apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-09-26 20:41:53 +08:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

fix typos (#3951)

Browse Source
This commit is contained in:
co63oc
2025-09-08 15:22:41 +08:00
committed by GitHub
parent 2033450391
commit aadd6a94d8
7 changed files with 26 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
custom_ops/cpu_ops/set_value_by_flags.cc
View File

@@ -14,7 +14,7 @@
#include "paddle/extension.h"
void set_value_by_flag_and_id(const bool *stop_flags,
void set_value_by_flags_and_idx(const bool *stop_flags,
int64_t *pre_ids_all,
const int64_t *input_ids,
const int *seq_lens_encoder,
@@ -50,7 +50,7 @@ void SetValueByFlagsAndIdx(const paddle::Tensor &pre_ids_all,
int length = pre_ids_all_shape[1];
int length_input_ids = input_ids.shape()[1];
set_value_by_flag_and_id(stop_flags.data<bool>(),
set_value_by_flags_and_idx(stop_flags.data<bool>(),
const_cast<int64_t *>(pre_ids_all.data<int64_t>()),
input_ids.data<int64_t>(),
seq_lens_encoder.data<int>(),

4
custom_ops/cpu_ops/update_inputs.cc
View File

@@ -46,7 +46,7 @@ void update_inputs_kernel(bool *not_need_stop,
not_need_stop[0] = stop_sum < stop_nums[0];
}
void UpdateInputes(const paddle::Tensor &stop_flags,
void UpdateInputs(const paddle::Tensor &stop_flags,
const paddle::Tensor &not_need_stop,
const paddle::Tensor &seq_lens_this_time,
const paddle::Tensor &seq_lens_encoder,
@@ -90,4 +90,4 @@ PD_BUILD_STATIC_OP(update_inputs_cpu)
{"seq_lens_encoder", "seq_lens_encoder_out"},
{"seq_lens_decoder", "seq_lens_decoder_out"},
{"input_ids", "input_ids_out"}})
.SetKernelFn(PD_KERNEL(UpdateInputes));
.SetKernelFn(PD_KERNEL(UpdateInputs));

4
custom_ops/gpu_ops/append_attn/encoder_write_cache_with_rope_impl.cuh
View File

@@ -449,8 +449,8 @@ __global__ void GQAVariableLengthRotaryQKNormKernel(
const int half_lastdim = last_dim / 2;
const int offset = (q_num_head + kv_num_head) * last_dim;
const int all_head_num = elem_cnt / last_dim;
for (int gloabl_hi = global_warp_idx; gloabl_hi < all_head_num; gloabl_hi += all_warp_num) {
int64_t linear_index = gloabl_hi * last_dim + threadIdx.x * VecSize;
for (int global_hi = global_warp_idx; global_hi < all_head_num; global_hi += all_warp_num) {
int64_t linear_index = global_hi * last_dim + threadIdx.x * VecSize;
const int token_idx = linear_index / offset;
const int ori_bi = batch_id_per_token[token_idx];
if (seq_lens[ori_bi] == 0) continue;

