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/LlamaWeight.h

@@ -40,7 +40,7 @@ struct LlamaWeight {
 
     ~LlamaWeight();
 
-    LlamaWeight(const LlamaWeight& other) = delete;
+    LlamaWeight(const LlamaWeight& other)            = delete;
     LlamaWeight& operator=(const LlamaWeight& other) = delete;
 
     void loadModel(std::string dir_path);
@@ -67,4 +67,4 @@ private:
     size_t     tensor_para_rank_;
 };
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/Request.h

@@ -25,8 +25,7 @@ struct Request {
     using Callback = std::function<void(std::unordered_map<std::string, Tensor>*)>;
     Callback stream_cb;
 
-    enum
-    {
+    enum {
         kInvalid  = 1,
         kConflict = 2,
         kBusy     = 3,
@@ -88,4 +87,4 @@ private:
     std::condition_variable              cv_;
 };
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/fused_multi_head_attention/llama_flash_attention_kernel.cu

@@ -10,7 +10,7 @@
 #include <cutlass/half.h>
 #include <cutlass/platform/platform.h>
 
-// modifiy from:
+// modified from:
 // https://github.com/NVIDIA/cutlass/blob/main/examples/41_fused_multi_head_attention/kernel_forward.h
 
 namespace fastertransformer {

src/fastertransformer/models/llama/llama_decoder_kernels.cu

@@ -11,7 +11,8 @@ namespace cg = cooperative_groups;
 namespace fastertransformer {
 
 template<typename T>
-struct res_norm_ops_t {};
+struct res_norm_ops_t {
+};
 
 template<typename T>
 struct res_norm_t {
@@ -144,7 +145,7 @@ __global__ void fusedAddBiasResidualNorm(T* __restrict__ r_data,
 
 template<typename T>
 void invokeFusedAddBiasResidualRMSNorm(
-    T* residual, T* inout, const T* bias, const T* scale, float eps, int batch_size, int n_dims, cudaStream_t stream)
+    T* residual, T* in_out, const T* bias, const T* scale, float eps, int batch_size, int n_dims, cudaStream_t stream)
 {
     constexpr int PACK_DIM = sizeof(uint4) / sizeof(T);
     FT_CHECK(n_dims % PACK_DIM == 0);
@@ -154,7 +155,7 @@ void invokeFusedAddBiasResidualRMSNorm(
     n_threads           = (n_threads + 31) / 32 * 32;      // round up to the nearest multiple of warp size
 
     fusedAddBiasResidualNorm<<<batch_size, n_threads, 0, stream>>>(
-        residual, inout, bias, scale, eps, batch_size, n_dims);
+        residual, in_out, bias, scale, eps, batch_size, n_dims);
 }
 
 template void

src/fastertransformer/models/llama/llama_decoder_kernels.h

@@ -6,6 +6,6 @@ namespace fastertransformer {
 
 template<typename T>
 void invokeFusedAddBiasResidualRMSNorm(
-    T* residual, T* inout, const T* bias, const T* scale, float eps, int batch_size, int n_dims, cudaStream_t stream);
+    T* residual, T* in_out, const T* bias, const T* scale, float eps, int batch_size, int n_dims, cudaStream_t stream);
 
 }  // namespace fastertransformer

src/fastertransformer/models/llama/llama_kernels.cu

 // Copyright (c) OpenMMLab. All rights reserved.
 
+#include "src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h"
 #include "src/fastertransformer/kernels/reduce_kernel_utils.cuh"
 #include "src/fastertransformer/models/llama/llama_kernels.h"
 #include "src/fastertransformer/models/llama/llama_utils.h"
 #include "src/fastertransformer/utils/cuda_type_utils.cuh"
-#include "src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h"
 
 namespace fastertransformer {
 
@@ -293,37 +293,40 @@ inline __device__ float2 float2div(float a, float2 b)
     return c;
 }
 
