apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-12-24 13:28:13 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

fix typos (#3684)

Browse Source
This commit is contained in:
co63oc
2025-09-01 17:50:17 +08:00
committed by GitHub
parent 0513a78ecc
commit d6369b4d51
67 changed files with 85 additions and 85 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
custom_ops/gpu_ops/cpp_extensions.cc
View File

@@ -980,7 +980,7 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
m.def("per_token_quant_padding", &PerTokenQuantPadding, py::arg("input"),
py::arg("block_size"),
"per token per block quant and padding tranpose scale");
"per token per block quant and padding transpose scale");
m.def("masked_per_token_quant", &MaskedPerTokenQuant, py::arg("input"),
py::arg("recv_expert_count"), py::arg("block_size"),

4
custom_ops/gpu_ops/cutlass_extensions/gemm/threadblock/wint2x_mma_base.h
View File

@@ -89,11 +89,11 @@ public:
GemmShape<Shape::kM / WarpGemm::kM, Shape::kN / WarpGemm::kN,
Shape::kK / WarpGemm::kK>;
/// Number of warp-level GEMM oeprations
/// Number of warp-level GEMM operations
static int const kWarpGemmIterations =
(WarpGemm::kK / Operator::Policy::MmaShape::kK);
/// Number of warp-level GEMM oeprations per load for B
/// Number of warp-level GEMM operations per load for B
static constexpr int kWarpGemmIterationsPerLoadForB =
Operator::IteratorB::InstructionShape::kRow / Operator::InstructionShape::kK;
static_assert(!(kWarpGemmIterations % kWarpGemmIterationsPerLoadForB), "");

2
custom_ops/gpu_ops/cutlass_kernels/fp8_gemm_fused/dual_gemm/thread/left_gelu_and_mul.h
View File

@@ -117,7 +117,7 @@ class LeftGELUAndMul {
CUTLASS_HOST_DEVICE
FragmentOutput operator()(FragmentAccumulator const &lhs,
FragmentAccumulator const &rhs) const {
// Convert source to interal compute numeric type
// Convert source to internal compute numeric type
NumericArrayConverter<ElementCompute, ElementAccumulator, kCount, Round>
accumulator_to_compute;

2
custom_ops/gpu_ops/cutlass_kernels/fp8_gemm_fused/dual_gemm/thread/left_silu_and_mul.h
View File

@@ -117,7 +117,7 @@ class LeftSiLUAndMul {
CUTLASS_HOST_DEVICE
FragmentOutput operator()(FragmentAccumulator const &lhs,
FragmentAccumulator const &rhs) const {
// Convert source to interal compute numeric type
// Convert source to internal compute numeric type
NumericArrayConverter<ElementCompute, ElementAccumulator, kCount, Round>
accumulator_to_compute;

2
custom_ops/gpu_ops/cutlass_kernels/fp8_gemm_fused/dual_gemm/threadblock/dual_mma_base.h
View File

@@ -92,7 +92,7 @@ class DualMmaBase {
Shape::kN / WarpGemm::kN,
Shape::kK / WarpGemm::kK>;
/// Number of warp-level GEMM oeprations
/// Number of warp-level GEMM operations
static int const kWarpGemmIterations =
(WarpGemm::kK / Operator0::Policy::MmaShape::kK);

2
custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/cutlass_extensions/epilogue/threadblock/epilogue_per_row_per_col_scale_nf4.h
View File

@@ -219,7 +219,7 @@ class EpilogueVisitorPerRowPerColNf4 {
iterator_C_.clear_mask();
}
// NOTE(wangbojun) Currently, this kernel don't hanve implantention for
// adding elementwise beta, we keep this here for future useage beta_ =
// adding elementwise beta, we keep this here for future usage beta_ =
// (params.elementwise.beta_ptr ? *params.elementwise.beta_ptr :
// params.elementwise.beta); if (beta_ == ElementAccumulator()) {
// iterator_C_.clear_mask();

2
custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/cutlass_extensions/epilogue/threadblock/epilogue_tensor_op_int32.h
View File

@@ -176,7 +176,7 @@ struct Nf4DefaultIteratorsTensorOp<cutlass::bfloat16_t,
///
/// Satisfies: ReadableTileIterator
///
template <typename ThreadMap_ ///< Thread map (conept: OutputTileThreadMap)
template <typename ThreadMap_ ///< Thread map (concept: OutputTileThreadMap)
>
class SharedLoadIteratorMixed<ThreadMap_, int32_t, 32, 16, 8, 8> {
public:

2
custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/cutlass_extensions/gemm/threadblock/default_mma_nf4_int8_interleaved.h
View File

@@ -64,7 +64,7 @@ template <
typename InstructionShape_,
/// Number of stages used in the pipelined mainloop
int Stages,
/// Operation perfomed by GEMM
/// Operation performed by GEMM
typename Operator,
/// Store the accumulators in row major or column major. Row major is used
/// when output layout is interleaved.