24
custom_ops/gpu_ops/append_attn/gqa_rope_write_cache.cu
View File

@@ -217,7 +217,7 @@ __global__ void append_cache_kv_c16(
// load k_smem 64 rows 128 cols
for (int fz = 0; fz < 4; fz++) { // 4 rows pre warp once, 16 rows all 4 warps once, need 4 iter
for (int fy = 0; fy < 2; fy++) { // 8 * 128b = 64 * bf16 noce, need 2 iter
for (int fy = 0; fy < 2; fy++) { // 8 * 128b = 64 * bf16 once, need 2 iter
k_smem.load_128b_async<SharedMemFillMode::kNoFill>(
k_smem_offset_w, cur_cache_k + k_read_idx, end_idx > 0);
k_smem_offset_w =
@@ -235,7 +235,7 @@ __global__ void append_cache_kv_c16(
// deal k_smem 64 rows 128 cols
for (int fz = 0; fz < 1; fz++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 1 iter
uint32_t row_idx = wid * 16 + tid / 4;
for (int fy = 0; fy < 8; fy++) { // 2 * 128b = 16 * bf16 noce, need 8 iter
for (int fy = 0; fy < 8; fy++) { // 2 * 128b = 16 * bf16 once, need 8 iter
uint32_t col_idx = fy * 16 + tid % 4 * 2;
k_smem.ldmatrix_m8n8x4(k_smem_offset_r, kv_frag);
// layout
@@ -278,7 +278,7 @@ __global__ void append_cache_kv_c16(
// load v_smem 64 rows 128 cols
for (int fz = 0; fz < 4; fz++) { // // 4 rows pre warp once, 16 rows all 4 warps once, need 4 iter
for (int fy = 0; fy < 2; fy++) { // 8 * 128b = 64 * bf16 noce, need 2 iter
for (int fy = 0; fy < 2; fy++) { // 8 * 128b = 64 * bf16 once, need 2 iter
v_smem.load_128b_async<SharedMemFillMode::kNoFill>(
v_smem_offset_w, cur_cache_v + v_read_idx, end_idx > 0);
v_smem_offset_w =
@@ -296,7 +296,7 @@ __global__ void append_cache_kv_c16(
// deal v_smem 64 rows 128 cols
for (int fz = 0; fz < 1; fz++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 1 iter
uint32_t row_idx = wid * 16 + tid / 4;
for (int fy = 0; fy < 8; fy++) { // 2 * 128b = 16 * bf16 noce, need 8 iter
for (int fy = 0; fy < 8; fy++) { // 2 * 128b = 16 * bf16 once, need 8 iter
uint32_t col_idx = fy * 16 + tid % 4 * 2;
v_smem.ldmatrix_m8n8x4(v_smem_offset_r, kv_frag);
// layout
@@ -400,7 +400,7 @@ __global__ void append_cache_kv_c8(
// load v_smem 64 rows, 128 cols
for (int fz = 0; fz < 4; fz++) { // 4 rows pre warp once, 16 rows all 4 warps once, need 4 iter
for (int fy = 0; fy < 1; fy++) { // 8 * 128b = 128 * uint8 noce, need 1 iter
for (int fy = 0; fy < 1; fy++) { // 8 * 128b = 128 * uint8 once, need 1 iter
k_smem.load_128b_async<SharedMemFillMode::kNoFill>(
k_smem_offset_w, cur_cache_k + k_read_idx, end_idx > 0);
k_smem_offset_w =
@@ -418,7 +418,7 @@ __global__ void append_cache_kv_c8(
// deal k_smem 64 rows, 128 cols
for (int fz = 0; fz < 1; fz++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 1 iter
uint32_t row_idx = wid * 16 + tid / 4;
for (int fy = 0; fy < 4; fy++) { // 2 * 128b = 32 * uint8 noce, need 4 iter
for (int fy = 0; fy < 4; fy++) { // 2 * 128b = 32 * uint8 once, need 4 iter
uint32_t col_idx = fy * 32 + tid % 4 * 2;
k_smem.ldmatrix_m8n8x4(k_smem_offset_r, k_frag);
// layout
@@ -466,7 +466,7 @@ __global__ void append_cache_kv_c8(
tid % 4 * num_elems_per_128b<CacheT>();
// load v_smem 128 rows 64 cols
for (int fy = 0; fy < 4; fy++) { // 8 rows pre warp once, 32 rows all 4 warps once, need 4 iter
for (int fz = 0; fz < 1; fz++) { // 4 * 128b = 64 * uint8 noce, need 1 iter
for (int fz = 0; fz < 1; fz++) { // 4 * 128b = 64 * uint8 once, need 1 iter
v_smem.load_128b_async<SharedMemFillMode::kNoFill>(
v_smem_offset_w, cur_cache_v + v_read_idx, end_idx > 0);
v_smem_offset_w =
@@ -485,7 +485,7 @@ __global__ void append_cache_kv_c8(
// deal v_smem 128 rows 64 cols
for (int fy = 0; fy < 2; fy++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 2 iter
uint32_t dim_idx = fy * NUM_WARPS * 16 + wid * 16 + tid / 4;
for (int fz = 0; fz < 2; fz++) { // 2 * 128b = 32 * uint8 noce, need 2 iter
for (int fz = 0; fz < 2; fz++) { // 2 * 128b = 32 * uint8 once, need 2 iter
uint32_t kv_idx = fz * 32 + tid % 4 * 2;
v_smem.ldmatrix_m8n8x4(v_smem_offset_r, v_frag);
// layout
@@ -614,7 +614,7 @@ __global__ void append_cache_kv_c4(
// load k_smem 64 rows 128 cols
for (int fz = 0; fz < 2; fz++) { // 4 rows pre warp once, 16 rows all 4 warps once, need 4 iter
for (int fy = 0; fy < 1; fy++) { // 4 * 128b = 128 * int4 noce, need 1 iter
for (int fy = 0; fy < 1; fy++) { // 4 * 128b = 128 * int4 once, need 1 iter
k_smem.load_128b_async<SharedMemFillMode::kNoFill>(
k_smem_offset_w, cur_cache_k + k_read_idx, end_idx > 0);
k_smem_offset_w =
@@ -632,7 +632,7 @@ __global__ void append_cache_kv_c4(
// deal k_smem 64 rows 128 cols
for (int fz = 0; fz < 1; fz++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 1 iter
uint32_t row_idx = wid * 16 + tid / 4;
for (int fy = 0; fy < 2; fy++) { // 2 * 128b = 64 * int4 noce, need 2 iter
for (int fy = 0; fy < 2; fy++) { // 2 * 128b = 64 * int4 once, need 2 iter
uint32_t col_idx = fy * 64 + tid % 4 * 2;
k_smem.ldmatrix_m8n8x4(k_smem_offset_r, k_frag);
@@ -685,7 +685,7 @@ __global__ void append_cache_kv_c4(
tid % 2 * num_elems_per_128b<CacheT>();
// load v_smem 128 rows 64 rows
for (int fy = 0; fy < 2; fy++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 2 iter
for (int fz = 0; fz < 1; fz++) { // 2 * 128b = 64 * int4 noce, need 1 iter
for (int fz = 0; fz < 1; fz++) { // 2 * 128b = 64 * int4 once, need 1 iter
v_smem.load_128b_async<SharedMemFillMode::kNoFill>(
v_smem_offset_w, cur_cache_v + v_read_idx, end_idx > 0);
v_smem_offset_w =
@@ -704,7 +704,7 @@ __global__ void append_cache_kv_c4(
// deal v_smem 128 rows 64 cols
for (int fy = 0; fy < 2; fy++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 2 iter
uint32_t dim_idx = fy * NUM_WARPS * 16 + wid * 16 + tid / 4;
for (int fz = 0; fz < 1; fz++) { // 2 * 128b = 64 * int4 noce, need 1 iter
for (int fz = 0; fz < 1; fz++) { // 2 * 128b = 64 * int4 once, need 1 iter
uint32_t kv_idx = fz * 64 + tid % 4 * 2;
v_smem.ldmatrix_m8n8x4(v_smem_offset_r, v_frag);
// layout