-static inline __device__ half4 char4_scale_to_half4(char4 value, const float scale) {
-  half4 dst;
-  dst.x = __float2half(value.x * scale);
-  dst.y = __float2half(value.y * scale);
-  dst.z = __float2half(value.z * scale);
-  dst.w = __float2half(value.w * scale);
-  return dst;
+static inline __device__ half4 char4_scale_to_half4(char4 value, const float scale)
+{
+    half4 dst;
+    dst.x = __float2half(value.x * scale);
+    dst.y = __float2half(value.y * scale);
+    dst.z = __float2half(value.z * scale);
+    dst.w = __float2half(value.w * scale);
+    return dst;
 }
 
-static inline __device__ uint32_t float4_to_char4(float x,
-                                                  float y,
-                                                  float z,
-                                                  float w) {
-  uint32_t dst;
+static inline __device__ uint32_t float4_to_char4(float x, float y, float z, float w)
+{
+    uint32_t dst;
 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 720
-  uint32_t a; asm volatile("cvt.rni.sat.s32.f32 %0, %1;\n" : "=r"(a) : "f"(x));
-  uint32_t b; asm volatile("cvt.rni.sat.s32.f32 %0, %1;\n" : "=r"(b) : "f"(y));
-  uint32_t c; asm volatile("cvt.rni.sat.s32.f32 %0, %1;\n" : "=r"(c) : "f"(z));
-  uint32_t d; asm volatile("cvt.rni.sat.s32.f32 %0, %1;\n" : "=r"(d) : "f"(w));
-
-  asm volatile("cvt.pack.sat.s8.s32.b32 %0, %1, %2,  0;\n" : "=r"(dst) : "r"(d), "r"(c));
-  asm volatile("cvt.pack.sat.s8.s32.b32 %0, %1, %2, %0;\n" : "+r"(dst) : "r"(b), "r"(a));
+    uint32_t a;
+    asm volatile("cvt.rni.sat.s32.f32 %0, %1;\n" : "=r"(a) : "f"(x));
+    uint32_t b;
+    asm volatile("cvt.rni.sat.s32.f32 %0, %1;\n" : "=r"(b) : "f"(y));
+    uint32_t c;
+    asm volatile("cvt.rni.sat.s32.f32 %0, %1;\n" : "=r"(c) : "f"(z));
+    uint32_t d;
+    asm volatile("cvt.rni.sat.s32.f32 %0, %1;\n" : "=r"(d) : "f"(w));
+
+    asm volatile("cvt.pack.sat.s8.s32.b32 %0, %1, %2,  0;\n" : "=r"(dst) : "r"(d), "r"(c));
+    asm volatile("cvt.pack.sat.s8.s32.b32 %0, %1, %2, %0;\n" : "+r"(dst) : "r"(b), "r"(a));
 #else
-  char4 tmp;
-  tmp.x = x;
-  tmp.y = y;
-  tmp.z = z;
-  tmp.w = w;
-  dst = reinterpret_cast<const uint32_t&>(tmp);
+    char4 tmp;
+    tmp.x = x;
+    tmp.y = y;
+    tmp.z = z;
+    tmp.w = w;
+    dst   = reinterpret_cast<const uint32_t&>(tmp);
 #endif
-  return dst;
+    return dst;
 }
 
 template<typename T>
@@ -380,7 +383,6 @@ __global__ void extend_value_cache_int8(int8_t**     v_dst,
     }
 }
 
-
 template<typename T>
 void invokeExtendKVCache(T**          k_dst,
                          T**          v_dst,
@@ -404,18 +406,48 @@ void invokeExtendKVCache(T**          k_dst,
     dim3 grid((max_q_len * size_per_head / x + block_sz - 1) / block_sz, local_batch_size, local_head_num);
 