2
custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/cutlass_extensions/gemm/threadblock/int8_mma_base.h
View File

@@ -133,7 +133,7 @@ public:
/// Shape describing the number of warps filling the CTA
using WarpCount = GemmShape<Shape::kM / WarpGemm::kM, Shape::kN / WarpGemm::kN, Shape::kK / WarpGemm::kK>;
/// Number of warp-level GEMM oeprations
/// Number of warp-level GEMM operations
static int const kWarpGemmIterations = (WarpGemm::kK / Operator::Policy::MmaShape::kK);
static_assert(Operator::IteratorB::InstructionShape::kRow>=Operator::InstructionShape::kK,"");
static constexpr int kNumKIterationsPerWarpBLoad =

2
custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/cutlass_extensions/gemm/threadblock/int8_mma_multistage.h
View File

@@ -509,7 +509,7 @@ public:
this->warp_tile_iterator_B_.load(warp_frag_B[(warp_tileB_k_load_offset + 1) % 2]);
++this->warp_tile_iterator_B_;
}
// TOOD(wangbojun) lds_converter can be remove for int8 B input
// TODO(wangbojun) lds_converter can be remove for int8 B input
typename TransformBAfterLDS::result_type converted_frag_B =
lds_converter(warp_frag_B[warp_tileB_k_load_offset % 2]);

2
custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/cutlass_extensions/gemm/threadblock/nf4_int8_mma_base.h
View File

@@ -96,7 +96,7 @@ public:
/// Shape describing the number of warps filling the CTA
using WarpCount = GemmShape<Shape::kM / WarpGemm::kM, Shape::kN / WarpGemm::kN, Shape::kK / WarpGemm::kK>;
/// Number of warp-level GEMM oeprations
/// Number of warp-level GEMM operations
static int const kWarpGemmIterations = (WarpGemm::kK / Operator::Policy::MmaShape::kK);
static_assert(Operator::IteratorB::InstructionShape::kRow>=Operator::InstructionShape::kK,"");
static constexpr int kNumKIterationsPerWarpBLoad =

2
custom_ops/gpu_ops/cutlass_kernels/w4a8_moe/cutlass_extensions/gemm/threadblock/nf4_int8_mma_multistage.h
View File

@@ -646,7 +646,7 @@ public:
// );
// }
}
// TOOD(wangbojun) lds_converter can be remove for int8 B input
// TODO(wangbojun) lds_converter can be remove for int8 B input
// int4
// typename TransformBAfterLDS::result_type converted_frag_B =
// lds_converter(warp_frag_B[warp_tileB_k_load_offset % 2]);

2
custom_ops/gpu_ops/int8_gemm_with_cutlass/epilogue_tensor_op_int32.h
View File

@@ -171,7 +171,7 @@ struct DefaultIteratorsTensorOp<cutlass::bfloat16_t,
///
/// Satisfies: ReadableTileIterator
///
template <typename ThreadMap_ ///< Thread map (conept: OutputTileThreadMap)
template <typename ThreadMap_ ///< Thread map (concept: OutputTileThreadMap)
>
class SharedLoadIteratorMixed<ThreadMap_, int32_t, 32, 16, 8, 8> {
public:

2
custom_ops/gpu_ops/moe/moe_dispatch.cu
View File

@@ -80,7 +80,7 @@ void MoeDispatchKernel(
if (group_moe) {
paddle::Tensor softmax_max_prob_tensor =
GetEmptyTensor({num_rows, moe_topk}, paddle::DataType::FLOAT32, place);
// (TODO: check fill sucess ?)
// (TODO: check fill success ?)
paddle::experimental::fill(softmax_max_prob_tensor, 0.f);
softmax_max_prob = softmax_max_prob_tensor.data<float>();
}

2
custom_ops/gpu_ops/save_output_msg_with_topk.cc
View File

@@ -75,7 +75,7 @@ void SaveOutMmsgTopK(const paddle::Tensor& x,
std::string inference_msg_id_env_str(inference_msg_id_env_p);
inference_msg_id_from_env = std::stoi(inference_msg_id_env_str);
if (inference_msg_id_from_env == 2) {
// 2 and -2 is perserve for no-output indication.
// 2 and -2 is preserve for no-output indication.
throw std::runtime_error(
" INFERENCE_MSG_ID cannot be 2, please use other number.");
}

