[kernel] fix sliding window in prefix prefill Triton kernel #4405

mmoskal · 2024-04-26T22:40:12Z

This adds support for the sliding window in prefix prefill kernel.

I had to use a large negative value instead of -inf for masking, since otherwise in some situations we get '-inf - -inf' in softmax which leads to NaNs.

Added tests comparing with xformers.

Also added a bunch of comments with tensor shapes etc.

FIX #4057

CC @rkooo567 @cadedaniel @simon-mo

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!

zhuohan123 · 2024-04-26T23:45:04Z

CC @caoshiyi Can you help take a look at this?

rkooo567 · 2024-04-27T00:59:10Z

yep! it is Sat here, but I will take a look at it very soon! @mmoskal thanks for the amazing contribution!

rkooo567

Looks pretty good! Also thanks for the comment on shapes. Some comments regarding tests

rkooo567 · 2024-04-30T06:46:30Z

tests/kernels/test_prefix_prefill.py

@@ -15,18 +15,21 @@
 CUDA_DEVICES = [
    f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)
 ]
+SLIDING_WINDOW = [0, 512]


can you try various values? (16, 64, 128, 256, 512, 2048). 2048 is bigger than max seq, but just for sanity check

Also curious if we can do e2e test against mistral or other model that has sliding window attn enabled..

I'll try different sizes. I'm trying to get the sliding window to work with v2 block manager (somewhat based on #3967) which should exercise this.

sg. The test itself will be done in this PR right?

yes, adding more sliding window parameters now

rkooo567 · 2024-04-30T06:54:04Z

vllm/attention/ops/prefix_prefill.py

+                # exp().
+                qk = tl.where((cur_batch_ctx_len + offs_m[:, None]) -
+                              (start_n + offs_n[None, :]) < SLIDING_WINDOW, qk,
+                              -10000)


Is -10000 small enough? Maybe consider even smaller values? like -10000000

-65000 or so is the smallest value for f16. Anyway, this gets exp()ed so -10k should enough.

Actually good you point this out - I found that in one place I do sm_scale before and in the other after masking with -10000; it doesn't matter much (smallest scale is 1/16 or so, and exp(-10000/16) is still zero) but better to always scale and then mask so we get consistent mask values - fixed

vllm/attention/ops/prefix_prefill.py

rkooo567 · 2024-04-30T07:31:43Z

vllm/attention/ops/prefix_prefill.py

+                # This then makes m_ij contain -inf, which causes NaNs in
+                # exp().
+                qk = tl.where((cur_batch_ctx_len + offs_m[:, None]) -
+                              (start_n + offs_n[None, :]) < SLIDING_WINDOW, qk,


so this means

cur_batch_ctx_len + offs_m[:, None] == end of q start_n + offs_n[None, :] == end of k

so q-k length is within slinding window, attend it, is this correct?

yes; I added a comment to that effect

Co-authored-by: SangBin Cho <rkooo567@gmail.com>

rkooo567

LGTM given it passes with more variations in unit test!

mmoskal · 2024-05-02T00:57:32Z

@rkooo567 all tests passed, should be good

rkooo567 · 2024-05-02T01:36:01Z

cc @simon-mo to merge!

mmoskal · 2024-05-02T17:51:30Z

@simon-mo can we get this merged? I need it to rebase #4545

…ject#4405) Co-authored-by: SangBin Cho <rkooo567@gmail.com>

mmoskal added 2 commits April 26, 2024 22:34

start on sliding window in prefix_prefill

538b2f0

fix nans, tests pass

8c7abb1

mmoskal mentioned this pull request Apr 27, 2024

[Model] Adding sliding window support for block table [#3665] #3967

Closed

rkooo567 self-assigned this Apr 27, 2024

fix argument order

20d5e13

rkooo567 reviewed Apr 30, 2024

View reviewed changes

rkooo567 added the action-required label May 1, 2024

mmoskal and others added 3 commits May 2, 2024 02:01

Update vllm/attention/ops/prefix_prefill.py

c25d929

Co-authored-by: SangBin Cho <rkooo567@gmail.com>

apply scale before sliding window masking

c045c6e

add more tests per PR feedback

a7406e9

rkooo567 approved these changes May 2, 2024

View reviewed changes

mmoskal added 3 commits May 2, 2024 00:16

explain masking

9bc5f2e

fix comment formatting and wording

01e2b82

fix whitespace

7610903

simon-mo approved these changes May 2, 2024

View reviewed changes

simon-mo merged commit 32881f3 into vllm-project:main May 2, 2024
48 checks passed

mmoskal deleted the triton_sliding_window branch May 2, 2024 22:42

robertgshaw2-redhat pushed a commit to neuralmagic/nm-vllm that referenced this pull request May 6, 2024

[kernel] fix sliding window in prefix prefill Triton kernel (vllm-pro…

673e4eb

…ject#4405) Co-authored-by: SangBin Cho <rkooo567@gmail.com>

z103cb pushed a commit to z103cb/opendatahub_vllm that referenced this pull request May 7, 2024

[kernel] fix sliding window in prefix prefill Triton kernel (vllm-pro…

df04c10

…ject#4405) Co-authored-by: SangBin Cho <rkooo567@gmail.com>

dtrifiro pushed a commit to opendatahub-io/vllm that referenced this pull request May 7, 2024

[kernel] fix sliding window in prefix prefill Triton kernel (vllm-pro…

9001e2f

…ject#4405) Co-authored-by: SangBin Cho <rkooo567@gmail.com>

dtrifiro mentioned this pull request May 15, 2024

bump ubi base image tag opendatahub-io/vllm#24

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[kernel] fix sliding window in prefix prefill Triton kernel #4405

[kernel] fix sliding window in prefix prefill Triton kernel #4405

mmoskal commented Apr 26, 2024

zhuohan123 commented Apr 26, 2024

rkooo567 commented Apr 27, 2024

rkooo567 left a comment

rkooo567 Apr 30, 2024

rkooo567 Apr 30, 2024

mmoskal May 1, 2024

rkooo567 May 1, 2024

mmoskal May 2, 2024

rkooo567 Apr 30, 2024

mmoskal May 1, 2024

mmoskal May 2, 2024

rkooo567 Apr 30, 2024

mmoskal May 2, 2024

rkooo567 left a comment

mmoskal commented May 2, 2024

rkooo567 commented May 2, 2024

mmoskal commented May 2, 2024

[kernel] fix sliding window in prefix prefill Triton kernel #4405

[kernel] fix sliding window in prefix prefill Triton kernel #4405

Conversation

mmoskal commented Apr 26, 2024

PR Title and Classification

Code Quality

Notes for Large Changes

What to Expect for the Reviews

Thank You

zhuohan123 commented Apr 26, 2024

rkooo567 commented Apr 27, 2024

rkooo567 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

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

rkooo567 left a comment

Choose a reason for hiding this comment

mmoskal commented May 2, 2024

rkooo567 commented May 2, 2024

mmoskal commented May 2, 2024