Paged Attention¶

Currently, vLLM utilizes its own implementation of a multi-head query attention kernel (csrc/attention/attention_kernels.cu). This kernel is designed to be compatible with vLLM's paged KV caches, where the key and value cache are stored in separate blocks (note that this block concept differs from the GPU thread block. So in a later document, I will refer to vLLM paged attention block as "block", while refer to GPU thread block as "thread block").

To achieve high performance, this kernel relies on a specially designed memory layout and access method, specifically when threads read data from global memory to shared memory. The purpose of this document is to provide a high-level explanation of the kernel implementation step by step, aiding those who wish to learn about the vLLM multi-head query attention kernel. After going through this document, users will likely have a better understanding and feel easier to follow the actual implementation.

Please note that this document may not cover all details, such as how to calculate the correct index for the corresponding data or the dot multiplication implementation. However, after reading this document and becoming familiar with the high-level logic flow, it should be easier for you to read the actual code and understand the details.

Inputs¶

The kernel function takes a list of arguments for the current thread to perform its assigned work. The three most important arguments are the input pointers q, k_cache, and v_cache, which point to query, key, and value data on global memory that need to be read and processed. The output pointer out points to global memory where the result should be written. These four pointers actually refer to multi-dimensional arrays, but each thread only accesses the portion of data assigned to it. I have omitted all other runtime parameters here for simplicity.

template<typename scalar_t, int HEAD_SIZE, int BLOCK_SIZE, int NUM_THREADS, int PARTITION_SIZE = 0>
__device__ void paged_attention_kernel(
    ... // Other side args.
    const scalar_t* __restrict__ out,       // [num_seqs, num_heads, max_num_partitions, head_size]
    const scalar_t* __restrict__ q,         // [num_seqs, num_heads, head_size]
    const scalar_t* __restrict__ k_cache,   // [num_blocks, num_kv_heads, head_size/x, block_size, x]
    const scalar_t* __restrict__ v_cache,   // [num_blocks, num_kv_heads, head_size, block_size]
    ... // Other side args.
)

There are also a list of template arguments above the function signature that are determined during compilation time. scalar_t represents the data type of the query, key, and value data elements, such as FP16. HEAD_SIZE indicates the number of elements in each head. BLOCK_SIZE refers to the number of tokens in each block. NUM_THREADS denotes the number of threads in each thread block. PARTITION_SIZE represents the number of tensor parallel GPUs (For simplicity, we assume this is 0 and tensor parallel is disabled).

With these arguments, we need to perform a sequence of preparations. This includes calculating the current head index, block index, and other necessary variables. However, for now, we can ignore these preparations and proceed directly to the actual calculations. It will be easier to understand them once we grasp the entire flow.

Concepts¶

Just before we dive into the calculation flow, I want to describe a few concepts that are needed for later sections. However, you may skip this section and return later if you encounter any confusing terminologies.

• Sequence: A sequence represents a client request. For example, the data pointed to by q has a shape of [num_seqs, num_heads, head_size]. That represents there are total num_seqs of query sequence data are pointed by q. Since this kernel is a single query attention kernel, each sequence only has one query token. Hence, the num_seqs equals the total number of tokens that are processed in the batch.
• Context: The context consists of the generated tokens from the sequence. For instance, ["What", "is", "your"] are the context tokens, and the input query token is "name". The model might generate the token "?".
• Vec: The vec is a list of elements that are fetched and calculated together. For query and key data, the vec size (VEC_SIZE) is determined so that each thread group can fetch and calculate 16 bytes of data at a time. For value data, the vec size (V_VEC_SIZE) is determined so that each thread can fetch and calculate 16 bytes of data at a time. For example, if the scalar_t is FP16 (2 bytes) and THREAD_GROUP_SIZE is 2, the VEC_SIZE will be 4, while the V_VEC_SIZE will be 8.
• Thread group: The thread group is a small group of threads(THREAD_GROUP_SIZE) that fetches and calculates one query token and one key token at a time. Each thread handles only a portion of the token data. The total number of elements processed by one thread group is referred as x. For example, if the thread group contains 2 threads and the head size is 8, then thread 0 handles the query and key elements at index 0, 2, 4, 6, while thread 1 handles the elements at index 1, 3, 5, 7.
• Block: The key and value cache data in vLLM are split into blocks. Each block stores data for a fixed number(BLOCK_SIZE) of tokens at one head. Each block may contain only a portion of the whole context tokens. For example, if the block size is 16 and the head size is 128, then for one head, one block can store 16 * 128 = 2048 elements.
• Warp: A warp is a group of 32 threads(WARP_SIZE) that execute simultaneously on a stream multiprocessor (SM). In this kernel, each warp processes the calculation between one query token and key tokens of one entire block at a time (it may process multiple blocks in multiple iterations). For example, if there are 4 warps and 6 blocks for one context, the assignment would be like warp 0 handles the 0th, 4th blocks, warp 1 handles the 1st, 5th blocks, warp 2 handles the 2nd block and warp 3 handles the 3rd block.
• Thread block: A thread block is a group of threads(NUM_THREADS) that can access the same shared memory. Each thread block contains multiple warps(NUM_WARPS), and in this kernel, each thread block processes the calculation between one query token and key tokens of a whole context.
• Grid: A grid is a collection of thread blocks and defines the shape of the collection. In this kernel, the shape is (num_heads, num_seqs, max_num_partitions). Therefore, each thread block only handles the calculation for one head, one sequence, and one partition.

