Skip to content

GitLab

Commit 0c70376b authored by zhuwenwen's avatar zhuwenwen
Browse files

add pa tc

parent fe1ec8c5
Hide whitespace changes
Inline Side-by-side
Showing with 614 additions and 583 deletions
csrc/attention/attention_kernels_opt.cu
View file @ 0c70376b
......@@ -20,12 +20,30 @@ typedef __hip_bfloat16 __nv_bfloat16;
#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))
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.
}
namespace vllm {
// Utility function for attention softmax.
......@@ -64,16 +82,63 @@ inline __device__ float block_sum(float* red_smem, float sum) {
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 = 1,
bool odd_nheads = false,
int PARTITION_SIZE = 0> // Zero means no partitioning.
__device__ void paged_attention_kernel_opt(
bool IS_BLOCK_SPARSE,int REUSE_KV_TIMES,bool use_vmac,int PARTITION_SIZE = 0> // Zero means no partitioning.
__device__ void paged_attention_kernel_TC(
float* __restrict__ exp_sums, // [num_seqs, num_heads, max_num_partitions]
float* __restrict__ max_logits, // [num_seqs, num_heads,
// max_num_partitions]
......@@ -84,8 +149,8 @@ __device__ void paged_attention_kernel_opt(
// 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 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]
......@@ -99,285 +164,170 @@ __device__ void paged_attention_kernel_opt(
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];
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;
}
if constexpr (sizeof(scalar_t)==2){
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 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;
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;
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 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_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 warp_idx = __builtin_amdgcn_readfirstlane(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 rowid = lane%16;
const int rows = lane/16;
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
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<REUSE_KV_TIMES;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.
float* logits = reinterpret_cast<float*>(shared_mem);
scalar_t* logits = reinterpret_cast<scalar_t*>(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
__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;
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]);
// 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;
// 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;
for (int block_idx = start_block_idx + warp_idx; block_idx < end_block_idx;
block_idx += NUM_WARPS) {
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);
}
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;
}
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]);
}
}
__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;
__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;
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();
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[reuse_kv_idx] = lane < NUM_WARPS ? red_smem[reuse_kv_idx][lane] : -FLT_MAX;
// 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[reuse_kv_idx] = fmaxf(qk_max[reuse_kv_idx], VLLM_SHFL_XOR_SYNC(qk_max[reuse_kv_idx], mask));
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[reuse_kv_idx] = VLLM_SHFL_SYNC(qk_max[reuse_kv_idx], 0);
qk_max_tmp = VLLM_SHFL_SYNC(qk_max_tmp, 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;
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[reuse_kv_idx] = block_sum<NUM_WARPS>(&red_smem[reuse_kv_idx][NUM_WARPS], exp_sum[reuse_kv_idx]);
exp_sum = block_sum<NUM_WARPS>(&red_smem[NUM_WARPS], exp_sum);
// Compute softmax.
const float inv_sum = __fdividef(1.f, exp_sum[reuse_kv_idx] + 1e-6f);
const float inv_sum = __fdividef(1.f, exp_sum + 1e-6f);
for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) {
logits[(reuse_kv_idx * partition_size) + i] *= inv_sum;
from_float(logits[(reuse_kv_idx * partition_size) + i] ,to_float(logits[(reuse_kv_idx * partition_size) + i])*inv_sum);
}
__syncthreads();
......@@ -385,222 +335,212 @@ __device__ void paged_attention_kernel_opt(
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];
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[reuse_kv_idx];
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[REUSE_KV_TIMES][NUM_ROWS_PER_THREAD];
#pragma unroll
for(int reuse_kv_idx=0; reuse_kv_idx<REUSE_KV_TIMES; reuse_kv_idx++) {
constexpr int NUM_ROWS_PER_THREAD =DIVIDE_ROUND_UP(HEAD_SIZE, WARP_SIZE);//2
if constexpr(REUSE_KV_TIMES<=2&&(NUM_WARPS>64||USE_PARTITIONING)){
float accs[REUSE_KV_TIMES][NUM_ROWS_PER_THREAD];
#pragma unroll
for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
accs[reuse_kv_idx][i] = 0.f;
#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 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;
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);
}
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 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];
}
}
}
}
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);
}
}
__syncthreads();
// Perform reduction across warps.
