Skip to content

GitLab

Commit 44e3ca68 authored by 王敏's avatar 王敏
Browse files

[feat]优化medusa代码，通过VLLM_TREE_DECODING环境变量控制是否采用tree-style解码，计算逻辑主干隔离

parent 54b92ba4
Show whitespace changes
Inline Side-by-side
Showing with 3801 additions and 312 deletions
CMakeLists.txt
View file @ 44e3ca68
......@@ -198,7 +198,10 @@ set(VLLM_EXT_SRC
"csrc/cuda_utils_kernels.cu"
"csrc/moe_align_block_size_kernels.cu"
"csrc/prepare_inputs/advance_step.cu"
"csrc/torch_bindings.cpp")
"csrc/torch_bindings.cpp"
"csrc/attention/attention_with_mask_kernels.cu"
"csrc/attention/attention_with_mask_kernels_opt.cu"
"csrc/attention/attention_with_mask_kernels_opt_tc.cu")
if(VLLM_GPU_LANG STREQUAL "CUDA")
SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
......
csrc/attention/attention_kernels.cu
View file @ 44e3ca68
......@@ -107,8 +107,7 @@ __device__ void paged_attention_kernel(
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
const int seq_idx = blockIdx.y;
const int partition_idx = blockIdx.z;
const int max_num_partitions = gridDim.z;
......@@ -297,14 +296,6 @@ __device__ void paged_attention_kernel(
// Add the ALiBi bias if slopes are given.
qk += (alibi_slope != 0) ? alibi_slope * (token_idx - seq_len + 1) : 0;
// used for tree-style attention
if (attn_masks != nullptr) {
const int* attn_masks_ptr = attn_masks + seq_idx * attn_masks_stride;
if (attn_masks_ptr[token_idx] == 0) {
qk = -FLT_MAX;
}
}
if (thread_group_offset == 0) {
// Store the partial reductions to shared memory.
// NOTE(woosuk): It is required to zero out the masked logits.
......@@ -524,8 +515,7 @@ __global__ void paged_attention_v1_kernel(
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
KV_DTYPE, IS_BLOCK_SPARSE>(
/* exp_sums */ nullptr, /* max_logits */ nullptr, out, q, k_cache,
......@@ -533,7 +523,7 @@ __global__ void paged_attention_v1_kernel(
max_num_blocks_per_seq, alibi_slopes, q_stride, kv_block_stride,
kv_head_stride, k_scale, v_scale, tp_rank, blocksparse_local_blocks,
blocksparse_vert_stride, blocksparse_block_size,
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
blocksparse_head_sliding_step);
}
// Grid: (num_heads, num_seqs, max_num_partitions).
......@@ -561,15 +551,14 @@ __global__ void paged_attention_v2_kernel(
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
KV_DTYPE, IS_BLOCK_SPARSE, PARTITION_SIZE>(
exp_sums, max_logits, tmp_out, q, k_cache, v_cache, num_kv_heads, scale,
block_tables, seq_lens, max_num_blocks_per_seq, alibi_slopes, q_stride,
kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank,
blocksparse_local_blocks, blocksparse_vert_stride, blocksparse_block_size,
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
blocksparse_head_sliding_step);
}
// Grid: (num_heads, num_seqs).
......@@ -695,8 +684,7 @@ __global__ void paged_attention_v2_reduce_kernel(
alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride, \
k_scale, v_scale, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step, attn_masks_ptr, \
attn_masks_stride);
blocksparse_head_sliding_step);
// TODO(woosuk): Tune NUM_THREADS.
template <typename T, typename CACHE_T, int BLOCK_SIZE,
......@@ -709,9 +697,7 @@ void paged_attention_v1_launcher(
const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
float v_scale, const int tp_rank, const int blocksparse_local_blocks,
const int blocksparse_vert_stride, const int blocksparse_block_size,
const int blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks,
const int64_t attn_masks_stride=0) {
const int blocksparse_head_sliding_step) {
int num_seqs = query.size(0);
int num_heads = query.size(1);
int head_size = query.size(2);
......@@ -736,12 +722,6 @@ void paged_attention_v1_launcher(
int* block_tables_ptr = block_tables.data_ptr<int>();
int* seq_lens_ptr = seq_lens.data_ptr<int>();
// NOTE: attn_masks is optional.
const int* attn_masks_ptr =
attn_masks
? attn_masks.value().data_ptr<int>()
: nullptr;
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
int padded_max_seq_len =
DIVIDE_ROUND_UP(max_seq_len, BLOCK_SIZE) * BLOCK_SIZE;
......@@ -798,8 +778,7 @@ void paged_attention_v1_launcher(
out, query, key_cache, value_cache, num_kv_heads, scale, block_tables, \
seq_lens, max_seq_len, alibi_slopes, k_scale, v_scale, tp_rank, \
blocksparse_local_blocks, blocksparse_vert_stride, \
blocksparse_block_size, blocksparse_head_sliding_step, \
attn_masks, attn_masks_stride);
blocksparse_block_size, blocksparse_head_sliding_step);
#define CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE) \
switch (is_block_sparse) { \
......@@ -845,9 +824,7 @@ void paged_attention_v1(
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride) {
const int64_t blocksparse_head_sliding_step) {
const bool is_block_sparse = (blocksparse_vert_stride > 1);
DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
......@@ -864,8 +841,7 @@ void paged_attention_v1(
seq_lens_ptr, max_num_blocks_per_seq, alibi_slopes_ptr, q_stride, \
kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank, \
blocksparse_local_blocks, blocksparse_vert_stride, \
blocksparse_block_size, blocksparse_head_sliding_step, \
attn_masks_ptr, attn_masks_stride); \
blocksparse_block_size, blocksparse_head_sliding_step); \
vllm::paged_attention_v2_reduce_kernel<T, HEAD_SIZE, NUM_THREADS, \
PARTITION_SIZE> \
<<<reduce_grid, block, reduce_shared_mem_size, stream>>>( \
......@@ -883,9 +859,7 @@ void paged_attention_v2_launcher(
const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
float v_scale, const int tp_rank, const int blocksparse_local_blocks,
const int blocksparse_vert_stride, const int blocksparse_block_size,
const int blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks,
const int attn_masks_stride) {
const int blocksparse_head_sliding_step) {
int num_seqs = query.size(0);
int num_heads = query.size(1);
int head_size = query.size(2);
......@@ -913,10 +887,6 @@ void paged_attention_v2_launcher(
int* block_tables_ptr = block_tables.data_ptr<int>();
int* seq_lens_ptr = seq_lens.data_ptr<int>();
// NOTE: attn_masks is optional.
const int* attn_masks_ptr =
attn_masks ? attn_masks.value().data_ptr<int>() : nullptr;
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
int max_num_partitions = DIVIDE_ROUND_UP(max_seq_len, PARTITION_SIZE);
int logits_size = PARTITION_SIZE * sizeof(float);
......@@ -976,7 +946,7 @@ void paged_attention_v2_launcher(
num_kv_heads, scale, block_tables, seq_lens, max_seq_len, alibi_slopes, \
k_scale, v_scale, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
blocksparse_head_sliding_step);
#define CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE) \
switch (is_block_sparse) { \
......@@ -1026,9 +996,7 @@ void paged_attention_v2(
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride) {
const int64_t blocksparse_head_sliding_step) {
const bool is_block_sparse = (blocksparse_vert_stride > 1);
DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
CALL_V2_LAUNCHER_BLOCK_SIZE)
......
csrc/attention/attention_kernels_opt.cu
View file @ 44e3ca68
......@@ -94,8 +94,7 @@ __device__ void paged_attention_kernel_opt(
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
const int seq_idx = blockIdx.z;
const int partition_idx = blockIdx.y;
const int max_num_partitions = gridDim.y;
......@@ -328,25 +327,11 @@ __device__ void paged_attention_kernel_opt(
// Add the ALiBi bias if slopes are given.
qk += (alibi_slope != 0) ? alibi_slope * (token_idx - seq_len + 1) : 0;
// used for tree-style attention
if (attn_masks != nullptr) {
const int* attn_masks_ptr = attn_masks + seq_idx * attn_masks_stride;
if (attn_masks_ptr[token_idx] == 0) {
qk = -FLT_MAX;
}
}
__builtin_amdgcn_sched_barrier(0);
if (thread_group_offset == 0) {
// Store the partial reductions to shared memory.
// NOTE(woosuk): It is required to zero out the masked logits.
const bool mask = token_idx >= seq_len;
// used for tree-style attention
/*if (attn_masks != nullptr) {
const int* attn_masks_ptr = attn_masks + seq_idx * attn_masks_stride;
mask |= attn_masks_ptr[token_idx] == 0;
}*/
logits[(reuse_kv_idx * partition_size) + (token_idx - start_token_idx)] = mask ? 0.f : qk;
// Update the max value.
qk_max[reuse_kv_idx] = mask ? qk_max[reuse_kv_idx] : fmaxf(qk_max[reuse_kv_idx], qk);
......@@ -627,8 +612,7 @@ __global__ __launch_bounds__(256,1) void paged_attention_v1_kernel_opt(
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
paged_attention_kernel_opt<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
KV_DTYPE, IS_BLOCK_SPARSE, REUSE_KV_TIMES, odd_nheads>(
/* exp_sums */ nullptr, /* max_logits */ nullptr, out, q, k_cache,
......@@ -636,7 +620,7 @@ __global__ __launch_bounds__(256,1) void paged_attention_v1_kernel_opt(
max_num_blocks_per_seq, alibi_slopes, q_stride, kv_block_stride,
kv_head_stride, k_scale, v_scale, tp_rank, blocksparse_local_blocks,
blocksparse_vert_stride, blocksparse_block_size,
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
blocksparse_head_sliding_step);
}
// Grid: (num_heads, num_seqs, max_num_partitions).
......@@ -668,15 +652,14 @@ __global__ __launch_bounds__(256,1) void paged_attention_v2_kernel_opt(
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
paged_attention_kernel_opt<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
KV_DTYPE, IS_BLOCK_SPARSE, REUSE_KV_TIMES, odd_nheads, PARTITION_SIZE>(
exp_sums, max_logits, tmp_out, q, k_cache, v_cache, num_heads, num_kv_heads, scale,
block_tables, seq_lens, max_num_blocks_per_seq, alibi_slopes, q_stride,
kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank,
blocksparse_local_blocks, blocksparse_vert_stride, blocksparse_block_size,
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
blocksparse_head_sliding_step);
}
// Grid: (num_heads, num_seqs).
......@@ -802,8 +785,7 @@ __global__ __launch_bounds__(256,1) void paged_attention_v2_reduce_kernel_opt(
alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride, \
k_scale, v_scale, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step, attn_masks_ptr, \
attn_masks_stride);
blocksparse_head_sliding_step);
// #define LAUNCH_PAGED_ATTENTION_V1(HEAD_SIZE) \
// vllm::paged_attention_v1_kernel<T, CACHE_T, HEAD_SIZE, BLOCK_SIZE, \
......@@ -826,9 +808,7 @@ void paged_attention_v1_launcher(
const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
float v_scale, const int tp_rank, const int blocksparse_local_blocks,
const int blocksparse_vert_stride, const int blocksparse_block_size,
const int blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks,
const int64_t attn_masks_stride=0) {
const int blocksparse_head_sliding_step) {
int num_seqs = query.size(0);
int num_heads = query.size(1);
int head_size = query.size(2);
......@@ -857,12 +837,6 @@ void paged_attention_v1_launcher(
int* block_tables_ptr = block_tables.data_ptr<int>();
int* seq_lens_ptr = seq_lens.data_ptr<int>();
// NOTE: attn_masks is optional.
const int* attn_masks_ptr =
attn_masks
? attn_masks.value().data_ptr<int>()
: nullptr;
int padded_max_seq_len = DIVIDE_ROUND_UP(max_seq_len, BLOCK_SIZE) * BLOCK_SIZE;
REUSEKV_SWITCH_V1(num_heads * num_seqs , [&] {
BOOL_SWITCH((num_heads/num_kv_heads % REUSE_KV_TIMES != 0), odd_nheads, [&] {
......@@ -896,8 +870,7 @@ void paged_attention_v1_launcher(
out, query, key_cache, value_cache, num_kv_heads, scale, block_tables, \
seq_lens, max_seq_len, alibi_slopes, k_scale, v_scale, tp_rank, \
blocksparse_local_blocks, blocksparse_vert_stride, \
blocksparse_block_size, blocksparse_head_sliding_step, \
attn_masks, attn_masks_stride);
blocksparse_block_size, blocksparse_head_sliding_step);
#define CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE) \
switch (is_block_sparse) { \
......@@ -943,9 +916,7 @@ void paged_attention_v1_opt(
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride) {
const int64_t blocksparse_head_sliding_step) {
const bool is_block_sparse = (blocksparse_vert_stride > 1);
DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
......@@ -962,8 +933,7 @@ void paged_attention_v1_opt(
seq_lens_ptr, max_num_blocks_per_seq, alibi_slopes_ptr, q_stride, \
kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank, \
blocksparse_local_blocks, blocksparse_vert_stride, \
blocksparse_block_size, blocksparse_head_sliding_step, \
attn_masks_ptr, attn_masks_stride); \
blocksparse_block_size, blocksparse_head_sliding_step); \
hipLaunchKernelGGL(( vllm::paged_attention_v2_reduce_kernel_opt<T, HEAD_SIZE, NUM_THREADS, \
PARTITION_SIZE>) \
, dim3(reduce_grid), dim3(block), reduce_shared_mem_size, stream, \
......@@ -981,9 +951,7 @@ void paged_attention_v2_launcher(
const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
float v_scale, const int tp_rank, const int blocksparse_local_blocks,
const int blocksparse_vert_stride, const int blocksparse_block_size,
const int blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks,
const int attn_masks_stride) {
const int blocksparse_head_sliding_step) {
int num_seqs = query.size(0);
int num_heads = query.size(1);
int head_size = query.size(2);
......@@ -1011,10 +979,6 @@ void paged_attention_v2_launcher(
int* block_tables_ptr = block_tables.data_ptr<int>();
int* seq_lens_ptr = seq_lens.data_ptr<int>();
// NOTE: attn_masks is optional.
const int* attn_masks_ptr =
attn_masks ? attn_masks.value().data_ptr<int>() : nullptr;
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
int max_num_partitions = DIVIDE_ROUND_UP(max_seq_len, PARTITION_SIZE);
REUSEKV_SWITCH(num_heads * max_num_partitions * num_seqs , [&] {
......@@ -1053,7 +1017,7 @@ void paged_attention_v2_launcher(
num_kv_heads, scale, block_tables, seq_lens, max_seq_len, alibi_slopes, \
k_scale, v_scale, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
blocksparse_head_sliding_step);
#define CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE) \
switch (is_block_sparse) { \
......@@ -1103,9 +1067,7 @@ void paged_attention_v2_opt(
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride) {
const int64_t blocksparse_head_sliding_step) {
const bool is_block_sparse = (blocksparse_vert_stride > 1);
DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
CALL_V2_LAUNCHER_BLOCK_SIZE)
......
csrc/attention/attention_kernels_opt_tc.cu
View file @ 44e3ca68
......@@ -168,8 +168,7 @@ __device__ void paged_attention_kernel_TC(
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
const int seq_idx = blockIdx.z;
const int partition_idx = blockIdx.y;
const int max_num_partitions = gridDim.y;
......@@ -293,14 +292,6 @@ __device__ void paged_attention_kernel_TC(
qk_vec[i] = alibi;
}
// used for tree-style attention
if (attn_masks != nullptr) {
const int* attn_masks_ptr = attn_masks + seq_idx * attn_masks_stride;
if (attn_masks_ptr[token_idx] == 0) {
qk_vec[i] = -FLT_MAX;
}
}
const bool mask = (token_idx >= seq_len);
if(mask){
from_float(logits[partition_size*reuse_kv_idx+token_idx - start_token_idx] , 0.f);
......@@ -565,8 +556,7 @@ __global__ void paged_attention_v1_kernel_TC(
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
#if defined(__gfx936__) || defined(__gfx928__)
paged_attention_kernel_TC<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
KV_DTYPE, IS_BLOCK_SPARSE,REUSE_KV_TIMES,use_vmac>(
......@@ -575,7 +565,7 @@ __global__ void paged_attention_v1_kernel_TC(
max_num_blocks_per_seq, alibi_slopes, q_stride, kv_block_stride,
kv_head_stride, k_scale, v_scale, tp_rank, blocksparse_local_blocks,
blocksparse_vert_stride, blocksparse_block_size,
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
blocksparse_head_sliding_step);
#endif
}
......@@ -605,8 +595,7 @@ __global__ __launch_bounds__(256, 1) void paged_attention_v2_kernel_TC(
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
#if defined(__gfx936__) || defined(__gfx928__)
paged_attention_kernel_TC<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
KV_DTYPE, IS_BLOCK_SPARSE, REUSE_KV_TIMES,use_vmac,
......@@ -615,7 +604,7 @@ __global__ __launch_bounds__(256, 1) void paged_attention_v2_kernel_TC(
num_kv_heads, scale, block_tables, seq_lens, max_num_blocks_per_seq,
alibi_slopes, q_stride, kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank,
blocksparse_local_blocks, blocksparse_vert_stride, blocksparse_block_size,
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
blocksparse_head_sliding_step);
#endif
}
......@@ -742,8 +731,7 @@ __global__ __launch_bounds__(256, 1) void paged_attention_v2_reduce_kernel_opt_t
alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride, \
k_scale, v_scale, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step, attn_masks_ptr, \
attn_masks_stride);
blocksparse_head_sliding_step);
void get_numberthread_and_reuse_kv_v1(int& num_thread,int& reusekv,int batchsize,int seq,int qheads,int kvheads){
//mha
......@@ -809,9 +797,7 @@ void paged_attention_v1_launcher_opt_tc(
const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
float v_scale, const int tp_rank, const int blocksparse_local_blocks,
const int blocksparse_vert_stride, const int blocksparse_block_size,
const int blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks,
const int64_t attn_masks_stride=0) {
const int blocksparse_head_sliding_step) {
int num_seqs = query.size(0);
int num_heads = query.size(1);
int head_size = query.size(2);
......@@ -840,12 +826,6 @@ void paged_attention_v1_launcher_opt_tc(
int* block_tables_ptr = block_tables.data_ptr<int>();
int* seq_lens_ptr = seq_lens.data_ptr<int>();
// NOTE: attn_masks is optional.
const int* attn_masks_ptr =
attn_masks
? attn_masks.value().data_ptr<int>()
: nullptr;
int padded_max_seq_len = DIVIDE_ROUND_UP(max_seq_len, BLOCK_SIZE) * BLOCK_SIZE;
const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
......@@ -880,8 +860,7 @@ void paged_attention_v1_launcher_opt_tc(
out, query, key_cache, value_cache, num_kv_heads, scale, block_tables, \
seq_lens, max_seq_len, alibi_slopes, k_scale, v_scale, tp_rank, \
blocksparse_local_blocks, blocksparse_vert_stride, \
blocksparse_block_size, blocksparse_head_sliding_step, \
attn_masks, attn_masks_stride);
blocksparse_block_size, blocksparse_head_sliding_step);
#define CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE) \
switch (is_block_sparse) { \
......@@ -927,9 +906,7 @@ void paged_attention_v1_opt(
const std::string& kv_cache_dtype, double k_scale, double v_scale, const int64_t tp_rank,
const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride);
const int64_t blocksparse_head_sliding_step);
void paged_attention_v1_opt_tc(
torch::Tensor& out, // [num_seqs, num_heads, head_size]
......@@ -947,17 +924,14 @@ void paged_attention_v1_opt_tc(
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride) {
const int64_t blocksparse_head_sliding_step) {
const bool is_block_sparse = (blocksparse_vert_stride > 1);
if(kv_cache_dtype != "auto"||query.dtype() == at::ScalarType::Float||is_block_sparse||
block_size!=16||query.size(2)!=128||(get_device_name()!="gfx928" && get_device_name()!="gfx936")){
paged_attention_v1_opt(out,query,key_cache,value_cache,num_kv_heads,
scale,block_tables,seq_lens,block_size,max_seq_len,alibi_slopes,kv_cache_dtype,
k_scale,v_scale,tp_rank,blocksparse_local_blocks,blocksparse_vert_stride,
blocksparse_block_size,blocksparse_head_sliding_step,
attn_masks, attn_masks_stride);
blocksparse_block_size,blocksparse_head_sliding_step);
}
else{
DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
......@@ -976,8 +950,7 @@ void paged_attention_v1_opt_tc(
max_num_blocks_per_seq, alibi_slopes_ptr, q_stride, kv_block_stride, \
kv_head_stride, k_scale, v_scale, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step, \
attn_masks_ptr, attn_masks_stride); \
blocksparse_head_sliding_step); \
hipLaunchKernelGGL( \
(vllm::paged_attention_v2_reduce_kernel_opt_tc<T, HEAD_SIZE, NUM_THREADS, \
PARTITION_SIZE>), \
......@@ -1028,9 +1001,7 @@ void paged_attention_v2_launcher_opt_tc(
const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
float v_scale, const int tp_rank, const int blocksparse_local_blocks,
const int blocksparse_vert_stride, const int blocksparse_block_size,
const int blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks,
const int attn_masks_stride) {
const int blocksparse_head_sliding_step) {
int num_seqs = query.size(0);
int num_heads = query.size(1);
int head_size = query.size(2);
......@@ -1058,10 +1029,6 @@ void paged_attention_v2_launcher_opt_tc(
int* block_tables_ptr = block_tables.data_ptr<int>();
int* seq_lens_ptr = seq_lens.data_ptr<int>();
// NOTE: attn_masks is optional.
const int* attn_masks_ptr =
attn_masks ? attn_masks.value().data_ptr<int>() : nullptr;
const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
dim3 reduce_grid(num_heads, num_seqs);
......@@ -1103,7 +1070,7 @@ void paged_attention_v2_launcher_opt_tc(
num_kv_heads, scale, block_tables, seq_lens, max_seq_len, alibi_slopes, \
k_scale, v_scale, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
blocksparse_head_sliding_step);
#define CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE) \
switch (is_block_sparse) { \
......@@ -1153,9 +1120,7 @@ void paged_attention_v2_opt(
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride);
const int64_t blocksparse_head_sliding_step);
void paged_attention_v2_opt_tc(
torch::Tensor& out, // [num_seqs, num_heads, head_size]
......@@ -1177,17 +1142,14 @@ void paged_attention_v2_opt_tc(
const std::string& kv_cache_dtype, double k_scale, double v_scale, const int64_t tp_rank,
const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride) {
const int64_t blocksparse_head_sliding_step) {
const bool is_block_sparse = (blocksparse_vert_stride > 1);
if(kv_cache_dtype != "auto"||query.dtype() == at::ScalarType::Float||is_block_sparse||
block_size!=16||query.size(2)!=128||(get_device_name()!="gfx928" && get_device_name()!="gfx936")){
paged_attention_v2_opt(out,exp_sums,max_logits,tmp_out,query,key_cache,value_cache,num_kv_heads,
scale,block_tables,seq_lens,block_size,max_seq_len,alibi_slopes,kv_cache_dtype,
k_scale,v_scale,tp_rank,blocksparse_local_blocks,blocksparse_vert_stride,
blocksparse_block_size,blocksparse_head_sliding_step, attn_masks,
attn_masks_stride);
blocksparse_block_size,blocksparse_head_sliding_step);
}
else{
DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
......
csrc/attention/attention_with_mask_kernels.cu 0 → 100644
View file @ 44e3ca68
/*
* Adapted from
* https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
* Copyright (c) 2023, The vLLM team.
* Copyright (c) 2020-2023, NVIDIA CORPORATION. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include <torch/all.h>
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <algorithm>
#include "attention_dtypes.h"
#include "attention_utils.cuh"
#ifdef USE_ROCM
#include <hip/hip_bf16.h>
#include "../quantization/fp8/amd/quant_utils.cuh"
typedef __hip_bfloat16 __nv_bfloat16;
#else
#include "../quantization/fp8/nvidia/quant_utils.cuh"
#endif
#ifndef USE_ROCM
#define WARP_SIZE 32
#else
#define WARP_SIZE warpSize
#endif
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define DIVIDE_ROUND_UP(a, b) (((a) + (b) - 1) / (b))
namespace vllm {
// Utility function for attention softmax.
template <int NUM_WARPS>
inline __device__ float block_sum(float* red_smem, float sum) {
// Decompose the thread index into warp / lane.
int warp = threadIdx.x / WARP_SIZE;
int lane = threadIdx.x % WARP_SIZE;
// Compute the sum per warp.
#pragma unroll
for (int mask = WARP_SIZE / 2; mask >= 1; mask /= 2) {
sum += VLLM_SHFL_XOR_SYNC(sum, mask);
}
// Warp leaders store the data to shared memory.
if (lane == 0) {
red_smem[warp] = sum;
}
// Make sure the data is in shared memory.
__syncthreads();
// The warps compute the final sums.
if (lane < NUM_WARPS) {
sum = red_smem[lane];
}
// Parallel reduction inside the warp.
#pragma unroll
for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
sum += VLLM_SHFL_XOR_SYNC(sum, mask);
}
// Broadcast to other threads.
return VLLM_SHFL_SYNC(sum, 0);
}
// TODO(woosuk): Merge the last two dimensions of the grid.