     if (quant & QuantPolicy::kCacheKVInt8) {
-        extend_value_cache_int8<<<grid, block_sz, 0, stream>>>(
-            reinterpret_cast<int8_t**>(k_dst), dst_offset, k_src, local_head_num, size_per_head, query_length, history_length, max_q_len, max_seq_len, kv_scale[0]);
-
-        extend_value_cache_int8<<<grid, block_sz, 0, stream>>>(
-            reinterpret_cast<int8_t**>(v_dst), dst_offset, v_src, local_head_num, size_per_head, query_length, history_length, max_q_len, max_seq_len, kv_scale[1]);
-    
-    } else {
-        extend_value_cache<<<grid, block_sz, 0, stream>>>(
-            k_dst, dst_offset, k_src, local_head_num, size_per_head, query_length, history_length, max_q_len, max_seq_len);
-
-        extend_value_cache<<<grid, block_sz, 0, stream>>>(
-            v_dst, dst_offset, v_src, local_head_num, size_per_head, query_length, history_length, max_q_len, max_seq_len);
+        extend_value_cache_int8<<<grid, block_sz, 0, stream>>>(reinterpret_cast<int8_t**>(k_dst),
+                                                               dst_offset,
+                                                               k_src,
+                                                               local_head_num,
+                                                               size_per_head,
+                                                               query_length,
+                                                               history_length,
+                                                               max_q_len,
+                                                               max_seq_len,
+                                                               kv_scale[0]);
+
+        extend_value_cache_int8<<<grid, block_sz, 0, stream>>>(reinterpret_cast<int8_t**>(v_dst),
+                                                               dst_offset,
+                                                               v_src,
+                                                               local_head_num,
+                                                               size_per_head,
+                                                               query_length,
+                                                               history_length,
+                                                               max_q_len,
+                                                               max_seq_len,
+                                                               kv_scale[1]);
+    }
+    else {
+        extend_value_cache<<<grid, block_sz, 0, stream>>>(k_dst,
+                                                          dst_offset,
+                                                          k_src,
+                                                          local_head_num,
+                                                          size_per_head,
+                                                          query_length,
+                                                          history_length,
+                                                          max_q_len,
+                                                          max_seq_len);
+
+        extend_value_cache<<<grid, block_sz, 0, stream>>>(v_dst,
+                                                          dst_offset,
+                                                          v_src,
+                                                          local_head_num,
+                                                          size_per_head,
+                                                          query_length,
+                                                          history_length,
+                                                          max_q_len,
+                                                          max_seq_len);
     }
 }
 
@@ -492,17 +524,16 @@ __global__ void transpose_value_cache(T*           v_dst,  //
     }
 }
 
-
 template<typename T>
-__global__ void transpose_value_cache_int8(T*           v_dst,  //
-                                      const int8_t**    v_src,
-                                      const size_t src_offset,
-                                      const int    head_num,
-                                      const int    size_per_head,
-                                      const int*   seq_length,
-                                      const int    max_kv_len,
-                                      const int    max_seq_len,
-                                      const float  v_scale)
+__global__ void transpose_value_cache_int8(T*             v_dst,  //
+                                           const int8_t** v_src,
+                                           const size_t   src_offset,
+                                           const int      head_num,
+                                           const int      size_per_head,
+                                           const int*     seq_length,
+                                           const int      max_kv_len,
+                                           const int      max_seq_len,
+                                           const float    v_scale)
 {
     const int     batch_id = blockIdx.y;
     const int     head_id  = blockIdx.z;
@@ -533,7 +564,7 @@ __global__ void transpose_value_cache_int8(T*           v_dst,  //
 
         // int8x8 -> fp16x8
         const auto from_ptr = reinterpret_cast<const char4*>(val_src + src_idx);
-        auto to_ptr = reinterpret_cast<half4*>(val_dst + dst_idx);
+        auto       to_ptr   = reinterpret_cast<half4*>(val_dst + dst_idx);
 
         to_ptr[0] = char4_scale_to_half4(from_ptr[0], v_scale);
         to_ptr[1] = char4_scale_to_half4(from_ptr[1], v_scale);
@@ -562,13 +593,27 @@ void invokeTransposeKVCache(T*           key_cache_trans,
     dim3 grid((max_kv_len * size_per_head / x + block_sz - 1) / block_sz, batch_size, head_num);
 