for(int reuse_kv_idx=0; reuse_kv_idx<q_boundary; reuse_kv_idx++) {
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 + i * WARP_SIZE;
dst[row_idx] = accs[reuse_kv_idx][i];
}
}
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;
__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 + i * WARP_SIZE;
accs[reuse_kv_idx][i] += src[row_idx];
}
}
// 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);
__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]);
}
}
}
}
// 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);
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;
}
}
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
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++) {
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];
#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];
}
}
}
}
}
}
__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];
// Perform reduction across warps.
for(int reuse_kv_idx=0; reuse_kv_idx<GROUPS; reuse_kv_idx++) {
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 + i * WARP_SIZE;
dst[row_idx] = accs[reuse_kv_idx][i];
}
}
__syncthreads();
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 + i * WARP_SIZE;
accs[reuse_kv_idx][i] += src[row_idx];
}
}
__syncthreads();
}
// Write the final output.
}
__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]);
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]);
}
}
}
}
}
}
}
}
// 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_kernel_opt(
bool IS_BLOCK_SPARSE,int REUSE_KV_TIMES,bool use_vmac>
__global__ void paged_attention_v1_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, // [num_heads]
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]
......@@ -608,28 +548,27 @@ __global__ __launch_bounds__(256,1) void paged_attention_v1_kernel_opt(
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 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) {
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,
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);
}
#ifdef __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>(
/* 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);
#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,
int PARTITION_SIZE,
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_kernel_opt(
__global__ __launch_bounds__(256, 1) void paged_attention_v2_kernel_TC(
float* __restrict__ exp_sums, // [num_seqs, num_heads, max_num_partitions]
float* __restrict__ max_logits, // [num_seqs, num_heads,
// max_num_partitions]
......@@ -640,7 +579,7 @@ __global__ __launch_bounds__(256,1) void paged_attention_v2_kernel_opt(
// 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_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]
......@@ -648,23 +587,24 @@ __global__ __launch_bounds__(256,1) void paged_attention_v2_kernel_opt(
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 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) {
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);
#ifdef __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,
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);
#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(
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]
......@@ -717,7 +657,7 @@ __global__ __launch_bounds__(256,1) void paged_attention_v2_reduce_kernel_opt(
// Get the global max logit.
// Reduce within the warp.
#pragma unroll
#pragma unroll
for (int mask = WARP_SIZE / 2; mask >= 1; mask /= 2) {
max_logit = fmaxf(max_logit, VLLM_SHFL_XOR_SYNC(max_logit, mask));
}
......@@ -727,7 +667,7 @@ __global__ __launch_bounds__(256,1) void paged_attention_v2_reduce_kernel_opt(
__syncthreads();
// Reduce across warps.
max_logit = lane < NUM_WARPS ? red_smem[lane] : -FLT_MAX;
#pragma unroll
#pragma unroll
for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
max_logit = fmaxf(max_logit, VLLM_SHFL_XOR_SYNC(max_logit, mask));
}
......@@ -757,7 +697,7 @@ __global__ __launch_bounds__(256,1) void paged_attention_v2_reduce_kernel_opt(
head_idx * max_num_partitions * HEAD_SIZE;
scalar_t* out_ptr =
out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
#pragma unroll
#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) {
......@@ -770,37 +710,27 @@ __global__ __launch_bounds__(256,1) void paged_attention_v2_reduce_kernel_opt(
} // namespace vllm
#define LAUNCH_PAGED_ATTENTION_V1(HEAD_SIZE) \
#define LAUNCH_PAGED_ATTENTION_V1_TC(HEAD_SIZE) \
VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize( \
((void*)vllm::paged_attention_v1_kernel_opt<T, CACHE_T, HEAD_SIZE, \
((void*)vllm::paged_attention_v1_kernel_TC<T, CACHE_T, HEAD_SIZE, \
BLOCK_SIZE, NUM_THREADS, \
KV_DTYPE, REUSE_KV_TIMES, IS_BLOCK_SPARSE, odd_nheads>), \
KV_DTYPE, IS_BLOCK_SPARSE,REUSE_KV_TIMES,use_vmac>), \
shared_mem_size); \
hipLaunchKernelGGL(( vllm::paged_attention_v1_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, \
vllm::paged_attention_v1_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, \
k_scale, v_scale, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step);
// #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(
void paged_attention_v1_launcher_opt(
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,
......@@ -816,11 +746,11 @@ void paged_attention_v1_launcher(
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;
// 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);
[[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
assert(head_size % thread_group_size == 0);
// NOTE: alibi_slopes is optional.