// Grid: (num_heads, num_seqs, max_num_partitions).
template <typename scalar_t, typename cache_t, int HEAD_SIZE, int BLOCK_SIZE,
int NUM_THREADS, vllm::Fp8KVCacheDataType KV_DTYPE,
bool IS_BLOCK_SPARSE,
int PARTITION_SIZE = 0> // Zero means no partitioning.
__device__ void paged_attention_with_mask_kernel(
float* __restrict__ exp_sums, // [num_seqs, num_heads, max_num_partitions]
float* __restrict__ max_logits, // [num_seqs, num_heads,
// max_num_partitions]
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 cache_t* __restrict__ k_cache, // [num_blocks, num_kv_heads,
// head_size/x, block_size, x]
const cache_t* __restrict__ v_cache, // [num_blocks, num_kv_heads,
// head_size, block_size]
const int num_kv_heads, // [num_heads]
const float scale,
const int* __restrict__ block_tables, // [num_seqs, max_num_blocks_per_seq]
const int* __restrict__ seq_lens, // [num_seqs]
const int max_num_blocks_per_seq,
const float* __restrict__ alibi_slopes, // [num_heads]
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
const int seq_idx = blockIdx.y;
const int partition_idx = blockIdx.z;
const int max_num_partitions = gridDim.z;
constexpr bool USE_PARTITIONING = PARTITION_SIZE > 0;
const int seq_len = seq_lens[seq_idx];
if (USE_PARTITIONING && partition_idx * PARTITION_SIZE >= seq_len) {
// No work to do. Terminate the thread block.
return;
}
const int num_seq_blocks = DIVIDE_ROUND_UP(seq_len, BLOCK_SIZE);
const int num_blocks_per_partition =
USE_PARTITIONING ? PARTITION_SIZE / BLOCK_SIZE : num_seq_blocks;
// [start_block_idx, end_block_idx) is the range of blocks to process.
const int start_block_idx =
USE_PARTITIONING ? partition_idx * num_blocks_per_partition : 0;
const int end_block_idx =
MIN(start_block_idx + num_blocks_per_partition, num_seq_blocks);
const int num_blocks = end_block_idx - start_block_idx;
// [start_token_idx, end_token_idx) is the range of tokens to process.
const int start_token_idx = start_block_idx * BLOCK_SIZE;
const int end_token_idx =
MIN(start_token_idx + num_blocks * BLOCK_SIZE, seq_len);
const int num_tokens = end_token_idx - start_token_idx;
constexpr int THREAD_GROUP_SIZE = MAX(WARP_SIZE / BLOCK_SIZE, 1);
constexpr int NUM_THREAD_GROUPS =
NUM_THREADS / THREAD_GROUP_SIZE; // Note: This assumes THREAD_GROUP_SIZE
// divides NUM_THREADS
assert(NUM_THREADS % THREAD_GROUP_SIZE == 0);
constexpr int NUM_TOKENS_PER_THREAD_GROUP =
DIVIDE_ROUND_UP(BLOCK_SIZE, WARP_SIZE);
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
const int thread_idx = threadIdx.x;
const int warp_idx = thread_idx / WARP_SIZE;
const int lane = thread_idx % WARP_SIZE;
const int head_idx = blockIdx.x;
const int num_heads = gridDim.x;
const int num_queries_per_kv = num_heads / num_kv_heads;
const int kv_head_idx = head_idx / num_queries_per_kv;
const float alibi_slope =
alibi_slopes == nullptr ? 0.f : alibi_slopes[head_idx];
// A vector type to store a part of a key or a query.
// The vector size is configured in such a way that the threads in a thread
// group fetch or compute 16 bytes at a time. For example, if the size of a
// thread group is 4 and the data type is half, then the vector size is 16 /
// (4 * sizeof(half)) == 2.
constexpr int VEC_SIZE = MAX(16 / (THREAD_GROUP_SIZE * sizeof(scalar_t)), 1);
using K_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
using Q_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
using Quant_vec = typename Vec<cache_t, VEC_SIZE>::Type;
constexpr int NUM_ELEMS_PER_THREAD = HEAD_SIZE / THREAD_GROUP_SIZE;
constexpr int NUM_VECS_PER_THREAD = NUM_ELEMS_PER_THREAD / VEC_SIZE;
const int thread_group_idx = thread_idx / THREAD_GROUP_SIZE;
const int thread_group_offset = thread_idx % THREAD_GROUP_SIZE;
// Load the query to registers.
// Each thread in a thread group has a different part of the query.
// For example, if the the thread group size is 4, then the first thread in
// the group has 0, 4, 8, ... th vectors of the query, and the second thread
// has 1, 5, 9, ... th vectors of the query, and so on. NOTE(woosuk): Because
// q is split from a qkv tensor, it may not be contiguous.
const scalar_t* q_ptr = q + seq_idx * q_stride + head_idx * HEAD_SIZE;
__shared__ Q_vec q_vecs[THREAD_GROUP_SIZE][NUM_VECS_PER_THREAD];
#pragma unroll
for (int i = thread_group_idx; i < NUM_VECS_PER_THREAD;
i += NUM_THREAD_GROUPS) {
const int vec_idx = thread_group_offset + i * THREAD_GROUP_SIZE;
q_vecs[thread_group_offset][i] =
*reinterpret_cast<const Q_vec*>(q_ptr + vec_idx * VEC_SIZE);
}
__syncthreads(); // TODO(naed90): possible speedup if this is replaced with a
// memory wall right before we use q_vecs
// Memory planning.
extern __shared__ char shared_mem[];
// NOTE(woosuk): We use FP32 for the softmax logits for better accuracy.
float* logits = reinterpret_cast<float*>(shared_mem);
// Workspace for reduction.
__shared__ float red_smem[2 * NUM_WARPS];
// x == THREAD_GROUP_SIZE * VEC_SIZE
// Each thread group fetches x elements from the key at a time.
constexpr int x = 16 / sizeof(cache_t);
float qk_max = -FLT_MAX;
// Iterate over the key blocks.
// Each warp fetches a block of keys for each iteration.
// Each thread group in a warp fetches a key from the block, and computes
// dot product with the query.
const int* block_table = block_tables + seq_idx * max_num_blocks_per_seq;
// blocksparse specific vars
int bs_block_offset;
int q_bs_block_id;
if constexpr (IS_BLOCK_SPARSE) {
// const int num_blocksparse_blocks = DIVIDE_ROUND_UP(seq_len,
// blocksparse_block_size);
q_bs_block_id = (seq_len - 1) / blocksparse_block_size;
if (blocksparse_head_sliding_step >= 0)
// sliding on q heads
bs_block_offset =
(tp_rank * num_heads + head_idx) * blocksparse_head_sliding_step + 1;
else
// sliding on kv heads
bs_block_offset = (tp_rank * num_kv_heads + kv_head_idx) *
(-blocksparse_head_sliding_step) +
1;
}
for (int block_idx = start_block_idx + warp_idx; block_idx < end_block_idx;
block_idx += NUM_WARPS) {
// NOTE(woosuk): The block number is stored in int32. However, we cast it to
// int64 because int32 can lead to overflow when this variable is multiplied
// by large numbers (e.g., kv_block_stride).
// For blocksparse attention: skip computation on blocks that are not
// attended
if constexpr (IS_BLOCK_SPARSE) {
const int k_bs_block_id = block_idx * BLOCK_SIZE / blocksparse_block_size;
const bool is_remote =
((k_bs_block_id + bs_block_offset) % blocksparse_vert_stride == 0);
const bool is_local =
(k_bs_block_id > q_bs_block_id - blocksparse_local_blocks);
if (!is_remote && !is_local) {
for (int i = 0; i < NUM_TOKENS_PER_THREAD_GROUP; i++) {
const int physical_block_offset =
(thread_group_idx + i * WARP_SIZE) % BLOCK_SIZE;
const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
if (thread_group_offset == 0) {
// NOTE(linxihui): assign very large number to skipped tokens to
// avoid contribution to the sumexp softmax normalizer. This will
// not be used at computing sum(softmax*v) as the blocks will be
// skipped.
logits[token_idx - start_token_idx] = -FLT_MAX;
}
}
continue;
}
}
const int64_t physical_block_number =
static_cast<int64_t>(block_table[block_idx]);
// Load a key to registers.
// Each thread in a thread group has a different part of the key.
// For example, if the the thread group size is 4, then the first thread in
// the group has 0, 4, 8, ... th vectors of the key, and the second thread
// has 1, 5, 9, ... th vectors of the key, and so on.
for (int i = 0; i < NUM_TOKENS_PER_THREAD_GROUP; i++) {
const int physical_block_offset =
(thread_group_idx + i * WARP_SIZE) % BLOCK_SIZE;
const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
K_vec k_vecs[NUM_VECS_PER_THREAD];
#pragma unroll
for (int j = 0; j < NUM_VECS_PER_THREAD; j++) {
const cache_t* k_ptr =
k_cache + physical_block_number * kv_block_stride +
kv_head_idx * kv_head_stride + physical_block_offset * x;
const int vec_idx = thread_group_offset + j * THREAD_GROUP_SIZE;
const int offset1 = (vec_idx * VEC_SIZE) / x;
const int offset2 = (vec_idx * VEC_SIZE) % x;
if constexpr (KV_DTYPE == Fp8KVCacheDataType::kAuto) {
k_vecs[j] = *reinterpret_cast<const K_vec*>(
k_ptr + offset1 * BLOCK_SIZE * x + offset2);
} else {
// Vector conversion from Quant_vec to K_vec.
Quant_vec k_vec_quant = *reinterpret_cast<const Quant_vec*>(
k_ptr + offset1 * BLOCK_SIZE * x + offset2);
k_vecs[j] = fp8::scaled_convert<K_vec, Quant_vec, KV_DTYPE>(
k_vec_quant, k_scale);
}
}
// Compute dot product.
// This includes a reduction across the threads in the same thread group.
float qk = scale * Qk_dot<scalar_t, THREAD_GROUP_SIZE>::dot(
q_vecs[thread_group_offset], k_vecs);
// Add the ALiBi bias if slopes are given.
qk += (alibi_slope != 0) ? alibi_slope * (token_idx - seq_len + 1) : 0;
// used for tree-style attention
if (attn_masks != nullptr) {
const int* attn_masks_ptr = attn_masks + seq_idx * attn_masks_stride;
if (attn_masks_ptr[token_idx] == 0) {
qk = -FLT_MAX;
}
}
if (thread_group_offset == 0) {
// Store the partial reductions to shared memory.
// NOTE(woosuk): It is required to zero out the masked logits.
const bool mask = token_idx >= seq_len;
logits[token_idx - start_token_idx] = mask ? 0.f: qk;
// Update the max value.
qk_max = mask ? qk_max : fmaxf(qk_max, qk);
}
}
}
// Perform reduction across the threads in the same warp to get the
// max qk value for each "warp" (not across the thread block yet).
// The 0-th thread of each thread group already has its max qk value.
#pragma unroll
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;
}
__syncthreads();
// TODO(woosuk): Refactor this part.
// Get the max qk value for the sequence.
qk_max = lane < NUM_WARPS ? red_smem[lane] : -FLT_MAX;
#pragma unroll
for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
qk_max = fmaxf(qk_max, VLLM_SHFL_XOR_SYNC(qk_max, mask));
}
// Broadcast the max qk value to all threads.
qk_max = VLLM_SHFL_SYNC(qk_max, 0);
// Get the sum of the exp values.
float exp_sum = 0.f;
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);
// Compute softmax.
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;
}
__syncthreads();
// If partitioning is enabled, store the max logit and exp_sum.
if (USE_PARTITIONING && thread_idx == 0) {
float* max_logits_ptr = max_logits +
seq_idx * num_heads * max_num_partitions +
head_idx * max_num_partitions + partition_idx;
*max_logits_ptr = qk_max;
float* exp_sums_ptr = exp_sums + seq_idx * num_heads * max_num_partitions +
head_idx * max_num_partitions + partition_idx;
*exp_sums_ptr = exp_sum;
}
// Each thread will fetch 16 bytes from the value cache at a time.
constexpr int V_VEC_SIZE = MIN(16 / sizeof(scalar_t), BLOCK_SIZE);
using V_vec = typename Vec<scalar_t, V_VEC_SIZE>::Type;
using L_vec = typename Vec<scalar_t, V_VEC_SIZE>::Type;
using V_quant_vec = typename Vec<cache_t, V_VEC_SIZE>::Type;
using Float_L_vec = typename FloatVec<L_vec>::Type;
constexpr int NUM_V_VECS_PER_ROW = BLOCK_SIZE / V_VEC_SIZE;
constexpr int NUM_ROWS_PER_ITER = WARP_SIZE / NUM_V_VECS_PER_ROW;
constexpr int NUM_ROWS_PER_THREAD =
DIVIDE_ROUND_UP(HEAD_SIZE, NUM_ROWS_PER_ITER);
// NOTE(woosuk): We use FP32 for the accumulator for better accuracy.
float accs[NUM_ROWS_PER_THREAD];
#pragma unroll
for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
accs[i] = 0.f;
}
scalar_t zero_value;
zero(zero_value);
for (int block_idx = start_block_idx + warp_idx; block_idx < end_block_idx;
block_idx += NUM_WARPS) {
// NOTE(woosuk): The block number is stored in int32. However, we cast it to
// int64 because int32 can lead to overflow when this variable is multiplied
// by large numbers (e.g., kv_block_stride).
// For blocksparse attention: skip computation on blocks that are not
// attended
if constexpr (IS_BLOCK_SPARSE) {
int v_bs_block_id = block_idx * BLOCK_SIZE / blocksparse_block_size;
if (!((v_bs_block_id + bs_block_offset) % blocksparse_vert_stride == 0) &&
!((v_bs_block_id > q_bs_block_id - blocksparse_local_blocks))) {
continue;
}
}
const int64_t physical_block_number =
static_cast<int64_t>(block_table[block_idx]);
const int physical_block_offset = (lane % NUM_V_VECS_PER_ROW) * V_VEC_SIZE;
const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
L_vec logits_vec;
from_float(logits_vec, *reinterpret_cast<Float_L_vec*>(logits + token_idx -
start_token_idx));
const cache_t* v_ptr = v_cache + physical_block_number * kv_block_stride +
kv_head_idx * kv_head_stride;
#pragma unroll
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) {
const int offset = row_idx * BLOCK_SIZE + physical_block_offset;
V_vec v_vec;
if constexpr (KV_DTYPE == Fp8KVCacheDataType::kAuto) {
v_vec = *reinterpret_cast<const V_vec*>(v_ptr + offset);
} else {
V_quant_vec v_quant_vec =
*reinterpret_cast<const V_quant_vec*>(v_ptr + offset);
// Vector conversion from V_quant_vec to V_vec.
v_vec = fp8::scaled_convert<V_vec, V_quant_vec, KV_DTYPE>(v_quant_vec,
v_scale);
}
if (block_idx == num_seq_blocks - 1) {
// NOTE(woosuk): When v_vec contains the tokens that are out of the
// context, we should explicitly zero out the values since they may
// contain NaNs. See
// https://github.com/vllm-project/vllm/issues/641#issuecomment-1682544472
scalar_t* v_vec_ptr = reinterpret_cast<scalar_t*>(&v_vec);
#pragma unroll
for (int j = 0; j < V_VEC_SIZE; j++) {
v_vec_ptr[j] = token_idx + j < seq_len ? v_vec_ptr[j] : zero_value;
}
}
accs[i] += dot(logits_vec, v_vec);
}
}
}
// Perform reduction within each warp.
#pragma unroll
for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
float acc = accs[i];
#pragma unroll
for (int mask = NUM_V_VECS_PER_ROW / 2; mask >= 1; mask /= 2) {
acc += VLLM_SHFL_XOR_SYNC(acc, mask);
}
accs[i] = acc;
}
// NOTE(woosuk): A barrier is required because the shared memory space for
// logits is reused for the output.
__syncthreads();
// Perform reduction across warps.
float* out_smem = reinterpret_cast<float*>(shared_mem);
#pragma unroll
for (int i = NUM_WARPS; i > 1; i /= 2) {
int mid = i / 2;
// Upper warps write to shared memory.
if (warp_idx >= mid && warp_idx < i) {
float* dst = &out_smem[(warp_idx - mid) * HEAD_SIZE];
#pragma unroll
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) {
dst[row_idx] = accs[i];
}
}
}
__syncthreads();
// Lower warps update the output.
if (warp_idx < mid) {
const float* src = &out_smem[warp_idx * HEAD_SIZE];
#pragma unroll
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) {
accs[i] += src[row_idx];
}
}
}
__syncthreads();
}
// Write the final output.
if (warp_idx == 0) {
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;
#pragma unroll
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]);
}
}
}
}
// Grid: (num_heads, num_seqs, 1).
template <typename scalar_t, typename cache_t, int HEAD_SIZE, int BLOCK_SIZE,
int NUM_THREADS, vllm::Fp8KVCacheDataType KV_DTYPE,
bool IS_BLOCK_SPARSE>
__global__ void paged_attention_v1_with_mask_kernel(
scalar_t* __restrict__ out, // [num_seqs, num_heads, head_size]
const scalar_t* __restrict__ q, // [num_seqs, num_heads, head_size]
const cache_t* __restrict__ k_cache, // [num_blocks, num_kv_heads,
// head_size/x, block_size, x]
const cache_t* __restrict__ v_cache, // [num_blocks, num_kv_heads,
// head_size, block_size]
const int num_kv_heads, // [num_heads]
const float scale,
const int* __restrict__ block_tables, // [num_seqs, max_num_blocks_per_seq]
const int* __restrict__ seq_lens, // [num_seqs]
const int max_num_blocks_per_seq,
const float* __restrict__ alibi_slopes, // [num_heads]
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
paged_attention_with_mask_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
KV_DTYPE, IS_BLOCK_SPARSE>(
/* exp_sums */ nullptr, /* max_logits */ nullptr, out, q, k_cache,
v_cache, num_kv_heads, scale, block_tables, seq_lens,
max_num_blocks_per_seq, alibi_slopes, q_stride, kv_block_stride,
kv_head_stride, k_scale, v_scale, tp_rank, blocksparse_local_blocks,
blocksparse_vert_stride, blocksparse_block_size,
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
}
// Grid: (num_heads, num_seqs, max_num_partitions).
template <typename scalar_t, typename cache_t, int HEAD_SIZE, int BLOCK_SIZE,
int NUM_THREADS, vllm::Fp8KVCacheDataType KV_DTYPE,
bool IS_BLOCK_SPARSE,
int PARTITION_SIZE>
__global__ void paged_attention_v2_with_mask_kernel(
float* __restrict__ exp_sums, // [num_seqs, num_heads, max_num_partitions]
float* __restrict__ max_logits, // [num_seqs, num_heads,
// max_num_partitions]
scalar_t* __restrict__ tmp_out, // [num_seqs, num_heads,
// max_num_partitions, head_size]
const scalar_t* __restrict__ q, // [num_seqs, num_heads, head_size]
const cache_t* __restrict__ k_cache, // [num_blocks, num_kv_heads,
// head_size/x, block_size, x]
const cache_t* __restrict__ v_cache, // [num_blocks, num_kv_heads,
// head_size, block_size]
const int num_kv_heads, // [num_heads]
const float scale,
const int* __restrict__ block_tables, // [num_seqs, max_num_blocks_per_seq]
const int* __restrict__ seq_lens, // [num_seqs]
const int max_num_blocks_per_seq,
const float* __restrict__ alibi_slopes, // [num_heads]
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
paged_attention_with_mask_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
KV_DTYPE, IS_BLOCK_SPARSE, PARTITION_SIZE>(
exp_sums, max_logits, tmp_out, q, k_cache, v_cache, num_kv_heads, scale,
block_tables, seq_lens, max_num_blocks_per_seq, alibi_slopes, q_stride,
kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank,
blocksparse_local_blocks, blocksparse_vert_stride, blocksparse_block_size,
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
}
// Grid: (num_heads, num_seqs).
template <typename scalar_t, int HEAD_SIZE, int NUM_THREADS,
int PARTITION_SIZE>
__global__ void paged_attention_v2_reduce_kernel(
scalar_t* __restrict__ out, // [num_seqs, num_heads, head_size]
const float* __restrict__ exp_sums, // [num_seqs, num_heads,
// max_num_partitions]
const float* __restrict__ max_logits, // [num_seqs, num_heads,
// max_num_partitions]
const scalar_t* __restrict__ tmp_out, // [num_seqs, num_heads,
// max_num_partitions, head_size]
const int* __restrict__ seq_lens, // [num_seqs]
const int max_num_partitions) {
const int num_heads = gridDim.x;
const int head_idx = blockIdx.x;
const int seq_idx = blockIdx.y;
const int seq_len = seq_lens[seq_idx];
const int num_partitions = DIVIDE_ROUND_UP(seq_len, PARTITION_SIZE);
if (num_partitions == 1) {
// No need to reduce. Only copy tmp_out to out.
scalar_t* out_ptr =
out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
const scalar_t* tmp_out_ptr =
tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
head_idx * max_num_partitions * HEAD_SIZE;
for (int i = threadIdx.x; i < HEAD_SIZE; i += blockDim.x) {
out_ptr[i] = tmp_out_ptr[i];
}
// Terminate the thread block.
return;
}
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
const int warp_idx = threadIdx.x / WARP_SIZE;
const int lane = threadIdx.x % WARP_SIZE;
// Size: 2 * num_partitions.
extern __shared__ char shared_mem[];
// Workspace for reduction.
__shared__ float red_smem[2 * NUM_WARPS];
// Load max logits to shared memory.
float* shared_max_logits = reinterpret_cast<float*>(shared_mem);
const float* max_logits_ptr = max_logits +
seq_idx * num_heads * max_num_partitions +
head_idx * max_num_partitions;
float max_logit = -FLT_MAX;
for (int i = threadIdx.x; i < num_partitions; i += blockDim.x) {
const float l = max_logits_ptr[i];
shared_max_logits[i] = l;
max_logit = fmaxf(max_logit, l);
}
__syncthreads();
// Get the global max logit.
// Reduce within the warp.