     if (quant & QuantPolicy::kCacheKVInt8) {
-        transpose_value_cache_int8<<<grid, block_sz, 0, stream>>>(
-            key_cache_trans, reinterpret_cast<const int8_t**>(key_cache), src_offset, head_num, size_per_head, key_length, max_kv_len, max_seq_len, kv_scale[0]);
-
-        transpose_value_cache_int8<<<grid, block_sz, 0, stream>>>(
-            val_cache_trans, reinterpret_cast<const int8_t**>(val_cache), src_offset, head_num, size_per_head, key_length, max_kv_len, max_seq_len, kv_scale[1]);
-    
-    } else {
+        transpose_value_cache_int8<<<grid, block_sz, 0, stream>>>(key_cache_trans,
+                                                                  reinterpret_cast<const int8_t**>(key_cache),
+                                                                  src_offset,
+                                                                  head_num,
+                                                                  size_per_head,
+                                                                  key_length,
+                                                                  max_kv_len,
+                                                                  max_seq_len,
+                                                                  kv_scale[0]);
+
+        transpose_value_cache_int8<<<grid, block_sz, 0, stream>>>(val_cache_trans,
+                                                                  reinterpret_cast<const int8_t**>(val_cache),
+                                                                  src_offset,
+                                                                  head_num,
+                                                                  size_per_head,
+                                                                  key_length,
+                                                                  max_kv_len,
+                                                                  max_seq_len,
+                                                                  kv_scale[1]);
+    }
+    else {
         transpose_value_cache<<<grid, block_sz, 0, stream>>>(
             key_cache_trans, key_cache, src_offset, head_num, size_per_head, key_length, max_kv_len, max_seq_len);
 
@@ -577,10 +622,34 @@ void invokeTransposeKVCache(T*           key_cache_trans,
     }
 }
 
-template void invokeTransposeKVCache(
-    float*, float*, const float**, const float**, size_t, int, const int*, int, int, int, int, cudaStream_t stream, int, const float*);
-template void invokeTransposeKVCache(
-    half*, half*, const half**, const half**, size_t, int, const int*, int, int, int, int, cudaStream_t stream, int, const float*);
+template void invokeTransposeKVCache(float*,
+                                     float*,
+                                     const float**,
+                                     const float**,
+                                     size_t,
+                                     int,
+                                     const int*,
+                                     int,
+                                     int,
+                                     int,
+                                     int,
+                                     cudaStream_t stream,
+                                     int,
+                                     const float*);
+template void invokeTransposeKVCache(half*,
+                                     half*,
+                                     const half**,
+                                     const half**,
+                                     size_t,
+                                     int,
+                                     const int*,
+                                     int,
+                                     int,
+                                     int,
+                                     int,
+                                     cudaStream_t stream,
+                                     int,
+                                     const float*);
 
 __global__ void gatherOutput(int*       output_ids,
                              const int* ids,
@@ -619,4 +688,4 @@ void invokeGatherOutput(int*         output_ids,
         output_ids, ids, context_length, max_context_len, max_gen_step, max_output_len, batch_size);
 }
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/llama_kernels.h

@@ -163,4 +163,4 @@ inline void dump_sequence_len(int* d_seq_len, int step, int tp_rank, cudaStream_
     FT_LOG_ERROR("--------> rank = %d, step = %d, seq_len = %d <--------", tp_rank, step, h_seq_len);
 }
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/llama_utils.cu

@@ -157,4 +157,4 @@ bool isDebug()
     return is_debug;
 }
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/llama_utils.h

@@ -18,8 +18,7 @@ enum QuantPolicy {
     kCacheKVInt8 = 0x04,
 };
 