......@@ -835,39 +765,48 @@ void paged_attention_v1_launcher(
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>();
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);
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){
// if(head_size==128&&get_device_name()=="gfx928"){
REUSEKV_SWITCH_V1([&] {
constexpr int HEAD_SIZE=128;
// constexpr int REUSE_KV_TIMES=8;
int num_thread=64;
if(REUSE_KV_TIMES>1){
if(padded_max_seq_len>1024||num_heads * num_seqs/REUSE_KV_TIMES<600)num_thread=256;
else num_thread=128;
}
else if(num_heads * num_seqs<800)num_thread=128;
NUM_THREADS_SWITCH(num_thread , [&] {
constexpr static int use_vmac = false;
//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 = REUSE_KV_TIMES * (NUM_WARPS / 2) * head_size * sizeof(float);
if(REUSE_KV_TIMES==1)outputs_size=0;
// 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;
// printf("REUSE_KV_TIMES=%d,use_vmac=%d\n",REUSE_KV_TIMES,(int)use_vmac);
dim3 grid((num_heads/num_kv_heads + REUSE_KV_TIMES - 1) / REUSE_KV_TIMES*num_kv_heads, 1,num_seqs);
dim3 block(NUM_THREADS);
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<T, CACHE_T, BLOCK_SIZE, KV_DTYPE, \
paged_attention_v1_launcher_opt<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, \
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);
......@@ -899,6 +838,24 @@ void paged_attention_v1_launcher(
break; \
}
void paged_attention_v1(
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);
void paged_attention_v1_opt(
torch::Tensor& out, // [num_seqs, num_heads, head_size]
torch::Tensor& query, // [num_seqs, num_heads, head_size]
......@@ -912,37 +869,46 @@ void paged_attention_v1_opt(
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 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 bool is_block_sparse = (blocksparse_vert_stride > 1);
DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
CALL_V1_LAUNCHER_BLOCK_SIZE)
if(kv_cache_dtype != "auto"||query.dtype() == at::ScalarType::Float||is_block_sparse||
block_size!=16||query.size(2)!=128||get_device_name()!="gfx928"){
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);
}
else{
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_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); \
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);
#define LAUNCH_PAGED_ATTENTION_V2_TC(HEAD_SIZE) \
hipLaunchKernelGGL( \
(vllm::paged_attention_v2_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); \
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(
vllm::Fp8KVCacheDataType KV_DTYPE, bool IS_BLOCK_SPARSE, int PARTITION_SIZE = 512>
void paged_attention_v2_launcher_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, int num_kv_heads, float scale,
......@@ -958,8 +924,8 @@ void paged_attention_v2_launcher(
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);
// 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.
......@@ -977,40 +943,46 @@ void paged_attention_v2_launcher(
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>();
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
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);
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);
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;
REUSEKV_SWITCH_V2([&] {
int num_thread;
if(REUSE_KV_TIMES>1){
if(num_seqs<16)num_thread=256;
else if(max_num_partitions*num_seqs*num_heads/REUSE_KV_TIMES>4000)num_thread=64;
else num_thread=128;
}
else{
if(num_seqs<16&&max_num_partitions<10)num_thread=256;
else num_thread=64;
}
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 = REUSE_KV_TIMES*(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);
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<T, CACHE_T, BLOCK_SIZE, KV_DTYPE, \
paged_attention_v2_launcher_opt<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, \
......@@ -1046,7 +1018,7 @@ void paged_attention_v2_launcher(
break; \
}
void paged_attention_v2_opt(
void paged_attention_v2(
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]
......@@ -1063,13 +1035,44 @@ void paged_attention_v2_opt(
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 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, // [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 bool is_block_sparse = (blocksparse_vert_stride > 1);
DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
CALL_V2_LAUNCHER_BLOCK_SIZE)
if(kv_cache_dtype != "auto"||query.dtype() == at::ScalarType::Float||is_block_sparse||
block_size!=16||query.size(2)!=128||get_device_name()!="gfx928"){
paged_attention_v2(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);
}
else{
DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
CALL_V2_LAUNCHER_BLOCK_SIZE)
}
}
#undef WARP_SIZE
......
