[Hardware][Intel] Add CPU inference backend #3634

bigPYJ1151 · 2024-03-26T15:16:57Z

This PR adds a new CPU backend to vLLM and supports the basic model inference feature, with BF16 and FP32 dtype. FP16 support and TP support will be added in the future.

Changes to vLLM:

Added VLLM_TARGET_DEVICE ENV to specify backend explicitily.
Added CPUExecutor to isolate CPU backend with others.
Added TorchSDPABackend to support MHA on CPU.
Added _C related kernels on CPU.
Forwarded DeviceConfig to CacheEngine to avoid cuda hardcoded device memory allocation.
Added documents with install instructions.

RFC: #3654

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!

WoosukKwon · 2024-03-26T17:37:47Z

Hi @bigPYJ1151 Thanks for updating the PR! It looks really nice.

Just for other people's understanding, could you write an RFC about the overall design, supported features, key technical decisions, and integration plan? I think this should be easy since you already wrote most of them in the previous PR. Please check out #3620 and #1866 for reference.

ProExpertProg

Had a couple of CMake questions/nits

cmake/cpu_extension.cmake

bigPYJ1151 · 2024-03-27T07:46:52Z

@WoosukKwon Sure, please refer to #3654

hustnn · 2024-03-27T09:18:00Z

Is there initial performance result for cpu reference?

zhouyuan · 2024-03-27T13:47:37Z

Hi @hustnn ,

In general the performance number on CPU is not as good as GPU, for both latency and throughput. However we do find there are two value proposition for vLLM w/ CPU based on our initial tests:

much higher throughput vs naïve/static batching solutions(TGI), this is due to the great throughput oriented design of vLLM
for near-offline inference cases, vLLM w/ CPU throughput performance is competitive vs. entry-level GPU, mostly due to the much larger KV cache space(vs. GPU). The cost for CPU based solution may also be lower.

thanks,
-yuan

mgoin

Overall nice job keeping the complexity in check and working with recent changes. Left a few comments

Dockerfile.cpu

csrc/cpu/pybind.cpp

mgoin · 2024-03-27T20:26:35Z

docs/source/getting_started/cpu-installation.rst

+
+.. code-block:: console
+
+    $ VLLM_TARGET_DEVICE=cpu python setup.py install


Include an example of how to enable bf16 support i.e. through enabling the VLLM_CPU_AVX512BF16 env var

BF16 is always supported in the cpu backend. AVX512BF16 is an extension ISA. The build script will check the host CPU flag to determine whether to enable it. VLLM_CPU_AVX512BF16 is used for the cross-compilation requirement. To clarify this question, I have updated some information in the doc.

hustnn · 2024-03-28T02:49:26Z

In general the performance number on CPU is not as good as GPU, for both latency and throughput. However we do find there are two value proposition for vLLM w/ CPU based on our initial tests:

much higher throughput vs naïve/static batching solutions(TGI), this is due to the great throughput oriented design of vLLM

for near-offline inference cases, vLLM w/ CPU throughput performance is competitive vs. entry-level GPU, mostly due to the much larger KV cache space(vs. GPU). The cost for CPU based solution may also be lower.

@zhouyuan Thanks for your reply, it is very helpful. These 2 points match our requirement quite well. We are planning to integrate a inference operator into a olap database, we care more about throughput compared to latency since we are targeting some offline analysis scenario.

Do you have any suggestion on how should we start with some experiment? Should we wait for these MR to be merged? We want to get some initial number on the throughput and see it is acceptable or any improvement we can further do from DB's aspect.

I also found a article from intel, is it the result and method consistent with your testing?
https://medium.com/@NeuralCompressor/llm-performance-of-intel-extension-for-transformers-f7d061556176

zhouyuan · 2024-03-28T07:04:34Z

In general the performance number on CPU is not as good as GPU, for both latency and throughput. However we do find there are two value proposition for vLLM w/ CPU based on our initial tests:

much higher throughput vs naïve/static batching solutions(TGI), this is due to the great throughput oriented design of vLLM

for near-offline inference cases, vLLM w/ CPU throughput performance is competitive vs. entry-level GPU, mostly due to the much larger KV cache space(vs. GPU). The cost for CPU based solution may also be lower.

@zhouyuan Thanks for your reply, it is very helpful. These 2 points match our requirement quite well. We are planning to integrate a inference operator into a olap database, we care more about throughput compared to latency since we are targeting some offline analysis scenario.

