[Lora] Support long context lora #4787

rkooo567 · 2024-05-13T13:25:45Z

Currently we need to call rotary embedding kernel for each LoRA, which makes it hard to serve multiple long context length LoRA. Add batched rotary embedding kernel and pipe it through.

It replaces the rotary embedding layer to the one that is aware of multiple cos-sin-cache per scaling factors.

Follow up of https://github.com/vllm-project/vllm/pull/3095/files

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!

rkooo567 · 2024-05-13T13:33:41Z

vllm/lora/layers.py

@@ -1193,3 +1203,204 @@ def can_replace_layer(cls, source_layer: nn.Module,
 model_config: Optional[PretrainedConfig]) -> bool:
 # Special handling for the LogitsProcessor.
 return False
+
+
+class LoRALinearScalingRotaryEmbedding(LinearScalingRotaryEmbedding):


rkooo567 · 2024-05-13T13:33:44Z

vllm/lora/layers.py

+ return query, key
+
+
+class LoRAPagedAttentionWithRoPE(LoRALayer):


rkooo567 · 2024-05-15T14:53:03Z

tests/lora/conftest.py

@@ -21,6 +21,41 @@
 from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
 from vllm.model_executor.model_loader import get_model

+LONG_LORA_INFOS = [


Will be moved to hf hub

rkooo567 · 2024-05-15T14:55:32Z

tests/lora/test_long_context.py

+ )
+ return lora_llm
+
+ def test_batched_rope_kernel(self, long_context_infos):


currently this test has illegal memory access

rkooo567 · 2024-05-17T03:52:11Z

The PR is ready to be reviewed

richardliaw · 2024-05-17T06:24:44Z

tests failing?

rkooo567 · 2024-05-17T07:19:16Z

that' expected because we need to move fine tuned lora model to hf hub so that CI can access it.

will be done soon. The test succeeds locally.

rkooo567 · 2024-05-17T15:19:03Z

Use hf instead of s3 now

rkooo567

@Yard1 compared to internal version;

clean up on some of variables and logic
Added a test to test_layers.py
Replaced rotary embedding instead of self.attn
fixed a bug where rotary emb is only replacing the first layer
download lora fine tuned models from hf hub I created

Yard1

Thanks, looks good! Some small comments

Yard1 · 2024-05-17T18:01:13Z

vllm/core/scheduler.py

+ prompt_limit = (seq_group.lora_request.long_lora_max_len
+ if seq_group.lora_request
+ and seq_group.lora_request.long_lora_max_len else
+ self.prompt_limit)


let's put it in a method/function

removed self.prompt_limit btw because in this case, it doesn't make much sense to have it.

vllm/engine/arg_utils.py

vllm/engine/output_processor/stop_checker.py

vllm/lora/layers.py

vllm/transformers_utils/tokenizer_group/tokenizer_group.py

Yard1 · 2024-05-17T18:15:25Z

tests/lora/test_lora_checkpoints.py

@@ -28,6 +28,7 @@ def test_load_checkpoints(
 # and the test should pass.
 LoRAModel.from_local_checkpoint(
 baichuan_lora_files,
+ 4096,


use kwarg instead

Default None == meaning it will use scaling factor 1. Also added docstring

tests/lora/test_lora_manager.py

rkooo567 · 2024-05-18T00:17:55Z

@Yard1 everything addressed

Currently we need to call rotary embedding kernel for each LoRA, which makes it hard to serve multiple long context length LoRA. Add batched rotary embedding kernel and pipe it through. It replaces the rotary embedding layer to the one that is aware of multiple cos-sin-cache per scaling factors. Follow up of https://github.com/vllm-project/vllm/pull/3095/files

AllenDou · 2024-05-21T10:47:40Z

May I ask a question? Only LinearScalingRotaryEmbedding requires the 'withlora' version. Do DynamicNTKScalingRotaryEmbedding/YaRNScalingRotaryEmbedding/Phi3SuScaledRotaryEmbedding also require it?

rkooo567 · 2024-05-21T12:35:50Z

I think those are not working with long context multi lora. In order to get it working, I think other rotary embedding should also support multi scaling factors like we are doing for LinearScalingRotaryEmbedding .

