[Typos] Fix typos

csrc/gpu/cutlass_kernels/fp8_gemm_fused/dual_gemm/thread/left_gelu_and_mul.h

@@ -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;
 

csrc/gpu/cutlass_kernels/fp8_gemm_fused/dual_gemm/thread/left_silu_and_mul.h

@@ -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;
 

csrc/gpu/cutlass_kernels/fp8_gemm_fused/dual_gemm/threadblock/dual_mma_base.h

@@ -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);
 

csrc/gpu/dequant_int8.cu

@@ -158,7 +158,7 @@ std::vector<paddle::DataType> DequantInt8Dtype(const paddle::DataType& input_dty
 }
 
 PD_BUILD_OP(dequant_int8)
-    .Inputs({"intput","out_scale"})
+    .Inputs({"input","out_scale"})
     .Outputs({"output"})
     .Attrs({"dtype: std::string"})
     .SetKernelFn(PD_KERNEL(DequantInt8))

csrc/gpu/int8_gemm_with_cutlass/epilogue_tensor_op_int32.h

@@ -176,7 +176,7 @@ 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:

csrc/gpu/quant_int8.cu

@@ -271,7 +271,7 @@ std::vector<paddle::DataType> QuantInt8Dtype(const paddle::DataType& input_dtype
 }
 
 PD_BUILD_OP(quant_int8)
-    .Inputs({"intput", paddle::Optional("shift"),paddle::Optional("smooth") })
+    .Inputs({"input", paddle::Optional("shift"),paddle::Optional("smooth") })
     .Outputs({"output"})
     .Attrs({"scale: float","round_type: int","max_bound: float", "min_bound: float"})
     .SetKernelFn(PD_KERNEL(QuantInt8))

csrc/gpu/set_preids_token_penalty_multi_scores.cu

@@ -46,13 +46,13 @@ __global__ void set_preids_token_penalty_multi_scores_kernel(const bool *stop_fl
         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
 
         const int step_idx_now = step_idx[bi];
         if (tid == 0 && step_idx_now >= 0) {
             if (seq_len_enc > 0) { // encoder, get last token accord to seq_lens_encoder
                 pre_ids_now[step_idx_now] = input_ids_now[seq_len_enc - 1];
-            } else { // decoedr, get first token
+            } else { // decoder, get first token
                 pre_ids_now[step_idx_now] = input_ids_now[0];
             }
         }

csrc/gpu/set_value_by_flags_v2.cu

@@ -29,11 +29,11 @@ __global__ void set_value_by_flag_and_id_v2(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_dec == 0) { // encoder, get last token accord to seq_lens_encoder
                 pre_ids_all_now[step_idx[tid]] = input_ids_now[seq_len_enc - 1];
-            } else { // decoedr, get first token
+            } else { // decoder, get first token
                 pre_ids_all_now[step_idx[tid]] = input_ids_now[0];
             }
         }

csrc/gpu/speculate_decoding_kernels/speculate_set_value_by_flags.cu

@@ -30,7 +30,7 @@ __global__ void speculate_set_value_by_flag_and_id(int64_t *pre_ids_all,
         const int64_t *accept_tokens_now = 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
         if (step_idx[tid] >= 0) {
             for (int i = 0; i < accept_num[tid]; i++) {
                 pre_ids_all_now[step_idx[tid] - i] = accept_tokens_now[accept_num[tid] - 1 - i];