#pragma unroll
for (int mask = WARP_SIZE / 2; mask >= 1; mask /= 2) {
max_logit = fmaxf(max_logit, VLLM_SHFL_XOR_SYNC(max_logit, mask));
}
if (lane == 0) {
red_smem[warp_idx] = max_logit;
}
__syncthreads();
// Reduce across warps.
max_logit = lane < NUM_WARPS ? red_smem[lane] : -FLT_MAX;
#pragma unroll
for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
max_logit = fmaxf(max_logit, VLLM_SHFL_XOR_SYNC(max_logit, mask));
}
// Broadcast the max value to all threads.
max_logit = VLLM_SHFL_SYNC(max_logit, 0);
// Load rescaled exp sums to shared memory.
float* shared_exp_sums =
reinterpret_cast<float*>(shared_mem + sizeof(float) * num_partitions);
const float* exp_sums_ptr = exp_sums +
seq_idx * num_heads * max_num_partitions +
head_idx * max_num_partitions;
float global_exp_sum = 0.0f;
for (int i = threadIdx.x; i < num_partitions; i += blockDim.x) {
float l = shared_max_logits[i];
float rescaled_exp_sum = exp_sums_ptr[i] * expf(l - max_logit);
global_exp_sum += rescaled_exp_sum;
shared_exp_sums[i] = rescaled_exp_sum;
}
__syncthreads();
global_exp_sum = block_sum<NUM_WARPS>(&red_smem[NUM_WARPS], global_exp_sum);
const float inv_global_exp_sum = __fdividef(1.0f, global_exp_sum + 1e-6f);
// Aggregate tmp_out to out.
const scalar_t* tmp_out_ptr =
tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
head_idx * max_num_partitions * HEAD_SIZE;
scalar_t* out_ptr =
out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
#pragma unroll
for (int i = threadIdx.x; i < HEAD_SIZE; i += NUM_THREADS) {
float acc = 0.0f;
for (int j = 0; j < num_partitions; ++j) {
acc += to_float(tmp_out_ptr[j * HEAD_SIZE + i]) * shared_exp_sums[j] *
inv_global_exp_sum;
}
from_float(out_ptr[i], acc);
}
}
} // namespace vllm
#define LAUNCH_PAGED_ATTENTION_V1(HEAD_SIZE) \
VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize( \
((void*)vllm::paged_attention_v1_with_mask_kernel<T, CACHE_T, HEAD_SIZE, \
BLOCK_SIZE, NUM_THREADS, \
KV_DTYPE, IS_BLOCK_SPARSE>), \
shared_mem_size); \
vllm::paged_attention_v1_with_mask_kernel<T, CACHE_T, HEAD_SIZE, BLOCK_SIZE, \
NUM_THREADS, KV_DTYPE, IS_BLOCK_SPARSE> \
<<<grid, block, shared_mem_size, stream>>>( \
out_ptr, query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, \
scale, block_tables_ptr, seq_lens_ptr, max_num_blocks_per_seq, \
alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride, \
k_scale, v_scale, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step, attn_masks_ptr, \
attn_masks_stride);
// TODO(woosuk): Tune NUM_THREADS.
template <typename T, typename CACHE_T, int BLOCK_SIZE,
vllm::Fp8KVCacheDataType KV_DTYPE, bool IS_BLOCK_SPARSE,
int NUM_THREADS = 128>
void paged_attention_v1_launcher(
torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int num_kv_heads, float scale,
torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
float v_scale, const int tp_rank, const int blocksparse_local_blocks,
const int blocksparse_vert_stride, const int blocksparse_block_size,
const int blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks,
const int64_t attn_masks_stride=0) {
int num_seqs = query.size(0);
int num_heads = query.size(1);
int head_size = query.size(2);
int max_num_blocks_per_seq = block_tables.size(1);
int q_stride = query.stride(0);
int kv_block_stride = key_cache.stride(0);
int kv_head_stride = key_cache.stride(1);
[[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
assert(head_size % thread_group_size == 0);
// NOTE: alibi_slopes is optional.
const float* alibi_slopes_ptr =
alibi_slopes
? reinterpret_cast<const float*>(alibi_slopes.value().data_ptr())
: nullptr;
T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
CACHE_T* key_cache_ptr = reinterpret_cast<CACHE_T*>(key_cache.data_ptr());
CACHE_T* value_cache_ptr = reinterpret_cast<CACHE_T*>(value_cache.data_ptr());
int* block_tables_ptr = block_tables.data_ptr<int>();
int* seq_lens_ptr = seq_lens.data_ptr<int>();
// NOTE: attn_masks is optional.
const int* attn_masks_ptr =
attn_masks
? attn_masks.value().data_ptr<int>()
: nullptr;
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
int padded_max_seq_len =
DIVIDE_ROUND_UP(max_seq_len, BLOCK_SIZE) * BLOCK_SIZE;
int logits_size = padded_max_seq_len * sizeof(float);
int outputs_size = (NUM_WARPS / 2) * head_size * sizeof(float);
// Python-side check in vllm.worker.worker._check_if_can_support_max_seq_len
// Keep that in sync with the logic here!
int shared_mem_size = std::max(logits_size, outputs_size);
dim3 grid(num_heads, num_seqs, 1);
dim3 block(NUM_THREADS);
const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
switch (head_size) {
// NOTE(woosuk): To reduce the compilation time, we only compile for the
// head sizes that we use in the model. However, we can easily extend this
// to support any head size which is a multiple of 16.
case 64:
LAUNCH_PAGED_ATTENTION_V1(64);
break;
case 80:
LAUNCH_PAGED_ATTENTION_V1(80);
break;
case 96:
LAUNCH_PAGED_ATTENTION_V1(96);
break;
case 112:
LAUNCH_PAGED_ATTENTION_V1(112);
break;
case 120:
LAUNCH_PAGED_ATTENTION_V1(120);
break;
case 128:
LAUNCH_PAGED_ATTENTION_V1(128);
break;
case 160:
LAUNCH_PAGED_ATTENTION_V1(160);
break;
case 192:
LAUNCH_PAGED_ATTENTION_V1(192);
break;
case 256:
LAUNCH_PAGED_ATTENTION_V1(256);
break;
default:
TORCH_CHECK(false, "Unsupported head size: ", head_size);
break;
}
}
#define CALL_V1_LAUNCHER(T, CACHE_T, BLOCK_SIZE, KV_DTYPE, IS_BLOCK_SPARSE) \
paged_attention_v1_launcher<T, CACHE_T, BLOCK_SIZE, KV_DTYPE, \
IS_BLOCK_SPARSE>( \
out, query, key_cache, value_cache, num_kv_heads, scale, block_tables, \
seq_lens, max_seq_len, alibi_slopes, k_scale, v_scale, tp_rank, \
blocksparse_local_blocks, blocksparse_vert_stride, \
blocksparse_block_size, blocksparse_head_sliding_step, \
attn_masks, attn_masks_stride);
#define CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE) \
switch (is_block_sparse) { \
case true: \
CALL_V1_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE, true); \
break; \
case false: \
CALL_V1_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE, false); \
break; \
}
// NOTE(woosuk): To reduce the compilation time, we omitted block sizes
// 1, 2, 4, 64, 128, 256.
#define CALL_V1_LAUNCHER_BLOCK_SIZE(T, CACHE_T, KV_DTYPE) \
switch (block_size) { \
case 8: \
CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, 8, KV_DTYPE); \
break; \
case 16: \
CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, 16, KV_DTYPE); \
break; \
case 32: \
CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, 32, KV_DTYPE); \
break; \
default: \
TORCH_CHECK(false, "Unsupported block size: ", block_size); \
break; \
}
void paged_attention_v1_with_mask(
torch::Tensor& out, // [num_seqs, num_heads, head_size]
torch::Tensor& query, // [num_seqs, num_heads, head_size]
torch::Tensor&
key_cache, // [num_blocks, num_heads, head_size/x, block_size, x]
torch::Tensor&
value_cache, // [num_blocks, num_heads, head_size, block_size]
int64_t num_kv_heads, // [num_heads]
double scale,
torch::Tensor& block_tables, // [num_seqs, max_num_blocks_per_seq]
torch::Tensor& seq_lens, // [num_seqs]
int64_t block_size, int64_t max_seq_len,
const c10::optional<torch::Tensor>& alibi_slopes,
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride) {
const bool is_block_sparse = (blocksparse_vert_stride > 1);
DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
CALL_V1_LAUNCHER_BLOCK_SIZE)
}
#define LAUNCH_PAGED_ATTENTION_V2(HEAD_SIZE) \
vllm::paged_attention_v2_with_mask_kernel<T, CACHE_T, HEAD_SIZE, BLOCK_SIZE, \
NUM_THREADS, KV_DTYPE, IS_BLOCK_SPARSE, \
PARTITION_SIZE> \
<<<grid, block, shared_mem_size, stream>>>( \
exp_sums_ptr, max_logits_ptr, tmp_out_ptr, query_ptr, key_cache_ptr, \
value_cache_ptr, num_kv_heads, scale, block_tables_ptr, \
seq_lens_ptr, max_num_blocks_per_seq, alibi_slopes_ptr, q_stride, \
kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank, \
blocksparse_local_blocks, blocksparse_vert_stride, \
blocksparse_block_size, blocksparse_head_sliding_step, \
attn_masks_ptr, attn_masks_stride); \
vllm::paged_attention_v2_reduce_kernel<T, HEAD_SIZE, NUM_THREADS, \
PARTITION_SIZE> \
<<<reduce_grid, block, reduce_shared_mem_size, stream>>>( \
out_ptr, exp_sums_ptr, max_logits_ptr, tmp_out_ptr, seq_lens_ptr, \
max_num_partitions);
template <typename T, typename CACHE_T, int BLOCK_SIZE,
vllm::Fp8KVCacheDataType KV_DTYPE, bool IS_BLOCK_SPARSE,
int NUM_THREADS = 128, int PARTITION_SIZE = 512>
void paged_attention_v2_launcher(
torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int num_kv_heads, float scale,
torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
float v_scale, const int tp_rank, const int blocksparse_local_blocks,
const int blocksparse_vert_stride, const int blocksparse_block_size,
const int blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks,
const int attn_masks_stride) {
int num_seqs = query.size(0);
int num_heads = query.size(1);
int head_size = query.size(2);
int max_num_blocks_per_seq = block_tables.size(1);
int q_stride = query.stride(0);
int kv_block_stride = key_cache.stride(0);
int kv_head_stride = key_cache.stride(1);
[[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
assert(head_size % thread_group_size == 0);
// NOTE: alibi_slopes is optional.
const float* alibi_slopes_ptr =
alibi_slopes
? reinterpret_cast<const float*>(alibi_slopes.value().data_ptr())
: nullptr;
T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
float* exp_sums_ptr = reinterpret_cast<float*>(exp_sums.data_ptr());
float* max_logits_ptr = reinterpret_cast<float*>(max_logits.data_ptr());
T* tmp_out_ptr = reinterpret_cast<T*>(tmp_out.data_ptr());
T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
CACHE_T* key_cache_ptr = reinterpret_cast<CACHE_T*>(key_cache.data_ptr());
CACHE_T* value_cache_ptr = reinterpret_cast<CACHE_T*>(value_cache.data_ptr());
int* block_tables_ptr = block_tables.data_ptr<int>();
int* seq_lens_ptr = seq_lens.data_ptr<int>();
// NOTE: attn_masks is optional.
const int* attn_masks_ptr =
attn_masks ? attn_masks.value().data_ptr<int>() : nullptr;
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
int max_num_partitions = DIVIDE_ROUND_UP(max_seq_len, PARTITION_SIZE);
int logits_size = PARTITION_SIZE * sizeof(float);
int outputs_size = (NUM_WARPS / 2) * head_size * sizeof(float);
// For paged attention v2 kernel.
dim3 grid(num_heads, num_seqs, max_num_partitions);
int shared_mem_size = std::max(logits_size, outputs_size);
// For paged attention v2 reduce kernel.
dim3 reduce_grid(num_heads, num_seqs);
int reduce_shared_mem_size = 2 * max_num_partitions * sizeof(float);
dim3 block(NUM_THREADS);
const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
switch (head_size) {
// NOTE(woosuk): To reduce the compilation time, we only compile for the
// head sizes that we use in the model. However, we can easily extend this
// to support any head size which is a multiple of 16.
case 64:
LAUNCH_PAGED_ATTENTION_V2(64);
break;
case 80:
LAUNCH_PAGED_ATTENTION_V2(80);
break;
case 96:
LAUNCH_PAGED_ATTENTION_V2(96);
break;
case 112:
LAUNCH_PAGED_ATTENTION_V2(112);
break;
case 120:
LAUNCH_PAGED_ATTENTION_V2(120);
break;
case 128:
LAUNCH_PAGED_ATTENTION_V2(128);
break;
case 160:
LAUNCH_PAGED_ATTENTION_V2(160);
break;
case 192:
LAUNCH_PAGED_ATTENTION_V2(192);
break;
case 256:
LAUNCH_PAGED_ATTENTION_V2(256);
break;
default:
TORCH_CHECK(false, "Unsupported head size: ", head_size);
break;
}
}
#define CALL_V2_LAUNCHER(T, CACHE_T, BLOCK_SIZE, KV_DTYPE, IS_BLOCK_SPARSE) \
paged_attention_v2_launcher<T, CACHE_T, BLOCK_SIZE, KV_DTYPE, \
IS_BLOCK_SPARSE>( \
out, exp_sums, max_logits, tmp_out, query, key_cache, value_cache, \
num_kv_heads, scale, block_tables, seq_lens, max_seq_len, alibi_slopes, \
k_scale, v_scale, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
#define CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE) \
switch (is_block_sparse) { \
case true: \
CALL_V2_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE, true); \
break; \
case false: \
CALL_V2_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE, false); \
break; \
}
// NOTE(woosuk): To reduce the compilation time, we omitted block sizes
// 1, 2, 4, 64, 128, 256.
#define CALL_V2_LAUNCHER_BLOCK_SIZE(T, CACHE_T, KV_DTYPE) \
switch (block_size) { \
case 8: \
CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, 8, KV_DTYPE); \
break; \
case 16: \
CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, 16, KV_DTYPE); \
break; \
case 32: \
CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, 32, KV_DTYPE); \
break; \
default: \
TORCH_CHECK(false, "Unsupported block size: ", block_size); \
break; \
}
void paged_attention_v2_with_mask(
torch::Tensor& out, // [num_seqs, num_heads, head_size]
torch::Tensor& exp_sums, // [num_seqs, num_heads, max_num_partitions]
torch::Tensor& max_logits, // [num_seqs, num_heads, max_num_partitions]
torch::Tensor&
tmp_out, // [num_seqs, num_heads, max_num_partitions, head_size]
torch::Tensor& query, // [num_seqs, num_heads, head_size]
torch::Tensor&
key_cache, // [num_blocks, num_heads, head_size/x, block_size, x]
torch::Tensor&
value_cache, // [num_blocks, num_heads, head_size, block_size]
int64_t num_kv_heads, // [num_heads]
double scale,
torch::Tensor& block_tables, // [num_seqs, max_num_blocks_per_seq]
torch::Tensor& seq_lens, // [num_seqs]
int64_t block_size, int64_t max_seq_len,
const c10::optional<torch::Tensor>& alibi_slopes,
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride) {
const bool is_block_sparse = (blocksparse_vert_stride > 1);
DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
CALL_V2_LAUNCHER_BLOCK_SIZE)
}
#undef WARP_SIZE
#undef MAX
#undef MIN
#undef DIVIDE_ROUND_UP
\ No newline at end of file
csrc/attention/attention_with_mask_kernels_opt.cu 0 → 100644
View file @ 44e3ca68
#include <torch/all.h>
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <algorithm>
#include "attention_dtypes.h"
#include "attention_utils.cuh"
#ifdef USE_ROCM
#include <hip/hip_bf16.h>
#include "../quantization/fp8/amd/quant_utils.cuh"
typedef __hip_bfloat16 __nv_bfloat16;
#else
#include "../quantization/fp8/nvidia/quant_utils.cuh"
#endif
#ifndef USE_ROCM
#define WARP_SIZE 32
#else
#define WARP_SIZE warpSize
#endif
#include "static_switch.h"
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define DIVIDE_ROUND_UP(a, b) (((a) + (b) - 1) / (b))
namespace vllm {
// Utility function for attention softmax.
template <int NUM_WARPS>
inline __device__ float block_sum(float* red_smem, float sum) {
// Decompose the thread index into warp / lane.
int warp = threadIdx.x / WARP_SIZE;
int lane = threadIdx.x % WARP_SIZE;
// Compute the sum per warp.
#pragma unroll
for (int mask = WARP_SIZE / 2; mask >= 1; mask /= 2) {
sum += VLLM_SHFL_XOR_SYNC(sum, mask);
}
// Warp leaders store the data to shared memory.
if (lane == 0) {
red_smem[warp] = sum;
}
// Make sure the data is in shared memory.
__syncthreads();
// The warps compute the final sums.
if (lane < NUM_WARPS) {
sum = red_smem[lane];
}
// Parallel reduction inside the warp.
#pragma unroll
for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
sum += VLLM_SHFL_XOR_SYNC(sum, mask);
}
// Broadcast to other threads.
return VLLM_SHFL_SYNC(sum, 0);
}
// TODO(woosuk): Merge the last two dimensions of the grid.
// Grid: (num_heads, num_seqs, max_num_partitions).
template <typename scalar_t, typename cache_t, int HEAD_SIZE, int BLOCK_SIZE,
int NUM_THREADS, vllm::Fp8KVCacheDataType KV_DTYPE,
bool IS_BLOCK_SPARSE,
int REUSE_KV_TIMES = 1,
bool odd_nheads = false,
int PARTITION_SIZE = 0> // Zero means no partitioning.
__device__ void paged_attention_with_mask_kernel_opt(
float* __restrict__ exp_sums, // [num_seqs, num_heads, max_num_partitions]
float* __restrict__ max_logits, // [num_seqs, num_heads,
// max_num_partitions]
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 cache_t* __restrict__ k_cache, // [num_blocks, num_kv_heads,
// head_size/x, block_size, x]
const cache_t* __restrict__ v_cache, // [num_blocks, num_kv_heads,
// head_size, block_size]
const int num_heads, // [num_heads]
const int num_kv_heads, // [num_kv_heads]
const float scale,
const int* __restrict__ block_tables, // [num_seqs, max_num_blocks_per_seq]
const int* __restrict__ seq_lens, // [num_seqs]
const int max_num_blocks_per_seq,
const float* __restrict__ alibi_slopes, // [num_heads]
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
const int seq_idx = blockIdx.z;
const int partition_idx = blockIdx.y;
const int max_num_partitions = gridDim.y;
constexpr bool USE_PARTITIONING = PARTITION_SIZE > 0;
const int seq_len = seq_lens[seq_idx];
if (USE_PARTITIONING && partition_idx * PARTITION_SIZE >= seq_len) {
// No work to do. Terminate the thread block.
return;
}
if constexpr (sizeof(scalar_t)==2){
const int num_seq_blocks = DIVIDE_ROUND_UP(seq_len, BLOCK_SIZE);
const int num_blocks_per_partition =
USE_PARTITIONING ? PARTITION_SIZE / BLOCK_SIZE : num_seq_blocks;
const int partition_size = USE_PARTITIONING ? PARTITION_SIZE : num_seq_blocks * BLOCK_SIZE;
// [start_block_idx, end_block_idx) is the range of blocks to process.
const int start_block_idx =
USE_PARTITIONING ? partition_idx * num_blocks_per_partition : 0;
const int end_block_idx =
MIN(start_block_idx + num_blocks_per_partition, num_seq_blocks);
const int num_blocks = end_block_idx - start_block_idx;
// [start_token_idx, end_token_idx) is the range of tokens to process.
const int start_token_idx = start_block_idx * BLOCK_SIZE;
const int end_token_idx =
MIN(start_token_idx + num_blocks * BLOCK_SIZE, seq_len);
const int num_tokens = end_token_idx - start_token_idx;
constexpr int THREAD_GROUP_SIZE = MAX(WARP_SIZE / BLOCK_SIZE, 1);
constexpr int NUM_THREAD_GROUPS =
NUM_THREADS / THREAD_GROUP_SIZE; // Note: This assumes THREAD_GROUP_SIZE
// divides NUM_THREADS
assert(NUM_THREADS % THREAD_GROUP_SIZE == 0);
constexpr int NUM_TOKENS_PER_THREAD_GROUP =
DIVIDE_ROUND_UP(BLOCK_SIZE, WARP_SIZE);
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
const int thread_idx = threadIdx.x;
// const int warp_idx_vec = thread_idx / WARP_SIZE;
// int warp_idx =0;
// asm volatile("v_readfirstlane_b32 %0,%1"
// : "=s"(warp_idx)
// : "v"(warp_idx_vec)
// :);
// // const int warp_idx = thread_idx / WARP_SIZE;
// const int lane = thread_idx % WARP_SIZE;
//const int warp_idx = thread_idx / WARP_SIZE;
const int lane = thread_idx % WARP_SIZE;
int warp_id_vec = threadIdx.x / WARP_SIZE; //warp id in a block
int warp_idx =0;
asm volatile("v_readfirstlane_b32 %0,%1"
: "=s"(warp_idx)
: "v"(warp_id_vec)
:);
// const int head_idx = blockIdx.x;
// const int num_heads = gridDim.x;
const int num_queries_per_kv = num_heads / num_kv_heads;
// const float alibi_slope =
// alibi_slopes == nullptr ? 0.f : alibi_slopes[head_idx];
// A vector type to store a part of a key or a query.
// The vector size is configured in such a way that the threads in a thread
// group fetch or compute 16 bytes at a time. For example, if the size of a
// thread group is 4 and the data type is half, then the vector size is 16 /
// (4 * sizeof(half)) == 2.
constexpr int VEC_SIZE = MAX(32 / (THREAD_GROUP_SIZE * sizeof(scalar_t)), 1);
using K_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
using Q_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
using Quant_vec = typename Vec<cache_t, VEC_SIZE>::Type;
constexpr int NUM_ELEMS_PER_THREAD = HEAD_SIZE / THREAD_GROUP_SIZE;
constexpr int NUM_VECS_PER_THREAD = NUM_ELEMS_PER_THREAD / VEC_SIZE;
const int thread_group_idx = thread_idx / THREAD_GROUP_SIZE;
const int thread_group_offset = thread_idx % THREAD_GROUP_SIZE;
// Load the query to registers.
// Each thread in a thread group has a different part of the query.
// For example, if the the thread group size is 4, then the first thread in
// the group has 0, 4, 8, ... th vectors of the query, and the second thread
// has 1, 5, 9, ... th vectors of the query, and so on. NOTE(woosuk): Because
// q is split from a qkv tensor, it may not be contiguous.
// const scalar_t* q_ptr = q + seq_idx * q_stride;
const scalar_t* q_ptr_offset = q + seq_idx * q_stride;
__shared__ Q_vec q_vecs[REUSE_KV_TIMES * THREAD_GROUP_SIZE][NUM_VECS_PER_THREAD];
// #pragma unroll
// for (int i = thread_group_idx; i < NUM_VECS_PER_THREAD;
// i += NUM_THREAD_GROUPS) {
// const int vec_idx = thread_group_offset + i * THREAD_GROUP_SIZE;
// q_vecs[thread_group_offset][i] =
// *reinterpret_cast<const Q_vec*>(q_ptr + vec_idx * VEC_SIZE);
// }
// __syncthreads(); // TODO(naed90): possible speedup if this is replaced with a
// // memory wall right before we use q_vecs
// Memory planning.
extern __shared__ char shared_mem[];
// NOTE(woosuk): We use FP32 for the softmax logits for better accuracy.
float* logits = reinterpret_cast<float*>(shared_mem);
// Workspace for reduction.
__shared__ float red_smem[REUSE_KV_TIMES][2 * NUM_WARPS];
// float (*red_smem)[2 * NUM_WARPS] = reinterpret_cast<float(*)[2 * NUM_WARPS]>(&shared_mem[10*1024]);
// __shared__ char shared_mem[12 * 1024];
// float* logits = reinterpret_cast<float*>(shared_mem);
// __shared__ float red_smem[REUSE_KV_TIMES][2 * NUM_WARPS];
// x == THREAD_GROUP_SIZE * VEC_SIZE
// Each thread group fetches x elements from the key at a time.
constexpr int x = 16 / sizeof(cache_t);
float qk_max[REUSE_KV_TIMES];
for(int reuse_kv_idx=0; reuse_kv_idx<REUSE_KV_TIMES; reuse_kv_idx++) {
qk_max[reuse_kv_idx] = -FLT_MAX;
}
const int num_blocks_per_kv = ((num_queries_per_kv + REUSE_KV_TIMES -1) / REUSE_KV_TIMES);
const int head_idx_soffset = (blockIdx.x / num_blocks_per_kv) * num_queries_per_kv + (blockIdx.x % num_blocks_per_kv) * REUSE_KV_TIMES;
const int kv_head_idx = head_idx_soffset / num_queries_per_kv;
const int q_boundary = (kv_head_idx + 1)* num_queries_per_kv;
#pragma unroll
for(int reuse_kv_idx=0; reuse_kv_idx<REUSE_KV_TIMES; reuse_kv_idx++) {
const int head_idx = head_idx_soffset + reuse_kv_idx;//blockIdx.x * REUSE_KV_TIMES + reuse_kv_idx;
const scalar_t* q_ptr = q_ptr_offset + head_idx * HEAD_SIZE;
#pragma unroll
for (int i = thread_group_idx; i < NUM_VECS_PER_THREAD; i += NUM_THREAD_GROUPS) {
const int vec_idx = thread_group_offset + i * THREAD_GROUP_SIZE;
q_vecs[reuse_kv_idx*THREAD_GROUP_SIZE + thread_group_offset][i] = *reinterpret_cast<const Q_vec*>(q_ptr + vec_idx * VEC_SIZE);
}
}
__syncthreads(); // TODO(naed90): possible speedup if this is replaced with a memory wall right before we use q_vecs
// Iterate over the key blocks.
// Each warp fetches a block of keys for each iteration.
// Each thread group in a warp fetches a key from the block, and computes
// dot product with the query.
const int* block_table = block_tables + seq_idx * max_num_blocks_per_seq;
for (int block_idx = start_block_idx + warp_idx; block_idx < end_block_idx; block_idx += NUM_WARPS) {
// NOTE(woosuk): The block number is stored in int32. However, we cast it to
// int64 because int32 can lead to overflow when this variable is multiplied
// by large numbers (e.g., kv_block_stride).
// For blocksparse attention: skip computation on blocks that are not
// attended
for(int reuse_kv_idx=0; reuse_kv_idx<REUSE_KV_TIMES; reuse_kv_idx++) {
const int head_idx = head_idx_soffset + reuse_kv_idx;//blockIdx.x * REUSE_KV_TIMES + reuse_kv_idx;
if(!odd_nheads || head_idx < q_boundary) {
// blocksparse specific vars
int bs_block_offset;
int q_bs_block_id;
if constexpr (IS_BLOCK_SPARSE) {
// const int num_blocksparse_blocks = DIVIDE_ROUND_UP(seq_len,
// blocksparse_block_size);
q_bs_block_id = (seq_len - 1) / blocksparse_block_size;
if (blocksparse_head_sliding_step >= 0)
// sliding on q heads
bs_block_offset =
(tp_rank * num_heads + head_idx) * blocksparse_head_sliding_step + 1;
else
// sliding on kv heads
bs_block_offset = (tp_rank * num_kv_heads + kv_head_idx) *
(-blocksparse_head_sliding_step) +
1;
}
if constexpr (IS_BLOCK_SPARSE) {
const int k_bs_block_id = block_idx * BLOCK_SIZE / blocksparse_block_size;
const bool is_remote =
((k_bs_block_id + bs_block_offset) % blocksparse_vert_stride == 0);
const bool is_local =
(k_bs_block_id > q_bs_block_id - blocksparse_local_blocks);
if (!is_remote && !is_local) {
for (int i = 0; i < NUM_TOKENS_PER_THREAD_GROUP; i++) {
const int physical_block_offset =
(thread_group_idx + i * WARP_SIZE) % BLOCK_SIZE;
const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
if (thread_group_offset == 0) {
// NOTE(linxihui): assign very large number to skipped tokens to
// avoid contribution to the sumexp softmax normalizer. This will
// not be used at computing sum(softmax*v) as the blocks will be
// skipped.
logits[token_idx - start_token_idx] = -FLT_MAX;
}
}
continue;
}
}
const float alibi_slope = alibi_slopes == nullptr ? 0.f : alibi_slopes[head_idx];
const int64_t physical_block_number = static_cast<int64_t>(block_table[block_idx]);
// Load a key to registers.
