Add lint action (#32)

* temp

* fix lint

* csrc->src

* remove clang-format

* skip .rst

* skip doc

* clang-format

version

version

* mat_B

src/fastertransformer/models/llama/LlamaCacheManager.cc

@@ -189,4 +189,4 @@ bool LlamaCacheManager::contains(uint64_t id) const noexcept
     return it != device_cache_.end();
 }
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaCacheManager.h

@@ -99,4 +99,4 @@ private:
     std::vector<Sequence> device_cache_;
 };
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaContextAttentionLayer.cc

@@ -199,7 +199,6 @@ inline void LlamaContextAttentionLayer<T>::forward(TensorMap*
                         quant_policy_,
                         weights->past_kv_scale.data());
 
-
     sync_check_cuda_error();
     if (use_fmha_) {
         fusedMultiHeadAttention(k_cache_ptrs,
@@ -226,8 +225,6 @@ inline void LlamaContextAttentionLayer<T>::forward(TensorMap*
                                   max_seq_len,
                                   quant_policy_,
                                   weights->past_kv_scale.data());
-
-        
     }
 
     //////////////////////////////////////////////
@@ -240,8 +237,6 @@ inline void LlamaContextAttentionLayer<T>::forward(TensorMap*
         sync_check_cuda_error();
     }
 
- 
-
     if (is_free_buffer_after_forward_ == true) {
         freeBuffer();
     }
@@ -302,19 +297,19 @@ void LlamaContextAttentionLayer<T>::fusedMultiHeadAttention(T**    key_cache_ptr
 }
 
 template<typename T>
-void LlamaContextAttentionLayer<T>::unfusedMultiHeadAttention(T**        key_cache_ptrs,
-                                                              T**        val_cache_ptrs,
-                                                              size_t     cache_layer_offset,
-                                                              const T*   attention_mask,
-                                                              const int* padding_offset,
-                                                              const int* context_length,
-                                                              int        batch_size,
-                                                              int        num_token,
-                                                              int        max_q_len,
-                                                              int        max_k_len,
-                                                              int        max_seq_len,
-                                                              int           quant,
-                                                              const float*     kv_scale)
+void LlamaContextAttentionLayer<T>::unfusedMultiHeadAttention(T**          key_cache_ptrs,
+                                                              T**          val_cache_ptrs,
+                                                              size_t       cache_layer_offset,
+                                                              const T*     attention_mask,
+                                                              const int*   padding_offset,
+                                                              const int*   context_length,
+                                                              int          batch_size,
+                                                              int          num_token,
+                                                              int          max_q_len,
+                                                              int          max_k_len,
+                                                              int          max_seq_len,
+                                                              int          quant,
+                                                              const float* kv_scale)
 {
     // key_cache [B, H, S[:t+s], D/x, x] -> [B, H, t+s, D]
     // val_cache [B, H, S[:t+s], D/x, x] -> [B, H, t+s, D]
@@ -408,4 +403,4 @@ void LlamaContextAttentionLayer<T>::unfusedMultiHeadAttention(T**        key_cac
 template class LlamaContextAttentionLayer<float>;
 template class LlamaContextAttentionLayer<half>;
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaContextAttentionLayer.h

@@ -15,8 +15,9 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
- 
-// Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/layers/attention_layers/GptContextAttentionLayer.h
+
+// Modified from
+// https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/layers/attention_layers/GptContextAttentionLayer.h
 
 #pragma once
 
@@ -100,7 +101,7 @@ private:
     const bool neox_rotary_style_;
 
     const bool use_fmha_;
-    const int quant_policy_;
+    const int  quant_policy_;
 
     NcclParam tensor_para_;
 
@@ -123,4 +124,4 @@ private:
     bool is_allocate_buffer_ = false;
 };
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaContextDecoder.cc

@@ -283,4 +283,4 @@ void LlamaContextDecoder<T>::forward(std::unordered_map<std::string, Tensor>*
 template class LlamaContextDecoder<float>;
 template class LlamaContextDecoder<half>;
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaContextDecoder.h

@@ -15,7 +15,8 @@
  * limitations under the License.
  */
 
-// Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/models/multi_gpu_gpt/ParallelGptContextDecoder.h
+// Modified from
+// https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/models/multi_gpu_gpt/ParallelGptContextDecoder.h
 
 #pragma once
 
@@ -111,4 +112,4 @@ public:
                          const std::vector<LlamaDecoderLayerWeight<T>*>* decoder_layer_weights);
 };
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaDecoder.h

