncLLMEngine rather than LLM

…-zmq-messages

github-actions · 2024-09-02T20:54:45Z

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which consists a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of default ones by unblocking the steps in your fast-check build on Buildkite UI.

Once the PR is approved and ready to go, please make sure to run full CI as it is required to merge (or just use auto-merge).

To run full CI, you can do one of these:

Comment /ready on the PR
Add ready label to the PR
Enable auto-merge.

🚀

robertgshaw2-neuralmagic · 2024-09-02T21:59:24Z

cc @KuntaiDu

robertgshaw2-neuralmagic · 2024-09-03T03:03:31Z

vllm/engine/multiprocessing/mp_client.py

+                request=RPCUtilityRequest.DO_LOG_STATS,
+                socket=self.input_socket)
+
+    @property


todo: actually set these

robertgshaw2-neuralmagic · 2024-09-03T03:03:55Z

vllm/engine/multiprocessing/mp_client.py

+        await self._send_one_way_rpc_request(
+            request=RPCUtilityRequest.CHECK_HEALTH, socket=self.input_socket)
+
+        # Await acknowledgement from MPLLMEngine.


I think this is a genuinely open question?

It indeed seems not ideal. Either of the following approaches make more sense to me (but we probably don't have to fix it in this PR).

Make sure the response is corresponding to the right sender by matching request ID.

Cache the health check result in an interval (e.g., 10 seconds) and decouple client health check requests and the actual health checks.

robertgshaw2-neuralmagic · 2024-09-03T03:04:42Z

vllm/engine/multiprocessing/mp_llm_engine.py

+            ray_utils.assert_ray_available()
+
+        # TODO: better abstraction?
+        executor_class = AsyncLLMEngine._get_executor_cls(engine_config)


TODO: make this generic

robertgshaw2-neuralmagic · 2024-09-03T03:05:01Z

vllm/engine/multiprocessing/mp_llm_engine.py

+
+        if not self.engine_use_ray:
+            engine_class = self._engine_class
+        elif self.worker_use_ray:


what is worker_use_ray?

it is the same as --distributed_executor_backend=ray

robertgshaw2-neuralmagic · 2024-09-03T03:05:25Z

vllm/engine/multiprocessing/mp_llm_engine.py

+            raise NotImplementedError("Not supported yet!")
+        return engine_class(*args, **kwargs)
+
+    def run_background_loop(self):


Inner loop?

How should we make this propogate exceptions and do things like have the BackgroundLoopDeadError

Is it even worth pulling over the LoopDead logic from the async llm engine?

IIUC that's all there to prevent misconfiguration errors from putting the engine into a state where it only ever responds with an exception, but it looks like it's done pretty bluntly where any exception at all from the model executor will kill the loop. If we want to keep that behavior, we could simply raise from here and exit (after notifying the clients of the exception), and let the frontend die as well.

robertgshaw2-neuralmagic · 2024-09-03T03:07:21Z

vllm/engine/multiprocessing/mp_llm_engine.py

+                        response = VLLM_RPC_SUCCESS_STR
+                    elif request == RPCStartupRequest.CLIENT_IS_READY:
+                        response = VLLM_RPC_SUCCESS_STR
+                        # Breakout of loop once client is ready.


this should be a break ?

We shouldn't break here because we need to response the client in the next 2 lines? We'll get out of the while loop anyways due to client_is_ready = True

robertgshaw2-neuralmagic · 2024-09-03T03:07:40Z

vllm/engine/multiprocessing/mp_llm_engine.py

+            self.stream_outputs(request_outputs)
+
+    def wait_for_new_input(self):
+        while self.input_socket.poll(timeout=POLLING_TIMEOUT_MS) == 0:


should we do something else here?

youkaichao · 2024-09-03T06:32:16Z

vllm/engine/multiprocessing/mp_llm_engine.py

+            # Block until there is a new request.
+            if not self.engine.has_unfinished_requests():
+                self.wait_for_new_input()
+
+            # Handle any new input from the input socket.
+            self.maybe_handle_new_input()
+
+            # Engine step.
+            request_outputs = self.engine.step()
+
+            # Stream results to output socket.
+            self.stream_outputs(request_outputs)


to handle error propogatation, maybe we should add a try-except here, and send the exception back?

alexm-neuralmagic · 2024-09-03T13:26:52Z

@robertgshaw2-neuralmagic @njhill great find about the asyncio overheads being the root cause!

alexm-neuralmagic

Great work, some small comments

alexm-neuralmagic · 2024-09-03T13:27:34Z

examples/openai_completion_client.py

    stream=stream,
-    logprobs=3)
+    max_tokens=100)


Why did you need to change this?

