[Kernel] Add prefix-caching support for phi-3-small-8k/128k model triton kernel #8345

congcongchen123 · 2024-09-10T21:44:38Z

This PR is a feature enrichment of #4799 which introduces blocksparse flash attention, Microsoft Phi-3-Small-8K and Phi-3-Small-128K models.

This PR modifies the block-sparse attention prefill Triton kernel to add prefix-caching support.

Tested utilizing tool: benchmarks/benchmark_prefix_caching.py --model=microsoft/Phi-3-small-8k-instruct --enable-prefix-caching with different prompts, and verified that the output is correct.
Added unit test.
Benchmark phi3-small 8k/128k model with prefix-caching enabled, and verified that it has achieved significant TTFT latency gain with different prefix-cache hit rates.

Thanks for offline discussion @linxihui, @wschin !

FIX #xxxx (link existing issues this PR will resolve)

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!

github-actions · 2024-09-10T21:44:50Z

👋 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 starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

Add ready label to the PR
Enable auto-merge.

🚀

wschin · 2024-09-12T19:31:57Z

@comaniac, would you mind find someone to review this PR? This is an important model. If reviewer feels reviewing kernel takes too long, maybe requesting more numerical tests and profiling result? Many thanks.

comaniac · 2024-09-12T20:29:55Z

@comaniac, would you mind find someone to review this PR? This is an important model. If reviewer feels reviewing kernel takes too long, maybe requesting more numerical tests and profiling result? Many thanks.

I'll review it ASAP. @mgoin @tlrmchlsmth @pcmoritz could any of you also help review the kernel implementation?
Meanwhile, numerical tests are definitely required, and do you need to tune the kernel configs offline like we did for the MoE kernels?

wschin · 2024-09-12T23:16:43Z

vllm/attention/ops/blocksparse_attention/blocksparse_attention_kernel.py

@@ -3,6 +3,407 @@
 import triton.language as tl


+@triton.jit
+def _context_fwd_kernel_inner(


It's better to write done the math formulation implemented by this kernel. The same for other kernel functions.

wschin · 2024-09-12T23:18:25Z

vllm/attention/ops/blocksparse_attention/blocksparse_attention_kernel.py

+
+
+@torch.inference_mode()
+def context_blocksparse_flash_attn_varlen_fwd(


If math equation is added, people can improve and replace it easier in the future.

tlrmchlsmth

I made a pass through the PR. I left a couple of superficial comments and had a question on CUDA graphs.

Also congcongchen123 looks like you need to run bash format.sh

tlrmchlsmth · 2024-09-12T21:01:15Z

vllm/attention/ops/blocksparse_attention/interface.py

@@ -49,6 +51,10 @@ def __init__(
        self.use_spda = use_spda
        self.dtype = dtype
        self.device = device
+        # block size used for blocksparse attention, used in `local_blocks`,
+        # `vert_stride`.
+        # It is different from kv_cache block size which is he size of of a


nit: "he size of of" -> "the size of"

Actually can we avoid use block_size to reduce confusion? And how this size usually be set? Is it set via LLM engine config or a predefined constant?

tlrmchlsmth · 2024-09-13T14:25:27Z

vllm/attention/ops/blocksparse_attention/blocksparse_attention_kernel.py

+    query_start_block_id = (context_lens_tensor // sparse_block_size).cpu()
+    query_end_block_id = ((seq_lens_tensor - 1) // sparse_block_size).cpu()


Does moving these to the CPU break CUDA graphs?

tlrmchlsmth · 2024-09-13T14:28:20Z

vllm/attention/ops/blocksparse_attention/blocksparse_attention_kernel.py

+        k_cache.stride(2),
+        k_cache.stride(3),
+        k_cache.stride(
+            4),  #[num_blocks, num_kv_heads, head_size/x, block_size, x]


nit: Could you take a look at the comments here? It looks like the formatter made this a little hard to read

tlrmchlsmth · 2024-09-13T17:35:16Z

vllm/attention/ops/blocksparse_attention/blocksparse_attention_kernel.py

+                context_len,
+                offs_n)
+    # flash-attn 2
+    # m_i += tl.math.log2(l_i)


is this debug cruft?

tlrmchlsmth · 2024-09-13T17:44:08Z

vllm/attention/ops/blocksparse_attention/blocksparse_attention_kernel.py

+            k = tl.load(K_cache + off_k,
+                        mask=(start_n + offs_n[None, :]) < context_len,
+                        other=0.0)  # [D,N]


based on a lack of masking along N here, looks like we require that the sparse block size divides the hidden dim? I think we should assert that somewhere if we don't already. I acknowledge this is paranoia.