Query¶

This section will introduce how query data is stored in memory and fetched by each thread. As mentioned above, each thread group fetches one query token data, while each thread itself only handles a part of one query token data. Within each warp, every thread group will fetch the same query token data, but will multiply it with different key token data.

const scalar_t* q_ptr = q + seq_idx * q_stride + head_idx * HEAD_SIZE;

Each thread defines its own q_ptr which points to the assigned query token data on global memory. For example, if VEC_SIZE is 4 and HEAD_SIZE is 128, the q_ptr points to data that contains total of 128 elements divided into 128 / 4 = 32 vecs.

__shared__ Q_vec q_vecs[THREAD_GROUP_SIZE][NUM_VECS_PER_THREAD];

Next, we need to read the global memory data pointed to by q_ptr into shared memory as q_vecs. It is important to note that each vecs is assigned to a different row. For example, if the THREAD_GROUP_SIZE is 2, thread 0 will handle the 0th row vecs, while thread 1 handles the 1st row vecs. By reading the query data in this way, neighboring threads like thread 0 and thread 1 can read neighbor memory, achieving the memory coalescing to improve performance.

Key¶

Similar to the "Query" section, this section introduces memory layout and assignment for keys. While each thread group only handle one query token one kernel run, it may handle multiple key tokens across multiple iterations. Meanwhile, each warp will process multiple blocks of key tokens in multiple iterations, ensuring that all context tokens are processed by the entire thread group after the kernel run. In this context, "handle" refers to performing the dot multiplication between query data and key data.

const scalar_t* k_ptr = k_cache + physical_block_number * kv_block_stride
                    + kv_head_idx * kv_head_stride
                    + physical_block_offset * x;

Unlike to q_ptr, k_ptr in each thread will point to different key token at different iterations. As shown above, that k_ptr points to key token data based on k_cache at assigned block, assigned head and assigned token.

The diagram above illustrates the memory layout for key data. It assumes that the BLOCK_SIZE is 16, HEAD_SIZE is 128, x is 8, THREAD_GROUP_SIZE is 2, and there are a total of 4 warps. Each rectangle represents all the elements for one key token at one head, which will be processed by one thread group. The left half shows the total 16 blocks of key token data for warp 0, while the right half represents the remaining key token data for other warps or iterations. Inside each rectangle, there are a total 32 vecs (128 elements for one token) that will be processed by 2 threads (one thread group) separately.

K_vec k_vecs[NUM_VECS_PER_THREAD]

Next, we need to read the key token data from k_ptr and store them on register memory as k_vecs. We use register memory for k_vecs because it will only be accessed by one thread once, whereas q_vecs will be accessed by multiple threads multiple times. Each k_vecs will contain multiple vectors for later calculation. Each vec will be set at each inner iteration. The assignment of vecs allows neighboring threads in a warp to read neighboring memory together, which again promotes the memory coalescing. For instance, thread 0 will read vec 0, while thread 1 will read vec 1. In the next inner loop, thread 0 will read vec 2, while thread 1 will read vec 3, and so on.

You may still be a little confused about the overall flow. Don't worry, please keep reading the next "QK" section. It will illustrate the query and key calculation flow in a clearer and higher-level manner.