Will revert before merge. I just use this as a simple test example for myself in dev

alexm-neuralmagic · 2024-09-03T13:33:27Z

vllm/engine/multiprocessing/mp_llm_engine.py

+    def cleanup(self):
+        """Cleanup zeromq state on shutdown."""
+        self.input_socket.close()
+        self.output_socket.close()


Do we need also to close the health socket here?

alexm-neuralmagic · 2024-09-03T14:02:25Z

vllm/engine/multiprocessing/mp_llm_engine.py

+                                          usage_context=usage_context,
+                                          ipc_path=ipc_path)
+
+    engine.run_background_loop()


As I understand from the code, the loop that it is running here is not actually run in "the background".

mp_llm_engine runs in the background. I could change the name of the loop

vllm/entrypoints/openai/api_server.py

vllm/engine/multiprocessing/mp_client.py

comaniac · 2024-09-03T17:30:38Z

vllm/engine/multiprocessing/mp_client.py

+        await self._send_one_way_rpc_request(
+            request=RPCUtilityRequest.CHECK_HEALTH, socket=self.input_socket)
+
+        # Await acknowledgement from MPLLMEngine.


It indeed seems not ideal. Either of the following approaches make more sense to me (but we probably don't have to fix it in this PR).

Make sure the response is corresponding to the right sender by matching request ID.

Cache the health check result in an interval (e.g., 10 seconds) and decouple client health check requests and the actual health checks.

comaniac · 2024-09-03T17:34:43Z

vllm/engine/multiprocessing/mp_llm_engine.py

+                        response = VLLM_RPC_SUCCESS_STR
+                    elif request == RPCStartupRequest.CLIENT_IS_READY:
+                        response = VLLM_RPC_SUCCESS_STR
+                        # Breakout of loop once client is ready.


We shouldn't break here because we need to response the client in the next 2 lines? We'll get out of the while loop anyways due to client_is_ready = True

vllm/engine/multiprocessing/mp_llm_engine.py

comaniac · 2024-09-03T17:43:28Z

vllm/engine/multiprocessing/mp_llm_engine.py

+                    socket.send_multipart((identity, pickle.dumps(e)),
+                                          copy=False)
+
+    def run_engine_loop(self) -> None:


Is it possible and beneficial to use async functions for all the function calls in this function?

What I gathered from this PR description is that this new engine implementation is running in a process without the use of any asyncio event loops, it's just a serial loop taking inputs over zmq and passing them to an LLMEngine

no, there is no asyncio at all in this design

I know. I'm just wondering whether making the following functions async could further help with performance:

self.maybe_handle_new_input() request_outputs = self.engine.step() self.stream_outputs(request_outputs)

It is a good point. I have a follow up PR #8125 that makes maybe_handle_new_input() and stream_outputs() async with GPU forward pass and it indeed helps. Maybe we will integrate the changes there immediately into this PR, but not sure yet.

I have added the async behavior of sockets to this PR (and closed 8125)

benchmarks/benchmark_throughput_async.py

# Conflicts: # vllm/entrypoints/openai/api_server.py # vllm/entrypoints/openai/rpc/client.py

DarkLight1337 · 2024-09-04T06:11:08Z

vllm/entrypoints/openai/api_server.py

@@ -21,6 +21,9 @@
 from vllm.config import ModelConfig
 from vllm.engine.arg_utils import AsyncEngineArgs
 from vllm.engine.async_llm_engine import AsyncLLMEngine
+# yapf: enable


You accidentally copied this line over.

Good catch, thanks for pointing this.

alexm-neuralmagic · 2024-09-04T19:09:05Z

This PR is transferred to #8157

alexm-neuralmagic · 2024-09-04T19:18:06Z

We are addressing all review comments from here in here #8157

njhill and others added 6 commits August 28, 2024 12:10

[Benchmark] Add async throughput benchmark

a7a6e43

Like benchmark_throughput but using AsyncLLMEngine rather than LLM

wip

ce7d159

Merge remote-tracking branch 'njhill/async-llm-eng-bench' into reduce…

569cd43

…-zmq-messages

stash

d99ce6f

remove proxy

8d6b2e9

stash

14f3637

added mp_llm_engine

3b8311b

robertgshaw2-neuralmagic requested a review from njhill September 2, 2024 21:03

robertgshaw2-neuralmagic added 9 commits September 2, 2024 21:05

fixed

5e2eb74

format

aa62f2e

cleanup

863081b

revert asyncllmengine

965b97a

fix nit

8fd72f6

format

ddeb7c6

Merge branch 'main' into reduce-asyncio-oh

6539e10

clean

4b111e4

fix

a5ffd2c

robertgshaw2-neuralmagic requested a review from simon-mo September 2, 2024 21:42

robertgshaw2-neuralmagic mentioned this pull request Sep 2, 2024

[Bug]: OpenAI server errors out with "ZMQError Too many open files" under heavy load #7920