comaniac · 2024-09-13T18:20:24Z

vllm/attention/ops/blocksparse_attention/interface.py

@@ -49,6 +51,10 @@ def __init__(
        self.use_spda = use_spda
        self.dtype = dtype
        self.device = device
+        # block size used for blocksparse attention, used in `local_blocks`,
+        # `vert_stride`.
+        # It is different from kv_cache block size which is he size of of a


Actually can we avoid use block_size to reduce confusion? And how this size usually be set? Is it set via LLM engine config or a predefined constant?

comaniac · 2024-09-13T18:20:47Z

vllm/attention/ops/blocksparse_attention/interface.py

+        context_lens_tensor: torch.Tensor,
+        sm_scale: float,
+    ):
+        """


Can you also add descriptions about what this function does?

comaniac · 2024-09-13T18:21:49Z

vllm/attention/ops/blocksparse_attention/interface.py

+        query, key, value: shape = (num_tokens, num_heads_q/kv, head_size).
+                Support grouped attention, with `q[:, i*r:(i*r + r)]`
+                is correspondent to `k[:, i]`, where `r` is the q/k ratio.
+        key_cache: shape = (num_blocks, num_kv_heads, 
+                            head_size // x, kv_cache_block_size, x),
+                            where x is defined in paged_attn.py.
+        value_cache: shape = (num_blocks, num_kv_heads, head_size, 
+                             kv_cache_block_size).
+        block_tables: shape = (batch_size, num_blocks).
+        query_start_loc: shape = (batch_size + 1,). 
+                        The cumulative subquery lengths of the sequences in 
+                        the batch, used to index into subquery. E.g., if 
+                        the subquery length is [4, 6], it is [0, 4, 10].
+        seq_lens_tensor: shape=(batch_size + 1,).
+                         The sequence length per sequence.
+        context_lens_tensor: shape=(batch_size,).
+                            The context length per sequence (tokens that are 
+                            computed so far).
+        sm_scale: softmax scale, default to 1/sqrt(head_size).
+
+        return: tensor of shape as q.


Suggested change

query, key, value: shape = (num_tokens, num_heads_q/kv, head_size).

Support grouped attention, with `q[:, i*r:(i*r + r)]`

is correspondent to `k[:, i]`, where `r` is the q/k ratio.

key_cache: shape = (num_blocks, num_kv_heads,

head_size // x, kv_cache_block_size, x),

where x is defined in paged_attn.py.

value_cache: shape = (num_blocks, num_kv_heads, head_size,

kv_cache_block_size).

block_tables: shape = (batch_size, num_blocks).

query_start_loc: shape = (batch_size + 1,).

The cumulative subquery lengths of the sequences in

the batch, used to index into subquery. E.g., if

the subquery length is [4, 6], it is [0, 4, 10].

seq_lens_tensor: shape=(batch_size + 1,).

The sequence length per sequence.

context_lens_tensor: shape=(batch_size,).

The context length per sequence (tokens that are

computed so far).

sm_scale: softmax scale, default to 1/sqrt(head_size).

return: tensor of shape as q.

Args:

query, key, value: shape = (num_tokens, num_heads_q/kv, head_size).

Support grouped attention, with `q[:, i*r:(i*r + r)]`

is correspondent to `k[:, i]`, where `r` is the q/k ratio.

key_cache: shape = (num_blocks, num_kv_heads,

head_size // x, kv_cache_block_size, x),

where x is defined in paged_attn.py.

value_cache: shape = (num_blocks, num_kv_heads, head_size,

kv_cache_block_size).

block_tables: shape = (batch_size, num_blocks).

query_start_loc: shape = (batch_size + 1,).

The cumulative subquery lengths of the sequences in

the batch, used to index into subquery. E.g., if

the subquery length is [4, 6], it is [0, 4, 10].

seq_lens_tensor: shape=(batch_size + 1,).

The sequence length per sequence.

context_lens_tensor: shape=(batch_size,).

The context length per sequence (tokens that are

computed so far).

sm_scale: softmax scale, default to 1/sqrt(head_size).