-enum CmpMode
-{
+enum CmpMode {
     kCmpNone,
     kCmpRead,
     kCmpWrite,
@@ -65,4 +64,4 @@ size_t curandStateGetSize();
 
 bool isDebug();
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file

src/fastertransformer/models/llama/prefix_cache.h

@@ -6,4 +6,4 @@ template<typename T>
 void invokeInsertKeyCache(T* key_cache, const T* src, int L, int H, int Dx, int s, int X, int S, cudaStream_t st);
 
 template<typename T>
-void invokeInsertValueCache(T* value_cache, const T* src, int L, int H, int s, int D, int S, cudaStream_t st);
+void invokeInsertValueCache(T* value_cache, const T* src, int L, int H, int s, int D, int S, cudaStream_t st);
\ No newline at end of file

src/fastertransformer/triton_backend/libfastertransformer.cc

@@ -25,7 +25,8 @@
 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
-// Modified from https://github.com/triton-inference-server/fastertransformer_backend/blob/main/src/libfastertransformer.cc
+// Modified from
+// https://github.com/triton-inference-server/fastertransformer_backend/blob/main/src/libfastertransformer.cc
 
 #include <stdint.h>
 
@@ -1399,7 +1400,7 @@ void ModelInstanceState::SetInputTensors(
             auto batch_input_name = batch_input.TargetNames()[0];
 
             // we only take care of the ragged input_ids
-            // Assume the first dimention (length) are different and others are the
+            // Assume the first dimension (length) are different and others are the
             // same BATCH_ITEM_SHAPE [num_requests (batches), num_dims (excluding
             // batch dimension)]
             if (batch_input_kind == BatchInput::Kind::BATCH_ITEM_SHAPE
@@ -1464,7 +1465,7 @@ void ModelInstanceState::SetInputTensors(
                                                     param.batch_input_ptr + param.batch_intput_size,
                                                     [&](int x) { return x != param.batch_input_ptr[0]; });
 
-                // calculate statics of elements
+                // calculate statistics of elements
                 if (param.is_input_ragged) {
                     param.max_elements_per_seq =
                         *std::max_element(param.batch_input_ptr, param.batch_input_ptr + param.batch_intput_size);

src/fastertransformer/triton_backend/libtriton_fastertransformer.ldscript

-# Copyright (c) 2021-2022, NVIDIA CORPORATION. All rights reserved.
+#Copyright(c) 2021 - 2022, NVIDIA CORPORATION.All rights reserved.
 #
-# Redistribution and use in source and binary forms, with or without
-# modification, are permitted provided that the following conditions
-# are met:
-#  * Redistributions of source code must retain the above copyright
-#    notice, this list of conditions and the following disclaimer.
-#  * Redistributions in binary form must reproduce the above copyright
-#    notice, this list of conditions and the following disclaimer in the
-#    documentation and/or other materials provided with the distribution.
-#  * Neither the name of NVIDIA CORPORATION nor the names of its
-#    contributors may be used to endorse or promote products derived
-#    from this software without specific prior written permission.
+#Redistribution and use in source and binary forms, with or without
+#modification, are permitted provided that the following conditions
+#are met:
+#* Redistributions of source code must retain the above copyright
+#notice, this list of conditions and the following disclaimer.
+#* Redistributions in binary form must reproduce the above copyright
+#notice, this list of conditions and the following disclaimer in the
+#documentation and / or other materials provided with the distribution.
+#* Neither the name of NVIDIA CORPORATION nor the names of its
+#contributors may be used to endorse or promote products derived
+#from this software without specific prior written       permission.
 #
-# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
-# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
-# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
-# PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
-# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
-# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
-# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
-# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
-# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+#THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+#EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+#IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+#PURPOSE ARE DISCLAIMED.IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+#CONTRIBUTORS BE LIABLE FOR ANY                               DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+#EXEMPLARY, OR CONSEQUENTIAL DAMAGES(INCLUDING, BUT NOT LIMITED TO,
+#PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+#PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+#OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+#(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+#OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 {
-  global:
+global:
     TRITONBACKEND_*;
-  local: *;
+local:
+    *;
 };

