youngkingdom/vllm
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[FP8][Kernel] Dynamic kv cache scaling factors computation (#11906)

Browse Source 
Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
Co-authored-by: Micah Williamson <micah.williamson@amd.com>
This commit is contained in:
Gregory Shtrasberg
2025-01-23 13:04:03 -05:00
committed by GitHub
parent 6e650f56a1
commit e97f802b2d
60 changed files with 276 additions and 1365 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
csrc/attention/attention_kernels.cuh
View File

@ -105,7 +105,7 @@ __device__ void paged_attention_kernel(
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 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 seq_idx = blockIdx.y;
@ -285,7 +285,7 @@ __device__ void paged_attention_kernel(
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);
k_vec_quant, *k_scale);
}
}
@ -415,7 +415,7 @@ __device__ void paged_attention_kernel(
*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);
*v_scale);
}
if (block_idx == num_seq_blocks - 1) {
// NOTE(woosuk): When v_vec contains the tokens that are out of the
@ -513,7 +513,7 @@ __global__ void paged_attention_v1_kernel(
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 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<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
@ -549,7 +549,7 @@ __global__ void paged_attention_v2_kernel(
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 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<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,

17
csrc/attention/paged_attention_v1.cu
View File

@ -41,7 +41,7 @@
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, \
k_scale_ptr, v_scale_ptr, tp_rank, blocksparse_local_blocks, \
blocksparse_vert_stride, blocksparse_block_size, \
blocksparse_head_sliding_step);
@ -53,10 +53,10 @@ 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 std::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 std::optional<torch::Tensor>& alibi_slopes, torch::Tensor& k_scale,
torch::Tensor& 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) {
int num_seqs = query.size(0);
int num_heads = query.size(1);
int head_size = query.size(2);
@ -80,6 +80,8 @@ 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>();
const float* k_scale_ptr = reinterpret_cast<const float*>(k_scale.data_ptr());
const float* v_scale_ptr = reinterpret_cast<const float*>(v_scale.data_ptr());
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
int padded_max_seq_len =
@ -177,8 +179,9 @@ void paged_attention_v1(
torch::Tensor& seq_lens, // [num_seqs]
int64_t block_size, int64_t max_seq_len,
const std::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 std::string& kv_cache_dtype, torch::Tensor& k_scale,
torch::Tensor& 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);

17
csrc/attention/paged_attention_v2.cu
View File

@ -37,7 +37,7 @@
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, \
kv_block_stride, kv_head_stride, k_scale_ptr, v_scale_ptr, tp_rank, \
blocksparse_local_blocks, blocksparse_vert_stride, \
blocksparse_block_size, blocksparse_head_sliding_step); \
vllm::paged_attention_v2_reduce_kernel<T, HEAD_SIZE, NUM_THREADS, \
@ -54,10 +54,10 @@ void paged_attention_v2_launcher(
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 std::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 std::optional<torch::Tensor>& alibi_slopes, torch::Tensor& k_scale,
torch::Tensor& 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) {
int num_seqs = query.size(0);
int num_heads = query.size(1);
int head_size = query.size(2);
@ -84,6 +84,8 @@ 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>();
const float* k_scale_ptr = reinterpret_cast<const float*>(k_scale.data_ptr());
const float* v_scale_ptr = reinterpret_cast<const float*>(v_scale.data_ptr());
constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
int max_num_partitions = DIVIDE_ROUND_UP(max_seq_len, PARTITION_SIZE);
@ -188,8 +190,9 @@ void paged_attention_v2(
torch::Tensor& seq_lens, // [num_seqs]
int64_t block_size, int64_t max_seq_len,
const std::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 std::string& kv_cache_dtype, torch::Tensor& k_scale,
torch::Tensor& 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);