Return: tensor of shape as q.

comaniac · 2024-09-13T18:23:08Z

vllm/attention/ops/blocksparse_attention/interface.py

+
+        return: tensor of shape as q.
+        """
+        assert (not self.use_spda), "forward_prefix does not support spda"


Suggested change

assert (not self.use_spda), "forward_prefix does not support spda"

if self.use_spda:

raise ValueError("Prefix caching with block sparse attention does not support SPDA")

comaniac · 2024-09-13T18:25:15Z

vllm/attention/ops/blocksparse_attention/interface.py

+        assert (
+            self.block_size == self.q_block_size
+            or self.q_block_size is None), "Different block size not supported"


Suggested change

assert (

self.block_size == self.q_block_size

or self.q_block_size is None), "Different block size not supported"

if self.block_size != self.q_block_size and self.q_block_size is not None:

raise NotImplementedError("Different block sizes are not supported")

comaniac · 2024-09-13T18:26:49Z

tests/kernels/test_blocksparse_attention.py

+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(0)
+    torch.set_default_device(device)
+
+    # Need this, otherwise when we capture the graph the process
+    # for GPU 1 would run on both GPU0 and GPU1 and things would hang
+    #
+    # see also similar issue: https://github.com/Dao-AILab/flash-attention/issues/523
+    torch.cuda.set_device(device)


Do we have a better for to deal with this in tests and CI?
cc @youkaichao

comaniac · 2024-09-13T18:50:06Z

tests/kernels/test_blocksparse_attention.py

+@pytest.mark.parametrize("blocksparse_block_size", BLOCKSPARSE_BLOCK_SIZES)
+@pytest.mark.parametrize("blocksparse_homo_heads", BLOCKSPARSE_HOMO_HEADS)
+@torch.inference_mode()
+def test_contexted_kv_attention(


Can you measure how long does this test take roughly?

mergify · 2024-11-26T05:50:16Z

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @congcongchen123.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

Your Name added 2 commits September 10, 2024 13:16

[Kernel] Add prefix-caching support for phi-3-small-8k/128k model

46643c2

apply format

2d1870d

congcongchen123 marked this pull request as ready for review September 10, 2024 21:46

wschin reviewed Sep 12, 2024

View reviewed changes

tlrmchlsmth reviewed Sep 13, 2024

View reviewed changes

comaniac reviewed Sep 13, 2024

View reviewed changes

simon-mo requested a review from WoosukKwon as a code owner November 26, 2024 05:49

mergify bot added the needs-rebase label Nov 26, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Kernel] Add prefix-caching support for phi-3-small-8k/128k model triton kernel #8345

[Kernel] Add prefix-caching support for phi-3-small-8k/128k model triton kernel #8345

congcongchen123 commented Sep 10, 2024 •

edited

Loading

github-actions bot commented Sep 10, 2024

wschin commented Sep 12, 2024

comaniac commented Sep 12, 2024

wschin Sep 12, 2024

wschin Sep 12, 2024

tlrmchlsmth left a comment

tlrmchlsmth Sep 12, 2024

comaniac Sep 13, 2024

tlrmchlsmth Sep 13, 2024

tlrmchlsmth Sep 13, 2024

tlrmchlsmth Sep 13, 2024

tlrmchlsmth Sep 13, 2024

comaniac Sep 13, 2024

comaniac Sep 13, 2024

comaniac Sep 13, 2024

comaniac Sep 13, 2024

comaniac Sep 13, 2024

comaniac Sep 13, 2024

comaniac Sep 13, 2024

mergify bot commented Nov 26, 2024



		@torch.inference_mode()
		def context_blocksparse_flash_attn_varlen_fwd(

		query_start_block_id = (context_lens_tensor // sparse_block_size).cpu()
		query_end_block_id = ((seq_lens_tensor - 1) // sparse_block_size).cpu()

	assert (not self.use_spda), "forward_prefix does not support spda"
	if self.use_spda:
	raise ValueError("Prefix caching with block sparse attention does not support SPDA")

[Kernel] Add prefix-caching support for phi-3-small-8k/128k model triton kernel #8345

Are you sure you want to change the base?

[Kernel] Add prefix-caching support for phi-3-small-8k/128k model triton kernel #8345

Conversation

congcongchen123 commented Sep 10, 2024 • edited Loading

PR Title and Classification

Code Quality

Notes for Large Changes

What to Expect for the Reviews

Thank You

github-actions bot commented Sep 10, 2024

wschin commented Sep 12, 2024

comaniac commented Sep 12, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tlrmchlsmth 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

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mergify bot commented Nov 26, 2024

congcongchen123 commented Sep 10, 2024 •

edited

Loading