// Each thread in a thread group has a different part of the key.
// For example, if the the thread group size is 4, then the first thread in
// the group has 0, 4, 8, ... th vectors of the key, and the second thread
// has 1, 5, 9, ... th vectors of the key, and so on.
for (int i = 0; i < NUM_TOKENS_PER_THREAD_GROUP; i++) {
const int physical_block_offset = (thread_group_idx + i * WARP_SIZE) % BLOCK_SIZE;
const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
K_vec k_vecs[NUM_VECS_PER_THREAD];
if(reuse_kv_idx == 0) {
#pragma unroll
for (int j = 0; j < NUM_VECS_PER_THREAD; j++) {
const cache_t* k_ptr =
k_cache + physical_block_number * kv_block_stride +
kv_head_idx * kv_head_stride + physical_block_offset * x;
const int vec_idx = thread_group_offset + j * THREAD_GROUP_SIZE;
const int offset1 = (vec_idx * VEC_SIZE) / x;
const int offset2 = (vec_idx * VEC_SIZE) % x;
if constexpr (KV_DTYPE == Fp8KVCacheDataType::kAuto) {
k_vecs[j] = *reinterpret_cast<const K_vec*>(
k_ptr + offset1 * BLOCK_SIZE * x + offset2);
} else {
// Vector conversion from Quant_vec to K_vec.
Quant_vec k_vec_quant = *reinterpret_cast<const Quant_vec*>(
k_ptr + offset1 * BLOCK_SIZE * x + offset2);
k_vecs[j] = fp8::scaled_convert<K_vec, Quant_vec, KV_DTYPE>(
k_vec_quant, k_scale);
}
}
}
__builtin_amdgcn_sched_barrier(0);
// Compute dot product.
// This includes a reduction across the threads in the same thread group.
float qk = scale * Qk_dot<scalar_t, THREAD_GROUP_SIZE>::dot(q_vecs[reuse_kv_idx*THREAD_GROUP_SIZE + thread_group_offset], k_vecs);
// Add the ALiBi bias if slopes are given.
qk += (alibi_slope != 0) ? alibi_slope * (token_idx - seq_len + 1) : 0;
// used for tree-style attention
if (attn_masks != nullptr) {
const int* attn_masks_ptr = attn_masks + seq_idx * attn_masks_stride;
if (attn_masks_ptr[token_idx] == 0) {
qk = -FLT_MAX;
}
}
__builtin_amdgcn_sched_barrier(0);
if (thread_group_offset == 0) {
// Store the partial reductions to shared memory.
// NOTE(woosuk): It is required to zero out the masked logits.
const bool mask = token_idx >= seq_len;
// used for tree-style attention
/*if (attn_masks != nullptr) {
const int* attn_masks_ptr = attn_masks + seq_idx * attn_masks_stride;
mask |= attn_masks_ptr[token_idx] == 0;
}*/
logits[(reuse_kv_idx * partition_size) + (token_idx - start_token_idx)] = mask ? 0.f : qk;
// Update the max value.
qk_max[reuse_kv_idx] = mask ? qk_max[reuse_kv_idx] : fmaxf(qk_max[reuse_kv_idx], qk);
}
}
}
}
}
// Get the sum of the exp values.
float exp_sum[REUSE_KV_TIMES] = {0.f};
// Perform reduction across the threads in the same warp to get the
// max qk value for each "warp" (not across the thread block yet).
// The 0-th thread of each thread group already has its max qk value.
for(int reuse_kv_idx=0; reuse_kv_idx<REUSE_KV_TIMES; reuse_kv_idx++) {
const int head_idx = head_idx_soffset + reuse_kv_idx;
if(!odd_nheads || head_idx < q_boundary) {
#pragma unroll
for (int mask = WARP_SIZE / 2; mask >= THREAD_GROUP_SIZE; mask /= 2) {
qk_max[reuse_kv_idx] = fmaxf(qk_max[reuse_kv_idx], VLLM_SHFL_XOR_SYNC(qk_max[reuse_kv_idx], mask));
}
if (lane == 0) {
red_smem[reuse_kv_idx][warp_idx] = qk_max[reuse_kv_idx];
}
__syncthreads();
// TODO(woosuk): Refactor this part.
// Get the max qk value for the sequence.
qk_max[reuse_kv_idx] = lane < NUM_WARPS ? red_smem[reuse_kv_idx][lane] : -FLT_MAX;
#pragma unroll
for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
qk_max[reuse_kv_idx] = fmaxf(qk_max[reuse_kv_idx], VLLM_SHFL_XOR_SYNC(qk_max[reuse_kv_idx], mask));
}
// Broadcast the max qk value to all threads.
qk_max[reuse_kv_idx] = VLLM_SHFL_SYNC(qk_max[reuse_kv_idx], 0);
for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) {
float val = __expf(logits[(reuse_kv_idx * partition_size) + i] - qk_max[reuse_kv_idx]);
logits[(reuse_kv_idx * partition_size) + i] = val;
exp_sum[reuse_kv_idx] += val;
}
exp_sum[reuse_kv_idx] = block_sum<NUM_WARPS>(&red_smem[reuse_kv_idx][NUM_WARPS], exp_sum[reuse_kv_idx]);
// Compute softmax.
const float inv_sum = __fdividef(1.f, exp_sum[reuse_kv_idx] + 1e-6f);
for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) {
logits[(reuse_kv_idx * partition_size) + i] *= inv_sum;
}
__syncthreads();
// If partitioning is enabled, store the max logit and exp_sum.
if (USE_PARTITIONING && thread_idx == 0) {
float* max_logits_ptr = max_logits +
seq_idx * num_heads * max_num_partitions +
head_idx * max_num_partitions + partition_idx;
*max_logits_ptr = qk_max[reuse_kv_idx];
float* exp_sums_ptr = exp_sums + seq_idx * num_heads * max_num_partitions +
head_idx * max_num_partitions + partition_idx;
*exp_sums_ptr = exp_sum[reuse_kv_idx];
}
}
}
// Each thread will fetch 16 bytes from the value cache at a time.
constexpr int V_VEC_SIZE = MIN(16 / sizeof(scalar_t), BLOCK_SIZE);
using V_vec = typename Vec<scalar_t, V_VEC_SIZE>::Type;
using L_vec = typename Vec<scalar_t, V_VEC_SIZE>::Type;
using V_quant_vec = typename Vec<cache_t, V_VEC_SIZE>::Type;
using Float_L_vec = typename FloatVec<L_vec>::Type;
constexpr int NUM_V_VECS_PER_ROW = BLOCK_SIZE / V_VEC_SIZE;
constexpr int NUM_ROWS_PER_ITER = WARP_SIZE / NUM_V_VECS_PER_ROW;
constexpr int NUM_ROWS_PER_THREAD =
DIVIDE_ROUND_UP(HEAD_SIZE, NUM_ROWS_PER_ITER);
// NOTE(woosuk): We use FP32 for the accumulator for better accuracy.
float accs[REUSE_KV_TIMES][NUM_ROWS_PER_THREAD];
#pragma unroll
for(int reuse_kv_idx=0; reuse_kv_idx<REUSE_KV_TIMES; reuse_kv_idx++) {
#pragma unroll
for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
accs[reuse_kv_idx][i] = 0.f;
}
}
scalar_t zero_value;
zero(zero_value);
for (int block_idx = start_block_idx + warp_idx; block_idx < end_block_idx;
block_idx += NUM_WARPS) {
const int64_t physical_block_number =
static_cast<int64_t>(block_table[block_idx]);
const int physical_block_offset = (lane % NUM_V_VECS_PER_ROW) * V_VEC_SIZE;
const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
L_vec logits_vec;
#pragma unroll
for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
V_vec v_vec;
for(int reuse_kv_idx=0; reuse_kv_idx<REUSE_KV_TIMES; reuse_kv_idx++) {
// NOTE(woosuk): The block number is stored in int32. However, we cast it to
// int64 because int32 can lead to overflow when this variable is multiplied
// by large numbers (e.g., kv_block_stride).
// For blocksparse attention: skip computation on blocks that are not
// attended
// blocksparse specific vars
const int head_idx = head_idx_soffset + reuse_kv_idx;
int bs_block_offset;
int q_bs_block_id;
if constexpr (IS_BLOCK_SPARSE) {
// const int num_blocksparse_blocks = DIVIDE_ROUND_UP(seq_len,
// blocksparse_block_size);
q_bs_block_id = (seq_len - 1) / blocksparse_block_size;
if (blocksparse_head_sliding_step >= 0)
// sliding on q heads
bs_block_offset =
(tp_rank * num_heads + head_idx) * blocksparse_head_sliding_step + 1;
else
// sliding on kv heads
bs_block_offset = (tp_rank * num_kv_heads + kv_head_idx) *
(-blocksparse_head_sliding_step) +
1;
}
if constexpr (IS_BLOCK_SPARSE) {
int v_bs_block_id = block_idx * BLOCK_SIZE / blocksparse_block_size;
if (!((v_bs_block_id + bs_block_offset) % blocksparse_vert_stride == 0) &&
!((v_bs_block_id > q_bs_block_id - blocksparse_local_blocks))) {
continue;
}
}
if(!odd_nheads || head_idx < q_boundary) {
const cache_t* v_ptr = v_cache + physical_block_number * kv_block_stride
+ kv_head_idx * kv_head_stride;
from_float(logits_vec, *reinterpret_cast<Float_L_vec*>(logits + (reuse_kv_idx * partition_size) + token_idx - start_token_idx));
// scalar_t* logits_vec_ptr = reinterpret_cast<scalar_t*>(&logits_vec);
// for(int i=0;i<8;++i){
// from_float(*(logits_vec_ptr+i), 1000);
// }
if(reuse_kv_idx==0) {
const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
if (row_idx < HEAD_SIZE) {
const int offset = row_idx * BLOCK_SIZE + physical_block_offset;
if constexpr (KV_DTYPE == Fp8KVCacheDataType::kAuto) {
v_vec = *reinterpret_cast<const V_vec*>(v_ptr + offset);
} else {
V_quant_vec v_quant_vec =
*reinterpret_cast<const V_quant_vec*>(v_ptr + offset);
// Vector conversion from V_quant_vec to V_vec.
v_vec = fp8::scaled_convert<V_vec, V_quant_vec, KV_DTYPE>(v_quant_vec,
v_scale);
}
if (block_idx == num_seq_blocks - 1) {
// NOTE(woosuk): When v_vec contains the tokens that are out of the
// context, we should explicitly zero out the values since they may
// contain NaNs. See
// https://github.com/vllm-project/vllm/issues/641#issuecomment-1682544472
scalar_t* v_vec_ptr = reinterpret_cast<scalar_t*>(&v_vec);
#pragma unroll
for (int j = 0; j < V_VEC_SIZE; j++) {
v_vec_ptr[j] = token_idx + j < seq_len ? v_vec_ptr[j] : zero_value;
}
}
// if(threadIdx.x==0){
// scalar_t* v_vec_ptr = reinterpret_cast<scalar_t*>(&v_vec);
// scalar_t* logits_vec_ptr = reinterpret_cast<scalar_t*>(&logits_vec);
// for(int i=0;i<8;++i){
// printf("v_vec[%d] = %f\n",i, half_to_float(v_vec_ptr[i]));
// // from_float(*(v_vec_ptr + i), 1000);
// }
// for(int i=0;i<8;++i){
// printf("logits_vec[%d] = %f\n",i,half_to_float(logits_vec_ptr[i]));
// // from_float(*(logits_vec_ptr + i), 1000);
// }
// }
// accs[reuse_kv_idx][i] += dot(logits_vec, v_vec);
}
}
accs[reuse_kv_idx][i] += dot(logits_vec, v_vec);
}
}
}
}
// Perform reduction within each warp.
#pragma unroll
for(int reuse_kv_idx=0; reuse_kv_idx<REUSE_KV_TIMES; reuse_kv_idx++) {
int head_idx = head_idx_soffset + reuse_kv_idx;
if(!odd_nheads || head_idx < q_boundary) {
#pragma unroll
for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
float acc = accs[reuse_kv_idx][i];
#pragma unroll
for (int mask = NUM_V_VECS_PER_ROW / 2; mask >= 1; mask /= 2) {
acc += VLLM_SHFL_XOR_SYNC(acc, mask);
}
accs[reuse_kv_idx][i] = acc;
}
// NOTE(woosuk): A barrier is required because the shared memory space for
// logits is reused for the output.
__syncthreads();
// Perform reduction across warps.
float* out_smem = reinterpret_cast<float*>(shared_mem);
#pragma unroll
for (int i = NUM_WARPS; i > 1; i /= 2) {
int mid = i / 2;
// Upper warps write to shared memory.
if (warp_idx >= mid && warp_idx < i) {
float* dst = &out_smem[(reuse_kv_idx * (NUM_WARPS / 2) * HEAD_SIZE) + (warp_idx - mid) * HEAD_SIZE];
#pragma unroll
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) {
dst[row_idx] = accs[reuse_kv_idx][i];
}
}
}
__syncthreads();
// Lower warps update the output.
if (warp_idx < mid) {
const float* src = &out_smem[(reuse_kv_idx * (NUM_WARPS / 2) * HEAD_SIZE) + warp_idx * HEAD_SIZE];
#pragma unroll
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) {
accs[reuse_kv_idx][i] += src[row_idx];
}
}
}
__syncthreads();
}
// Write the final output.
if (warp_idx == 0) {
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;
#pragma unroll
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[reuse_kv_idx][i]);
}
}
}
}
}
}
}
// Grid: (num_heads, num_seqs, 1).
template <typename scalar_t, typename cache_t, int HEAD_SIZE, int BLOCK_SIZE,
int NUM_THREADS, vllm::Fp8KVCacheDataType KV_DTYPE,
int REUSE_KV_TIMES = 1,
bool IS_BLOCK_SPARSE,
bool odd_nheads = false>
__global__ __launch_bounds__(256,1) void paged_attention_v1_with_mask_kernel_opt(
scalar_t* __restrict__ out, // [num_seqs, num_heads, head_size]
const scalar_t* __restrict__ q, // [num_seqs, num_heads, head_size]
const cache_t* __restrict__ k_cache, // [num_blocks, num_kv_heads,
// head_size/x, block_size, x]
const cache_t* __restrict__ v_cache, // [num_blocks, num_kv_heads,
// head_size, block_size]
const int num_heads, // [num_heads]
const int num_kv_heads, // [num_heads]
const float scale,
const int* __restrict__ block_tables, // [num_seqs, max_num_blocks_per_seq]
const int* __restrict__ seq_lens, // [num_seqs]
const int max_num_blocks_per_seq,
const float* __restrict__ alibi_slopes, // [num_heads]
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
paged_attention_with_mask_kernel_opt<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
KV_DTYPE, IS_BLOCK_SPARSE, REUSE_KV_TIMES, odd_nheads>(
/* exp_sums */ nullptr, /* max_logits */ nullptr, out, q, k_cache,
v_cache, num_heads, num_kv_heads, scale, block_tables, seq_lens,
max_num_blocks_per_seq, alibi_slopes, q_stride, kv_block_stride,
kv_head_stride, k_scale, v_scale, tp_rank, blocksparse_local_blocks,
blocksparse_vert_stride, blocksparse_block_size,
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
}
// Grid: (num_heads, num_seqs, max_num_partitions).
template <typename scalar_t, typename cache_t, int HEAD_SIZE, int BLOCK_SIZE,
int NUM_THREADS, vllm::Fp8KVCacheDataType KV_DTYPE,
bool IS_BLOCK_SPARSE,
int REUSE_KV_TIMES,
int PARTITION_SIZE,
bool odd_nheads = false>
__global__ __launch_bounds__(256,1) void paged_attention_v2_with_mask_kernel_opt(
float* __restrict__ exp_sums, // [num_seqs, num_heads, max_num_partitions]
float* __restrict__ max_logits, // [num_seqs, num_heads,
// max_num_partitions]
scalar_t* __restrict__ tmp_out, // [num_seqs, num_heads,
// max_num_partitions, head_size]
const scalar_t* __restrict__ q, // [num_seqs, num_heads, head_size]
const cache_t* __restrict__ k_cache, // [num_blocks, num_kv_heads,
// head_size/x, block_size, x]
const cache_t* __restrict__ v_cache, // [num_blocks, num_kv_heads,
// head_size, block_size]
const int num_heads, // [num_heads]
const int num_kv_heads, // [num_kv_heads]
const float scale,
const int* __restrict__ block_tables, // [num_seqs, max_num_blocks_per_seq]
const int* __restrict__ seq_lens, // [num_seqs]
const int max_num_blocks_per_seq,
const float* __restrict__ alibi_slopes, // [num_heads]
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
paged_attention_with_mask_kernel_opt<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
KV_DTYPE, IS_BLOCK_SPARSE, REUSE_KV_TIMES, odd_nheads, PARTITION_SIZE>(
exp_sums, max_logits, tmp_out, q, k_cache, v_cache, num_heads, num_kv_heads, scale,
block_tables, seq_lens, max_num_blocks_per_seq, alibi_slopes, q_stride,
kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank,
blocksparse_local_blocks, blocksparse_vert_stride, blocksparse_block_size,
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
}
// Grid: (num_heads, num_seqs).
template <typename scalar_t, int HEAD_SIZE, int NUM_THREADS,
int PARTITION_SIZE>
__global__ __launch_bounds__(256,1) void paged_attention_v2_reduce_kernel_opt(
scalar_t* __restrict__ out, // [num_seqs, num_heads, head_size]
const float* __restrict__ exp_sums, // [num_seqs, num_heads,
// max_num_partitions]
const float* __restrict__ max_logits, // [num_seqs, num_heads,
// max_num_partitions]
const scalar_t* __restrict__ tmp_out, // [num_seqs, num_heads,
// max_num_partitions, head_size]
const int* __restrict__ seq_lens, // [num_seqs]
const int max_num_partitions) {
const int num_heads = gridDim.x;
const int head_idx = blockIdx.x;
const int seq_idx = blockIdx.y;
const int seq_len = seq_lens[seq_idx];
const int num_partitions = DIVIDE_ROUND_UP(seq_len, PARTITION_SIZE);
if (num_partitions == 1) {
// No need to reduce. Only copy tmp_out to out.
scalar_t* out_ptr =
out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
const scalar_t* tmp_out_ptr =
tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
head_idx * max_num_partitions * HEAD_SIZE;
for (int i = threadIdx.x; i < HEAD_SIZE; i += blockDim.x) {
out_ptr[i] = tmp_out_ptr[i];
}
// Terminate the thread block.
return;
}
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
const int warp_idx = threadIdx.x / WARP_SIZE;
const int lane = threadIdx.x % WARP_SIZE;
// Size: 2 * num_partitions.
extern __shared__ char shared_mem[];
// Workspace for reduction.
__shared__ float red_smem[2 * NUM_WARPS];
// Load max logits to shared memory.
float* shared_max_logits = reinterpret_cast<float*>(shared_mem);
const float* max_logits_ptr = max_logits +
seq_idx * num_heads * max_num_partitions +
head_idx * max_num_partitions;
float max_logit = -FLT_MAX;
for (int i = threadIdx.x; i < num_partitions; i += blockDim.x) {
const float l = max_logits_ptr[i];
shared_max_logits[i] = l;
max_logit = fmaxf(max_logit, l);
}
__syncthreads();
// Get the global max logit.
// Reduce within the warp.
#pragma unroll
for (int mask = WARP_SIZE / 2; mask >= 1; mask /= 2) {
max_logit = fmaxf(max_logit, VLLM_SHFL_XOR_SYNC(max_logit, mask));
}
if (lane == 0) {
red_smem[warp_idx] = max_logit;
}
__syncthreads();
// Reduce across warps.
max_logit = lane < NUM_WARPS ? red_smem[lane] : -FLT_MAX;
#pragma unroll
for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
max_logit = fmaxf(max_logit, VLLM_SHFL_XOR_SYNC(max_logit, mask));
}
// Broadcast the max value to all threads.
max_logit = VLLM_SHFL_SYNC(max_logit, 0);
// Load rescaled exp sums to shared memory.
float* shared_exp_sums =
reinterpret_cast<float*>(shared_mem + sizeof(float) * num_partitions);
const float* exp_sums_ptr = exp_sums +
seq_idx * num_heads * max_num_partitions +
head_idx * max_num_partitions;
float global_exp_sum = 0.0f;
for (int i = threadIdx.x; i < num_partitions; i += blockDim.x) {
float l = shared_max_logits[i];
float rescaled_exp_sum = exp_sums_ptr[i] * expf(l - max_logit);
global_exp_sum += rescaled_exp_sum;
shared_exp_sums[i] = rescaled_exp_sum;
}
__syncthreads();
global_exp_sum = block_sum<NUM_WARPS>(&red_smem[NUM_WARPS], global_exp_sum);
const float inv_global_exp_sum = __fdividef(1.0f, global_exp_sum + 1e-6f);
// Aggregate tmp_out to out.
const scalar_t* tmp_out_ptr =
tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
head_idx * max_num_partitions * HEAD_SIZE;
scalar_t* out_ptr =
out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
#pragma unroll
for (int i = threadIdx.x; i < HEAD_SIZE; i += NUM_THREADS) {
float acc = 0.0f;
for (int j = 0; j < num_partitions; ++j) {
acc += to_float(tmp_out_ptr[j * HEAD_SIZE + i]) * shared_exp_sums[j] *
inv_global_exp_sum;
}
from_float(out_ptr[i], acc);
}
}
} // namespace vllm
#define LAUNCH_PAGED_ATTENTION_V1(HEAD_SIZE) \
VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize( \
((void*)vllm::paged_attention_v1_with_mask_kernel_opt<T, CACHE_T, HEAD_SIZE, \
BLOCK_SIZE, NUM_THREADS, \
KV_DTYPE, REUSE_KV_TIMES, IS_BLOCK_SPARSE, odd_nheads>), \
shared_mem_size); \
hipLaunchKernelGGL(( vllm::paged_attention_v1_with_mask_kernel_opt<T, CACHE_T, HEAD_SIZE, BLOCK_SIZE, \
NUM_THREADS, KV_DTYPE, REUSE_KV_TIMES, IS_BLOCK_SPARSE, odd_nheads>) \
, dim3(grid), dim3(block), shared_mem_size, stream, \
out_ptr, query_ptr, key_cache_ptr, value_cache_ptr, num_heads, num_kv_heads, \
scale, block_tables_ptr, seq_lens_ptr, max_num_blocks_per_seq, \
alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride, \
k_scale, v_scale, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step, attn_masks_ptr, \
attn_masks_stride);
// #define LAUNCH_PAGED_ATTENTION_V1(HEAD_SIZE) \
// vllm::paged_attention_v1_kernel<T, CACHE_T, HEAD_SIZE, BLOCK_SIZE, \
// NUM_THREADS, KV_DTYPE, REUSE_KV_TIMES, IS_BLOCK_SPARSE, odd_nheads> \
// <<<dim3(grid), dim3(block)>>>( \
// out_ptr, query_ptr, key_cache_ptr, value_cache_ptr, num_heads, num_kv_heads, \
// scale, block_tables_ptr, seq_lens_ptr, max_num_blocks_per_seq, \
// alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride, \
// kv_scale, tp_rank, blocksparse_local_blocks, \
// blocksparse_vert_stride, blocksparse_block_size, \
// blocksparse_head_sliding_step);
// TODO(woosuk): Tune NUM_THREADS.
template <typename T, typename CACHE_T, int BLOCK_SIZE,
vllm::Fp8KVCacheDataType KV_DTYPE, bool IS_BLOCK_SPARSE>
void paged_attention_v1_launcher(
torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int num_kv_heads, float scale,
torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
float v_scale, const int tp_rank, const int blocksparse_local_blocks,
const int blocksparse_vert_stride, const int blocksparse_block_size,
const int blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks,
const int64_t attn_masks_stride=0) {
int num_seqs = query.size(0);
int num_heads = query.size(1);
int head_size = query.size(2);
int max_num_blocks_per_seq = block_tables.size(1);
int q_stride = query.stride(0);
int kv_block_stride = key_cache.stride(0);
int kv_head_stride = key_cache.stride(1);
int num_threads = 128;
if(num_heads!=num_kv_heads){
num_threads =256;
}
[[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
assert(head_size % thread_group_size == 0);
// NOTE: alibi_slopes is optional.
const float* alibi_slopes_ptr =
alibi_slopes
? reinterpret_cast<const float*>(alibi_slopes.value().data_ptr())
: nullptr;
T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
CACHE_T* key_cache_ptr = reinterpret_cast<CACHE_T*>(key_cache.data_ptr());
CACHE_T* value_cache_ptr = reinterpret_cast<CACHE_T*>(value_cache.data_ptr());
int* block_tables_ptr = block_tables.data_ptr<int>();
int* seq_lens_ptr = seq_lens.data_ptr<int>();
// NOTE: attn_masks is optional.
const int* attn_masks_ptr =
attn_masks
? attn_masks.value().data_ptr<int>()
: nullptr;
int padded_max_seq_len = DIVIDE_ROUND_UP(max_seq_len, BLOCK_SIZE) * BLOCK_SIZE;
REUSEKV_SWITCH_V1(num_heads * num_seqs , [&] {
BOOL_SWITCH((num_heads/num_kv_heads % REUSE_KV_TIMES != 0), odd_nheads, [&] {
HEADSIZE_SWITCH(head_size, [&] {
NUM_THREADS_SWITCH(num_threads, [&] {
OPT_SWITCH(num_heads == num_kv_heads, [&] {
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
int logits_size = REUSE_KV_TIMES*padded_max_seq_len * sizeof(float);
int outputs_size = REUSE_KV_TIMES*(NUM_WARPS / 2) * head_size * sizeof(float);
// Python-side check in vllm.worker.worker._check_if_can_support_max_seq_len
// Keep that in sync with the logic here!
int shared_mem_size = ::max(logits_size, outputs_size);
if(num_heads == num_kv_heads) shared_mem_size = ::max(12 * 1024, shared_mem_size);
// int shared_mem_size = ::max(31*1024, ::max(logits_size, outputs_size));
// std::cout<<"shared_mem_size = "<<shared_mem_size<<std::endl;
dim3 grid((num_heads/num_kv_heads + REUSE_KV_TIMES - 1) / REUSE_KV_TIMES*num_kv_heads, 1, num_seqs);
dim3 block(NUM_THREADS);
const at::hip::OptionalHIPGuardMasqueradingAsCUDA device_guard(device_of(query));
const hipStream_t stream = at::hip::getCurrentHIPStreamMasqueradingAsCUDA();
LAUNCH_PAGED_ATTENTION_V1(HEAD_SIZE);
});
});
});
});
});
}
#define CALL_V1_LAUNCHER(T, CACHE_T, BLOCK_SIZE, KV_DTYPE, IS_BLOCK_SPARSE) \
paged_attention_v1_launcher<T, CACHE_T, BLOCK_SIZE, KV_DTYPE, \
IS_BLOCK_SPARSE>( \
out, query, key_cache, value_cache, num_kv_heads, scale, block_tables, \
seq_lens, max_seq_len, alibi_slopes, k_scale, v_scale, tp_rank, \
blocksparse_local_blocks, blocksparse_vert_stride, \
blocksparse_block_size, blocksparse_head_sliding_step, \
attn_masks, attn_masks_stride);
#define CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE) \
switch (is_block_sparse) { \
case true: \
CALL_V1_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE, true); \
break; \
case false: \
CALL_V1_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE, false); \
break; \
}
// NOTE(woosuk): To reduce the compilation time, we omitted block sizes
// 1, 2, 4, 64, 128, 256.
