[Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel #7766

dsikka · 2024-08-22T01:12:50Z

Summary

Expands weight loading to support grouped and per channel weight quantization. Cleans-up fp8 MoE to use the updated weight loading
Adds Marlin Fused MoE Kernel for w4a16 by @ElizaWszola
Adds CompressedTensorsMoEMethod to support MoE w4a16 models from llm-compressor and compressed-tensors
Tested using 2 and 4 TP with Mixtral

Next Steps:

The CompressedTensorsMoEMethod is not leveraging the scheme structure in-place for compressed-tensors in order to keep the scope of this PR focused on the kernel + updated weight loading. Will be updated in a follow-up to use the scheme structure

co-authored by @ElizaWszola, from Neural Magic

github-actions · 2024-08-22T01:13:00Z

👋 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.

🚀

tests/quantization/test_compressed_tensors.py

dsikka · 2024-08-23T18:27:19Z

/ready

halexan · 2024-08-27T03:47:58Z

W4A16, the "4" is int4? or fp4?

dsikka · 2024-08-27T09:39:32Z

W4A16, the "4" is int4? or fp4?

Int4

FILL IN THE PR DESCRIPTION HERE FIX #xxxx (*link existing issues this PR will resolve*) **BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE** --- <details>  <summary><b> PR Checklist (Click to Expand) </b></summary> <p>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.</p> <h3>PR Title and Classification</h3> <p>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:</p> <ul> <li><code>[Bugfix]</code> for bug fixes.</li> <li><code>[CI/Build]</code> for build or continuous integration improvements.</li> <li><code>[Doc]</code> for documentation fixes and improvements.</li> <li><code>[Model]</code> for adding a new model or improving an existing model. Model name should appear in the title.</li> <li><code>[Frontend]</code> For changes on the vLLM frontend (e.g., OpenAI API server, <code>LLM</code> class, etc.) </li> <li><code>[Kernel]</code> for changes affecting CUDA kernels or other compute kernels.</li> <li><code>[Core]</code> for changes in the core vLLM logic (e.g., <code>LLMEngine</code>, <code>AsyncLLMEngine</code>, <code>Scheduler</code>, etc.)</li> <li><code>[Hardware][Vendor]</code> for hardware-specific changes. Vendor name should appear in the prefix (e.g., <code>[Hardware][AMD]</code>).</li> <li><code>[Misc]</code> for PRs that do not fit the above categories. Please use this sparingly.</li> </ul> <p><strong>Note:</strong> If the PR spans more than one category, please include all relevant prefixes.</p> <h3>Code Quality</h3> <p>The PR need to meet the following code quality standards:</p> <ul> <li>We adhere to <a href="https://google.github.io/styleguide/pyguide.html">Google Python style guide</a> and <a href="https://google.github.io/styleguide/cppguide.html">Google C++ style guide</a>.</li> <li>Pass all linter checks. Please use <a href="https://github.com/vllm-project/vllm/blob/main/format.sh"><code>format.sh</code></a> to format your code.</li> <li>The code need to be well-documented to ensure future contributors can easily understand the code.</li> <li>Include sufficient tests to ensure the project to stay correct and robust. This includes both unit tests and integration tests.</li> <li>Please add documentation to <code>docs/source/</code> if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.</li> </ul> <h3>Notes for Large Changes</h3> <p>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 <code>rfc-required</code> and might not go through the PR.</p> <h3>What to Expect for the Reviews</h3> <p>The goal of the vLLM team is to be a <i>transparent reviewing machine</i>. 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: </p> <ul> <li> 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.</li> <li> 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.</li> <li> After the review, the reviewer will put an <code> action-required</code> label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.</li> <li> 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. </li> </ul> <h3>Thank You</h3> <p> 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! </p> </details>

mgoin · 2024-08-27T18:27:08Z

tests/weight_loading/models.txt

+compressed-tensors, nm-testing/Mixtral-8x7B-Instruct-v0.1-W4A16-quantized, main
+compressed-tensors, nm-testing/Mixtral-8x7B-Instruct-v0.1-W4A16-channel-quantized, main
 awq, casperhansen/mixtral-instruct-awq, main
 awq_marlin, casperhansen/mixtral-instruct-awq, main


Future work: these mixtral models seem quite large to have in this test, maybe we should have a small and large test

Maximilianxu · 2024-08-28T08:47:20Z

@dsikka @mgoin Hi, is there a UT for this kernel? I saw the API has differences from Triton-based FusedMoE.

