[Core] Consolidate prompt arguments to LLM engines #4328

DarkLight1337 · 2024-04-24T08:55:24Z

Currently, LLM.generate (and similar methods in LLMEngine and AsyncLLMEngine) accept prompt, prompt_token_ids and multi_modal_data separately. This PR consolidates them into PromptInputs so that only a single argument has to be passed in, using type annotations to ensure a consistent format. This reduces the chance for the user to accidentally pass in different lengths of prompt, prompt_token_ids, and multi_modal_data (related checks have been removed to avoid redundant code). On the other hand, sampling_params remains untouched because it is common to only pass a single instance even for multiple prompts.

This would also make it easier to define the interface for processing the inputs using HuggingFace processor, as mentioned in #4194.

API changes

The APIs of LLM.generate and LLM.encode have been updated, where the parameters prompt, prompt_token_ids and multi_modal_data are replaced with inputs. Currently, we still maintain the old API but it may be deprecated in a future release. Users may update their code as follows:

Single prompt:

# No change required since the parameter is not referred by name
llm.generate("Hello, my name is")

- llm.generate(prompt="Hello, my name is")
+ llm.generate("Hello, my name is")

- llm.generate(prompt_token_ids=[1, 2, 3])
+ llm.generate({"prompt_token_ids": [1, 2, 3]})

# image is a tensor in NCHW format where N=1
- llm.generate("Hello, my name is", multi_modal_data=MultiModalData(type=..., data=image))
+ llm.generate({"prompt": "Hello, my name is", "multi_modal_data": MultiModalData(type=..., data=image)})

Multiple prompts:

# No change required since the parameter is not referred by name
llm.generate(["Hello, my name is", "The future of AI is"])

- llm.generate(prompt=["Hello, my name is", "The future of AI is"])
+ llm.generate(["Hello, my name is", "The future of AI is"])

- llm.generate(prompt_token_ids=[[1, 2, 3], [4, 5, 6]])
+ llm.generate([{"prompt_token_ids": [1, 2, 3]}, {"prompt_token_ids": [4, 5, 6]}])

# images is a tensor in NCHW format where N=len(prompts)
- prompts = ["Hello, my name is", "The future of AI is"]
- llm.generate(prompts, multi_modal_data=MultiModalData(type=..., data=images))
+ llm.generate([
+    {"prompt": prompt, "multi_modal_data": MultiModalData(type=..., data=images[i:i+1])}
+    for i, prompt in enumerate(prompts)
+ ])

Based on the examples in the documentation, most users should already prefer the first way of calling LLM.generate; those users need not make any changes.

Other changes

By setting gpu_memory_utilization to a smaller value, we can now run multiple LLM instances and OpenAI servers at the same time. To better manage GPU utilization between entrypoints tests, I have grouped the tests using pytest markers so we no longer have to refer to the file name directly. This makes it easier to edit the commands used to test multiple files in the same pytest session.

DarkLight1337 · 2024-04-27T12:45:56Z

@ywang96 Any thoughts about this?

ywang96 · 2024-04-29T05:16:21Z

Hey @DarkLight1337! Sorry I've been a bit busy lately, but I will surely take a look in the upcoming week! Apologies for the delay!

ywang96

Thank you very much for the contribution @DarkLight1337 and sorry for the delayed review. I left a few comments and questions. Please take a look!

vllm/entrypoints/llm.py

vllm/engine/llm_engine.py

ywang96 · 2024-05-03T00:06:08Z

vllm/engine/llm_engine.py

-        # Create the sequence group.
-        seq_group = SequenceGroup(request_id, [seq], sampling_params,
-                                  arrival_time, lora_request, multi_modal_data)
+        processed_inputs = self.encode_request(request_id=request_id,


Right before when we call encode_request is where I think we could combine prompt, prompt_token_ids and multi_modal_data to PromptInputs.

I think we should do this directly on the API class (LLM) so that we do not have to involve the internals when finally deprecating the old API.

vllm/inputs.py

vllm/engine/llm_engine.py

- To facilitate equality tests, `CompletionOutput` is now a dataclass

rkooo567 · 2024-05-23T23:42:57Z

I think we should keep prompts/prompt_tokens_ids input for backward compatibility?

DarkLight1337 · 2024-05-24T00:04:30Z

I think we should keep prompts/prompt_tokens_ids input for backward compatibility?