#define CALL_V1_LAUNCHER_BLOCK_SIZE(T, CACHE_T, KV_DTYPE) \
switch (block_size) { \
case 8: \
CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, 8, KV_DTYPE); \
break; \
case 16: \
CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, 16, KV_DTYPE); \
break; \
case 32: \
CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, 32, KV_DTYPE); \
break; \
default: \
TORCH_CHECK(false, "Unsupported block size: ", block_size); \
break; \
}
void paged_attention_v1_opt_with_mask(
torch::Tensor& out, // [num_seqs, num_heads, head_size]
torch::Tensor& query, // [num_seqs, num_heads, head_size]
torch::Tensor&
key_cache, // [num_blocks, num_heads, head_size/x, block_size, x]
torch::Tensor&
value_cache, // [num_blocks, num_heads, head_size, block_size]
int64_t num_kv_heads, // [num_heads]
double scale,
torch::Tensor& block_tables, // [num_seqs, max_num_blocks_per_seq]
torch::Tensor& seq_lens, // [num_seqs]
int64_t block_size, int64_t max_seq_len,
const c10::optional<torch::Tensor>& alibi_slopes,
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride) {
const bool is_block_sparse = (blocksparse_vert_stride > 1);
DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
CALL_V1_LAUNCHER_BLOCK_SIZE)
}
#define LAUNCH_PAGED_ATTENTION_V2(HEAD_SIZE) \
hipLaunchKernelGGL(( vllm::paged_attention_v2_with_mask_kernel_opt<T, CACHE_T, HEAD_SIZE, BLOCK_SIZE, \
NUM_THREADS, KV_DTYPE, IS_BLOCK_SPARSE, \
REUSE_KV_TIMES, PARTITION_SIZE, odd_nheads>) \
, dim3(grid), dim3(block), shared_mem_size, stream, \
exp_sums_ptr, max_logits_ptr, tmp_out_ptr, query_ptr, key_cache_ptr, \
value_cache_ptr, num_heads, num_kv_heads, scale, block_tables_ptr, \
seq_lens_ptr, max_num_blocks_per_seq, alibi_slopes_ptr, q_stride, \
kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank, \
blocksparse_local_blocks, blocksparse_vert_stride, \
blocksparse_block_size, blocksparse_head_sliding_step, \
attn_masks_ptr, attn_masks_stride); \
hipLaunchKernelGGL(( vllm::paged_attention_v2_reduce_kernel_opt<T, HEAD_SIZE, NUM_THREADS, \
PARTITION_SIZE>) \
, dim3(reduce_grid), dim3(block), reduce_shared_mem_size, stream, \
out_ptr, exp_sums_ptr, max_logits_ptr, tmp_out_ptr, seq_lens_ptr, \
max_num_partitions);
template <typename T, typename CACHE_T, int BLOCK_SIZE,
vllm::Fp8KVCacheDataType KV_DTYPE, bool IS_BLOCK_SPARSE,
int NUM_THREADS = 256, int PARTITION_SIZE = 512>
void paged_attention_v2_launcher(
torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int num_kv_heads, float scale,
torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
float v_scale, const int tp_rank, const int blocksparse_local_blocks,
const int blocksparse_vert_stride, const int blocksparse_block_size,
const int blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks,
const int attn_masks_stride) {
int num_seqs = query.size(0);
int num_heads = query.size(1);
int head_size = query.size(2);
int max_num_blocks_per_seq = block_tables.size(1);
int q_stride = query.stride(0);
int kv_block_stride = key_cache.stride(0);
int kv_head_stride = key_cache.stride(1);
[[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
assert(head_size % thread_group_size == 0);
// NOTE: alibi_slopes is optional.
const float* alibi_slopes_ptr =
alibi_slopes
? reinterpret_cast<const float*>(alibi_slopes.value().data_ptr())
: nullptr;
T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
float* exp_sums_ptr = reinterpret_cast<float*>(exp_sums.data_ptr());
float* max_logits_ptr = reinterpret_cast<float*>(max_logits.data_ptr());
T* tmp_out_ptr = reinterpret_cast<T*>(tmp_out.data_ptr());
T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
CACHE_T* key_cache_ptr = reinterpret_cast<CACHE_T*>(key_cache.data_ptr());
CACHE_T* value_cache_ptr = reinterpret_cast<CACHE_T*>(value_cache.data_ptr());
int* block_tables_ptr = block_tables.data_ptr<int>();
int* seq_lens_ptr = seq_lens.data_ptr<int>();
// NOTE: attn_masks is optional.
const int* attn_masks_ptr =
attn_masks ? attn_masks.value().data_ptr<int>() : nullptr;
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
int max_num_partitions = DIVIDE_ROUND_UP(max_seq_len, PARTITION_SIZE);
REUSEKV_SWITCH(num_heads * max_num_partitions * num_seqs , [&] {
BOOL_SWITCH((num_heads/num_kv_heads % REUSE_KV_TIMES != 0), odd_nheads, [&] {
HEADSIZE_SWITCH(head_size, [&] {
OPT_SWITCH(num_heads == num_kv_heads, [&] {
int logits_size = REUSE_KV_TIMES*PARTITION_SIZE * sizeof(float);
int outputs_size = REUSE_KV_TIMES*(NUM_WARPS / 2) * head_size * sizeof(float);
// For paged attention v2 kernel.
// dim3 grid(num_heads, max_num_partitions, num_seqs);
dim3 grid;
grid.x = (num_heads/num_kv_heads + REUSE_KV_TIMES -1)/REUSE_KV_TIMES * num_kv_heads;
grid.y = max_num_partitions;
grid.z = num_seqs;
// int shared_mem_size = ::max(1024*32, ::max(logits_size, outputs_size));
int shared_mem_size = ::max(logits_size, outputs_size);
// For paged attention v2 reduce kernel.
dim3 reduce_grid(num_heads, num_seqs);
int reduce_shared_mem_size = 2 * max_num_partitions * sizeof(float);
dim3 block(NUM_THREADS);
const at::hip::OptionalHIPGuardMasqueradingAsCUDA device_guard(device_of(query));
const hipStream_t stream = at::hip::getCurrentHIPStreamMasqueradingAsCUDA();
LAUNCH_PAGED_ATTENTION_V2(HEAD_SIZE);
});
});
});
});
}
#define CALL_V2_LAUNCHER(T, CACHE_T, BLOCK_SIZE, KV_DTYPE, IS_BLOCK_SPARSE) \
paged_attention_v2_launcher<T, CACHE_T, BLOCK_SIZE, KV_DTYPE, \
IS_BLOCK_SPARSE>( \
out, exp_sums, max_logits, tmp_out, query, key_cache, value_cache, \
num_kv_heads, scale, block_tables, seq_lens, max_seq_len, alibi_slopes, \
k_scale, v_scale, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
#define CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE) \
switch (is_block_sparse) { \
case true: \
CALL_V2_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE, true); \
break; \
case false: \
CALL_V2_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE, false); \
break; \
}
// NOTE(woosuk): To reduce the compilation time, we omitted block sizes
// 1, 2, 4, 64, 128, 256.
#define CALL_V2_LAUNCHER_BLOCK_SIZE(T, CACHE_T, KV_DTYPE) \
switch (block_size) { \
case 8: \
CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, 8, KV_DTYPE); \
break; \
case 16: \
CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, 16, KV_DTYPE); \
break; \
case 32: \
CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, 32, KV_DTYPE); \
break; \
default: \
TORCH_CHECK(false, "Unsupported block size: ", block_size); \
break; \
}
void paged_attention_v2_opt_with_mask(
torch::Tensor& out, // [num_seqs, num_heads, head_size]
torch::Tensor& exp_sums, // [num_seqs, num_heads, max_num_partitions]
torch::Tensor& max_logits, // [num_seqs, num_heads, max_num_partitions]
torch::Tensor&
tmp_out, // [num_seqs, num_heads, max_num_partitions, head_size]
torch::Tensor& query, // [num_seqs, num_heads, head_size]
torch::Tensor&
key_cache, // [num_blocks, num_heads, head_size/x, block_size, x]
torch::Tensor&
value_cache, // [num_blocks, num_heads, head_size, block_size]
int64_t num_kv_heads, // [num_heads]
double scale,
torch::Tensor& block_tables, // [num_seqs, max_num_blocks_per_seq]
torch::Tensor& seq_lens, // [num_seqs]
int64_t block_size, int64_t max_seq_len,
const c10::optional<torch::Tensor>& alibi_slopes,
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride) {
const bool is_block_sparse = (blocksparse_vert_stride > 1);
DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
CALL_V2_LAUNCHER_BLOCK_SIZE)
}
#undef WARP_SIZE
#undef MAX
#undef MIN
#undef DIVIDE_ROUND_UP
\ No newline at end of file
csrc/attention/attention_with_mask_kernels_opt_tc.cu 0 → 100644
View file @ 44e3ca68
#include <torch/all.h>
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <algorithm>
#include "attention_dtypes.h"
#include "attention_utils.cuh"
#ifdef USE_ROCM
#include <hip/hip_bf16.h>
#include "../quantization/fp8/amd/quant_utils.cuh"
typedef __hip_bfloat16 __nv_bfloat16;
#else
#include "../quantization/fp8/nvidia/quant_utils.cuh"
#endif
#ifndef USE_ROCM
#define WARP_SIZE 32
#else
#define WARP_SIZE warpSize
#endif
#include "static_switch_tc.h"
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define DIVIDE_ROUND_UP(a, b) (((a) + (b) - 1) / (b))
inline std::string get_device_name()
{
hipDeviceProp_t props{};
int device;
auto status = hipGetDevice(&device);
if(status != hipSuccess)
{
return std::string();
}
status = hipGetDeviceProperties(&props, device);
if(status != hipSuccess)
{
return std::string();
}
const std::string raw_name(props.gcnArchName);
return raw_name.substr(0, raw_name.find(':')); // str.substr(0, npos) returns str.
}
static inline int get_env_(const char *env_var) {
if (char *value = std::getenv(env_var)) {
return atoi(value);
}
return 0;
}
static const int PA_REUSE_KV_TIMES = get_env_("PA_REUSE_KV_TIMES");
static const int PA_BLOCK_SIZE = get_env_("PA_BLOCK_SIZE");
static const int PA_PRINT_PARAM = get_env_("PA_PRINT_PARAM");
namespace vllm {
// Utility function for attention softmax.
template <int NUM_WARPS>
inline __device__ float block_sum(float* red_smem, float sum) {
// Decompose the thread index into warp / lane.
int warp = threadIdx.x / WARP_SIZE;
int lane = threadIdx.x % WARP_SIZE;
// Compute the sum per warp.
#pragma unroll
for (int mask = WARP_SIZE / 2; mask >= 1; mask /= 2) {
sum += VLLM_SHFL_XOR_SYNC(sum, mask);
}
// Warp leaders store the data to shared memory.
if (lane == 0) {
red_smem[warp] = sum;
}
// Make sure the data is in shared memory.
__syncthreads();
// The warps compute the final sums.
if (lane < NUM_WARPS) {
sum = red_smem[lane];
}
// Parallel reduction inside the warp.
#pragma unroll
for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
sum += VLLM_SHFL_XOR_SYNC(sum, mask);
}
// Broadcast to other threads.
return VLLM_SHFL_SYNC(sum, 0);
}
using half4_t = __attribute__( (__vector_size__(4 * sizeof(_Float16)) )) _Float16;
using v4bh = __attribute__( (__vector_size__(4 * sizeof(short)) )) short;
using float4_t = __attribute__( (__vector_size__(4 * sizeof(float)) )) float;
struct half4x2{
half4_t data[2];
};
template<bool is_half>
inline __device__ void float4_2_half4(half4_t& dst,const float4_t& src)
{
if constexpr(is_half){
#pragma unroll
for(int i=0;i<4;i++){
dst[i]=src[i];
}
}
else{
__nv_bfloat16* out = reinterpret_cast<__nv_bfloat16 *>(&dst);
#pragma unroll
for(int i=0;i<4;i++){
out[i]=__float2bfloat16(src[i]);
}
}
}
template<bool is_half>
inline __device__ void v_mmac_f32_16x16x16_f16(const half4_t& reg_a, const half4_t& reg_b, float4_t& reg_c)
{
if constexpr (is_half){
asm volatile("v_mmac_f32_16x16x16_f16 %0, %1, %2, %0" :
"=v"(reg_c) : "v"(reg_a), "v"(reg_b), "0"(reg_c));
}
else{
asm volatile("v_mmac_f32_16x16x16_bf16 %0, %1, %2, %0" :
"=v"(reg_c) : "v"(reg_a), "v"(reg_b), "0"(reg_c));
}
}
template<bool is_half,bool use_vmac>
inline __device__ void builtin_amdgcn_mmac(const half4_t& reg_a, const half4_t& reg_b, float4_t& reg_c)
{
if constexpr (use_vmac){v_mmac_f32_16x16x16_f16<is_half>(reg_a,reg_b,reg_c);}
else{
if constexpr (is_half){reg_c=__builtin_amdgcn_mmac_f32_16x16x16f16(reg_a,reg_b,reg_c);}
else{
reg_c=__builtin_amdgcn_mmac_f32_16x16x16bf16(*(v4bh*)&reg_a,*(v4bh*)&reg_b,reg_c);
}
}
}
// TODO(woosuk): Merge the last two dimensions of the grid.
// Grid: (num_heads, num_seqs, max_num_partitions).
template <typename scalar_t, typename cache_t, int HEAD_SIZE, int BLOCK_SIZE,
int NUM_THREADS, vllm::Fp8KVCacheDataType KV_DTYPE,
bool IS_BLOCK_SPARSE,int REUSE_KV_TIMES,bool use_vmac,int PARTITION_SIZE = 0> // Zero means no partitioning.
__device__ void paged_attention_with_mask_kernel_TC(
float* __restrict__ exp_sums, // [num_seqs, num_heads, max_num_partitions]
float* __restrict__ max_logits, // [num_seqs, num_heads,
// max_num_partitions]
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 cache_t* __restrict__ k_cache, // [num_blocks, num_kv_heads,
// head_size/x, block_size, x]
const cache_t* __restrict__ v_cache, // [num_blocks, num_kv_heads,
// head_size, block_size]
const int num_heads,
const int num_kv_heads, // [num_heads]
const float scale,
const int* __restrict__ block_tables, // [num_seqs, max_num_blocks_per_seq]
const int* __restrict__ seq_lens, // [num_seqs]
const int max_num_blocks_per_seq,
const float* __restrict__ alibi_slopes, // [num_heads]
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
const int seq_idx = blockIdx.z;
const int partition_idx = blockIdx.y;
const int max_num_partitions = gridDim.y;
constexpr bool USE_PARTITIONING = PARTITION_SIZE > 0;
const int seq_len = __builtin_amdgcn_readfirstlane(seq_lens[seq_idx]);
if (USE_PARTITIONING && partition_idx * PARTITION_SIZE >= seq_len) {
// No work to do. Terminate the thread block.
return;
}
constexpr bool is_half = std::is_same<scalar_t, uint16_t>::value;
static_assert(HEAD_SIZE<=4*NUM_THREADS,"HEAD_SIZE<=4*NUM_THREADS");
const int num_seq_blocks = DIVIDE_ROUND_UP(seq_len, BLOCK_SIZE);
const int num_blocks_per_partition = USE_PARTITIONING ? PARTITION_SIZE / BLOCK_SIZE : num_seq_blocks;
const int partition_size = USE_PARTITIONING ? PARTITION_SIZE : num_seq_blocks * BLOCK_SIZE;
// [start_block_idx, end_block_idx) is the range of blocks to process.
const int start_block_idx = partition_idx * num_blocks_per_partition;//0,64,128…
const int end_block_idx =MIN(start_block_idx + num_blocks_per_partition, num_seq_blocks);//64,128,192…
const int num_blocks = end_block_idx - start_block_idx;//64 or 1-63
// [start_token_idx, end_token_idx) is the range of tokens to process.
const int start_token_idx = start_block_idx * BLOCK_SIZE;//0,1024,2048…
const int end_token_idx = MIN(start_token_idx + num_blocks * BLOCK_SIZE, seq_len);//1024,2048,3072…
const int num_tokens = end_token_idx - start_token_idx;//1024 or 1-1023
// divides NUM_THREADS
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;//4
constexpr int x = 16 / sizeof(cache_t);//8
const int thread_idx = threadIdx.x;
const int warp_idx = __builtin_amdgcn_readfirstlane(thread_idx / WARP_SIZE);
const int lane = thread_idx % WARP_SIZE;
const int rowid = lane%16;
const int rows = lane/16;
const int num_queries_per_kv = num_heads / num_kv_heads;
const int num_blocks_per_kv = ((num_queries_per_kv + REUSE_KV_TIMES -1) / REUSE_KV_TIMES);
const int odd_tg_round = (((blockIdx.z * gridDim.y * gridDim.x) + blockIdx.y * gridDim.x) / 128) % 2;
const int mid_x = gridDim.x / 2;
const int blockIdx_shift = (odd_tg_round | (gridDim.x & 1)) ? blockIdx.x : (blockIdx.x < mid_x ? (blockIdx.x + mid_x) : (blockIdx.x - mid_x));
const int head_idx = (blockIdx_shift / num_blocks_per_kv) * num_queries_per_kv + (blockIdx_shift % num_blocks_per_kv) * REUSE_KV_TIMES;
//const int head_idx=(blockIdx.x / num_blocks_per_kv) * num_queries_per_kv + (blockIdx.x % num_blocks_per_kv) * REUSE_KV_TIMES;
int q_boundary=REUSE_KV_TIMES;
if(num_heads < REUSE_KV_TIMES*gridDim.x && (num_blocks_per_kv-1)*REUSE_KV_TIMES == head_idx%num_queries_per_kv)
q_boundary=num_queries_per_kv-(num_blocks_per_kv-1)*REUSE_KV_TIMES;
const int kv_head_idx = head_idx / num_queries_per_kv;
constexpr int reuse_group=(REUSE_KV_TIMES-1)/4+1;
float alibi_slope[reuse_group]={0.f};
if(alibi_slopes != nullptr){
for(int i=0;i<reuse_group;i++){
int reuse_kv_idx=rows+i*4;
if(reuse_kv_idx<q_boundary) alibi_slope[i]=alibi_slopes[head_idx+reuse_kv_idx];
}
}
float qk_max[reuse_group];
for(int i=0;i<reuse_group;i++){
qk_max[i]=-FLT_MAX;
}
const scalar_t* q_ptr = q + seq_idx * q_stride + head_idx * HEAD_SIZE;
half4x2 q_vec;
q_vec.data[0]={0,0,0,0};
q_vec.data[1]={0,0,0,0};
__shared__ half4x2 q_vecs[REUSE_KV_TIMES][16];
//if(thread_idx==0)printf("blockIdx.x==%d,q_boundary=%d,head_idx=%d,kv_head_idx=%d\n",blockIdx.x,q_boundary,head_idx,kv_head_idx);
for(int i=0;i<q_boundary;i++){
if(thread_idx<16){
q_vecs[i][thread_idx]=*reinterpret_cast<const half4x2*>(q_ptr+i*HEAD_SIZE+thread_idx*8);
}
}
__syncthreads();
// Memory planning.
extern __shared__ char shared_mem[];
// NOTE(woosuk): We use FP32 for the softmax logits for better accuracy.
scalar_t* logits = reinterpret_cast<scalar_t*>(shared_mem);
// Workspace for reduction.
__shared__ float red_smem[2 * NUM_WARPS];
// Iterate over the key blocks.
// Each warp fetches a block of keys for each iteration.
// Each thread group in a warp fetches a key from the block, and computes
// dot product with the query.
const int* block_table = block_tables + seq_idx * max_num_blocks_per_seq;
// blocksparse specific vars
int bs_block_offset;
int q_bs_block_id;
const cache_t* k_ptr_base = k_cache+kv_head_idx * kv_head_stride+lane*8;
for (int block_idx = start_block_idx + warp_idx; block_idx < end_block_idx;
block_idx += NUM_WARPS) {
const int64_t physical_block_number = static_cast<int64_t>(block_table[block_idx]);
const cache_t* k_ptr=k_ptr_base + physical_block_number * kv_block_stride;
float4_t qk_vec={0,0,0,0};
half4x2 k_vec[2];
k_vec[0]=*reinterpret_cast<const half4x2*>(k_ptr);
#pragma unroll
for(int i=0;i<3;i++){
if(rowid<q_boundary)q_vec=q_vecs[rowid][i*4+rows];
k_vec[1-i%2]=*reinterpret_cast<const half4x2*>(k_ptr+(i+1)*512);
builtin_amdgcn_mmac<is_half,use_vmac>(k_vec[i%2].data[0],q_vec.data[0],qk_vec);
builtin_amdgcn_mmac<is_half,use_vmac>(k_vec[i%2].data[1],q_vec.data[1],qk_vec);
}
//tail
{
if(rowid<q_boundary)q_vec=q_vecs[rowid][3*4+rows];
builtin_amdgcn_mmac<is_half,use_vmac>(k_vec[1].data[0],q_vec.data[0],qk_vec);
v_mmac_f32_16x16x16_f16<is_half>(k_vec[1].data[1],q_vec.data[1],qk_vec);
}
#pragma unroll
for(int i=0;i<reuse_group;i++){
int reuse_kv_idx=rows+i*4;
if(reuse_kv_idx<REUSE_KV_TIMES){
if(reuse_kv_idx>=q_boundary)qk_vec[i]=0;
else qk_vec[i]*=scale;
const int token_idx = block_idx * BLOCK_SIZE+rowid;
if(alibi_slope[i] != 0){
float alibi=alibi_slope[i]* (token_idx - seq_len + 1);
qk_vec[i] = alibi;
}
// used for tree-style attention
if (attn_masks != nullptr) {
const int* attn_masks_ptr = attn_masks + seq_idx * attn_masks_stride;
if (attn_masks_ptr[token_idx] == 0) {
qk_vec[i] = -FLT_MAX;
}
}
const bool mask = (token_idx >= seq_len);
if(mask){
from_float(logits[partition_size*reuse_kv_idx+token_idx - start_token_idx] , 0.f);
}
else{
from_float(logits[partition_size*reuse_kv_idx+token_idx - start_token_idx] , qk_vec[i]);
qk_max[i] = fmaxf(qk_max[i], qk_vec[i]);
}
}
}
}
// if(blockIdx.x==0)printf("%d,qkmax=%f\n",threadIdx.x,qk_max[0]);
// Perform reduction across the threads in the same warp to get the
// max qk value for each "warp" (not across the thread block yet).