src/fastertransformer/triton_backend/llama/LlamaTritonModel.cc

@@ -309,7 +309,8 @@ std::string LlamaTritonModel<T>::toString()
        << "\nuse_context_fmha: " << use_context_fmha_ << "\nstart_id: " << start_id_
        << "\ntensor_para_size: " << tensor_para_size_ << "\npipeline_para_size: " << pipeline_para_size_
        << "\nenable_custom_all_reduce: " << enable_custom_all_reduce_ << "\nmodel_name: " << model_name_
-       << "\nprefix_cache_len: " << prefix_cache_len_ << "\nmodel_dir: " << model_dir_ << "\nquant_policy: " << quant_policy_ << std::endl;
+       << "\nprefix_cache_len: " << prefix_cache_len_ << "\nmodel_dir: " << model_dir_
+       << "\nquant_policy: " << quant_policy_ << std::endl;
 
     return ss.str();
 }

src/fastertransformer/triton_backend/llama/LlamaTritonModelInstance.cc

@@ -15,7 +15,8 @@
  * limitations under the License.
  */
 
- // Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/triton_backend/multi_gpu_gpt/ParallelGptTritonModel.h
+// Modified from
+// https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/triton_backend/multi_gpu_gpt/ParallelGptTritonModel.h
 
 #include "src/fastertransformer/triton_backend/llama/LlamaTritonModelInstance.h"
 #include "src/fastertransformer/triton_backend/transformer_triton_backend.hpp"

src/fastertransformer/triton_backend/llama/LlamaTritonModelInstance.h

@@ -15,7 +15,8 @@
  * limitations under the License.
  */
 
- // Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/triton_backend/multi_gpu_gpt/ParallelGptTritonModel.h
+// Modified from
+// https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/triton_backend/multi_gpu_gpt/ParallelGptTritonModel.h
 
 #pragma once
 

src/fastertransformer/triton_backend/transformer_triton_backend.cpp

@@ -15,7 +15,8 @@
  * limitations under the License.
  */
 
- // Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/triton_backend/transformer_triton_backend.cpp
+// Modified from
+// https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/triton_backend/transformer_triton_backend.cpp
 
 #include "src/fastertransformer/triton_backend/transformer_triton_backend.hpp"
 #include "src/fastertransformer/utils/nccl_utils.h"

src/fastertransformer/triton_backend/transformer_triton_backend.hpp

@@ -15,7 +15,8 @@
  * limitations under the License.
  */
 
-// Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/triton_backend/transformer_triton_backend.hpp
+// Modified from
+// https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/triton_backend/transformer_triton_backend.hpp
 
 #pragma once
 

src/fastertransformer/utils/CMakeLists.txt

@@ -76,7 +76,7 @@ if(ENABLE_FP8)
 add_library(cublasFP8MMWrapper STATIC cublasFP8MMWrapper.cu)
 set_property(TARGET cublasFP8MMWrapper PROPERTY POSITION_INDEPENDENT_CODE  ON)
 set_property(TARGET cublasFP8MMWrapper PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS  ON)
-target_link_libraries(cublasFP8MMWrapper PUBLIC -lcublasLt -lcudart -lcurand 
+target_link_libraries(cublasFP8MMWrapper PUBLIC -lcublasLt -lcudart -lcurand
                       cublasAlgoMap cublasMMWrapper nvtx_utils fp8_qgmma_1x1_utils)
 endif()
 

src/fastertransformer/utils/Tensor.h

@@ -35,8 +35,7 @@
 
 namespace fastertransformer {
 
-typedef enum datatype_enum
-{
+typedef enum datatype_enum {
     TYPE_INVALID,
     TYPE_BOOL,
     TYPE_UINT8,
@@ -99,8 +98,7 @@ DataType getTensorType()
     }
 }
 
-typedef enum memorytype_enum
-{
+typedef enum memorytype_enum {
     MEMORY_CPU,
     MEMORY_CPU_PINNED,
     MEMORY_GPU

src/fastertransformer/utils/activation_types.h

@@ -63,4 +63,4 @@ inline bool isGatedActivation(ActivationType activaiton_type)
            || activaiton_type == ActivationType::SiGLU;
 }
 
-}  // namespace fastertransformer
+}  // namespace fastertransformer
\ No newline at end of file