…m-project#7766) Co-authored-by: ElizaWszola <eliza@neuralmagic.com>

B-201 · 2024-08-29T03:52:58Z

@dsikka Hi! Thanks for your work. Do you have plans to support gptq models in the future?

dsikka · 2024-08-29T19:07:59Z

@dsikka Hi! Thanks for your work. Do you have plans to support gptq models in the future?

Yup, this is in-scope to be worked on

binxuan · 2024-08-29T21:32:44Z

Hi, thanks for the work! I am wondering does this MoE kernel work on A100 GPU?

mgoin · 2024-08-29T21:52:22Z

@binxuan Yes it is supported on A100 (SM 8.0 and up)

binxuan · 2024-08-29T23:04:40Z

Thanks for confirming. I tried the mainline code, but got following error. I think the device name returned from get_device_name somehow was bytes instead of str.

device_name = current_platform.get_device_name().replace(" ", "_")
TypeError: a bytes-like object is required, not 'str'

After fix this issue, got a second error from triton mentioning

AssertionError: fp8e4nv data type is not supported on CUDA arch < 89

fengyang95 · 2024-08-30T08:33:55Z

If I want to use it with deepseek-v2, I saw that it uses fused_moe by default. Do I need to swap it out to get it running?

dsikka · 2024-08-30T19:52:54Z

If I want to use it with deepseek-v2, I saw that it uses fused_moe by default. Do I need to swap it out to get it running?

Hi @fengyang95 - are you trying to run a W4A16 deepseek-v2 model?

fengyang95 · 2024-08-31T02:59:26Z

If I want to use it with deepseek-v2, I saw that it uses fused_moe by default. Do I need to swap it out to get it running?

Hi @fengyang95 - are you trying to run a W4A16 deepseek-v2 model?

@dsikka YES, I am using the latest code, which seems to be using marlin_moe, but I encountered the following error.

  final_hidden_states = self.quant_method.apply(
  File "/usr/local/lib/python3.9/dist-packages/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py", line 275, in apply
    return fused_marlin_moe(x,
  File "/usr/local/lib/python3.9/dist-packages/vllm/model_executor/layers/fused_moe/fused_moe.py", line 519, in fused_marlin_moe
    sorted_token_ids, _, _ = moe_align_block_size(topk_ids, block_size_m, E)
  File "/usr/local/lib/python3.9/dist-packages/vllm/model_executor/layers/fused_moe/fused_moe.py", line 228, in moe_align_block_size
    ops.moe_align_block_size(topk_ids, num_experts, block_size, sorted_ids,
  File "/usr/local/lib/python3.9/dist-packages/vllm/_custom_ops.py", line 29, in wrapper
    return fn(*args, **kwargs)
  File "/usr/local/lib/python3.9/dist-packages/vllm/_custom_ops.py", line 538, in moe_align_block_size
    torch.ops._C.moe_align_block_size(topk_ids, num_experts, block_size,
  File "/usr/local/lib/python3.9/dist-packages/torch/_ops.py", line 1061, in __call__
    return self_._op(*args, **(kwargs or {}))
RuntimeError: CUDA error: invalid argument
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

fengyang95 · 2024-09-03T01:35:47Z

Using H20 (sm_90) can start normally; is it also because it is currently not compatible with L40 (sm_89)? @dsikka

…m-project#7766) Co-authored-by: ElizaWszola <eliza@neuralmagic.com>

…m-project#7766) Co-authored-by: ElizaWszola <eliza@neuralmagic.com> Signed-off-by: Alvant <alvasian@yandex.ru>

…m-project#7766) Co-authored-by: ElizaWszola <eliza@neuralmagic.com>

dsikka marked this pull request as ready for review August 22, 2024 21:29

dsikka commented Aug 23, 2024

View reviewed changes

tests/quantization/test_compressed_tensors.py Outdated Show resolved Hide resolved

github-actions bot added the ready ONLY add when PR is ready to merge/full CI is needed label Aug 23, 2024

halexan mentioned this pull request Aug 25, 2024

MODEL REQUESTS vllm-project/llm-compressor#69

Closed

dsikka and others added 22 commits August 27, 2024 12:55

clean-up model weight loading; support channel wise weight scales

c2a48a8

update fp8 MoE to use updated weight loading

47c4bd2

add group quant weight load support

d3f63dd

clean-up

99f0630

more clean-up

905cda4

update condition for arch

defd71f

update deepseek and qwen to skip weight_shape loading

52a08f0

comments

f408a26

format

9c855af

fix issue on upstream-main

09881ac

fix imports for tests

7b7b36a

PR comments + fix import issue

40c8dc0

format

4df0b52

Michael's feedback

1465a2b

testing

9409f67

Assert no fp8 in fused moe

162e579

Don't compile marlin_moe_ops on ROCm

e5a6724

clean-up model weight loading; support channel wise weight scales

05f37be

update fp8 MoE to use updated weight loading

eff08ba

add group quant weight load support

2c310c0

dsikka and others added 7 commits August 27, 2024 13:15

move to weight_loading test to deal with CI cuda memory issues

a5420e4

Disable fused Mixtral on AMD

cfbc594

format

e31a76c

Bring back missing test

3b2fcff

Remove fallback

24a3b57

remove redundant test

58ca798

rebase fix

eb72c6a

dsikka force-pushed the compressed-tensors-moe-updated branch from e59a1bd to eb72c6a Compare August 27, 2024 13:42