// The 0-th thread of each thread group already has its max qk value.
for(int reuse_kv_idx=0; reuse_kv_idx<q_boundary; reuse_kv_idx++) {
const int head_idx_ = head_idx + reuse_kv_idx;
float qk_max_tmp=qk_max[reuse_kv_idx/4];
float exp_sum = 0.f;
#pragma unroll
for (int mask = 8; mask >= 1; mask /= 2) {
qk_max_tmp = fmaxf(qk_max_tmp, VLLM_SHFL_XOR_SYNC(qk_max_tmp, mask));
}
if (rowid==0 && reuse_kv_idx%4==rows) {
red_smem[warp_idx] = qk_max_tmp;
}
__syncthreads();
// TODO(woosuk): Refactor this part.
// Get the max qk value for the sequence.
qk_max_tmp = lane < NUM_WARPS ? red_smem[lane] : -FLT_MAX;
#pragma unroll
for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
qk_max_tmp = fmaxf(qk_max_tmp, VLLM_SHFL_XOR_SYNC(qk_max_tmp, mask));
}
// Broadcast the max qk value to all threads.
qk_max_tmp = VLLM_SHFL_SYNC(qk_max_tmp, 0);
for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) {
float val = __expf(to_float(logits[(reuse_kv_idx * partition_size) + i]) - qk_max_tmp);
from_float(logits[(reuse_kv_idx * partition_size) + i] , val);
exp_sum += val;
}
exp_sum = block_sum<NUM_WARPS>(&red_smem[NUM_WARPS], exp_sum);
// Compute softmax.
const float inv_sum = __fdividef(1.f, exp_sum + 1e-6f);
for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) {
from_float(logits[(reuse_kv_idx * partition_size) + i] ,to_float(logits[(reuse_kv_idx * partition_size) + i])*inv_sum);
}
__syncthreads();
// If partitioning is enabled, store the max logit and exp_sum.
if (USE_PARTITIONING && thread_idx == 0) {
float* max_logits_ptr = max_logits +
seq_idx * num_heads * max_num_partitions +
head_idx_ * max_num_partitions + partition_idx;
*max_logits_ptr = qk_max_tmp;
float* exp_sums_ptr = exp_sums + seq_idx * num_heads * max_num_partitions +
head_idx_ * max_num_partitions + partition_idx;
*exp_sums_ptr = exp_sum;
}
}
constexpr int NUM_ROWS_PER_THREAD =DIVIDE_ROUND_UP(HEAD_SIZE, WARP_SIZE);//2
if constexpr(REUSE_KV_TIMES<=2){
float accs[REUSE_KV_TIMES][NUM_ROWS_PER_THREAD];
#pragma unroll
for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
#pragma unroll
for(int k=0;k<REUSE_KV_TIMES;k++)
{
accs[k][i] = 0.f;
}
}
scalar_t zero_value;
zero(zero_value);
for (int block_idx = start_block_idx + warp_idx; block_idx < end_block_idx;
block_idx += NUM_WARPS) {
const int64_t physical_block_number =
static_cast<int64_t>(block_table[block_idx]);
const int token_idx = block_idx * BLOCK_SIZE +rows*4;
half4_t logits_vec={0,0,0,0};
if(rowid<4*q_boundary){
logits_vec=*reinterpret_cast<half4_t*>(logits + rowid/4 * partition_size+token_idx - start_token_idx);
}
const cache_t* v_ptr = v_cache + physical_block_number * kv_block_stride +
kv_head_idx * kv_head_stride + rows*4+rowid*16;
#pragma unroll
for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
#pragma unroll
for(int k=0;k<4;k++){
int offset=i*1024+k*256;
half4_t v_vec=*reinterpret_cast<const half4_t*>(v_ptr + offset);
if (block_idx == num_seq_blocks - 1) {
scalar_t* v_vec_ptr = reinterpret_cast<scalar_t*>(&v_vec);
#pragma unroll
for (int j = 0; j < 4; j++) {
v_vec_ptr[j] = token_idx + j < seq_len ? v_vec_ptr[j] : zero_value;
}
}
float4_t out_vec={0,0,0,0};
builtin_amdgcn_mmac<is_half,use_vmac>(v_vec,logits_vec,out_vec);
if(rows==k){
for(int resuseid=0;resuseid<REUSE_KV_TIMES;resuseid++){
accs[resuseid][i]+=out_vec[resuseid];
}
}
}
}
}
__syncthreads();
using floatV_t = __attribute__( (__vector_size__(NUM_ROWS_PER_THREAD * sizeof(float)) )) float;
// Perform reduction across warps.
for(int reuse_kv_idx=0; reuse_kv_idx<q_boundary; reuse_kv_idx++) {
if constexpr (NUM_THREADS>64){
floatV_t* out_smem = reinterpret_cast<floatV_t*>(shared_mem);
#pragma unroll
for (int i = NUM_WARPS; i > 1; i /= 2) {
int mid = i / 2;
// Upper warps write to shared memory.
if (warp_idx >= mid && warp_idx < i) {
out_smem[(warp_idx - mid) * 64+lane]=*(floatV_t*)(accs[reuse_kv_idx]);
}
__syncthreads();
// Lower warps update the output.
if (warp_idx < mid) {
floatV_t tmp=out_smem[thread_idx];
#pragma unroll
for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
accs[reuse_kv_idx][i] += tmp[i];
}
}
__syncthreads();
}
}
// Write the final output.
if (warp_idx == 0) {
scalar_t* out_ptr =
out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
(head_idx+reuse_kv_idx) * max_num_partitions * HEAD_SIZE + partition_idx * HEAD_SIZE;
#pragma unroll
for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
const int row_idx = lane + i * WARP_SIZE;
from_float(*(out_ptr + row_idx), accs[reuse_kv_idx][i]);
}
}
}
}
else{
constexpr int GROUPS=reuse_group*4;
// NOTE(woosuk): We use FP32 for the accumulator for better accuracy.
float accs[GROUPS][NUM_ROWS_PER_THREAD];
#pragma unroll
for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
#pragma unroll
for(int k=0;k<GROUPS;k++)
{
accs[k][i] = 0.f;
}
}
scalar_t zero_value;
zero(zero_value);
for (int block_idx = start_block_idx + warp_idx; block_idx < end_block_idx;
block_idx += NUM_WARPS) {
const int64_t physical_block_number =
static_cast<int64_t>(block_table[block_idx]);
const int token_idx = block_idx * BLOCK_SIZE +rows*4;
half4_t logits_vec={0,0,0,0};
if(rowid<q_boundary){
logits_vec=*reinterpret_cast<half4_t*>(logits + rowid * partition_size+token_idx - start_token_idx);
}
const cache_t* v_ptr = v_cache + physical_block_number * kv_block_stride +
kv_head_idx * kv_head_stride + rows*4+rowid*16;
#pragma unroll
for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
#pragma unroll
for(int k=0;k<4;k++){
int offset=i*1024+k*256;
half4_t v_vec=*reinterpret_cast<const half4_t*>(v_ptr + offset);
if (block_idx == num_seq_blocks - 1) {
scalar_t* v_vec_ptr = reinterpret_cast<scalar_t*>(&v_vec);
#pragma unroll
for (int j = 0; j < 4; j++) {
v_vec_ptr[j] = token_idx + j < seq_len ? v_vec_ptr[j] : zero_value;
}
}
float4_t out_vec={0,0,0,0};
builtin_amdgcn_mmac<is_half,use_vmac>(v_vec,logits_vec,out_vec);
for(int g=0;g<reuse_group;g++){
accs[g*4+k][i]+=out_vec[g];
}
}
}
}
if constexpr (NUM_THREADS>64){
__syncthreads();
using floatV_t = __attribute__( (__vector_size__(NUM_ROWS_PER_THREAD * sizeof(float)) )) float;
// Perform reduction across warps.
for(int reuse_kv_idx=0; reuse_kv_idx<GROUPS; reuse_kv_idx++) {
floatV_t* out_smem = reinterpret_cast<floatV_t*>(shared_mem);
#pragma unroll
for (int i = NUM_WARPS; i > 1; i /= 2) {
int mid = i / 2;
// Upper warps write to shared memory.
if (warp_idx >= mid && warp_idx < i) {
out_smem[(warp_idx - mid) * 64+lane]=*(floatV_t*)(accs[reuse_kv_idx]);
}
__syncthreads();
// Lower warps update the output.
if (warp_idx < mid) {
floatV_t tmp=out_smem[thread_idx];
#pragma unroll
for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
accs[reuse_kv_idx][i] += tmp[i];
}
}
__syncthreads();
}
}
}
if (warp_idx == 0) {
for(int g=0;g<reuse_group;g++){
int reusekvid=g*4+rows;
if(reusekvid<q_boundary){
scalar_t* out_ptr =
out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
(head_idx+reusekvid) * max_num_partitions * HEAD_SIZE + partition_idx * HEAD_SIZE;
#pragma unroll
for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
for(int k=0;k<4;k++){
const int row_idx = rowid+16*k + i * WARP_SIZE;
from_float(*(out_ptr + row_idx), accs[g*4+k][i]);
}
}
}
}
}
}
}
template <typename scalar_t, typename cache_t, int HEAD_SIZE, int BLOCK_SIZE,
int NUM_THREADS, vllm::Fp8KVCacheDataType KV_DTYPE,
bool IS_BLOCK_SPARSE,int REUSE_KV_TIMES,bool use_vmac>
__global__ void paged_attention_v1_with_mask_kernel_TC(
scalar_t* __restrict__ out, // [num_seqs, num_heads, head_size]
const scalar_t* __restrict__ q, // [num_seqs, num_heads, head_size]
const cache_t* __restrict__ k_cache, // [num_blocks, num_kv_heads,
// head_size/x, block_size, x]
const cache_t* __restrict__ v_cache, // [num_blocks, num_kv_heads,
// head_size, block_size]
const int num_heads,
const int num_kv_heads, // [num_heads]
const float scale,
const int* __restrict__ block_tables, // [num_seqs, max_num_blocks_per_seq]
const int* __restrict__ seq_lens, // [num_seqs]
const int max_num_blocks_per_seq,
const float* __restrict__ alibi_slopes, // [num_heads]
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
#if defined(__gfx936__) || defined(__gfx928__)
paged_attention_with_mask_kernel_TC<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
KV_DTYPE, IS_BLOCK_SPARSE,REUSE_KV_TIMES,use_vmac>(
/* exp_sums */ nullptr, /* max_logits */ nullptr, out, q, k_cache,
v_cache, num_heads,num_kv_heads, scale, block_tables, seq_lens,
max_num_blocks_per_seq, alibi_slopes, q_stride, kv_block_stride,
kv_head_stride, k_scale, v_scale, tp_rank, blocksparse_local_blocks,
blocksparse_vert_stride, blocksparse_block_size,
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
#endif
}
// Grid: (num_heads, num_seqs, max_num_partitions).
template <typename scalar_t, typename cache_t, int HEAD_SIZE, int BLOCK_SIZE,
int NUM_THREADS, vllm::Fp8KVCacheDataType KV_DTYPE,
bool IS_BLOCK_SPARSE, int REUSE_KV_TIMES,bool use_vmac, int PARTITION_SIZE,
bool odd_nheads = false>
__global__ __launch_bounds__(256, 1) void paged_attention_v2_with_mask_kernel_TC(
float* __restrict__ exp_sums, // [num_seqs, num_heads, max_num_partitions]
float* __restrict__ max_logits, // [num_seqs, num_heads,
// max_num_partitions]
scalar_t* __restrict__ tmp_out, // [num_seqs, num_heads,
// max_num_partitions, head_size]
const scalar_t* __restrict__ q, // [num_seqs, num_heads, head_size]
const cache_t* __restrict__ k_cache, // [num_blocks, num_kv_heads,
// head_size/x, block_size, x]
const cache_t* __restrict__ v_cache, // [num_blocks, num_kv_heads,
// head_size, block_size]
const int num_heads, // [num_heads]
const int num_kv_heads, // [num_kv_heads]
const float scale,
const int* __restrict__ block_tables, // [num_seqs, max_num_blocks_per_seq]
const int* __restrict__ seq_lens, // [num_seqs]
const int max_num_blocks_per_seq,
const float* __restrict__ alibi_slopes, // [num_heads]
const int q_stride, const int kv_block_stride, const int kv_head_stride,
const float k_scale, const float v_scale, const int tp_rank,
const int blocksparse_local_blocks, const int blocksparse_vert_stride,
const int blocksparse_block_size, const int blocksparse_head_sliding_step,
const int* __restrict__ attn_masks=nullptr, const int attn_masks_stride=0) {
#if defined(__gfx936__) || defined(__gfx928__)
paged_attention_with_mask_kernel_TC<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
KV_DTYPE, IS_BLOCK_SPARSE, REUSE_KV_TIMES,use_vmac,
PARTITION_SIZE>(
exp_sums, max_logits, tmp_out, q, k_cache, v_cache, num_heads,
num_kv_heads, scale, block_tables, seq_lens, max_num_blocks_per_seq,
alibi_slopes, q_stride, kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank,
blocksparse_local_blocks, blocksparse_vert_stride, blocksparse_block_size,
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
#endif
}
// Grid: (num_heads, num_seqs).
template <typename scalar_t, int HEAD_SIZE, int NUM_THREADS, int PARTITION_SIZE>
__global__ __launch_bounds__(256, 1) void paged_attention_v2_reduce_kernel_opt_tc(
scalar_t* __restrict__ out, // [num_seqs, num_heads, head_size]
const float* __restrict__ exp_sums, // [num_seqs, num_heads,
// max_num_partitions]
const float* __restrict__ max_logits, // [num_seqs, num_heads,
// max_num_partitions]
const scalar_t* __restrict__ tmp_out, // [num_seqs, num_heads,
// max_num_partitions, head_size]
const int* __restrict__ seq_lens, // [num_seqs]
const int max_num_partitions) {
const int num_heads = gridDim.x;
const int head_idx = blockIdx.x;
const int seq_idx = blockIdx.y;
const int seq_len = seq_lens[seq_idx];
const int num_partitions = DIVIDE_ROUND_UP(seq_len, PARTITION_SIZE);
if (num_partitions == 1) {
// No need to reduce. Only copy tmp_out to out.
scalar_t* out_ptr =
out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
const scalar_t* tmp_out_ptr =
tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
head_idx * max_num_partitions * HEAD_SIZE;
for (int i = threadIdx.x; i < HEAD_SIZE; i += blockDim.x) {
out_ptr[i] = tmp_out_ptr[i];
}
// Terminate the thread block.
return;
}
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
const int warp_idx = threadIdx.x / WARP_SIZE;
const int lane = threadIdx.x % WARP_SIZE;
// Size: 2 * num_partitions.
extern __shared__ char shared_mem[];
// Workspace for reduction.
__shared__ float red_smem[2 * NUM_WARPS];
// Load max logits to shared memory.
float* shared_max_logits = reinterpret_cast<float*>(shared_mem);
const float* max_logits_ptr = max_logits +
seq_idx * num_heads * max_num_partitions +
head_idx * max_num_partitions;
float max_logit = -FLT_MAX;
for (int i = threadIdx.x; i < num_partitions; i += blockDim.x) {
const float l = max_logits_ptr[i];
shared_max_logits[i] = l;
max_logit = fmaxf(max_logit, l);
}
__syncthreads();
// Get the global max logit.
// Reduce within the warp.
#pragma unroll
for (int mask = WARP_SIZE / 2; mask >= 1; mask /= 2) {
max_logit = fmaxf(max_logit, VLLM_SHFL_XOR_SYNC(max_logit, mask));
}
if (lane == 0) {
red_smem[warp_idx] = max_logit;
}
__syncthreads();
// Reduce across warps.
max_logit = lane < NUM_WARPS ? red_smem[lane] : -FLT_MAX;
#pragma unroll
for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
max_logit = fmaxf(max_logit, VLLM_SHFL_XOR_SYNC(max_logit, mask));
}
// Broadcast the max value to all threads.
max_logit = VLLM_SHFL_SYNC(max_logit, 0);
// Load rescaled exp sums to shared memory.
float* shared_exp_sums =
reinterpret_cast<float*>(shared_mem + sizeof(float) * num_partitions);
const float* exp_sums_ptr = exp_sums +
seq_idx * num_heads * max_num_partitions +
head_idx * max_num_partitions;
float global_exp_sum = 0.0f;
for (int i = threadIdx.x; i < num_partitions; i += blockDim.x) {
float l = shared_max_logits[i];
float rescaled_exp_sum = exp_sums_ptr[i] * expf(l - max_logit);
global_exp_sum += rescaled_exp_sum;
shared_exp_sums[i] = rescaled_exp_sum;
}
__syncthreads();
global_exp_sum = block_sum<NUM_WARPS>(&red_smem[NUM_WARPS], global_exp_sum);
const float inv_global_exp_sum = __fdividef(1.0f, global_exp_sum + 1e-6f);
// Aggregate tmp_out to out.
const scalar_t* tmp_out_ptr =
tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
head_idx * max_num_partitions * HEAD_SIZE;
scalar_t* out_ptr =
out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
#pragma unroll
for (int i = threadIdx.x; i < HEAD_SIZE; i += NUM_THREADS) {
float acc = 0.0f;
for (int j = 0; j < num_partitions; ++j) {
acc += to_float(tmp_out_ptr[j * HEAD_SIZE + i]) * shared_exp_sums[j] *
inv_global_exp_sum;
}
from_float(out_ptr[i], acc);
}
}
} // namespace vllm
#define LAUNCH_PAGED_ATTENTION_V1_TC(HEAD_SIZE) \
VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize( \
((void*)vllm::paged_attention_v1_with_mask_kernel_TC<T, CACHE_T, HEAD_SIZE, \
BLOCK_SIZE, NUM_THREADS, \
KV_DTYPE, IS_BLOCK_SPARSE,REUSE_KV_TIMES,use_vmac>), \
shared_mem_size); \
vllm::paged_attention_v1_with_mask_kernel_TC<T, CACHE_T, HEAD_SIZE, BLOCK_SIZE, \
NUM_THREADS, KV_DTYPE, IS_BLOCK_SPARSE,REUSE_KV_TIMES,use_vmac> \
<<<grid, block, shared_mem_size, stream>>>( \
out_ptr, query_ptr, key_cache_ptr, value_cache_ptr, num_heads,num_kv_heads, \
scale, block_tables_ptr, seq_lens_ptr, max_num_blocks_per_seq, \
alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride, \
k_scale, v_scale, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step, attn_masks_ptr, \
attn_masks_stride);
void get_number_thread_and_reuse_kv_v1(int& num_thread,int& reusekv,int batchsize,int seq,int qheads,int kvheads){
//mha
reusekv=1;
if(qheads==kvheads){
//llama 7B ，其他模型未可知
if(seq<=16||batchsize>=32)num_thread=64;
else if(batchsize<=2)num_thread=256;
else if(batchsize<8)num_thread=128;
else num_thread=64;
return;
}
// mqa
if(qheads>kvheads*4){
if(seq<64){
if(batchsize<=64){reusekv=1;num_thread=64;}
else if(batchsize<128){reusekv=2;num_thread=64;}
else {reusekv=4;num_thread=64;}
}
else if(seq<=400){
if(batchsize<16){reusekv=1;num_thread=256;}
else if(batchsize<64){reusekv=2;num_thread=256;}
else if(batchsize<=128){
reusekv=4;
if(qheads%7==0)num_thread=64;//qwen7b
else num_thread=256;//llama70b
}
else {reusekv=8;num_thread=64;}
}
else if(seq<=1000){
if(batchsize<16){reusekv=1;num_thread=256;}
else if(qheads%7==0&&batchsize<=128){//qwen7b
if(batchsize<64){reusekv=4;num_thread=256;}
else{reusekv=4;num_thread=64;}
}
else if(batchsize<=64){reusekv=4;num_thread=256;}
else {reusekv=8;num_thread=128;}
}
else if(seq<3900) {reusekv=8;num_thread=256;}
else if(seq<7800) {reusekv=4;num_thread=256;}
else {reusekv=2;num_thread=256;}
return;
}
if(qheads/kvheads >4 && seq<3900)reusekv=8;
else if(qheads/kvheads >2 && seq<7800)reusekv=4;
else if(qheads/kvheads >=2 && seq<15600)reusekv=2;
if(seq<=64){
num_thread=64;
if(batchsize<=64)reusekv=1;
}
else num_thread=256;
}
// TODO(woosuk): Tune NUM_THREADS.
template <typename T, typename CACHE_T, int BLOCK_SIZE,
vllm::Fp8KVCacheDataType KV_DTYPE, bool IS_BLOCK_SPARSE>
void paged_attention_v1_launcher_opt_tc(
torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int num_kv_heads, float scale,
torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
float v_scale, const int tp_rank, const int blocksparse_local_blocks,
const int blocksparse_vert_stride, const int blocksparse_block_size,
const int blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks,
const int64_t attn_masks_stride=0) {
int num_seqs = query.size(0);
int num_heads = query.size(1);
int head_size = query.size(2);
int max_num_blocks_per_seq = block_tables.size(1);
int q_stride = query.stride(0);
int kv_block_stride = key_cache.stride(0);
int kv_head_stride = key_cache.stride(1);
int num_threads = 128;
// printf("paged_attention_v1\n");
if (num_heads != num_kv_heads) {
num_threads = 256;
}
[[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
assert(head_size % thread_group_size == 0);
// NOTE: alibi_slopes is optional.
const float* alibi_slopes_ptr =
alibi_slopes
? reinterpret_cast<const float*>(alibi_slopes.value().data_ptr())
: nullptr;
T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
CACHE_T* key_cache_ptr = reinterpret_cast<CACHE_T*>(key_cache.data_ptr());
CACHE_T* value_cache_ptr = reinterpret_cast<CACHE_T*>(value_cache.data_ptr());
int* block_tables_ptr = block_tables.data_ptr<int>();
int* seq_lens_ptr = seq_lens.data_ptr<int>();
// NOTE: attn_masks is optional.
const int* attn_masks_ptr =
attn_masks
? attn_masks.value().data_ptr<int>()
: nullptr;
int padded_max_seq_len = DIVIDE_ROUND_UP(max_seq_len, BLOCK_SIZE) * BLOCK_SIZE;
const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
if constexpr(BLOCK_SIZE==16 && IS_BLOCK_SPARSE==false && sizeof(T)==2 && KV_DTYPE==vllm::Fp8KVCacheDataType::kAuto){
constexpr int HEAD_SIZE=128;
constexpr static int use_vmac = false;
int reusekv, num_thread;
get_number_thread_and_reuse_kv_v1(num_thread,reusekv,num_seqs,padded_max_seq_len,num_heads,num_kv_heads);
if(PA_REUSE_KV_TIMES!=0&&num_heads>num_kv_heads)reusekv=PA_REUSE_KV_TIMES;
if(PA_BLOCK_SIZE!=0)num_thread=PA_BLOCK_SIZE;
REUSEKV_SWITCH(reusekv,[&] {
NUM_THREADS_SWITCH(num_thread , [&] {
//constexpr int NUM_THREADS = WARP_SIZE * REUSE_KV_TIMES;
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
int logits_size = REUSE_KV_TIMES * padded_max_seq_len * 2;
int outputs_size = (NUM_WARPS / 2) * head_size * sizeof(float);
if (NUM_WARPS==64)outputs_size=0;
int shared_mem_size = ::max(logits_size, outputs_size);
dim3 grid((num_heads/num_kv_heads + REUSE_KV_TIMES - 1) / REUSE_KV_TIMES*num_kv_heads, 1,num_seqs);
dim3 block(NUM_THREADS);
if(PA_PRINT_PARAM)printf("reusekv=%d,num_thread=%d,grid={%d,%d,%d},qhead=%d,kvhead=%d,seq=%d,batch=%d\n",
reusekv,num_thread,grid.x,grid.y,grid.z,num_heads,num_kv_heads,max_seq_len,num_seqs);
LAUNCH_PAGED_ATTENTION_V1_TC(HEAD_SIZE);
});
});
}
}
#define CALL_V1_LAUNCHER(T, CACHE_T, BLOCK_SIZE, KV_DTYPE, IS_BLOCK_SPARSE) \
paged_attention_v1_launcher_opt_tc<T, CACHE_T, BLOCK_SIZE, KV_DTYPE, \
IS_BLOCK_SPARSE>( \
out, query, key_cache, value_cache, num_kv_heads, scale, block_tables, \
seq_lens, max_seq_len, alibi_slopes, k_scale, v_scale, tp_rank, \
blocksparse_local_blocks, blocksparse_vert_stride, \
blocksparse_block_size, blocksparse_head_sliding_step, \
attn_masks, attn_masks_stride);
#define CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE) \
switch (is_block_sparse) { \
case true: \
CALL_V1_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE, true); \
break; \
case false: \
CALL_V1_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE, false); \
break; \
}
// NOTE(woosuk): To reduce the compilation time, we omitted block sizes
// 1, 2, 4, 64, 128, 256.
