build turbomind (#35)

* build turbomind

* change namespace fastertransformer to turbomind

* change logger name

src/turbomind/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_128.cu

@@ -14,10 +14,10 @@
  * limitations under the License.
  */
 
-#include "src/fastertransformer/kernels/decoder_masked_multihead_attention.h"
-#include "src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h"
-#include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
-#include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/turbomind/kernels/decoder_masked_multihead_attention.h"
+#include "src/turbomind/kernels/decoder_masked_multihead_attention_utils.h"
+#include "src/turbomind/utils/cuda_bf16_wrapper.h"
+#include "src/turbomind/utils/cuda_utils.h"
 #include <assert.h>
 #include <float.h>
 #include <type_traits>

src/turbomind/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.cuh

@@ -15,12 +15,12 @@
  */
 #pragma once
 
-#include "src/fastertransformer/kernels/decoder_masked_multihead_attention.h"
-#include "src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h"
-#include "src/fastertransformer/models/llama/llama_utils.h"
-#include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
-#include "src/fastertransformer/utils/cuda_fp8_utils.h"
-#include "src/fastertransformer/utils/cuda_type_utils.cuh"
+#include "src/turbomind/kernels/decoder_masked_multihead_attention.h"
+#include "src/turbomind/kernels/decoder_masked_multihead_attention_utils.h"
+#include "src/turbomind/models/llama/llama_utils.h"
+#include "src/turbomind/utils/cuda_bf16_wrapper.h"
+#include "src/turbomind/utils/cuda_fp8_utils.h"
+#include "src/turbomind/utils/cuda_type_utils.cuh"
 #include <assert.h>
 #include <float.h>
 #include <type_traits>

src/turbomind/kernels/decoder_masked_multihead_attention_utils.h

@@ -16,12 +16,12 @@
 
 #pragma once
 
-#include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
-#include "src/fastertransformer/utils/cuda_fp8_utils.h"
-#include "src/fastertransformer/utils/cuda_type_utils.cuh"
+#include "src/turbomind/utils/cuda_bf16_wrapper.h"
+#include "src/turbomind/utils/cuda_fp8_utils.h"
+#include "src/turbomind/utils/cuda_type_utils.cuh"
 #include <stdint.h>
 
-using namespace fastertransformer;
+using namespace turbomind;
 
 namespace mmha {
 

src/turbomind/kernels/decoding_kernels.cu

@@ -14,12 +14,12 @@
  * limitations under the License.
  */
 
-#include "src/fastertransformer/kernels/decoding_kernels.h"
-#include "src/fastertransformer/kernels/reduce_kernel_utils.cuh"
-#include "src/fastertransformer/utils/cuda_type_utils.cuh"
-#include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/turbomind/kernels/decoding_kernels.h"
+#include "src/turbomind/kernels/reduce_kernel_utils.cuh"
+#include "src/turbomind/utils/cuda_type_utils.cuh"
+#include "src/turbomind/utils/cuda_utils.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 // static const float HALF_FLT_MAX = 65504.F;
 
@@ -803,4 +803,4 @@ void invokeFinalize(int*         output_ids,
         max_seq_len);
 }
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/kernels/decoding_kernels.h

@@ -20,7 +20,7 @@
 #include <cuda_fp16.h>
 #include <cuda_runtime.h>
 
-namespace fastertransformer {
+namespace turbomind {
 
 template<typename T>
 void invokeDecodingInitialize(bool*        finished,
@@ -167,4 +167,4 @@ void invokeFinalize(int*         output_ids,
                     const int    batch_size,
                     cudaStream_t stream);
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/kernels/gen_relative_pos_bias.cu

@@ -17,11 +17,11 @@
 #include "cublas_v2.h"
 #include "gen_relative_pos_bias.h"
 #include "reduce_kernel_utils.cuh"
-#include "src/fastertransformer/kernels/activation_kernels.h"
-#include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/turbomind/kernels/activation_kernels.h"
+#include "src/turbomind/utils/cuda_utils.h"
 #include <cstdio>
 