Do you have any suggestion on how should we start with some experiment? Should we wait for these MR to be merged? We want to get some initial number on the throughput and see it is acceptable or any improvement we can further do from DB's aspect.

Hi @hustnn
The dockerfile in this patch maybe a good start to check:
https://github.com/vllm-project/vllm/blob/384623538c081ed621b04c1eec107132920e5045/Dockerfile.cpu

If build successfully, the docker image should be enough to run some benchmarks via the scripts provided in vLLM:
https://github.com/vllm-project/vllm/tree/main/benchmarks

Please note you may need to set some params to do NUMA binding as this may impact the performance for vLLM w/ CPU

For real deployment, vLLM provides the several methods to expose the service endpoint:
https://docs.vllm.ai/en/latest/serving/deploying_with_docker.html
You may then connect your application to the vLLM endpoint via langchain or other soltuons.

I also found a article from intel, is it the result and method consistent with your testing? https://medium.com/@NeuralCompressor/llm-performance-of-intel-extension-for-transformers-f7d061556176

Yes, the performance is improved if using INT4 quantization from Intel extension for transformers.
Intel PyTorch extensions is also a good refence. Here's the link to the project:
https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/performance.html

thanks,
-yuan

WoosukKwon · 2024-03-29T04:34:48Z

@bigPYJ1151 @zhouyuan QQ: Can we use torch.compile to auto-generate the custom C++ kernels except PagedAttention? This would increase the maintainability of the code a lot. I'm wondering how torch.compile performs on Intel CPUs.

bigPYJ1151 · 2024-03-29T05:24:39Z

@WoosukKwon Agree, I think this might be a good direction to try. For these element-wise operations and normalization operations, using torch.compile would unify the front-end to Python code and use different device backends to apply optimizations and generate binary code.

TorchInductor has two IR lowering path:

PyTorch → TorchDynamo → TorchInductor → Triton → NVIDIA GPU
PyTorch → TorchDynamo → TorchInductor → OpenMP (C++) → CPU

The second path is designed for CPU, and is under active development and evolution. Here is a blog contains some examples for your reference.

We need to further check the current development status and any gaps to utilize it to vLLM.

WoosukKwon

@bigPYJ1151 Thanks for the great PR! And apologies for the delay in the review.

I'm happy to see that the code is super clean and fits well into our new abstractions. I also appreciate that you provided detailed documents and instructions. My experience of building and running the CPU backend was indeed super smooth.

The only major design decision I'd like to discuss on this PR is whether we want to share ModelRunner between the GPU and CPU backends. While I do understand that sharing will save code duplication substantially, I'm a bit worried that any change in ModelRunner can break CPU backend. Also, sharing can over-complicate the development of the CPU backend since the CPU backend doesn't use CUDA graph, prefix caching, etc. That being said, for faster initial integration, we can merge the current PR and change this in the future PR. WDYT?

Dockerfile.cpu

vllm/attention/backends/torch_sdpa.py

docs/source/getting_started/cpu-installation.rst

vllm/executor/cpu_executor.py

vllm/worker/cache_engine.py

vllm/worker/cpu_worker.py

WoosukKwon · 2024-03-31T22:56:52Z

vllm/worker/cpu_worker.py

+logger = init_logger(__name__)
+
+
+class CPUModelRunner(ModelRunner):


Now I understand the context of the previous PR that changed the hardcoded "cuda" to self.device. I actually thought the CPU backend would have its own model runner. And I still believe it can be a better design since it isolates the two backends from each other.

What do you think? As CUDA graph is not used for CPUs, I believe the amount of the code duplicated when defining a new model runner for CPUs would not be that large.

Yes, separating CPUModelRunner entirely will avoid all potential code breaks. It needs to move hundreds of lines codes, we will do it in the future PRs.

Could you please create an issue about this?

Sure, refer to #3776

bigPYJ1151 · 2024-04-01T06:40:13Z

@WoosukKwon Thanks for your comments! I have fixed most of them.
For CPUModelRunner, yes, you are right, isolate it with ModelRunner will avoid potential code breaks completely. We can do it in the future to reduce the PR size.

docs/source/getting_started/cpu-installation.rst

vllm/worker/worker.py

vllm/worker/cpu_worker.py

vllm/executor/cpu_executor.py

WoosukKwon · 2024-04-01T08:35:36Z

vllm/engine/arg_utils.py

@@ -363,6 +364,10 @@ def add_cli_args(
            default=False,
            help='If True, the prefill requests can be chunked based on the '
            'max_num_batched_tokens')
+        parser.add_argument('--cpu-kvcache-space',


QQ: Can we somehow make this an argument only for CPU backends?