AllenDou · 2024-05-21T12:52:42Z

ok, thanks for your response :)

Currently we need to call rotary embedding kernel for each LoRA, which makes it hard to serve multiple long context length LoRA. Add batched rotary embedding kernel and pipe it through. It replaces the rotary embedding layer to the one that is aware of multiple cos-sin-cache per scaling factors. Follow up of https://github.com/vllm-project/vllm/pull/3095/files

wip

acfe6a3

rkooo567 commented May 13, 2024

View reviewed changes

Yard1 self-requested a review May 13, 2024 17:18

rkooo567 added 5 commits May 13, 2024 16:08

.

b1411c8

Merge branch 'main' into support-long-context-lora

ea99d53

ip

9a1ac3b

Merge branch 'main' into support-long-context-lora

1ca2b69

wip

9e06086

rkooo567 changed the title ~~[WIP] Support long context lora~~ [Lora] Support long context lora May 15, 2024

rkooo567 commented May 15, 2024

View reviewed changes

rkooo567 added 13 commits May 15, 2024 09:13

wip

b2feb0a

Merge branch 'main' into support-long-context-lora

2800c69

wip

c5a1424

.

09f0050

.

0e5f477

lora works

14d1b65

ip

774d18e

ip

34add71

working except long lora

9baeb76

.

ae22f0a

Merge branch 'main' into support-long-context-lora

d313e33

done

6157c0e

working e2e

bb1f2fa

rkooo567 added 2 commits May 16, 2024 20:59

Merge branch 'main' into support-long-context-lora

6eecae8

remove unnecessary file

7ee80e0

rkooo567 added 2 commits May 17, 2024 04:05

Merge branch 'main' into support-long-context-lora

c1a4590

working e2e

3a9aaa9

lint

8320001

rkooo567 commented May 17, 2024

View reviewed changes

rkooo567 added 3 commits May 17, 2024 09:13

fix

042fdbe

fix pipeline

e1228a7

lint

53bba7d

Yard1 reviewed May 17, 2024

View reviewed changes

rkooo567 added 2 commits May 17, 2024 16:24

Merge branch 'main' into support-long-context-lora

c875ea0

done

0f77b38

Yard1 approved these changes May 18, 2024

View reviewed changes

rkooo567 merged commit 2e9a222 into vllm-project:main May 18, 2024
61 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Lora] Support long context lora #4787

[Lora] Support long context lora #4787

rkooo567 commented May 13, 2024 •

edited

rkooo567 May 13, 2024

rkooo567 May 13, 2024

rkooo567 May 15, 2024

rkooo567 May 15, 2024

rkooo567 commented May 17, 2024

richardliaw commented May 17, 2024

rkooo567 commented May 17, 2024

rkooo567 commented May 17, 2024

rkooo567 left a comment

Yard1 left a comment

Yard1 May 17, 2024

rkooo567 May 17, 2024

Yard1 May 17, 2024

rkooo567 May 18, 2024

rkooo567 commented May 18, 2024

AllenDou commented May 21, 2024

rkooo567 commented May 21, 2024

AllenDou commented May 21, 2024

		return query, key


		class LoRAPagedAttentionWithRoPE(LoRALayer):

[Lora] Support long context lora #4787

[Lora] Support long context lora #4787

Conversation

rkooo567 commented May 13, 2024 • edited

PR Title and Classification

Code Quality

Notes for Large Changes

What to Expect for the Reviews

Thank You

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

rkooo567 commented May 17, 2024

richardliaw commented May 17, 2024

rkooo567 commented May 17, 2024

rkooo567 commented May 17, 2024

rkooo567 left a comment

Choose a reason for hiding this comment

Yard1 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

rkooo567 commented May 18, 2024

AllenDou commented May 21, 2024

rkooo567 commented May 21, 2024

AllenDou commented May 21, 2024

rkooo567 commented May 13, 2024 •

edited