4
custom_ops/gpu_ops/moba_attn/moba_decoder_attn/moba_decoder_attn.cu
View File

@@ -383,7 +383,7 @@ __global__ __launch_bounds__(Kernel_traits::kNThreads) void moba_decoder_attenti
template<typename Kernel_traits, typename ParamType>
inline __device__ float caluate_logit_scale(const int partition_num, const int pack_max_partition_num, ParamType &params, char * shared_mem, const int seq_len, const int *qk_gate_topk_idx_ptr) {
inline __device__ float calculate_logit_scale(const int partition_num, const int pack_max_partition_num, ParamType &params, char * shared_mem, const int seq_len, const int *qk_gate_topk_idx_ptr) {
constexpr int32_t kNFloatPacksize = 16 / sizeof(float);
constexpr int32_t kNReduceThreads = Kernel_traits::kNReduceThreads;
const int32_t bi = blockIdx.z;
@@ -524,7 +524,7 @@ __global__ void __launch_bounds__(Kernel_traits::kNReduceThreads) moba_decoder_a
const int kv_head_idx = head_idx / Kernel_traits::kGqaGroupSize;
const int * qk_gate_topk_idx_ptr = params.qk_gate_topk_idx_ptr + (bi * params.kv_head_num + kv_head_idx) * Kernel_traits::kMaxN;
float inv_global_exp_sum = caluate_logit_scale<Kernel_traits>(partition_num, pack_max_partition_num, params, shared_mem, seq_len, qk_gate_topk_idx_ptr);
float inv_global_exp_sum = calculate_logit_scale<Kernel_traits>(partition_num, pack_max_partition_num, params, shared_mem, seq_len, qk_gate_topk_idx_ptr);
using T_vec = Vec<cuteType, kNReducePacksize>;