#define CALL_V1_LAUNCHER_BLOCK_SIZE(T, CACHE_T, KV_DTYPE) \
switch (block_size) { \
case 8: \
CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, 8, KV_DTYPE); \
break; \
case 16: \
CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, 16, KV_DTYPE); \
break; \
case 32: \
CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, 32, KV_DTYPE); \
break; \
default: \
TORCH_CHECK(false, "Unsupported block size: ", block_size); \
break; \
}
void paged_attention_v1_opt_with_mask(
torch::Tensor& out, // [num_seqs, num_heads, head_size]
torch::Tensor& query, // [num_seqs, num_heads, head_size]
torch::Tensor&
key_cache, // [num_blocks, num_heads, head_size/x, block_size, x]
torch::Tensor&
value_cache, // [num_blocks, num_heads, head_size, block_size]
int64_t num_kv_heads, // [num_heads]
double scale,
torch::Tensor& block_tables, // [num_seqs, max_num_blocks_per_seq]
torch::Tensor& seq_lens, // [num_seqs]
int64_t block_size, int64_t max_seq_len,
const c10::optional<torch::Tensor>& alibi_slopes,
const std::string& kv_cache_dtype, double k_scale, double v_scale, const int64_t tp_rank,
const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride);
void paged_attention_v1_opt_tc_with_mask(
torch::Tensor& out, // [num_seqs, num_heads, head_size]
torch::Tensor& query, // [num_seqs, num_heads, head_size]
torch::Tensor&
key_cache, // [num_blocks, num_heads, head_size/x, block_size, x]
torch::Tensor&
value_cache, // [num_blocks, num_heads, head_size, block_size]
int64_t num_kv_heads, // [num_heads]
double scale,
torch::Tensor& block_tables, // [num_seqs, max_num_blocks_per_seq]
torch::Tensor& seq_lens, // [num_seqs]
int64_t block_size, int64_t max_seq_len,
const c10::optional<torch::Tensor>& alibi_slopes,
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride) {
const bool is_block_sparse = (blocksparse_vert_stride > 1);
if(kv_cache_dtype != "auto"||query.dtype() == at::ScalarType::Float||is_block_sparse||
block_size!=16||query.size(2)!=128||(get_device_name()!="gfx928" && get_device_name()!="gfx936")){
paged_attention_v1_opt_with_mask(out,query,key_cache,value_cache,num_kv_heads,
scale,block_tables,seq_lens,block_size,max_seq_len,alibi_slopes,kv_cache_dtype,
k_scale,v_scale,tp_rank,blocksparse_local_blocks,blocksparse_vert_stride,
blocksparse_block_size,blocksparse_head_sliding_step,
attn_masks, attn_masks_stride);
}
else{
DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
CALL_V1_LAUNCHER_BLOCK_SIZE)
}
}
#define LAUNCH_PAGED_ATTENTION_V2_TC(HEAD_SIZE) \
hipLaunchKernelGGL( \
(vllm::paged_attention_v2_with_mask_kernel_TC< \
T, CACHE_T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS, KV_DTYPE, \
IS_BLOCK_SPARSE, REUSE_KV_TIMES,use_vmac, PARTITION_SIZE>), \
dim3(grid), dim3(block), shared_mem_size, stream, exp_sums_ptr, \
max_logits_ptr, tmp_out_ptr, query_ptr, key_cache_ptr, value_cache_ptr, \
num_heads, num_kv_heads, scale, block_tables_ptr, seq_lens_ptr, \
max_num_blocks_per_seq, alibi_slopes_ptr, q_stride, kv_block_stride, \
kv_head_stride, k_scale, v_scale, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step, \
attn_masks_ptr, attn_masks_stride); \
hipLaunchKernelGGL( \
(vllm::paged_attention_v2_reduce_kernel_opt_tc<T, HEAD_SIZE, NUM_THREADS, \
PARTITION_SIZE>), \
dim3(reduce_grid), dim3(block), reduce_shared_mem_size, stream, out_ptr, \
exp_sums_ptr, max_logits_ptr, tmp_out_ptr, seq_lens_ptr, \
max_num_partitions);
void get_number_thread_and_reuse_kv_v2(int& num_thread,int& reusekv,int batchsize,int max_num_partitions,int qheads,int kvheads){
reusekv=1;
int blocks=batchsize*qheads*max_num_partitions;
if(qheads==kvheads){
if(blocks<=80||blocks>8000){num_thread=256;}
else if(blocks<=160){num_thread=128;}
else num_thread=64;
return;
}
if(qheads/kvheads>8&&blocks>4000){
reusekv=16;
if(blocks>40000)num_thread=64;
else num_thread=128;
}
else if(qheads/kvheads==5||qheads/kvheads==7){
if(blocks<=160){reusekv=1;num_thread=256;}
else if(blocks<640/5*qheads/kvheads){reusekv=4;num_thread=256;}
else if(blocks<1920){reusekv=8;num_thread=128;}
else {reusekv=8;num_thread=64;}
}
else if(qheads>kvheads*4){
if(blocks<=128){reusekv=1;num_thread=256;}
else if(blocks<1536){reusekv=4;num_thread=256;}
else if(blocks<6144){reusekv=8;num_thread=128;}
else {reusekv=8;num_thread=64;}
}
else {
if(blocks<=128){reusekv=1;num_thread=256;}
else if(blocks<3000){reusekv=4;num_thread=256;}
else {reusekv=4;num_thread=64;}
}
}
template <typename T, typename CACHE_T, int BLOCK_SIZE,
vllm::Fp8KVCacheDataType KV_DTYPE, bool IS_BLOCK_SPARSE, int PARTITION_SIZE = 512>
void paged_attention_v2_launcher_opt_tc(
torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int num_kv_heads, float scale,
torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
float v_scale, const int tp_rank, const int blocksparse_local_blocks,
const int blocksparse_vert_stride, const int blocksparse_block_size,
const int blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks,
const int attn_masks_stride) {
int num_seqs = query.size(0);
int num_heads = query.size(1);
int head_size = query.size(2);
int max_num_blocks_per_seq = block_tables.size(1);
int q_stride = query.stride(0);
int kv_block_stride = key_cache.stride(0);
int kv_head_stride = key_cache.stride(1);
// printf("paged_attention_v2\n");
int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
assert(head_size % thread_group_size == 0);
// NOTE: alibi_slopes is optional.
const float* alibi_slopes_ptr =
alibi_slopes
? reinterpret_cast<const float*>(alibi_slopes.value().data_ptr())
: nullptr;
T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
float* exp_sums_ptr = reinterpret_cast<float*>(exp_sums.data_ptr());
float* max_logits_ptr = reinterpret_cast<float*>(max_logits.data_ptr());
T* tmp_out_ptr = reinterpret_cast<T*>(tmp_out.data_ptr());
T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
CACHE_T* key_cache_ptr = reinterpret_cast<CACHE_T*>(key_cache.data_ptr());
CACHE_T* value_cache_ptr = reinterpret_cast<CACHE_T*>(value_cache.data_ptr());
int* block_tables_ptr = block_tables.data_ptr<int>();
int* seq_lens_ptr = seq_lens.data_ptr<int>();
// NOTE: attn_masks is optional.
const int* attn_masks_ptr =
attn_masks ? attn_masks.value().data_ptr<int>() : nullptr;
const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
dim3 reduce_grid(num_heads, num_seqs);
int max_num_partitions = DIVIDE_ROUND_UP(max_seq_len, PARTITION_SIZE);
int reduce_shared_mem_size = 2 * max_num_partitions * sizeof(float);
if constexpr(BLOCK_SIZE==16 && IS_BLOCK_SPARSE==false && sizeof(T)==2 && KV_DTYPE==vllm::Fp8KVCacheDataType::kAuto){
//if(head_size==128&&get_device_name()=="gfx928"){
constexpr int HEAD_SIZE=128;
constexpr static int use_vmac = false;
int reusekv, num_thread;
get_number_thread_and_reuse_kv_v2(num_thread,reusekv,num_seqs,max_num_partitions,num_heads,num_kv_heads);
if(PA_REUSE_KV_TIMES!=0&&num_heads>num_kv_heads)reusekv=PA_REUSE_KV_TIMES;
if(PA_BLOCK_SIZE!=0)num_thread=PA_BLOCK_SIZE;
REUSEKV_SWITCH(reusekv,[&] {
NUM_THREADS_SWITCH(num_thread , [&] {
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
int logits_size = REUSE_KV_TIMES*PARTITION_SIZE * 2;
int outputs_size = (NUM_WARPS / 2) * head_size * sizeof(float);
dim3 grid;
grid.x = (num_heads/num_kv_heads + REUSE_KV_TIMES -1)/REUSE_KV_TIMES * num_kv_heads;
grid.y = max_num_partitions;
grid.z = num_seqs;
dim3 block(NUM_THREADS);
int shared_mem_size = ::max(logits_size, outputs_size);
if(PA_PRINT_PARAM)printf("reusekv=%d,num_thread=%d,grid={%d,%d,%d},qhead=%d,kvhead=%d,seq=%d,batch=%d\n",
reusekv,num_thread,grid.x,grid.y,grid.z,num_heads,num_kv_heads,max_seq_len,num_seqs);
LAUNCH_PAGED_ATTENTION_V2_TC(HEAD_SIZE);
});
});
}
//}
}
#define CALL_V2_LAUNCHER(T, CACHE_T, BLOCK_SIZE, KV_DTYPE, IS_BLOCK_SPARSE) \
paged_attention_v2_launcher_opt_tc<T, CACHE_T, BLOCK_SIZE, KV_DTYPE, \
IS_BLOCK_SPARSE>( \
out, exp_sums, max_logits, tmp_out, query, key_cache, value_cache, \
num_kv_heads, scale, block_tables, seq_lens, max_seq_len, alibi_slopes, \
k_scale, v_scale, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step, attn_masks, attn_masks_stride);
#define CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE) \
switch (is_block_sparse) { \
case true: \
CALL_V2_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE, true); \
break; \
case false: \
CALL_V2_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE, false); \
break; \
}
// NOTE(woosuk): To reduce the compilation time, we omitted block sizes
// 1, 2, 4, 64, 128, 256.
#define CALL_V2_LAUNCHER_BLOCK_SIZE(T, CACHE_T, KV_DTYPE) \
switch (block_size) { \
case 8: \
CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, 8, KV_DTYPE); \
break; \
case 16: \
CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, 16, KV_DTYPE); \
break; \
case 32: \
CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, 32, KV_DTYPE); \
break; \
default: \
TORCH_CHECK(false, "Unsupported block size: ", block_size); \
break; \
}
void paged_attention_v2_opt_with_mask(
torch::Tensor& out, // [num_seqs, num_heads, head_size]
torch::Tensor& exp_sums, // [num_seqs, num_heads, max_num_partitions]
torch::Tensor& max_logits, // [num_seqs, num_heads, max_num_partitions]
torch::Tensor&
tmp_out, // [num_seqs, num_heads, max_num_partitions, head_size]
torch::Tensor& query, // [num_seqs, num_heads, head_size]
torch::Tensor&
key_cache, // [num_blocks, num_heads, head_size/x, block_size, x]
torch::Tensor&
value_cache, // [num_blocks, num_heads, head_size, block_size]
int64_t num_kv_heads, // [num_heads]
double scale,
torch::Tensor& block_tables, // [num_seqs, max_num_blocks_per_seq]
torch::Tensor& seq_lens, // [num_seqs]
int64_t block_size, int64_t max_seq_len,
const c10::optional<torch::Tensor>& alibi_slopes,
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride);
void paged_attention_v2_opt_tc_with_mask(
torch::Tensor& out, // [num_seqs, num_heads, head_size]
torch::Tensor& exp_sums, // [num_seqs, num_heads, max_num_partitions]
torch::Tensor& max_logits, // [num_seqs, num_heads, max_num_partitions]
torch::Tensor&
tmp_out, // [num_seqs, num_heads, max_num_partitions, head_size]
torch::Tensor& query, // [num_seqs, num_heads, head_size]
torch::Tensor&
key_cache, // [num_blocks, num_heads, head_size/x, block_size, x]
torch::Tensor&
value_cache, // [num_blocks, num_heads, head_size, block_size]
int64_t num_kv_heads, // [num_heads]
double scale,
torch::Tensor& block_tables, // [num_seqs, max_num_blocks_per_seq]
torch::Tensor& seq_lens, // [num_seqs]
int64_t block_size, int64_t max_seq_len,
const c10::optional<torch::Tensor>& alibi_slopes,
const std::string& kv_cache_dtype, double k_scale, double v_scale, const int64_t tp_rank,
const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step,
const c10::optional<torch::Tensor>& attn_masks, // [num_seqs, max_seq_len]
const int64_t attn_masks_stride) {
const bool is_block_sparse = (blocksparse_vert_stride > 1);
if(kv_cache_dtype != "auto"||query.dtype() == at::ScalarType::Float||is_block_sparse||
block_size!=16||query.size(2)!=128||(get_device_name()!="gfx928" && get_device_name()!="gfx936")){
paged_attention_v2_opt_with_mask(out,exp_sums,max_logits,tmp_out,query,key_cache,value_cache,num_kv_heads,
scale,block_tables,seq_lens,block_size,max_seq_len,alibi_slopes,kv_cache_dtype,
k_scale,v_scale,tp_rank,blocksparse_local_blocks,blocksparse_vert_stride,
blocksparse_block_size,blocksparse_head_sliding_step, attn_masks,
attn_masks_stride);
}
else{
DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
CALL_V2_LAUNCHER_BLOCK_SIZE)
}
}
#undef WARP_SIZE
#undef MAX
#undef MIN
#undef DIVIDE_ROUND_UP
csrc/ops.h
View file @ 44e3ca68
......@@ -6,6 +6,71 @@
#include "core/scalar_type.hpp"
void paged_attention_v1(
torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
int64_t max_seq_len, const c10::optional<torch::Tensor>& alibi_slopes,
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step);
void paged_attention_v2(
torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
int64_t max_seq_len, const c10::optional<torch::Tensor>& alibi_slopes,
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step);
void paged_attention_v1_opt(
torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
int64_t max_seq_len, const c10::optional<torch::Tensor>& alibi_slopes,
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step);
void paged_attention_v2_opt(
torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
int64_t max_seq_len, const c10::optional<torch::Tensor>& alibi_slopes,
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step);
void paged_attention_v1_opt_tc(
torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
int64_t max_seq_len, const c10::optional<torch::Tensor>& alibi_slopes,
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step);
void paged_attention_v2_opt_tc(
torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
int64_t max_seq_len, const c10::optional<torch::Tensor>& alibi_slopes,
const std::string& kv_cache_dtype, double k_scale, double v_scale,
const int64_t tp_rank, const int64_t blocksparse_local_blocks,
const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
const int64_t blocksparse_head_sliding_step);
// paged_attention with attn_masks
void paged_attention_v1_with_mask(
torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
......@@ -17,7 +82,7 @@ void paged_attention_v1(
const c10::optional<torch::Tensor>& attn_masks,
const int64_t attn_masks_stride=0);
void paged_attention_v2(
void paged_attention_v2_with_mask(
torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
......@@ -30,7 +95,7 @@ void paged_attention_v2(
const c10::optional<torch::Tensor>& attn_masks,
const int64_t attn_masks_stride=0);
void paged_attention_v1_opt(
void paged_attention_v1_opt_with_mask(
torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
......@@ -42,7 +107,7 @@ void paged_attention_v1_opt(
const c10::optional<torch::Tensor>& attn_masks,
const int64_t attn_masks_stride=0);
void paged_attention_v2_opt(
void paged_attention_v2_opt_with_mask(
torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
......@@ -55,7 +120,7 @@ void paged_attention_v2_opt(
const c10::optional<torch::Tensor>& attn_masks,
const int64_t attn_masks_stride=0);
void paged_attention_v1_opt_tc(
void paged_attention_v1_opt_tc_with_mask(
torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
......@@ -67,7 +132,7 @@ void paged_attention_v1_opt_tc(
const c10::optional<torch::Tensor>& attn_masks,
const int64_t attn_masks_stride=0);
void paged_attention_v2_opt_tc(
void paged_attention_v2_opt_tc_with_mask(
torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
......@@ -80,6 +145,7 @@ void paged_attention_v2_opt_tc(
const c10::optional<torch::Tensor>& attn_masks,
const int64_t attn_masks_stride=0);
void rms_norm(torch::Tensor& out, torch::Tensor& input, torch::Tensor& weight,
double epsilon);
......
csrc/torch_bindings.cpp
View file @ 44e3ca68
......@@ -30,14 +30,98 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
" str kv_cache_dtype, float k_scale, float v_scale,"
" int tp_rank, int blocksparse_local_blocks,"
" int blocksparse_vert_stride, int blocksparse_block_size,"
" int blocksparse_head_sliding_step) -> ()");
ops.impl("paged_attention_v1", torch::kCUDA, &paged_attention_v1);
// PagedAttention V2.
ops.def(
"paged_attention_v2("
" Tensor! out, Tensor! exp_sums, Tensor! max_logits,"
" Tensor! tmp_out, Tensor query, Tensor key_cache,"
" Tensor value_cache, int num_kv_heads, float scale,"
" Tensor block_tables, Tensor seq_lens, int block_size,"
" int max_seq_len, Tensor? alibi_slopes,"
" str kv_cache_dtype, float k_scale, float v_scale,"
" int tp_rank, int blocksparse_local_blocks,"
" int blocksparse_vert_stride, int blocksparse_block_size,"
" int blocksparse_head_sliding_step) -> ()");
ops.impl("paged_attention_v2", torch::kCUDA, &paged_attention_v2);
// Compute the attention between an input query and the cached
// keys/values using PagedAttention. (opt)
ops.def(
"paged_attention_v1_opt("
" Tensor! out, Tensor query, Tensor key_cache,"
" Tensor value_cache, int num_kv_heads, float scale,"
" Tensor block_tables, Tensor seq_lens, int block_size,"
" int max_seq_len, Tensor? alibi_slopes,"
" str kv_cache_dtype, float k_scale, float v_scale,"
" int tp_rank, int blocksparse_local_blocks,"
" int blocksparse_vert_stride, int blocksparse_block_size,"
" int blocksparse_head_sliding_step) -> ()");
ops.impl("paged_attention_v1_opt", torch::kCUDA, &paged_attention_v1_opt);
// PagedAttention V2 (opt).
ops.def(
"paged_attention_v2_opt("
" Tensor! out, Tensor exp_sums, Tensor max_logits,"
" Tensor tmp_out, Tensor query, Tensor key_cache,"
" Tensor value_cache, int num_kv_heads, float scale,"
" Tensor block_tables, Tensor seq_lens, int block_size,"
" int max_seq_len, Tensor? alibi_slopes,"
" str kv_cache_dtype, float k_scale, float v_scale,"
" int tp_rank, int blocksparse_local_blocks,"
" int blocksparse_vert_stride, int blocksparse_block_size,"
" int blocksparse_head_sliding_step) -> ()");
ops.impl("paged_attention_v2_opt", torch::kCUDA, &paged_attention_v2_opt);
// Compute the attention between an input query and the cached
// keys/values using PagedAttention. (opt)
ops.def(
"paged_attention_v1_opt_tc("
" Tensor! out, Tensor query, Tensor key_cache,"
" Tensor value_cache, int num_kv_heads, float scale,"
" Tensor block_tables, Tensor seq_lens, int block_size,"
" int max_seq_len, Tensor? alibi_slopes,"
" str kv_cache_dtype, float k_scale, float v_scale,"
" int tp_rank, int blocksparse_local_blocks,"
" int blocksparse_vert_stride, int blocksparse_block_size,"
" int blocksparse_head_sliding_step) -> ()");
ops.impl("paged_attention_v1_opt_tc", torch::kCUDA, &paged_attention_v1_opt_tc);
// PagedAttention V2 (opt).
ops.def(
"paged_attention_v2_opt_tc("
" Tensor! out, Tensor exp_sums, Tensor max_logits,"
" Tensor tmp_out, Tensor query, Tensor key_cache,"
" Tensor value_cache, int num_kv_heads, float scale,"
" Tensor block_tables, Tensor seq_lens, int block_size,"
" int max_seq_len, Tensor? alibi_slopes,"
" str kv_cache_dtype, float k_scale, float v_scale,"
" int tp_rank, int blocksparse_local_blocks,"
" int blocksparse_vert_stride, int blocksparse_block_size,"
" int blocksparse_head_sliding_step) -> ()");
ops.impl("paged_attention_v2_opt_tc", torch::kCUDA, &paged_attention_v2_opt_tc);
// paged_attention with atth_masks
ops.def(
"paged_attention_v1_with_mask("
" Tensor! out, Tensor query, Tensor key_cache,"
" Tensor value_cache, int num_kv_heads, float scale,"
" Tensor block_tables, Tensor seq_lens, int block_size,"
" int max_seq_len, Tensor? alibi_slopes,"
" str kv_cache_dtype, float k_scale, float v_scale,"
" int tp_rank, int blocksparse_local_blocks,"
" int blocksparse_vert_stride, int blocksparse_block_size,"
" int blocksparse_head_sliding_step,"
" Tensor? attn_masks,"
" int attn_masks_stride) -> ()");
ops.impl("paged_attention_v1", torch::kCUDA, &paged_attention_v1);
ops.impl("paged_attention_v1_with_mask", torch::kCUDA, &paged_attention_v1_with_mask);
// PagedAttention V2.
ops.def(
"paged_attention_v2("
"paged_attention_v2_with_mask("
" Tensor! out, Tensor! exp_sums, Tensor! max_logits,"
" Tensor! tmp_out, Tensor query, Tensor key_cache,"
" Tensor value_cache, int num_kv_heads, float scale,"
......@@ -49,12 +133,12 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
" int blocksparse_head_sliding_step,"
" Tensor? attn_masks,"
" int attn_masks_stride) -> ()");
ops.impl("paged_attention_v2", torch::kCUDA, &paged_attention_v2);
ops.impl("paged_attention_v2_with_mask", torch::kCUDA, &paged_attention_v2_with_mask);
// Compute the attention between an input query and the cached
// keys/values using PagedAttention. (opt)
ops.def(
"paged_attention_v1_opt("
"paged_attention_v1_opt_with_mask("
" Tensor! out, Tensor query, Tensor key_cache,"
" Tensor value_cache, int num_kv_heads, float scale,"
" Tensor block_tables, Tensor seq_lens, int block_size,"
......@@ -65,11 +149,11 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
" int blocksparse_head_sliding_step,"
" Tensor? attn_masks,"
" int attn_masks_stride) -> ()");
ops.impl("paged_attention_v1_opt", torch::kCUDA, &paged_attention_v1_opt);
ops.impl("paged_attention_v1_opt_with_mask", torch::kCUDA, &paged_attention_v1_opt_with_mask);
// PagedAttention V2 (opt).
ops.def(
"paged_attention_v2_opt("
"paged_attention_v2_opt_with_mask("
" Tensor! out, Tensor exp_sums, Tensor max_logits,"
" Tensor tmp_out, Tensor query, Tensor key_cache,"
" Tensor value_cache, int num_kv_heads, float scale,"
......@@ -81,12 +165,12 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
" int blocksparse_head_sliding_step,"
" Tensor? attn_masks,"
" int attn_masks_stride) -> ()");
ops.impl("paged_attention_v2_opt", torch::kCUDA, &paged_attention_v2_opt);
ops.impl("paged_attention_v2_opt_with_mask", torch::kCUDA, &paged_attention_v2_opt_with_mask);
// Compute the attention between an input query and the cached
// keys/values using PagedAttention. (opt)
ops.def(
"paged_attention_v1_opt_tc("
"paged_attention_v1_opt_tc_with_mask("
" Tensor! out, Tensor query, Tensor key_cache,"
" Tensor value_cache, int num_kv_heads, float scale,"
" Tensor block_tables, Tensor seq_lens, int block_size,"
......@@ -97,11 +181,11 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
" int blocksparse_head_sliding_step,"
" Tensor? attn_masks,"
" int attn_masks_stride) -> ()");
ops.impl("paged_attention_v1_opt_tc", torch::kCUDA, &paged_attention_v1_opt_tc);
ops.impl("paged_attention_v1_opt_tc_with_mask", torch::kCUDA, &paged_attention_v1_opt_tc_with_mask);
// PagedAttention V2 (opt).
ops.def(
"paged_attention_v2_opt_tc("
"paged_attention_v2_opt_tc_with_mask("
" Tensor! out, Tensor exp_sums, Tensor max_logits,"
" Tensor tmp_out, Tensor query, Tensor key_cache,"
" Tensor value_cache, int num_kv_heads, float scale,"
......@@ -113,7 +197,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
" int blocksparse_head_sliding_step,"
" Tensor? attn_masks,"
" int attn_masks_stride) -> ()");
ops.impl("paged_attention_v2_opt_tc", torch::kCUDA, &paged_attention_v2_opt_tc);
ops.impl("paged_attention_v2_opt_tc_with_mask", torch::kCUDA, &paged_attention_v2_opt_tc_with_mask);
// Activation ops
// Activation function used in SwiGLU.
......
examples/medusa/README.md
View file @ 44e3ca68
# Medusa Decoding
本文说明如何使用vllm构建和运行medusa模型，目前medusa支持tree-style generation，target model和draft model均可多卡推理
本文说明如何使用vllm构建和运行medusa模型
## Overview
Medusa是一种大模型并行解码算法，除了支持官方提供的Top1-proposer,我们还支持tree-style并行解码，target model和draft model均可多卡推理
与其他模型不同，medusa解码需要一个base model和若干Medusa heads.