Right, this flag is too specific.

It is difficult to find a proper method using CLI arguments. I tried to change this parameter as an environment variable and highlight it using warning message when it is not set. Please look at the latest commit.

vllm/executor/cpu_executor.py

WoosukKwon · 2024-04-01T08:42:31Z

docs/source/getting_started/cpu-installation.rst

+
+    - AVX512_BF16 is an extension ISA provides native BF16 data type conversion and vector product instructions, will brings some performance improvement compared with pure AVX512. The CPU backend build script will check the host CPU flags to determine whether to enable AVX512_BF16. 
+
+    - If you want to force enable AVX512_BF16 for the cross-compilation, please set environment variable VLLM_CPU_AVX512BF16=1 before the building.    


Dumb question: Can we always set this env var? When should the user set this flag off?

No need. This is because the AVX512BF16 is not widely supported on older CPU models.

If users want to run vLLM on the host, the build script will check the host CPU flag to determine the compile flag.

If users want to build vLLM as a package on the host without AVX512BF16, and run vLLM on other machines with AVX512BF16, this env var should be enabled for this cross-compilation.

vllm/attention/backends/torch_sdpa.py

WoosukKwon · 2024-04-01T08:43:47Z

vllm/worker/cpu_worker.py

+logger = init_logger(__name__)
+
+
+class CPUModelRunner(ModelRunner):


Could you please create an issue about this?

vllm/worker/cpu_worker.py

bigPYJ1151 · 2024-04-01T10:51:06Z

Hi @WoosukKwon Thanks for your further comments. I have fixed them all, please check, thanks.

Co-authored-by: Kunshang Ji <kunshang.ji@intel.com> Co-authored-by: Yuan Zhou <yuan.zhou@intel.com>

hustnn · 2024-04-02T06:23:27Z

hi @zhouyuan @bigPYJ1151 , Thanks for your great work! Do you have plan to public some results on cpu inference with vllm?

zhouyuan · 2024-04-02T13:17:32Z

@bigPYJ1151 @zhouyuan @jikunshang LGTM! Huge thanks for the great work! Very excited to finally have the CPU backend.

Please continue to work on investigating torch.compile and separating out ModelRunner for CPUs. These will increase the maintainability of the backend. Thanks again for the great work!

@WoosukKwon thanks for the detailed review and much appreciated on your guidance! Sure, will follow up on the refactoring, features and performance optimizations.

thanks,
-yuan

zhouyuan · 2024-04-02T13:21:32Z

hi @zhouyuan @bigPYJ1151 , Thanks for your great work! Do you have plan to public some results on cpu inference with vllm?

Hi @hustnn,

We are now reviewing/seeking approval for the performance data, should be able to publish the perf data soon if everything goes well.

thanks,
-yuan

hustnn · 2024-04-02T13:57:39Z

hi @zhouyuan @bigPYJ1151 , Thanks for your great work! Do you have plan to public some results on cpu inference with vllm?

Hi @hustnn,

We are now reviewing/seeking approval for the performance data, should be able to publish the perf data soon if everything goes well.

thanks, -yuan

@zhouyuan Thanks and look forward to the performance report.

Using importlib version() function doesn't work in the docker image. Introduced in vllm-project#3634 Equivalent fix to vllm-project#3735

abhilash1910 · 2024-04-03T05:39:51Z

Great work @bigPYJ1151 @zhouyuan @jikunshang !
Looking forward to #3814 .
Thanks @WoosukKwon for awesome collaboration!.

markluofd · 2024-04-03T06:41:37Z

Will intel cpu backend be equipped with AsyncLLMEngine? api_server is currently using AsyncLLMEngine

bigPYJ1151 · 2024-04-03T09:06:21Z

Hi @markluofd the online inference of the CPU backend is still under tunning, we will enable it when it is ready.

markluofd · 2024-04-03T09:19:23Z

@bigPYJ1151 To imitate the code of https://github.com/vllm-project/vllm/pull/3814/files#diff-d1c5ec4ddd588e3b7cac13bde85a98ac5b20686dc16b9da3b1c324467c3be2b5 url, I added CPUExecutorAsync in cpu_executor.py. AsyncLLMEngine and api_server can work normally. I want to know if this method will affect the inference performance of the CPU.