-namespace fastertransformer {
+namespace turbomind {
 
 /*******************  invokeGenRelativePosBias  ***********************/
 // relative_position_bias_table is [(2*window_size-1)*(2*window_size-1), headNum]
@@ -301,4 +301,4 @@ template void invokeGenRelativePosBiasV2(half*          relative_position_bias,
                                          const int      cpb_mlp_out_dim,
                                          const int      head_num,
                                          cudaStream_t   stream);
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/kernels/gen_relative_pos_bias.h

@@ -16,13 +16,13 @@
 
 #pragma once
 
-#include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
+#include "src/turbomind/utils/cuda_bf16_wrapper.h"
 
 #include <assert.h>
 #include <cuda_runtime.h>
 #include <stdint.h>
 
-namespace fastertransformer {
+namespace turbomind {
 
 enum class PositionEmbeddingType {
     relative,
@@ -52,4 +52,4 @@ void invokeGenRelativePosBiasV2(T*            relative_position_bias,
                                 const int     cpb_mlp_out_dim,
                                 const int     head_num,
                                 cudaStream_t  stream);
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/kernels/gpt_kernels.cu

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "src/fastertransformer/utils/cuda_fp8_utils.h"
+#include "src/turbomind/utils/cuda_fp8_utils.h"
 #ifndef CUDART_VERSION
 #error CUDART_VERSION Undefined!
 #elif (CUDART_VERSION >= 11050)
@@ -22,10 +22,10 @@
 #else
 #include "3rdparty/cub/cub.cuh"
 #endif
-#include "src/fastertransformer/kernels/gpt_kernels.h"
-#include "src/fastertransformer/utils/memory_utils.h"
+#include "src/turbomind/kernels/gpt_kernels.h"
+#include "src/turbomind/utils/memory_utils.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 // PROMPT_SRC: 0 --> no prompts, 1 --> from loaded prompts, 2 --> from request prompts
 template<typename T, bool OUTPUT_ID, int PROMPT_SRC>
@@ -1108,4 +1108,4 @@ INSTANTIATE_INVOKE_SUM_LENGTH_DIMENSION(__nv_bfloat16);
 #endif
 #undef INSTANTIATE_INVOKE_SUM_LENGTH_DIMENSION
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/kernels/gpt_kernels.h

@@ -20,10 +20,10 @@
 #include <cuda_runtime.h>
 #include <unordered_map>
 
-#include "src/fastertransformer/utils/Tensor.h"
-#include "src/fastertransformer/utils/memory_utils.h"
+#include "src/turbomind/utils/Tensor.h"
+#include "src/turbomind/utils/memory_utils.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 template<typename T>
 struct inputIdsEmbeddingLookupPosEncodingSoftPromptParam {
@@ -238,4 +238,4 @@ void invokeSumLengthDimension(float*       out_buf,
                               const size_t hidden_dim,
                               cudaStream_t stream = 0);
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/kernels/logprob_kernels.cu

@@ -26,11 +26,11 @@
 #include "3rdparty/cub/cub.cuh"
 #endif
 
-#include "src/fastertransformer/kernels/logprob_kernels.h"
-#include "src/fastertransformer/kernels/reduce_kernel_utils.cuh"
-#include "src/fastertransformer/utils/logger.h"
+#include "src/turbomind/kernels/logprob_kernels.h"
+#include "src/turbomind/kernels/reduce_kernel_utils.cuh"
+#include "src/turbomind/utils/logger.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 template<typename T>
 __global__ void log_probs_kernel(float*       log_probs,
@@ -158,7 +158,7 @@ void invokeLogProbFromLogits(float*       cum_log_probs,
     // input_lengths: [batch_size]
     // workspace: workspace buffer of size at least sizeof(float) * max_input_length * batch_size.
 