These arguments are still being maintained. I just decorated the relevant methods so we can deprecate them simply by turning on the corresponding flag.

DarkLight1337 · 2024-05-24T00:55:24Z

Pretty sure the CI will pass now.

njhill

Thanks @DarkLight1337, I added some comments inline

njhill · 2024-05-24T23:58:12Z

vllm/engine/llm_engine.py

+        if isinstance(inputs, str):
+            inputs = {"prompt": inputs}
+
+        if "prompt_token_ids" not in inputs:


nit: better to do

prompt_token_ids = inputs.get("prompt_token_ids") if prompt_token_ids is None: # ...

Doing so actually fails the type checker (I'm using Pyright on VSCode), so I'm reluctant to change it to this form.

What does it complain about? Couldn't additional hints be added for that e.g.

prompt_token_ids: List[int] = inputs.get("prompt_token_ids")

njhill · 2024-05-25T00:24:08Z

vllm/engine/async_llm_engine.py

+        if isinstance(inputs, str):
+            inputs = {"prompt": inputs}
+
+        if "prompt_token_ids" not in inputs:


suggest

prompt_token_ids = inputs.get("prompt_token_ids") if prompt_tokens_ids is not None: # ...

njhill · 2024-05-25T00:27:42Z

vllm/engine/async_llm_engine.py

        arrival_time: Optional[float] = None,
        lora_request: Optional[LoRARequest] = None,
-        multi_modal_data: Optional[MultiModalData] = None,
    ) -> AsyncStream:
        if self.log_requests:


I guess not related to this PR but I feel like the body of this if should go into a separate function

This bothers me as well since it's otherwise redundant to pass both prompt and prompt_token_ids at the same time. Perhaps it would be better to move this to OpenAI server in another PR?

vllm/engine/llm_engine.py

njhill · 2024-05-25T00:33:02Z

vllm/engine/llm_engine.py

+            eos_token_id = self.tokenizer.get_lora_tokenizer(
+                lora_request).eos_token_id
+        else:
+            logger.warning("Use None for EOS token id because tokenizer is "


I'm not sure that we want to log a warning on every request. This will be a lot of log output for the no-tokenizer cases.

Actually this came from #3748, this PR only moved the code from add_request into _add_processed_request.

@ywang96 Since you reviewed that PR, can you explain the reason behind this/is it still necessary?

Hmm - this is probably an oversight from me. I thought this warning was added at engine initialization time, not inference time. IMO we should move it to the initialization time for sure.

It may not be a good idea to warn during initialization time if the user indeed intends the tokenizer to not be used.

I have updated the code so that the warning still occurs on request but is only emitted at most once during the lifetime of the engine.

It may not be a good idea to warn during initialization time if the user indeed intends the tokenizer to not be used.

I'm not sure that I understand this, I think a single warning is ok in this case, it's only a warning. And a single warning will still be logged with this latest change right? Just at the time of the first request rather than llm engine construction

It may not be a good idea to warn during initialization time if the user indeed intends the tokenizer to not be used.

I'm not sure that I understand this, I think a single warning is ok in this case, it's only a warning. And a single warning will still be logged with this latest change right? Just at the time of the first request rather than llm engine construction

If the user didn't intend to use the tokenizer, then the warning might cause some confusion during startup. True that it's not a big deal either way though.

I personally vote for moving this to engine initialization time since that will make our codebase cleaner. (no need for another attribute of engine just for warning count)

Upon inspection of the docs, it seems that avoiding repeated errors is actually unnecessary:

Warning messages are normally written to sys.stderr, but their disposition can be changed flexibly, from ignoring all warnings to turning them into exceptions. The disposition of warnings can vary based on the warning category, the text of the warning message, and the source location where it is issued. Repetitions of a particular warning for the same source location are typically suppressed.

Therefore, we can keep the warning in the code for handling the request.

vllm/entrypoints/llm.py

vllm/entrypoints/openai/serving_chat.py

vllm/engine/llm_engine.py

DarkLight1337 · 2024-05-25T02:42:39Z

@njhill I have responded to your comments.

ywang96 · 2024-05-25T19:56:31Z

@DarkLight1337 I just went though this PR again and made a change to move offline API reference to under developer doc. #4710 was a great addition, but I think we should have links in examples to developer doc instead of putting API reference there directly.

ywang96

I'm going to give this PR a greenlight, and thank you @DarkLight1337 for working on this PR and addressing the comments from reviewers.