@@ -16,8 +16,8 @@
  * limitations under the License.
  */
 
- // Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/models/multi_gpu_gpt/ParallelGptDecoder.h
-
+// Modified from
+// https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/models/multi_gpu_gpt/ParallelGptDecoder.h
 
 #include "src/fastertransformer/layers/BaseLayer.h"
 // #include "src/fastertransformer/layers/FfnLayer.h"
@@ -81,9 +81,8 @@ public:
                  IAllocator*      allocator,
                  bool             is_free_buffer_after_forward,
                  int              quant_policy),
-                 
 
-    ~LlamaDecoder() override;
+        ~LlamaDecoder() override;
 
     virtual void forward(std::unordered_map<std::string, Tensor>*        output_tensors,
                          const std::unordered_map<std::string, Tensor>*  input_tensors,
@@ -94,4 +93,4 @@ public:
                          const std::vector<LlamaDecoderLayerWeight<T>*>* decoder_layer_weights);
 };
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaDecoderLayerWeight.cc

@@ -161,12 +161,13 @@ void LlamaDecoderLayerWeight<T>::loadModel(std::string dir_path, FtCudaDataType
 
     // load kv_cache quant scale
     // if file not exist, get empty vector
-    std::string scale_path = dir_path + ".past_kv_scale." + rank_spec + ".weight";
-    std::ifstream  in(scale_path, std::ios::in);
+    std::string   scale_path = dir_path + ".past_kv_scale." + rank_spec + ".weight";
+    std::ifstream in(scale_path, std::ios::in);
     if (in.is_open()) {
         in.close();
         self_attn_weights.past_kv_scale = loadArrayFromBin({2}, scale_path);
-    } else {
+    }
+    else {
         self_attn_weights.past_kv_scale = {};
     }
 }

src/fastertransformer/models/llama/LlamaDecoderLayerWeight.h

@@ -35,7 +35,7 @@ public:
                             size_t     tensor_para_size,
                             size_t     tensor_para_rank);
     ~LlamaDecoderLayerWeight();
-    LlamaDecoderLayerWeight(const LlamaDecoderLayerWeight& other) = delete;
+    LlamaDecoderLayerWeight(const LlamaDecoderLayerWeight& other)            = delete;
     LlamaDecoderLayerWeight& operator=(const LlamaDecoderLayerWeight& other) = delete;
 
     void loadModel(std::string dir_path, FtCudaDataType model_file_type);
@@ -58,4 +58,4 @@ private:
     void mallocWeights();
 };
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaDecoderSelfAttentionLayer.cc

@@ -256,17 +256,17 @@ void LlamaDecoderSelfAttentionLayer<T>::forward(TensorMap*                     o
         0,        // max_input_length, not used w/o linear_bias_slopes
         input_tensors->getPtr<int>("total_padding_tokens", nullptr),
         step,
-        1.f,      // q_scaling
-        0,        // relative_attention_bias_stride
-        nullptr,  // linear_bias_slopes
-        nullptr,  //  masked_tokens_data,
-        nullptr,  // ia3_tasks
-        nullptr,  // ia3_key_weights
-        nullptr,  // ia3_value_weights
-        nullptr,  // qkv_scale_out
-        nullptr,  // attention_out_scale
-        quant_policy_,        // int8_mode
-        weights->past_kv_scale.data(), // attention kv scale
+        1.f,                            // q_scaling
+        0,                              // relative_attention_bias_stride
+        nullptr,                        // linear_bias_slopes
+        nullptr,                        //  masked_tokens_data,
+        nullptr,                        // ia3_tasks
+        nullptr,                        // ia3_key_weights
+        nullptr,                        // ia3_value_weights
+        nullptr,                        // qkv_scale_out
+        nullptr,                        // attention_out_scale
+        quant_policy_,                  // int8_mode
+        weights->past_kv_scale.data(),  // attention kv scale
         stream_);
     sync_check_cuda_error();
 
@@ -289,4 +289,4 @@ void LlamaDecoderSelfAttentionLayer<T>::forward(TensorMap*                     o
 template class LlamaDecoderSelfAttentionLayer<float>;
 template class LlamaDecoderSelfAttentionLayer<half>;
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaDecoderSelfAttentionLayer.h

@@ -15,7 +15,8 @@
  * limitations under the License.
  */
 
-// Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/layers/attention_layers/DecoderSelfAttentionLayer.h
+// Modified from
+// https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/layers/attention_layers/DecoderSelfAttentionLayer.h
 