6
custom_ops/gpu_ops/moba_attn/moba_encoder_attn/moba_encoder_write_cache.cu
View File

@@ -40,7 +40,7 @@ __global__ void write_encoder_cachekv_c16(
if (seq_len == 0) return;
const int ramian_tokens = seq_len - block_idx;
const int remain_tokens = seq_len - block_idx;
const int32_t *block_table_now = block_tables + bidb * max_blocks_per_seq;
const uint32_t physical_block_number = block_table_now[blockIdx.x + seq_len_decoder[bidb] / kBlockSize];
@@ -51,7 +51,7 @@ __global__ void write_encoder_cachekv_c16(
#pragma unroll
for (int i = row_idx; i < kBlockSize; i += 128 / (kHeadDim / kPackSize)) {
if (i < ramian_tokens) {
if (i < remain_tokens) {
*reinterpret_cast<float4*>(cache + i * kHeadDim) = *reinterpret_cast<const float4*>(k_input + base_load_idx + i * kv_head_num * kHeadDim);
}
}
@@ -62,7 +62,7 @@ __global__ void write_encoder_cachekv_c16(
#pragma unroll
for (int i = row_idx; i < kBlockSize; i += 128 / (kHeadDim / kPackSize)) {
if (i < ramian_tokens) {
if (i < remain_tokens) {
*reinterpret_cast<float4*>(cache + i * kHeadDim) = *reinterpret_cast<const float4*>(v_input + base_load_idx + i * kv_head_num * kHeadDim);
}
}

6
custom_ops/gpu_ops/moba_attn/moba_process/moba_get_kv_from_cache.cu
View File

@@ -50,14 +50,14 @@ __global__ void get_kv_from_cache_c16_kernel(
const int physical_block_number = block_tables[bidb * max_blocks_per_seq + block_idx];
const int ramian_tokens = seq_len - base_token_idx;
const int remain_tokens = seq_len - base_token_idx;
if (bidh < kv_head_num) {
const int cache_offset = physical_block_number * kv_head_num * kBlockSize * kHeadDim + bidh * kBlockSize * kHeadDim + col_idx;
const int base_store_idx = (base_token_idx + cu_seq_k[bidb]) * kv_head_num * kHeadDim + bidh * kHeadDim + col_idx;
#pragma unroll
for (int i = row_idx; i < kBlockSize; i += 128 / (kHeadDim / kPackSize)) {
if (i < ramian_tokens) {
if (i < remain_tokens) {
*reinterpret_cast<float4*>(k_input + base_store_idx + i * kv_head_num * kHeadDim) = *reinterpret_cast<const float4*>(cache_k + cache_offset + i * kHeadDim);
}
}
@@ -67,7 +67,7 @@ __global__ void get_kv_from_cache_c16_kernel(
const int base_store_idx = (base_token_idx + cu_seq_k[bidb]) * kv_head_num * kHeadDim + bidh * kHeadDim + col_idx;
#pragma unroll
for (int i = row_idx; i < kBlockSize; i += 128 / (kHeadDim / kPackSize)) {
if (i < ramian_tokens) {
if (i < remain_tokens) {
*reinterpret_cast<float4*>(v_input + base_store_idx + i * kv_head_num * kHeadDim) = *reinterpret_cast<const float4*>(cache_v + cache_offset + i * kHeadDim);
}
}