@njhill @rkooo567 If you have any other concern feel free to leave a comment, otherwise I'd appreciate an approval from either of you as well. Thanks!

(P.S. @DarkLight1337 in the future, it'll be nice to break a big PR like this even further down into small PRs, that way it's easier for us to review and get things merged in a progressive way.)

njhill

Thanks again @DarkLight1337 @ywang96!

Co-authored-by: Roger Wang <[email protected]>

DarkLight1337 changed the title ~~[Core] Combine prompt inputs to LLM engines~~ [Core] Combine prompt arguments to LLM engines Apr 24, 2024

Combine prompt inputs

5d42800

DarkLight1337 force-pushed the llm-inputs branch from d6e9aa2 to 5d42800 Compare April 24, 2024 08:59

DarkLight1337 mentioned this pull request Apr 24, 2024

[RFC]: Multi-modality Support on vLLM #4194

Open

73 tasks

DarkLight1337 added 4 commits April 25, 2024 01:51

Fix a bunch of tests

5db2c5e

Fix LLaVA test

74c5905

Merge branch 'upstream' into llm-inputs

cd8917b

Fix benchmark_latency test

b49aba7

DarkLight1337 changed the title ~~[Core] Combine prompt arguments to LLM engines~~ [Core] Consolidate prompt arguments to LLM engines Apr 25, 2024

DarkLight1337 added 2 commits April 25, 2024 07:14

Merge branch 'upstream' into llm-inputs

bfd7295

Merge branch 'upstream' into llm-inputs

45c7f23

Merge branch 'upstream' into llm-inputs

493e6ed

Merge branch 'upstream' into llm-inputs

20aeceb

ywang96 reviewed May 3, 2024

View reviewed changes

DarkLight1337 added 4 commits May 3, 2024 03:36

Merge branch 'upstream' into llm-inputs

0f46653

Clarify tokenizer usage

c4f3540

Rename encode_request -> process_model_inputs

ab8182c

Support old API in LLM.generate

eac33e1

DarkLight1337 mentioned this pull request May 3, 2024

[Frontend] Refactor prompt processing #4028

Merged

Add tests to ensure old API still works

703d318

- To facilitate equality tests, `CompletionOutput` is now a dataclass

DarkLight1337 force-pushed the llm-inputs branch 6 times, most recently from 0c57f28 to 37baf38 Compare May 3, 2024 16:59

Let all entrypoints tests be run at the same time

19d85f9

DarkLight1337 force-pushed the llm-inputs branch from 37baf38 to 19d85f9 Compare May 3, 2024 17:21

Remove unused parameter in _validate_and_add_requests and fix test

cc84f65

njhill reviewed May 25, 2024

View reviewed changes

DarkLight1337 added 2 commits May 25, 2024 02:37

Simplify code

6c5d4a6

Move attribute assignment outside _init_tokenizer

fd2da12

DarkLight1337 and others added 4 commits May 25, 2024 03:12

Only emit warning once

d78de94

Simplify assignment expression

8a86829

Place special case at the start

731ac0e

move API reference to under developer doc

2d1a0bc

DarkLight1337 added 2 commits May 26, 2024 01:32

Fix links in docs

7b8ce2c

Remove unnecessary code to avoid repeated warning

fff21a1

ywang96 approved these changes May 27, 2024

View reviewed changes

njhill approved these changes May 28, 2024

View reviewed changes

ywang96 merged commit 5ae5ed1 into vllm-project:main May 28, 2024
63 checks passed

DarkLight1337 deleted the llm-inputs branch May 29, 2024 01:18

This was referenced May 29, 2024

[Bugfix] Fix arguments passed to Sequence in stop checker test #5092

Merged

[Core] Avoid the need to pass None values to Sequence.inputs #5099

Merged

blinkbear pushed a commit to blinkbear/vllm that referenced this pull request May 29, 2024

[Core] Consolidate prompt arguments to LLM engines (vllm-project#4328)

2ea6fd8

Co-authored-by: Roger Wang <[email protected]>

dtrifiro pushed a commit to opendatahub-io/vllm that referenced this pull request May 31, 2024