-    FT_LOG_DEBUG(__PRETTY_FUNCTION__);
+    TM_LOG_DEBUG(__PRETTY_FUNCTION__);
     // block_size should be multiple of 32 to use warpReduceMax.
     const int block_size = vocab_size < 1024 ? (vocab_size + 31) / 32 * 32 : 1024;
     assert(block_size % 32 == 0);
@@ -207,4 +207,4 @@ template void invokeLogProbFromLogits(float*       cum_log_probs,
                                       const size_t workspace_size,
                                       cudaStream_t stream,
                                       const bool   batch_first);
-}  // end of namespace fastertransformer
+}  // end of namespace turbomind

src/turbomind/kernels/logprob_kernels.h

@@ -16,7 +16,7 @@
 
 #pragma once
 
-namespace fastertransformer {
+namespace turbomind {
 
 template<typename T>
 void invokeLogProbFromLogits(float*       cum_log_probs,
@@ -31,4 +31,4 @@ void invokeLogProbFromLogits(float*       cum_log_probs,
                              const size_t workspace_size,
                              cudaStream_t stream,
                              const bool   batch_first = false);
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/kernels/online_softmax_beamsearch_kernels.cu

@@ -22,11 +22,11 @@
 #include "3rdparty/cub/cub.cuh"
 #endif
 
-#include "src/fastertransformer/kernels/online_softmax_beamsearch_kernels.h"
-#include "src/fastertransformer/kernels/reduce_kernel_utils.cuh"
-#include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/turbomind/kernels/online_softmax_beamsearch_kernels.h"
+#include "src/turbomind/kernels/reduce_kernel_utils.cuh"
+#include "src/turbomind/utils/cuda_utils.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 #define DO_SPLIT_SMALL_TOP_K_SOFTMAX
 static const int SMALL_TOP_K_SOFTMAX_THREADBLOCK_SIZE = 256;
@@ -736,4 +736,4 @@ template void invokeTopkSoftMax<half>(const half*     log_probs,
                                       const float     length_penalty,
                                       cudaStream_t    stream);
 
-}  // end of namespace fastertransformer
+}  // end of namespace turbomind

src/turbomind/kernels/online_softmax_beamsearch_kernels.h

@@ -15,9 +15,9 @@
  */
 #pragma once
 
-#include "src/fastertransformer/kernels/beam_search_topk_kernels.h"
+#include "src/turbomind/kernels/beam_search_topk_kernels.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 template<typename T>
 void invokeTopkSoftMax(const T*        log_probs,
@@ -38,4 +38,4 @@ void invokeTopkSoftMax(const T*        log_probs,
                        const float     length_penalty,
                        cudaStream_t    stream);
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/kernels/penalty_types.h

@@ -19,9 +19,9 @@
 #include <string>
 #include <unordered_map>
 
-#include "src/fastertransformer/utils/string_utils.h"
+#include "src/turbomind/utils/string_utils.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 enum class RepetitionPenaltyType {
     Additive,        // the presence penalty
@@ -42,4 +42,4 @@ inline float getDefaultPenaltyValue(RepetitionPenaltyType penalty_type)
     return 0.0f;
 }
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/kernels/reduce_kernel_utils.cuh

@@ -21,8 +21,8 @@
 #else
 #include <cooperative_groups.h>
 #endif
-#include "src/fastertransformer/utils/cuda_bf16_wrapper.h"
-#include "src/fastertransformer/utils/cuda_type_utils.cuh"
+#include "src/turbomind/utils/cuda_bf16_wrapper.h"
+#include "src/turbomind/utils/cuda_type_utils.cuh"
 #include <cuda_fp16.h>
 #include <cuda_runtime.h>
 #include <curand_kernel.h>
@@ -31,7 +31,7 @@
 
 namespace cg = cooperative_groups;
 