 #pragma once
 
@@ -96,4 +97,4 @@ private:
     bool is_allocate_buffer_{};
 };
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaDenseWeight.h

@@ -15,7 +15,7 @@
  * limitations under the License.
  */
 
- // Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/layers/DenseWeight.h
+// Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/layers/DenseWeight.h
 
 #pragma once
 
@@ -25,8 +25,7 @@
 
 namespace fastertransformer {
 
-enum class WeightType : int
-{
+enum class WeightType : int {
     kFP32,
     kFP16,
     kFP8,  // not supported yet
@@ -66,7 +65,7 @@ template<typename T>
 struct LlamaAttentionWeight {
     LlamaDenseWeight<T> qkv;
     LlamaDenseWeight<T> output;
-    std::vector<float> past_kv_scale;
+    std::vector<float>  past_kv_scale;
 };
 
 template<typename T>
@@ -76,4 +75,4 @@ struct LlamaFfnWeight {
     LlamaDenseWeight<T> output;
 };
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaFfnLayer.cc

@@ -15,7 +15,7 @@
  * limitations under the License.
  */
 
- // Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/layers/FfnLayer.h
+// Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/layers/FfnLayer.h
 
 #include "src/fastertransformer/models/llama/LlamaFfnLayer.h"
 #include "src/fastertransformer/kernels/activation_kernels.h"
@@ -110,4 +110,4 @@ void LlamaFfnLayer<T>::forward(TensorMap*               output_tensors,
 template class LlamaFfnLayer<float>;
 template class LlamaFfnLayer<half>;
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaFfnLayer.h

@@ -15,7 +15,7 @@
  * limitations under the License.
  */
 
- // Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/layers/FfnLayer.cc
+// Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/layers/FfnLayer.cc
 
 #pragma once
 
@@ -82,4 +82,4 @@ private:
     bool is_allocate_buffer_{};
 };
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaInstanceComm.h

@@ -31,4 +31,4 @@ private:
     void*   ptr{};
 };
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaLinear.h

@@ -58,4 +58,4 @@ private:
     cudaStream_t     stream_{};
 };
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaNcclGuard.h

@@ -89,4 +89,4 @@ struct NcclGuard {
     std::unique_ptr<std::lock_guard<std::mutex>> global_nccl_lock_;
 };
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaV2.cc

-/* 
+/*
  * Copyright (c) OpenMMLab. All rights reserved.
  * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
  * Copyright (c) 2021, NAVER Corp.  Authored by CLOVA.
@@ -17,7 +17,8 @@
  * limitations under the License.
  */
 
- // Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/models/multi_gpu_gpt/ParallelGpt.cc
+// Modified from
+// https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/models/multi_gpu_gpt/ParallelGpt.cc
 
 #include "src/fastertransformer/models/llama/LlamaV2.h"
 #include "src/fastertransformer/kernels/decoding_kernels.h"
@@ -97,7 +98,8 @@ LlamaV2<T>::LlamaV2(size_t                       head_num,
             FT_LOG_ERROR("use_context_fmha not support int8");
             assert(0);
         }
-    } else {
+    }
+    else {
         elem_bits = sizeof(T) * 8;
     }
     kv_cache_mgr_ = std::make_unique<LlamaCacheManager>(num_layer_,
@@ -213,7 +215,7 @@ void LlamaV2<T>::contextDecode(T*         deocder_output,
                                              static_cast<T*>(nullptr),
                                              pPromptTuningParam<T>{},
                                              input_ids,
-                                             0,  // only used for postion encoding
+                                             0,  // only used for position encoding
                                              token_num,
                                              token_num,
                                              1,
@@ -592,4 +594,4 @@ void LlamaV2<T>::forward(std::unordered_map<std::string, Tensor>*       outputs,
 template class LlamaV2<half>;
 template class LlamaV2<float>;
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaV2.h

@@ -16,7 +16,8 @@
  * limitations under the License.
  */
 
-// Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/models/multi_gpu_gpt/ParallelGpt.h
+// Modified from
+// https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/models/multi_gpu_gpt/ParallelGpt.h
 
 #pragma once
 
@@ -182,4 +183,4 @@ private:
     std::thread internal_thread_;
 };
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/LlamaWeight.cc

@@ -121,4 +121,4 @@ void LlamaWeight<T>::loadModel(std::string dir_path)
 template struct LlamaWeight<float>;
 template struct LlamaWeight<half>;
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file