QK¶

As shown the pseudo code below, before the entire for loop block, we fetch the query data for one token and store it in q_vecs. Then, in the outer for loop, we iterate through different k_ptrs that point to different tokens and prepare the k_vecs in the inner for loop. Finally, we perform the dot multiplication between the q_vecs and each k_vecs.

q_vecs = ...
for ... {
    k_ptr = ...
    for ... {
        k_vecs[i] = ...
    }
    ...
    float qk = scale * Qk_dot<scalar_t, THREAD_GROUP_SIZE>::dot(q_vecs[thread_group_offset], k_vecs);
}

As mentioned before, for each thread, it only fetches part of the query and key token data at a time. However, there will be a cross thread group reduction happen in the Qk_dot<>::dot . So qk returned here is not just between part of the query and key token dot multiplication, but actually a full result between entire query and key token data.

For example, if the value of HEAD_SIZE is 128 and THREAD_GROUP_SIZE is 2, each thread's k_vecs will contain total 64 elements. However, the returned qk is actually the result of dot multiplication between 128 query elements and 128 key elements. If you want to learn more about the details of the dot multiplication and reduction, you may refer to the implementation of Qk_dot<>::dot. However, for the sake of simplicity, I will not cover it in this document.

Softmax¶

Next, we need to calculate the normalized softmax for all qks, as shown above, where each $x$ represents a qk. To do this, we must obtain the reduced value of qk_max($m(x)$) and the exp_sum($\ell(x)$) of all qks. The reduction should be performed across the entire thread block, encompassing results between the query token and all context key tokens.

$$\begin{gather*} m(x):=\max _i \quad x_i \\ \quad f(x):=\left[\begin{array}{lll}e^{x_1-m(x)} & \ldots & e^{x_B-m(x)}\end{array}\right]\\ \quad \ell(x):=\sum_i f(x)_i \\ \quad \operatorname{softmax}(x):=\frac{f(x)}{\ell(x)} \end{gather*}$$

qk_max and logits¶

Just right after we get the qk result, we can set the temporary logits result with qk (In the end, the logits should store the normalized softmax result). Also we can compare and collect the qk_max for all qks that are calculated by current thread group.

if (thread_group_offset == 0) {
    const bool mask = token_idx >= context_len;
    logits[token_idx - start_token_idx] = mask ? 0.f : qk;
    qk_max = mask ? qk_max : fmaxf(qk_max, qk);
}

Please note that the logits here is on shared memory, so each thread group will set the fields for its own assigned context tokens. Overall, the size of logits should be number of context tokens.

for (int mask = WARP_SIZE / 2; mask >= THREAD_GROUP_SIZE; mask /= 2) {
    qk_max = fmaxf(qk_max, VLLM_SHFL_XOR_SYNC(qk_max, mask));
}

if (lane == 0) {
    red_smem[warp_idx] = qk_max;
}

Then we need to get the reduced qk_max across each warp. The main idea is to make threads in warp to communicate with each other and get the final max qk .

for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
    qk_max = fmaxf(qk_max, VLLM_SHFL_XOR_SYNC(qk_max, mask));
}
qk_max = VLLM_SHFL_SYNC(qk_max, 0);

Finally, we can get the reduced qk_max from whole thread block by compare the qk_max from all warps in this thread block. Then we need to broadcast the final result to each thread.

exp_sum¶

Similar to qk_max, we need to get the reduced sum value from the entire thread block too.

for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) {
    float val = __expf(logits[i] - qk_max);
    logits[i] = val;
    exp_sum += val;
}
...
exp_sum = block_sum<NUM_WARPS>(&red_smem[NUM_WARPS], exp_sum);

Firstly, sum all exp values from each thread group, and meanwhile, convert each entry of logits from qk to exp(qk - qk_max). Please note, the qk_max here is already the max qk across the whole thread block. And then we can do reduction for exp_sum across whole thread block just like the qk_max.

const float inv_sum = __fdividef(1.f, exp_sum + 1e-6f);
for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) {
    logits[i] *= inv_sum;
}

Finally, with the reduced qk_max and exp_sum, we can obtain the final normalized softmax result as logits. This logits variable will be used for dot multiplication with the value data in later steps. Now, it should store the normalized softmax result of qk for all assigned context tokens.