mgoin approved these changes Aug 27, 2024

View reviewed changes

mgoin changed the title ~~[Kerne] Expand MoE weight loading + Add Fused Marlin MoE Kernel~~ [Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel Aug 27, 2024

WoosukKwon merged commit fc91188 into vllm-project:main Aug 27, 2024
60 of 64 checks passed

kushanam pushed a commit to kushanam/vllm that referenced this pull request Aug 28, 2024

[Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel (vll…

128a871

…m-project#7766) Co-authored-by: ElizaWszola <eliza@neuralmagic.com>

kushanam pushed a commit to kushanam/vllm that referenced this pull request Aug 28, 2024

[Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel (vll…

cc3cddc

…m-project#7766) Co-authored-by: ElizaWszola <eliza@neuralmagic.com>

jeejeelee mentioned this pull request Aug 30, 2024

[Model] Add Support for GPTQ Fused MOE #6502

Open

triple-Mu pushed a commit to triple-Mu/vllm_official that referenced this pull request Sep 4, 2024

[Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel (vll…

89ec134

…m-project#7766) Co-authored-by: ElizaWszola <eliza@neuralmagic.com>

Jeffwan pushed a commit to aibrix/vllm that referenced this pull request Sep 19, 2024

[Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel (vll…

8a5e53d

…m-project#7766) Co-authored-by: ElizaWszola <eliza@neuralmagic.com>

donpromax mentioned this pull request Sep 20, 2024

Qwen1.5-MoE-A2.7B-Chat w4a16 Quantization Failed vllm-project/llm-compressor#189

Closed

Alvant pushed a commit to compressa-ai/vllm that referenced this pull request Oct 26, 2024

[Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel (vll…

3cea323

…m-project#7766) Co-authored-by: ElizaWszola <eliza@neuralmagic.com> Signed-off-by: Alvant <alvasian@yandex.ru>

KuntaiDu pushed a commit to KuntaiDu/vllm that referenced this pull request Nov 20, 2024

[Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel (vll…

a675fca

…m-project#7766) Co-authored-by: ElizaWszola <eliza@neuralmagic.com>

AlpinDale mentioned this pull request Dec 21, 2024

feat: add fused Marlin MoE kernel aphrodite-engine/aphrodite-engine#934

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel #7766

[Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel #7766

dsikka commented Aug 22, 2024 •

edited

Loading

github-actions bot commented Aug 22, 2024

dsikka commented Aug 23, 2024

halexan commented Aug 27, 2024

dsikka commented Aug 27, 2024

mgoin Aug 27, 2024

Maximilianxu commented Aug 28, 2024

B-201 commented Aug 29, 2024

dsikka commented Aug 29, 2024

binxuan commented Aug 29, 2024

mgoin commented Aug 29, 2024

binxuan commented Aug 29, 2024

fengyang95 commented Aug 30, 2024 •

edited

Loading

dsikka commented Aug 30, 2024

fengyang95 commented Aug 31, 2024 •

edited

Loading

fengyang95 commented Sep 3, 2024

[Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel #7766

[Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel #7766

Conversation

dsikka commented Aug 22, 2024 • edited Loading

Summary

github-actions bot commented Aug 22, 2024

dsikka commented Aug 23, 2024

halexan commented Aug 27, 2024

dsikka commented Aug 27, 2024

mgoin Aug 27, 2024

Choose a reason for hiding this comment

Maximilianxu commented Aug 28, 2024

B-201 commented Aug 29, 2024

dsikka commented Aug 29, 2024

binxuan commented Aug 29, 2024

mgoin commented Aug 29, 2024

binxuan commented Aug 29, 2024

fengyang95 commented Aug 30, 2024 • edited Loading

dsikka commented Aug 30, 2024

fengyang95 commented Aug 31, 2024 • edited Loading

fengyang95 commented Sep 3, 2024

dsikka commented Aug 22, 2024 •

edited

Loading

fengyang95 commented Aug 30, 2024 •

edited

Loading

fengyang95 commented Aug 31, 2024 •

edited

Loading