2
custom_ops/gpu_ops/save_with_output_msg.cc
View File

@@ -45,7 +45,7 @@ void save_kernel(const paddle::Tensor& x,
std::string inference_msg_id_env_str(inference_msg_id_env_p);
inference_msg_id_from_env = std::stoi(inference_msg_id_env_str);
if (inference_msg_id_from_env == 2) {
// 2 and -2 is perserve for no-output indication.
// 2 and -2 is preserve for no-output indication.
throw std::runtime_error(
" INFERENCE_MSG_ID cannot be 2, please use other number.");
}

2
custom_ops/gpu_ops/set_value_by_flags.cu
View File

@@ -34,7 +34,7 @@ __global__ void set_value_by_flag_and_id(const bool *stop_flags,
const int64_t *input_ids_now = input_ids + tid * length_input_ids;
const int seq_len_dec = seq_lens_decoder[tid];
const int seq_len_enc = seq_lens_encoder[tid];
if (seq_len_dec == 0 && seq_len_enc == 0) return; // stoped
if (seq_len_dec == 0 && seq_len_enc == 0) return; // stopped
if (step_idx[tid] >= 0) {
if (seq_len_enc > 0) { // encoder, get last token accord to seq_lens_encoder
pre_ids_all_now[step_idx[tid]] = input_ids_now[seq_len_enc - 1];

2
custom_ops/gpu_ops/speculate_decoding/draft_model/eagle_get_base_model_hidden_states.cu
View File

@@ -63,7 +63,7 @@ __global__ void ComputeOrderKernel(
position_map[in_offset++] = out_offset++;
}
in_offset += cur_base_model_seq_lens_this_time - accept_num;
// (liuzichang): Temperary Reserved for debug
// (liuzichang): Temporary Reserved for debug
// if (accept_num <= actual_draft_token_num) /*Accept partial draft tokens*/ {
// #ifdef DEBUG_EAGLE_KERNEL
// printf("batch %d: accept_num <= actual_draft_token_num \n", i);

2
custom_ops/gpu_ops/speculate_decoding/speculate_set_value_by_flags.cu
View File

@@ -35,7 +35,7 @@ __global__ void speculate_set_value_by_flag_and_id(int64_t *pre_ids_all,
accept_tokens + tid * max_draft_tokens;
const int seq_len_dec = seq_lens_decoder[tid];
const int seq_len_enc = seq_lens_encoder[tid];
if (seq_len_dec == 0 && seq_len_enc == 0) return; // stoped
if (seq_len_dec == 0 && seq_len_enc == 0) return; // stopped
// printf("step_idx[tid] %d\n", step_idx[tid]);
if (step_idx[tid] >= 0) {
for (int i = 0; i < accept_num[tid]; i++) {

2
custom_ops/gpu_ops/speculate_decoding/speculate_step_reschedule.cu
View File

@@ -295,7 +295,7 @@ void SpeculateStepSchedule(const paddle::Tensor &stop_flags,
std::string inference_msg_id_env_str(inference_msg_id_env_p);
inference_msg_id_from_env = std::stoi(inference_msg_id_env_str);
if (inference_msg_id_from_env == 2) {
// 2 and -2 is perserve for no-output indication.
// 2 and -2 is preserve for no-output indication.
throw std::runtime_error(
" INFERENCE_MSG_ID cannot be 2, please use other number.");
}

2
custom_ops/gpu_ops/step_reschedule.cu
View File

@@ -283,7 +283,7 @@ void Schedule(const paddle::Tensor &stop_flags,
std::string inference_msg_id_env_str(inference_msg_id_env_p);
inference_msg_id_from_env = std::stoi(inference_msg_id_env_str);
if (inference_msg_id_from_env == 2) {
// 2 and -2 is perserve for no-output indication.
// 2 and -2 is preserve for no-output indication.
throw std::runtime_error(
" INFERENCE_MSG_ID cannot be 2, please use other number.");
}

2
custom_ops/gpu_ops/token_transfer.hpp
View File

@@ -58,7 +58,7 @@ class TokenTransfer {
}
// once copy: cpu --> cpu
// arrary length should be (1 + MAX_BATCH)
// array length should be (1 + MAX_BATCH)
bool GetBatchToken(int64_t *array) {
if (Empty()) {
return false;

6
custom_ops/gpu_ops/update_split_fuse_input.cu
View File

@@ -75,10 +75,10 @@ void UpdateSplitFuseInputes(const paddle::Tensor& split_fuse_seq_lens,
const int max_seq_len,
const int max_batch_size,
const int split_fuse_size) {
dim3 girds;
girds.x = max_batch_size;
dim3 grids;
grids.x = max_batch_size;
const int block_size = 128;
update_split_fuse_inputs_kernel<<<girds,
update_split_fuse_inputs_kernel<<<grids,
block_size,
0,
input_ids.stream()>>>(