bigPYJ1151 · 2024-04-03T09:34:05Z

@markluofd Yes, the performance may have some regression. Because the CPU inference thread pool(OpenMP), HTTP service thread pool, and tokenizer threads will scramble CPU cores.
We plan to isolate the inference thread pool from others to avoid this problem.

markluofd · 2024-04-03T09:40:50Z

@bigPYJ1151 ok, thanks！There is another question , the introduction says that only bf16 and fp32 are supported. I found that if the dtype is fp16, the CPU backend can also execute normally. I want to know whether it uses the fp16 or fp32 kernel ？(My machine is a 3rd generation cpu with no avx512_bf16 instruction)

bigPYJ1151 · 2024-04-03T09:46:15Z

@markluofd FP16 will be cast to BF16 right now. BF16 is always supported even if there is no avx512_bf16 ISA.
Pure FP16 support will be added soon, might be at the end of the month.

markluofd · 2024-04-03T09:55:59Z

@bigPYJ1151 I get it, thank you! I need to find a 4th generation CPU and test the performance of bf16. I also found that the overall utilization rate of the current 96-core CPU is about 16% (30 concurrent requests). I hope that subsequent features will be incorporated to bring higher CPU utilization.

ProExpertProg · 2024-04-03T21:41:57Z

@bigPYJ1151 @WoosukKwon I did not want to slow down the merging of this PR but I was wondering if there's a plan to decouple the CPU backend via an additional level of abstraction to allow for choosing the backend at runtime (or startup time) as opposed to the build time. I'm guessing all calls that go to the _c pytorch extension would have to go through the worker/executor but it doesn't seem like it would be too complex?

zhouyuan · 2024-04-04T01:07:30Z

@bigPYJ1151 I get it, thank you! I need to find a 4th generation CPU and test the performance of bf16. I also found that the overall utilization rate of the current 96-core CPU is about 16% (30 concurrent requests). I hope that subsequent features will be incorporated to bring higher CPU utilization.

Hi @markluofd
Thanks for reporting, besides the threading pool conflict issue posted by @bigPYJ1151, in CPU based env you may need to do several tunings to get better performance, especially on NUMA node access and OpenMP threads.

OMP_NUM_THREADS=32 numactl --physcpubind=0-31 --membind=0 python benchmark.py

In my 4th gen Xeon env, this tuning can bring ~30% perf improvement. I suppose this can also help to improve the resource utilization in your tests.

Here are some tuning recipe for CPU based env:
https://pytorch.org/tutorials/intermediate/torchserve_with_ipex.html
https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/performance_tuning/tuning_guide.html

thanks,
-yuan

bigPYJ1151 · 2024-04-04T07:23:26Z

Hi @ProExpertProg It is feasible to load different backends dylib at runtime. vLLM has multple backends with different dependencies and configurations, so it might be a lot of works to support the runtime binding.

Co-authored-by: Kunshang Ji <kunshang.ji@intel.com> Co-authored-by: Yuan Zhou <yuan.zhou@intel.com>

bigPYJ1151 force-pushed the cpu_pr branch 3 times, most recently from de3232c to 02ec530 Compare March 26, 2024 15:58

WoosukKwon added the x86 CPU label Mar 26, 2024

WoosukKwon self-assigned this Mar 26, 2024

ProExpertProg reviewed Mar 26, 2024

View reviewed changes

cmake/cpu_extension.cmake Outdated Show resolved Hide resolved

cmake/cpu_extension.cmake Show resolved Hide resolved

bigPYJ1151 mentioned this pull request Mar 27, 2024

[RFC] Initial Support for CPUs #3654

Open

4 tasks

mgoin reviewed Mar 27, 2024

View reviewed changes

zhouyuan mentioned this pull request Mar 29, 2024

[Feature] Prototype of vLLM execution on CPU-only devices #1028

Closed

WoosukKwon self-requested a review March 29, 2024 04:18

youkaichao mentioned this pull request Mar 29, 2024

[Installation]: failed on WSL windows 10 CPU machine #3721

Closed

WoosukKwon reviewed Mar 31, 2024

View reviewed changes

bigPYJ1151 force-pushed the cpu_pr branch 2 times, most recently from df6a6c1 to 153e239 Compare April 1, 2024 06:14

WoosukKwon reviewed Apr 1, 2024

View reviewed changes

WoosukKwon added the action-required label Apr 1, 2024

bigPYJ1151 mentioned this pull request Apr 1, 2024

[Misc]: Isolate CPUModelRunner and ModelRunner for better maintenance #3776

Closed

bigPYJ1151 and others added 3 commits April 1, 2024 11:51

Add CPU CMake extension.