csrc/attention/static_switch.h
View file @ 0c70376b
......@@ -9,25 +9,17 @@
} \
}()
#define OPT_SWITCH(COND, ...) \
[&] { \
if (COND) { \
constexpr static int opt = 1; \
return __VA_ARGS__(); \
} else { \
constexpr static int opt = 2; \
return __VA_ARGS__(); \
} \
}()
#define NUM_THREADS_SWITCH(NUM_THREAD, ...) \
[&] { \
if (NUM_THREAD == 256) { \
constexpr static int NUM_THREADS = 256; \
return __VA_ARGS__(); \
} else { \
}else if (NUM_THREAD == 128) { \
constexpr static int NUM_THREADS = 128; \
return __VA_ARGS__(); \
} else { \
constexpr static int NUM_THREADS = 64; \
return __VA_ARGS__(); \
} \
}()
......@@ -45,12 +37,12 @@
} else if (HEADDIM == 112) { \
constexpr static int HEAD_SIZE = 112; \
return __VA_ARGS__(); \
} else if (HEADDIM == 120) { \
constexpr static int HEAD_SIZE = 120; \
return __VA_ARGS__(); \
} else if (HEADDIM == 128) { \
constexpr static int HEAD_SIZE = 128; \
return __VA_ARGS__(); \
} else if (HEADDIM == 192) { \
constexpr static int HEAD_SIZE = 192; \
return __VA_ARGS__(); \
} else if (HEADDIM == 256) { \
constexpr static int HEAD_SIZE = 256; \
return __VA_ARGS__(); \
......@@ -74,14 +66,48 @@
} \
}()
#define REUSEKV_SWITCH_V1(num_blocks , ...) \
#define REUSEKV_SWITCH_V2( ...) \
[&] { \
if (num_heads / num_kv_heads > 8 ){ \
constexpr static int REUSE_KV_TIMES = 16; \
return __VA_ARGS__(); \
}else if (num_heads / num_kv_heads > 4 ){ \
constexpr static int REUSE_KV_TIMES = 8; \
return __VA_ARGS__(); \
}else if (num_heads / num_kv_heads > 2 ){ \
constexpr static int REUSE_KV_TIMES = 4; \
return __VA_ARGS__(); \
} else { \
constexpr static int REUSE_KV_TIMES = 1; \
return __VA_ARGS__(); \
} \
}()
#define REUSEKV_SWITCH_V1( ...) \
[&] { \
if (num_heads > num_kv_heads && num_blocks >= 1200){ \
if (num_heads/num_kv_heads >4 && padded_max_seq_len<3900){ \
constexpr static int REUSE_KV_TIMES = 8; \
return __VA_ARGS__(); \
}else if (num_heads/num_kv_heads >2 && padded_max_seq_len<7800){ \
constexpr static int REUSE_KV_TIMES = 4; \
return __VA_ARGS__(); \
}else if (num_heads/num_kv_heads ==2 && padded_max_seq_len<15600){ \
constexpr static int REUSE_KV_TIMES = 2; \
return __VA_ARGS__(); \
} else { \
}else { \
constexpr static int REUSE_KV_TIMES = 1; \
return __VA_ARGS__(); \
} \
}()
#define USEVMAC_SWITCH_V1(num_blocks , ...) \
[&] { \
if (REUSE_KV_TIMES==1&&(num_blocks >2500 || padded_max_seq_len > 2048)){ \
constexpr static int use_vmac = false; \
return __VA_ARGS__(); \
} else { \
constexpr static int use_vmac = true; \
return __VA_ARGS__(); \
} \
}()
\ No newline at end of file
requirements-rocm.txt
View file @ 0c70376b
......@@ -14,4 +14,4 @@ torch == 2.3.0
triton == 2.1.0
flash_attn == 2.6.1
xformers == 0.0.25
lmslim == 0.1.0
\ No newline at end of file
lmslim == 0.1.1
\ No newline at end of file
vllm/attention/ops/paged_attn.py
View file @ 0c70376b
......@@ -124,8 +124,10 @@ class PagedAttention:
# to parallelize.
# TODO(woosuk): Tune this heuristic.
# For context len > 8192, use V2 kernel to avoid shared memory shortage.
use_v1 = (max_seq_len <= 8192
and (max_num_partitions == 1 or num_seqs * num_heads > 512))
# use_v1 = (max_seq_len <= 8192
# and (max_num_partitions == 1 or num_seqs * num_heads > 512))
use_v1 = (max_seq_len < 8192
and (max_seq_len<1000 or num_seqs * num_heads > (1024 if num_kv_heads < num_heads else 512)))
if use_v1:
# Run PagedAttention V1.
......
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