Closed

1 task

robertgshaw2-neuralmagic changed the title ~~[Core] Reduce Asyncio Overhead Protoype~~ [Core][Bugfix][Preformance] Asyncio-Free LLMEngine Sep 2, 2024

robertgshaw2-neuralmagic changed the title ~~[Core][Bugfix][Preformance] Asyncio-Free LLMEngine~~ [Core][Bugfix][Performance] Remove AsyncLLMEngine Sep 2, 2024

robertgshaw2-neuralmagic changed the title ~~[Core][Bugfix][Performance] Remove AsyncLLMEngine~~ [Core][Bugfix][Performance] Refactor OpenAI Server to Avoid Use of AsyncLLMEngine Sep 2, 2024

robertgshaw2-neuralmagic changed the title ~~[Core][Bugfix][Performance] Refactor OpenAI Server to Avoid Use of AsyncLLMEngine~~ [Core][Bugfix][Performance] Refactor Server to avoid AsyncLLMEngine Sep 2, 2024

robertgshaw2-neuralmagic changed the title ~~[Core][Bugfix][Performance] Refactor Server to avoid AsyncLLMEngine~~ [Core][Bugfix][Perf] Refactor Server to avoid AsyncLLMEngine Sep 2, 2024

stash

1395872

robertgshaw2-neuralmagic requested a review from KuntaiDu September 2, 2024 21:59

robertgshaw2-neuralmagic added 2 commits September 2, 2024 22:06

move files

938cf85

cleanup code

72d1d42

robertgshaw2-neuralmagic marked this pull request as ready for review September 3, 2024 02:46

github-actions bot added the ready ONLY add when PR is ready to merge/full CI is needed label Sep 3, 2024

robertgshaw2-neuralmagic added 7 commits September 3, 2024 02:50

awk -> ack

ae4564c

add better shutdown

f9ccecc

cleanup comment

89b730b

more awk --> ack

f3dc82b

use constant

ac97a9e

format

becd7ab

remove set to None

b7f49ed

robertgshaw2-neuralmagic commented Sep 3, 2024

View reviewed changes

youkaichao reviewed Sep 3, 2024

View reviewed changes

alexm-neuralmagic reviewed Sep 3, 2024

View reviewed changes

comaniac reviewed Sep 3, 2024

View reviewed changes

alexm-neuralmagic mentioned this pull request Sep 3, 2024

[Core][WIP] MPLLMEngine with async streaming (depends on 8090) #8125

Closed

njhill added 2 commits September 3, 2024 20:40

Merge remote-tracking branch 'origin/main' into reduce-asyncio-oh

58ae3b0

# Conflicts: # vllm/entrypoints/openai/api_server.py # vllm/entrypoints/openai/rpc/client.py

Remove redundant pass

d0f9641

DarkLight1337 reviewed Sep 4, 2024

View reviewed changes

youkaichao mentioned this pull request Sep 4, 2024

[Performance]: The impact of CPU on vLLM performance is significant. #8147

Open

Merge branch 'main' into reduce-asyncio-oh

aa64042

alexm-neuralmagic mentioned this pull request Sep 4, 2024

[Core][Bugfix][Perf] Introduce MQLLMEngine to avoid asyncio OH #8157

Merged

robertgshaw2-neuralmagic closed this Sep 7, 2024

[Core][Bugfix][Perf] Refactor Server to Avoid AsyncLLMEngine #8092

[Core][Bugfix][Perf] Refactor Server to Avoid AsyncLLMEngine #8092

Conversation

robertgshaw2-neuralmagic commented Sep 2, 2024 • edited Loading

Summary Performance vs Offline:

Multistep Performance

1xH100 PERFORMANCE BASELINE:

1xH100 SERVING PERFORMANCE

SUMMARY

Single-Step Performance

1xH100 PERFORMANCE BASELINE:

1xH100 SERVING PERFORMANCE

SUMMARY

PR Title and Classification

Code Quality

Notes for Large Changes

What to Expect for the Reviews

Thank You

github-actions bot commented Sep 2, 2024

robertgshaw2-neuralmagic commented Sep 2, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

robertgshaw2-neuralmagic Sep 3, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

alexm-neuralmagic commented Sep 3, 2024

alexm-neuralmagic left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

robertgshaw2-neuralmagic Sep 3, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

alexm-neuralmagic commented Sep 4, 2024

alexm-neuralmagic commented Sep 4, 2024

[Core][Bugfix][Perf] Refactor Server to Avoid `AsyncLLMEngine` #8092

[Core][Bugfix][Perf] Refactor Server to Avoid `AsyncLLMEngine` #8092

robertgshaw2-neuralmagic commented Sep 2, 2024 •

edited

Loading

robertgshaw2-neuralmagic Sep 3, 2024 •

edited

Loading

robertgshaw2-neuralmagic Sep 3, 2024 •

edited

Loading