-namespace fastertransformer {
+namespace turbomind {
 
 template<int VPT>
 struct BytesToType;
@@ -363,4 +363,4 @@ __device__ __forceinline__ __nv_bfloat16 clamp_inf_for_half(const float input)
 }
 #endif
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/kernels/sampling_penalty_kernels.cu

@@ -17,9 +17,9 @@
 #include <assert.h>
 #include <float.h>
 
-#include "src/fastertransformer/kernels/sampling_penalty_kernels.h"
+#include "src/turbomind/kernels/sampling_penalty_kernels.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 // TODO Add half2 implementation
 template<typename T>
@@ -534,4 +534,4 @@ template void invokeMinLengthPenalty(half*        logits,
                                      const int    vocab_size_padded,
                                      cudaStream_t stream);
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/kernels/sampling_penalty_kernels.h

@@ -17,10 +17,10 @@
 
 #include <cuda_fp16.h>
 
-#include "src/fastertransformer/kernels/penalty_types.h"
-#include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/turbomind/kernels/penalty_types.h"
+#include "src/turbomind/utils/cuda_utils.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 template<typename T>
 void invokeApplyRepetitionPenalty(T*                          logits,
@@ -78,4 +78,4 @@ void invokeMinLengthPenalty(T*           logits,
                             const int    vocab_size_padded,
                             cudaStream_t stream);
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/kernels/sampling_topk_kernels.cu

@@ -24,10 +24,10 @@
 #include "3rdparty/cub/cub.cuh"
 #endif
 
-#include "src/fastertransformer/kernels/reduce_kernel_utils.cuh"
-#include "src/fastertransformer/kernels/sampling_topk_kernels.h"
+#include "src/turbomind/kernels/reduce_kernel_utils.cuh"
+#include "src/turbomind/kernels/sampling_topk_kernels.h"
 
-namespace fastertransformer {
+namespace turbomind {
 
 __global__ void curandInitialize(curandState_t* state, const int size, const unsigned long long random_seed)
 {
@@ -575,4 +575,4 @@ template void invokeTopKTopPSampling(void*          workspace,
                                      const int*     end_ids,
                                      cudaStream_t   stream);
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/kernels/sampling_topk_kernels.h

@@ -16,9 +16,9 @@
  */
 #pragma once
 
-#include "src/fastertransformer/utils/logger.h"
+#include "src/turbomind/utils/logger.h"
 #include <curand_kernel.h>
-namespace fastertransformer {
+namespace turbomind {
 
 template<typename T>
 void invokeTopKSampling(void*          workspace,
@@ -95,4 +95,4 @@ void invokeTopKTopPSampling(void*          workspace,
                             const int*     end_ids,
                             cudaStream_t   stream);
 
-}  // namespace fastertransformer
+}  // namespace turbomind

src/turbomind/kernels/sampling_topp_kernels.cu

@@ -22,14 +22,14 @@
 #include "3rdparty/cub/cub.cuh"
 #endif
 
-#include "src/fastertransformer/kernels/reduce_kernel_utils.cuh"
-#include "src/fastertransformer/kernels/sampling_topp_kernels.h"
-#include "src/fastertransformer/utils/cuda_utils.h"
+#include "src/turbomind/kernels/reduce_kernel_utils.cuh"
+#include "src/turbomind/kernels/sampling_topp_kernels.h"
+#include "src/turbomind/utils/cuda_utils.h"
 
 constexpr int   ENABLE_SINGLE_PASS_TOP_P = 0;
 constexpr float SINGLE_PASS_THRESHOLD    = 0.9;
 
-namespace fastertransformer {
+namespace turbomind {
 
 namespace segmented_topp_impl {
 
@@ -1426,4 +1426,4 @@ void invokeComputeToppDecay(float*         runtime_top_p,
         runtime_top_p, runtime_initial_top_p, output_ids, top_p_decay, top_p_min, top_p_reset_ids, local_batch_size);
 }
 
-}  // namespace fastertransformer
+}  // namespace turbomind