46cd747

Integrate CPU module.

4060d26

Co-authored-by: Kunshang Ji <kunshang.ji@intel.com> Co-authored-by: Yuan Zhou <yuan.zhou@intel.com>

Refactor attention

4de2017

bigPYJ1151 added 3 commits April 2, 2024 08:07

Add CI

32f9521

Add CI

44df554

Add CI

738bd74

njhill mentioned this pull request Apr 2, 2024

[BugFix] Use different mechanism to get vllm version in is_cpu() #3804

Merged

xwu99 mentioned this pull request Apr 3, 2024

Install vLLM from upstream as CPU backend was merged. intel/llm-on-ray#172

Closed

jikunshang mentioned this pull request Apr 5, 2024

[Hardware][Intel GPU]Add Initial Intel GPU(XPU) inference backend #3814

Merged

This was referenced Apr 18, 2024

Any plan to support cpu only mode? #176

Closed

Inquiry Regarding vLLM Support for Mac Metal API #2081

Closed

[RFC] Speedup vLLM inference with Intel@ Extension for PyTorch* #2526

Open

skonto mentioned this pull request Apr 29, 2024

[Bug]: TypeError in XFormersMetadata #4399

Open

dtrifiro mentioned this pull request May 15, 2024

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

Merged

DamonFool mentioned this pull request Jun 12, 2024

[Hardware][Intel] Support CPU inference with AVX2 ISA #5452

Merged

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

[Hardware][Intel] Add CPU inference backend (vllm-project#3634)

67a2ece

Co-authored-by: Kunshang Ji <kunshang.ji@intel.com> Co-authored-by: Yuan Zhou <yuan.zhou@intel.com>


		.. code-block:: console

		$ VLLM_TARGET_DEVICE=cpu python setup.py install

		logger = init_logger(__name__)


		class CPUModelRunner(ModelRunner):


		- AVX512_BF16 is an extension ISA provides native BF16 data type conversion and vector product instructions, will brings some performance improvement compared with pure AVX512. The CPU backend build script will check the host CPU flags to determine whether to enable AVX512_BF16.

		- If you want to force enable AVX512_BF16 for the cross-compilation, please set environment variable VLLM_CPU_AVX512BF16=1 before the building.

[Hardware][Intel] Add CPU inference backend #3634

[Hardware][Intel] Add CPU inference backend #3634

Conversation

bigPYJ1151 commented Mar 26, 2024 • edited Loading

PR Title and Classification

Code Quality

Notes for Large Changes

What to Expect for the Reviews

Thank You

WoosukKwon commented Mar 26, 2024

ProExpertProg left a comment

Choose a reason for hiding this comment

bigPYJ1151 commented Mar 27, 2024

hustnn commented Mar 27, 2024

zhouyuan commented Mar 27, 2024

mgoin left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

hustnn commented Mar 28, 2024 • edited Loading

zhouyuan commented Mar 28, 2024

WoosukKwon commented Mar 29, 2024

bigPYJ1151 commented Mar 29, 2024

WoosukKwon 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

bigPYJ1151 commented Apr 1, 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

bigPYJ1151 commented Apr 1, 2024 • edited Loading

hustnn commented Apr 2, 2024

zhouyuan commented Apr 2, 2024

zhouyuan commented Apr 2, 2024

hustnn commented Apr 2, 2024

abhilash1910 commented Apr 3, 2024 • edited Loading

markluofd commented Apr 3, 2024

bigPYJ1151 commented Apr 3, 2024

markluofd commented Apr 3, 2024

bigPYJ1151 commented Apr 3, 2024

markluofd commented Apr 3, 2024

bigPYJ1151 commented Apr 3, 2024

markluofd commented Apr 3, 2024

ProExpertProg commented Apr 3, 2024

zhouyuan commented Apr 4, 2024

bigPYJ1151 commented Apr 4, 2024

bigPYJ1151 commented Mar 26, 2024 •

edited

Loading

hustnn commented Mar 28, 2024 •

edited

Loading

bigPYJ1151 commented Apr 1, 2024 •

edited

Loading

abhilash1910 commented Apr 3, 2024 •

edited

Loading