[Core] Consolidate prompt arguments to LLM engines (vllm-project#4328)

eab28b9

Co-authored-by: Roger Wang <[email protected]>

blinkbear pushed a commit to blinkbear/vllm that referenced this pull request Jun 6, 2024

[Core] Consolidate prompt arguments to LLM engines (vllm-project#4328)

b18e919

Co-authored-by: Roger Wang <[email protected]>

robertgshaw2-redhat pushed a commit to neuralmagic/nm-vllm that referenced this pull request Jun 8, 2024

[Core] Consolidate prompt arguments to LLM engines (vllm-project#4328)

8c9aab4

Co-authored-by: Roger Wang <[email protected]>

joerunde pushed a commit to joerunde/vllm that referenced this pull request Jun 17, 2024

[Core] Consolidate prompt arguments to LLM engines (vllm-project#4328)

85ed9d0

Co-authored-by: Roger Wang <[email protected]>

robertgshaw2-redhat pushed a commit to neuralmagic/nm-vllm that referenced this pull request Jul 14, 2024

[Core] Consolidate prompt arguments to LLM engines (vllm-project#4328)

1690706

Co-authored-by: Roger Wang <[email protected]>

Temirulan pushed a commit to Temirulan/vllm-whisper that referenced this pull request Sep 6, 2024

[Core] Consolidate prompt arguments to LLM engines (vllm-project#4328)

f621d50

Co-authored-by: Roger Wang <[email protected]>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Core] Consolidate prompt arguments to LLM engines #4328

[Core] Consolidate prompt arguments to LLM engines #4328

DarkLight1337 commented Apr 24, 2024 •

edited

Loading

DarkLight1337 commented Apr 27, 2024

ywang96 commented Apr 29, 2024

ywang96 left a comment

ywang96 May 3, 2024

DarkLight1337 May 3, 2024

rkooo567 commented May 23, 2024

DarkLight1337 commented May 24, 2024

DarkLight1337 commented May 24, 2024

njhill left a comment

njhill May 24, 2024

DarkLight1337 May 25, 2024

njhill May 25, 2024

njhill May 25, 2024

njhill May 25, 2024

DarkLight1337 May 25, 2024

njhill May 25, 2024

DarkLight1337 May 25, 2024 •

edited

Loading

njhill May 25, 2024

njhill May 25, 2024

DarkLight1337 May 25, 2024 •

edited

Loading

ywang96 May 25, 2024

DarkLight1337 May 25, 2024 •

edited

Loading

njhill May 25, 2024

DarkLight1337 May 25, 2024 •

edited

Loading

ywang96 May 25, 2024

DarkLight1337 May 26, 2024 •

edited

Loading

DarkLight1337 commented May 25, 2024

ywang96 commented May 25, 2024 •

edited

Loading

ywang96 left a comment •

edited

Loading

njhill left a comment

[Core] Consolidate prompt arguments to LLM engines #4328

[Core] Consolidate prompt arguments to LLM engines #4328

Conversation

DarkLight1337 commented Apr 24, 2024 • edited Loading

API changes

Other changes

DarkLight1337 commented Apr 27, 2024

ywang96 commented Apr 29, 2024

ywang96 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

rkooo567 commented May 23, 2024

DarkLight1337 commented May 24, 2024

DarkLight1337 commented May 24, 2024

njhill 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

DarkLight1337 May 25, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

DarkLight1337 May 25, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

DarkLight1337 May 25, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

DarkLight1337 May 25, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

DarkLight1337 May 26, 2024 • edited Loading

Choose a reason for hiding this comment

DarkLight1337 commented May 25, 2024

ywang96 commented May 25, 2024 • edited Loading

ywang96 left a comment • edited Loading

Choose a reason for hiding this comment

njhill left a comment

Choose a reason for hiding this comment

DarkLight1337 commented Apr 24, 2024 •

edited

Loading

DarkLight1337 May 25, 2024 •

edited

Loading

DarkLight1337 May 25, 2024 •

edited

Loading

DarkLight1337 May 25, 2024 •

edited

Loading

DarkLight1337 May 25, 2024 •

edited

Loading

DarkLight1337 May 26, 2024 •

edited

Loading

ywang96 commented May 25, 2024 •

edited

Loading

ywang96 left a comment •

edited

Loading