Value¶

Now we need to retrieve the value data and perform dot multiplication with logits. Unlike query and key, there is no thread group concept for value data. As shown in diagram, different from key token memory layout, elements from the same column correspond to the same value token. For one block of value data, there are HEAD_SIZE of rows and BLOCK_SIZE of columns that are split into multiple v_vecs.

Each thread always fetches V_VEC_SIZE elements from the same V_VEC_SIZE of tokens at a time. As a result, a single thread retrieves multiple v_vecs from different rows and the same columns through multiple inner iterations. For each v_vec, it needs to be dot multiplied with the corresponding logits_vec, which is also V_VEC_SIZE elements from logits. Overall, with multiple inner iterations, each warp will process one block of value tokens. And with multiple outer iterations, the whole context value tokens are processed

float accs[NUM_ROWS_PER_THREAD];
for ... { // Iteration over different blocks.
    logits_vec = ...
    for ... { // Iteration over different rows.
        v_vec = ...
        ...
        accs[i] += dot(logits_vec, v_vec);
    }
}

As shown in the above pseudo code, in the outer loop, similar to k_ptr, logits_vec iterates over different blocks and reads V_VEC_SIZE elements from logits. In the inner loop, each thread reads V_VEC_SIZE elements from the same tokens as a v_vec and performs dot multiplication. It is important to note that in each inner iteration, the thread fetches different head position elements for the same tokens. The dot result is then accumulated in accs. Therefore, each entry of accs is mapped to a head position assigned to the current thread.

For example, if BLOCK_SIZE is 16 and V_VEC_SIZE is 8, each thread fetches 8 value elements for 8 tokens at a time. Each element is from different tokens at the same head position. If HEAD_SIZE is 128 and WARP_SIZE is 32, for each inner loop, a warp needs to fetch WARP_SIZE * V_VEC_SIZE = 256 elements. This means there are a total of 128 * 16 / 256 = 8 inner iterations for a warp to handle a whole block of value tokens. And each accs in each thread contains 8 elements that accumulated at 8 different head positions. For the thread 0, the accs variable will have 8 elements, which are 0th, 32nd … 224th elements of a value head that are accumulated from all assigned 8 tokens.

LV¶

Now, we need to perform reduction for accs within each warp. This process allows each thread to accumulate the accs for the assigned head positions of all tokens in one block.

for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
    float acc = accs[i];
    for (int mask = NUM_V_VECS_PER_ROW / 2; mask >= 1; mask /= 2) {
        acc += VLLM_SHFL_XOR_SYNC(acc, mask);
    }
    accs[i] = acc;
}

Next, we perform reduction for accs across all warps, allowing each thread to have the accumulation of accs for the assigned head positions of all context tokens. Please note that each accs in every thread only stores the accumulation for a portion of elements of the entire head for all context tokens. However, overall, all results for output have been calculated but are just stored in different thread register memory.

float* out_smem = reinterpret_cast<float*>(shared_mem);
for (int i = NUM_WARPS; i > 1; i /= 2) {
    // Upper warps write to shared memory.
    ...
    float* dst = &out_smem[(warp_idx - mid) * HEAD_SIZE];
    for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
        ...
        dst[row_idx] = accs[i];
    }

    // Lower warps update the output.
    const float* src = &out_smem[warp_idx * HEAD_SIZE];
    for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
        ...
        accs[i] += src[row_idx];
    }

    // Write out the accs.
}

Output¶

Now we can write all of calculated result from local register memory to final output global memory.

scalar_t* out_ptr = out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE
                + head_idx * max_num_partitions * HEAD_SIZE
                + partition_idx * HEAD_SIZE;

First, we need to define the out_ptr variable, which points to the start address of the assigned sequence and assigned head.

for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
    const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
    if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
        from_float(*(out_ptr + row_idx), accs[i]);
    }
}

Finally, we need to iterate over different assigned head positions and write out the corresponding accumulated result based on the out_ptr.

Citation¶

@inproceedings{kwon2023efficient,
  title={Efficient Memory Management for Large Language Model Serving with PagedAttention},
  author={Woosuk Kwon and Zhuohan Li and Siyuan Zhuang and Ying Sheng and Lianmin Zheng and Cody Hao Yu and Joseph E. Gonzalez and Hao Zhang and Ion Stoica},
  booktitle={Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles},
  year={2023}
}