Vllm medusa model的实现在[vllm/model_executor/models/medusa.py]
......@@ -19,28 +20,43 @@ Vllm medusa model的实现在[vllm/model_executor/models/medusa.py]
```bash
python medusa_weight_converter.py --medusa_num_heads 4 --medusa_num_layers 1 --medusa_model_path /work/model.bin --vocab_size 152064 --hidden_size 8192 --output_dir /work/medusa/vllm-medusa-qwen2-72b-head-4 --medusa_choices="[(0), (0, 0), (0, 0, 0), (0, 1), (1), (1, 0), (0, 0, 0, 0), (0, 0, 1), (0, 2), (0, 1, 0), (2), (0, 0, 2), (0, 3), (1, 0, 0), (2, 0), (0, 2, 0), (0, 4), (0, 0, 3), (3), (0, 0, 0, 1), (0, 5), (0, 0, 1, 0), (0, 0, 4)]"
```
此处model.bin是训练后保存的medusa head权重
此处model.bin是训练后保存的medusa head权重，如果希望采用Top1-proposer，medusa_choices可以不设置
### Run
### Run tree-style generation server
```bash
python3 -m vllm.entrypoints.openai.api_server \
VLLM_TREE_DECODING=1 python3 -m vllm.entrypoints.openai.api_server \
--served-model-name qwen_medusa \
--model /models/Qwen2-72B-Instruct/ -tp 4 \
--max-model-len 1024 --max-num-seqs 8 --gpu-memory-utilization 0.8 \
--speculative-model /work/medusa/vllm-medusa-qwen2-72b-head-4 \
--speculative-draft-tensor-parallel-size 4 \
--speculative-disable-by-batch-size 4 \
--speculative-disable-by-batch-size 9 \
--use-v2-block-manager \
--spec-decoding-acceptance-method typical_acceptance_sampler \
--dtype float16 --trust-remote-code --port 8086\
--tree-style-spec-decoding True\
--num-speculative-heads 4 --num-speculative-tokens 24
```
注意：
num_speculative_tokens = len(medusa_choices) + 1
medusa_choices个数不能太多，否则多batch下会降低推理速度
speculative-disable-by-batch-size要大于max-num-seqs，否则当batch等于max-num-seqs时，不会走并行解码
merge-lora可以将lora权重和base model权重融合，提升整体推理速度，若对精度有严格要求，可不设置此参数
num-speculative-tokens和medusa choices的个数相关，num_speculative_tokens = len(medusa_choices) + 1
### Run Top1-proposer server
python3 -m vllm.entrypoints.openai.api_server \
--served-model-name qwen_medusa \
--model /models/Qwen2-72B-Instruct/ -tp 4 \
--max-model-len 1024 --max-num-seqs 8 --gpu-memory-utilization 0.8 \
--speculative-model /work/medusa/vllm-medusa-qwen2-72b-head-4 \
--speculative-draft-tensor-parallel-size 4 \
--speculative-disable-by-batch-size 9 \
--use-v2-block-manager \
--spec-decoding-acceptance-method typical_acceptance_sampler \
--dtype float16 --trust-remote-code --port 8086\
--num-speculative-tokens 4
注意:
使用Top1-proposer时，num-speculative-tokens就是medusa head的个数
# do request
```bash
......@@ -54,8 +70,14 @@ curl http://localhost:8086/v1/completions \
}'
```
### benchmark
python medusa_benchmark_throughput.py --model /data/llm-models/qwen2/Qwen2-72B-Instruct/ -tp 4 --dtype float16 --trust-remote-code --max-num-seqs 1 --dataset /work/test/medusa_benchmark_data.json --max-model-len 4096 --gpu-memory-utilization 0.9
### Run tree-style benchmark
```bash
VLLM_TREE_DECODING=1 python /work/test/medusa_benchmark_throughput.py --model /models/Qwen2-72B-Instruct/ -tp 4 --dtype float16 --trust-remote-code --max-num-seqs 4 --speculative-model /work/medusa/vllm-medusa1-qwen2-72b-head-4 --speculative-draft-tensor-parallel-size 4 --speculative-disable-by-batch-size 9 --use-v2-block-manager --spec-decoding-acceptance-method typical_acceptance_sampler --max-model-len 1024 --dataset /work/medusa_benchmark_data.json --num-speculative-heads 4 --num-speculative-tokens 24 --gpu-memory-utilization 0.95
```
### Run Top1-proposer benchmark
```bash
python /work/test/medusa_benchmark_throughput.py --model /models/Qwen2-72B-Instruct/ -tp 4 --dtype float16 --trust-remote-code --max-num-seqs 4 --speculative-model /work/medusa/vllm-medusa1-qwen2-72b-head-4 --speculative-draft-tensor-parallel-size 4 --speculative-disable-by-batch-size 9 --use-v2-block-manager --spec-decoding-acceptance-method typical_acceptance_sampler --max-model-len 1024 --dataset /work/medusa_benchmark_data.json --num-speculative-tokens 4 --gpu-memory-utilization 0.95
```
可设置max-num-seqs对不同的batch进行性能测试
examples/medusa/medusa_benchmark_throughput.py
View file @ 44e3ca68
......@@ -98,7 +98,6 @@ def run_vllm(
merge_lora: bool = False,
lora_extra_vocab_size: int = 0,
lora_target_modules: List[str] = None,
tree_style_spec_decoding: bool = False,
num_speculative_heads: int = 5,
num_speculative_tokens: int = 64,
use_new_beam_search_impl: bool = False,
......@@ -138,7 +137,6 @@ def run_vllm(
merge_lora=merge_lora,
lora_extra_vocab_size=lora_extra_vocab_size,
lora_target_modules=lora_target_modules,
tree_style_spec_decoding=tree_style_spec_decoding,
num_speculative_heads=num_speculative_heads,
num_speculative_tokens=num_speculative_tokens
)
......@@ -234,7 +232,6 @@ async def run_vllm_async(
merge_lora: bool = False,
lora_extra_vocab_size: int = 0,
lora_target_modules: List[str] = None,
tree_style_spec_decoding: bool = False,
num_speculative_heads: int = 5,
num_speculative_tokens: int = 64,
use_new_beam_search_impl: bool = False,
......@@ -276,7 +273,6 @@ async def run_vllm_async(
merge_lora=merge_lora,
lora_extra_vocab_size=lora_extra_vocab_size,
lora_target_modules=lora_target_modules,
tree_style_spec_decoding=tree_style_spec_decoding,
num_speculative_heads=num_speculative_heads,
num_speculative_tokens=num_speculative_tokens
)
......@@ -350,7 +346,7 @@ def main(args: argparse.Namespace):
args.speculative_model, args.speculative_draft_tensor_parallel_size,
args.speculative_disable_by_batch_size, args.spec_decoding_acceptance_method,
args.enable_lora, args.max_lora_rank, args.merge_lora, args.lora_extra_vocab_size,
args.lora_target_modules, args.tree_style_spec_decoding, args.num_speculative_heads,
args.lora_target_modules, args.num_speculative_heads,
args.num_speculative_tokens
]
else:
......@@ -368,7 +364,7 @@ def main(args: argparse.Namespace):
args.speculative_model, args.speculative_draft_tensor_parallel_size,
args.speculative_disable_by_batch_size, args.spec_decoding_acceptance_method,
args.enable_lora, args.max_lora_rank, args.merge_lora, args.lora_extra_vocab_size,
args.lora_target_modules, args.tree_style_spec_decoding, args.num_speculative_heads,
args.lora_target_modules, args.num_speculative_heads,
args.num_speculative_tokens
]
......@@ -625,11 +621,6 @@ if __name__ == "__main__":
default=None,
help='List of lora module name, If not specified, modules will be chosen according to the model architecture.')
parser.add_argument('--tree-style-spec-decoding',
type=bool,
default=False,
help='If set to True, tree-style generation will be activated.')
parser.add_argument(
'--num-speculative-heads',
type=int,
......
tests/kernels/test_attention.py
View file @ 44e3ca68
......@@ -225,7 +225,7 @@ def test_paged_attention(
opcheck(torch.ops._C.paged_attention_v1,
(output, query, key_cache, value_cache, num_kv_heads, scale,
block_tables, seq_lens, block_size, max_seq_len, alibi_slopes,
kv_cache_dtype, k_scale, v_scale, 0, 0, 0, 64, 0, None, 0),
kv_cache_dtype, k_scale, v_scale, 0, 0, 0, 64, 0),
cond=(head_size == HEAD_SIZES[0]
and block_size == BLOCK_SIZES[0]))
else:
......@@ -291,7 +291,7 @@ def test_paged_attention(
(output, exp_sums, max_logits, tmp_output, query,
key_cache, value_cache, num_kv_heads, scale, block_tables,
seq_lens, block_size, max_seq_len, alibi_slopes,
kv_cache_dtype, k_scale, v_scale, 0, 0, 0, 64, 0, None, 0),
kv_cache_dtype, k_scale, v_scale, 0, 0, 0, 64, 0),
cond=(head_size == HEAD_SIZES[0]
and block_size == BLOCK_SIZES[0]))
else:
......
tests/worker/test_model_input.py
View file @ 44e3ca68
......@@ -60,16 +60,6 @@ class MockAttentionBackend(AttentionBackend):
) -> None:
pass
@staticmethod
def move_cache(
kv_caches: List[torch.Tensor],
src_to_dists: torch.Tensor,
kv_cache_dtype: str,
num_kv_heads: int,
head_size: int,
) -> None:
NotImplementedError
def test_model_runner_input():
sampling_metadata = SamplingMetadata(
......
vllm/_custom_ops.py
View file @ 44e3ca68
......@@ -89,6 +89,66 @@ def gelu_quick(out: torch.Tensor, x: torch.Tensor) -> None:
# page attention ops
def paged_attention_v1(
out: torch.Tensor,
query: torch.Tensor,
key_cache: torch.Tensor,
value_cache: torch.Tensor,
num_kv_heads: int,
scale: float,
block_tables: torch.Tensor,
seq_lens: torch.Tensor,
block_size: int,
max_seq_len: int,
alibi_slopes: Optional[torch.Tensor],
kv_cache_dtype: str,
k_scale: float,
v_scale: float,
tp_rank: int = 0,
blocksparse_local_blocks: int = 0,
blocksparse_vert_stride: int = 0,
blocksparse_block_size: int = 64,
blocksparse_head_sliding_step: int = 0
) -> None:
torch.ops._C.paged_attention_v1(
out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
k_scale, v_scale, tp_rank, blocksparse_local_blocks,
blocksparse_vert_stride, blocksparse_block_size,
blocksparse_head_sliding_step)
def paged_attention_v2(
out: torch.Tensor,
exp_sum: torch.Tensor,
max_logits: torch.Tensor,
tmp_out: torch.Tensor,
query: torch.Tensor,
key_cache: torch.Tensor,
value_cache: torch.Tensor,
num_kv_heads: int,
scale: float,
block_tables: torch.Tensor,
seq_lens: torch.Tensor,
block_size: int,
max_seq_len: int,
alibi_slopes: Optional[torch.Tensor],
kv_cache_dtype: str,
k_scale: float,
v_scale: float,
tp_rank: int = 0,
blocksparse_local_blocks: int = 0,
blocksparse_vert_stride: int = 0,
blocksparse_block_size: int = 64,
blocksparse_head_sliding_step: int = 0
) -> None:
torch.ops._C.paged_attention_v2(
out, exp_sum, max_logits, tmp_out, query, key_cache, value_cache,
num_kv_heads, scale, block_tables, seq_lens, block_size, max_seq_len,
alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
blocksparse_local_blocks, blocksparse_vert_stride,
blocksparse_block_size, blocksparse_head_sliding_step)
def paged_attention_v1_with_mask(
out: torch.Tensor,
query: torch.Tensor,
key_cache: torch.Tensor,
......@@ -111,7 +171,7 @@ def paged_attention_v1(
attn_masks: Optional[torch.Tensor] = None,
attn_masks_stride: int = 0,
) -> None:
torch.ops._C.paged_attention_v1(
torch.ops._C.paged_attention_v1_with_mask(
out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
k_scale, v_scale, tp_rank, blocksparse_local_blocks,
......@@ -120,7 +180,7 @@ def paged_attention_v1(
attn_masks_stride)
def paged_attention_v2(
def paged_attention_v2_with_mask(
out: torch.Tensor,
exp_sum: torch.Tensor,
max_logits: torch.Tensor,
......@@ -146,7 +206,7 @@ def paged_attention_v2(
attn_masks: Optional[torch.Tensor] = None,
attn_masks_stride: int = 0,
) -> None:
torch.ops._C.paged_attention_v2(
torch.ops._C.paged_attention_v2_with_mask(
out, exp_sum, max_logits, tmp_out, query, key_cache, value_cache,
num_kv_heads, scale, block_tables, seq_lens, block_size, max_seq_len,
alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
......@@ -157,6 +217,67 @@ def paged_attention_v2(
# page attention ops (opt)
def paged_attention_v1_opt(
out: torch.Tensor,
query: torch.Tensor,
key_cache: torch.Tensor,
value_cache: torch.Tensor,
num_kv_heads: int,
scale: float,
block_tables: torch.Tensor,
seq_lens: torch.Tensor,
block_size: int,
max_seq_len: int,
alibi_slopes: Optional[torch.Tensor],
kv_cache_dtype: str,
k_scale: float,
v_scale: float,
tp_rank: int = 0,
blocksparse_local_blocks: int = 0,
blocksparse_vert_stride: int = 0,
blocksparse_block_size: int = 64,
blocksparse_head_sliding_step: int = 0
) -> None:
torch.ops._C.paged_attention_v1_opt(
out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
k_scale, v_scale, tp_rank, blocksparse_local_blocks,
blocksparse_vert_stride, blocksparse_block_size,
blocksparse_head_sliding_step)
def paged_attention_v2_opt(
out: torch.Tensor,
exp_sum: torch.Tensor,
max_logits: torch.Tensor,
tmp_out: torch.Tensor,
query: torch.Tensor,
key_cache: torch.Tensor,
value_cache: torch.Tensor,
num_kv_heads: int,
scale: float,
block_tables: torch.Tensor,
seq_lens: torch.Tensor,
block_size: int,
max_seq_len: int,
alibi_slopes: Optional[torch.Tensor],
kv_cache_dtype: str,
k_scale: float,
v_scale: float,
tp_rank: int = 0,
blocksparse_local_blocks: int = 0,
blocksparse_vert_stride: int = 0,
blocksparse_block_size: int = 64,
blocksparse_head_sliding_step: int = 0
) -> None:
torch.ops._C.paged_attention_v2_opt(
out, exp_sum, max_logits, tmp_out, query, key_cache, value_cache,
num_kv_heads, scale, block_tables, seq_lens, block_size, max_seq_len,
alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
blocksparse_local_blocks, blocksparse_vert_stride,
blocksparse_block_size, blocksparse_head_sliding_step)
def paged_attention_v1_opt_with_mask(
out: torch.Tensor,
query: torch.Tensor,
key_cache: torch.Tensor,
......@@ -179,7 +300,7 @@ def paged_attention_v1_opt(
attn_masks: Optional[torch.Tensor] = None,
attn_masks_stride: int = 0,
) -> None:
torch.ops._C.paged_attention_v1_opt(
torch.ops._C.paged_attention_v1_opt_with_mask(
out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
k_scale, v_scale, tp_rank, blocksparse_local_blocks,
......@@ -188,7 +309,7 @@ def paged_attention_v1_opt(
attn_masks_stride)
def paged_attention_v2_opt(
def paged_attention_v2_opt_with_mask(
out: torch.Tensor,
exp_sum: torch.Tensor,
max_logits: torch.Tensor,
......@@ -214,7 +335,7 @@ def paged_attention_v2_opt(
attn_masks: Optional[torch.Tensor] = None,
attn_masks_stride: int = 0,
) -> None:
torch.ops._C.paged_attention_v2_opt(
torch.ops._C.paged_attention_v2_opt_with_mask(
out, exp_sum, max_logits, tmp_out, query, key_cache, value_cache,
num_kv_heads, scale, block_tables, seq_lens, block_size, max_seq_len,
alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
......@@ -225,6 +346,67 @@ def paged_attention_v2_opt(
# page attention ops (opt)
def paged_attention_v1_opt_tc(
out: torch.Tensor,
query: torch.Tensor,
key_cache: torch.Tensor,
value_cache: torch.Tensor,
num_kv_heads: int,
scale: float,
block_tables: torch.Tensor,
seq_lens: torch.Tensor,
block_size: int,
max_seq_len: int,
alibi_slopes: Optional[torch.Tensor],
kv_cache_dtype: str,
k_scale: float,
v_scale: float,
tp_rank: int = 0,
blocksparse_local_blocks: int = 0,
blocksparse_vert_stride: int = 0,
blocksparse_block_size: int = 64,
blocksparse_head_sliding_step: int = 0
) -> None:
torch.ops._C.paged_attention_v1_opt_tc(
out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
k_scale, v_scale, tp_rank, blocksparse_local_blocks, blocksparse_vert_stride,
blocksparse_block_size, blocksparse_head_sliding_step)
def paged_attention_v2_opt_tc(
out: torch.Tensor,
exp_sum: torch.Tensor,
max_logits: torch.Tensor,
tmp_out: torch.Tensor,
query: torch.Tensor,
key_cache: torch.Tensor,
value_cache: torch.Tensor,
num_kv_heads: int,
scale: float,
block_tables: torch.Tensor,
seq_lens: torch.Tensor,
block_size: int,
max_seq_len: int,
alibi_slopes: Optional[torch.Tensor],
kv_cache_dtype: str,
k_scale: float,
v_scale: float,
tp_rank: int = 0,
blocksparse_local_blocks: int = 0,
blocksparse_vert_stride: int = 0,
blocksparse_block_size: int = 64,
blocksparse_head_sliding_step: int = 0
) -> None:
torch.ops._C.paged_attention_v2_opt_tc(
out, exp_sum, max_logits, tmp_out, query, key_cache, value_cache,
num_kv_heads, scale, block_tables, seq_lens, block_size, max_seq_len,
alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
blocksparse_local_blocks, blocksparse_vert_stride,
blocksparse_block_size, blocksparse_head_sliding_step)
# page attention ops (opt)
def paged_attention_v1_opt_tc_with_mask(
out: torch.Tensor,
query: torch.Tensor,
key_cache: torch.Tensor,
......@@ -247,7 +429,7 @@ def paged_attention_v1_opt_tc(
attn_masks: Optional[torch.Tensor] = None,
attn_masks_stride: int = 0,
) -> None:
torch.ops._C.paged_attention_v1_opt_tc(
torch.ops._C.paged_attention_v1_opt_tc_with_mask(
out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
k_scale, v_scale, tp_rank, blocksparse_local_blocks, blocksparse_vert_stride,
......@@ -255,7 +437,7 @@ def paged_attention_v1_opt_tc(
attn_masks, attn_masks_stride)
def paged_attention_v2_opt_tc(
def paged_attention_v2_opt_tc_with_mask(
out: torch.Tensor,
exp_sum: torch.Tensor,
max_logits: torch.Tensor,
......@@ -281,7 +463,7 @@ def paged_attention_v2_opt_tc(
attn_masks: Optional[torch.Tensor] = None,
attn_masks_stride: int = 0,
) -> None:
torch.ops._C.paged_attention_v2_opt_tc(
torch.ops._C.paged_attention_v2_opt_tc_with_mask(
out, exp_sum, max_logits, tmp_out, query, key_cache, value_cache,
num_kv_heads, scale, block_tables, seq_lens, block_size, max_seq_len,
alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
......
vllm/attention/backends/abstract.py
View file @ 44e3ca68
......@@ -83,17 +83,6 @@ class AttentionBackend(ABC):
) -> None:
raise NotImplementedError
@staticmethod
@abstractmethod
def move_cache(
kv_caches: List[torch.Tensor],
src_to_dists: torch.Tensor,
kv_cache_dtype: str,
num_kv_heads: int,
head_size: int,
) -> None:
raise NotImplementedError
def advance_step(self, model_input: "ModelRunnerInputBase",
sampled_token_ids: Optional[torch.Tensor],
block_size: int, num_seqs: int, num_queries: int) -> None:
......@@ -206,8 +195,7 @@ class AttentionMetadataBuilder(ABC, Generic[T]):
@abstractmethod
def build(self, seq_lens: List[int], query_lens: List[int],
cuda_graph_pad_size: int, batch_size: int,
tree_attention_masks_tensor: Optional[torch.Tensor] = None) -> T:
cuda_graph_pad_size: int, batch_size: int) -> T:
"""Build attention metadata with on-device tensors."""
raise NotImplementedError
......
vllm/attention/backends/blocksparse_attn.py
View file @ 44e3ca68
......@@ -129,50 +129,6 @@ class BlocksparseFlashAttentionBackend(AttentionBackend):
) -> None:
PagedAttention.copy_blocks(kv_caches, src_to_dists)
@staticmethod
def move_cache(
kv_caches: List[torch.Tensor],
src_to_dists: torch.Tensor,
kv_cache_dtype: str,
num_kv_heads: int,
head_size: int,
) -> None:
key_caches = []
value_caches = []
num_layers = len(kv_caches)
token_num = src_to_dists.shape[0]
tmp_store_kv = torch.empty(
(2, num_layers, token_num, num_kv_heads, head_size),
dtype=kv_caches[0].dtype, device=kv_caches[0].device)
keys = tmp_store_kv[0].contiguous()
values = tmp_store_kv[1].contiguous()
for kv_cache in kv_caches:
key_cache, value_cache = PagedAttention.split_kv_cache(
kv_cache, num_kv_heads, head_size)
key_caches.append(key_cache)
value_caches.append(value_cache)
ops.read_cache(
keys,
values,
key_caches,
value_caches,
src_to_dists[:, 0].contiguous(),
kv_cache_dtype
)
ops.write_cache_multi_layers(
keys,
values,
key_caches,
value_caches,
src_to_dists[:, 1].contiguous(),
kv_cache_dtype
)
@dataclass
class BlocksparseFlashAttentionMetadata(AttentionMetadata):
......@@ -236,8 +192,6 @@ class BlocksparseFlashAttentionMetadata(AttentionMetadata):
_cached_decode_metadata: Optional[
"BlocksparseFlashAttentionMetadata"] = None
tree_attention_masks_tensor: Optional[torch.Tensor] = None
block_tables_list: Optional[List[int]] = None
@property
......@@ -271,7 +225,6 @@ class BlocksparseFlashAttentionMetadata(AttentionMetadata):
context_lens_tensor=self.context_lens_tensor[:self.num_prefills],
block_tables=self.block_tables[:self.num_prefills],
use_cuda_graph=False,
tree_attention_masks_tensor=self.tree_attention_masks_tensor,
block_tables_list=self.block_tables_list
)
return self._cached_prefill_metadata
......@@ -301,7 +254,6 @@ class BlocksparseFlashAttentionMetadata(AttentionMetadata):
context_lens_tensor=None,
block_tables=self.block_tables[self.num_prefills:],
use_cuda_graph=self.use_cuda_graph,
tree_attention_masks_tensor=self.tree_attention_masks_tensor,
block_tables_list=self.block_tables_list
)
return self._cached_decode_metadata
......
vllm/attention/backends/flash_attn.py
View file @ 44e3ca68
......@@ -221,16 +221,6 @@ class FlashAttentionBackend(AttentionBackend):
value_caches = [kv_cache[1] for kv_cache in kv_caches]
ops.copy_blocks(key_caches, value_caches, src_to_dists)
@staticmethod
def move_cache(
kv_caches: List[torch.Tensor],
src_to_dists: torch.Tensor,
kv_cache_dtype: str,
num_kv_heads: int,
head_size: int,
) -> None:
NotImplementedError
@dataclass
class FlashAttentionMetadata(AttentionMetadata):
......
vllm/attention/backends/flashinfer.py
View file @ 44e3ca68
......@@ -93,16 +93,6 @@ class FlashInferBackend(AttentionBackend):
else:
raise ValueError(f"Unrecognized FP8 dtype: {kv_cache_dtype}")
@staticmethod
def move_cache(
kv_caches: List[torch.Tensor],
src_to_dists: torch.Tensor,
kv_cache_dtype: str,
num_kv_heads: int,
head_size: int,
) -> None:
NotImplementedError
class FlashInferState(AttentionState):
......
vllm/attention/backends/ipex_attn.py
View file @ 44e3ca68
......@@ -62,16 +62,6 @@ class IpexAttnBackend(AttentionBackend):
value_caches = [kv_cache[1] for kv_cache in kv_caches]
ops.copy_blocks(key_caches, value_caches, src_to_dists)
@staticmethod
def move_cache(
kv_caches: List[torch.Tensor],
src_to_dists: torch.Tensor,
kv_cache_dtype: str,
num_kv_heads: int,
head_size: int,
) -> None:
NotImplementedError
@dataclass
class IpexAttnMetadata(AttentionMetadata, PagedAttentionMetadata):
......
vllm/attention/backends/openvino.py
View file @ 44e3ca68
......@@ -62,16 +62,6 @@ class OpenVINOAttentionBackend(AttentionBackend):
key_cache.data[dst, :] = key_cache.data[src, :]
value_cache.data[dst, :] = value_cache.data[src, :]
@staticmethod
def move_cache(
kv_caches: List[torch.Tensor],
src_to_dists: torch.Tensor,
kv_cache_dtype: str,
num_kv_heads: int,
head_size: int,
) -> None:
NotImplementedError
@dataclass
class OpenVINOAttentionMetadata:
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment