[LLM INFER] Append attn Moved to https:/PaddlePaddle/PaddleNLP/pull/9244

csrc/gpu/append_attn/append_attention_bfloat16_bfloat16_kernel.cu

@@ -0,0 +1,51 @@
+// Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+#include "append_attention_kernel.h"
+
+template void CascadeAppendAttentionKernel<paddle::bfloat16, paddle::bfloat16>(
+ const paddle::Tensor& qkv, // [token_num, num_heads, head_dim]
+ const paddle::Tensor& cache_k, // [max_block_num, num_heads, block_size, head_dim]
+ const paddle::Tensor& cache_v, // [max_block_num, num_heads, head_dim, block_size]
+ const paddle::optional<paddle::Tensor>& attn_mask,
+ const paddle::optional<paddle::Tensor>& cache_k_scale, // [num_kv_heads, head_dim]
+ const paddle::optional<paddle::Tensor>& cache_v_scale, // [num_kv_heads, head_dim]
+ const paddle::optional<paddle::Tensor>& cache_k_zp, // [num_kv_heads, head_dim]
+ const paddle::optional<paddle::Tensor>& cache_v_zp, // [num_kv_heads, head_dim]
+ const paddle::optional<paddle::Tensor>& shift_bias, // [num_kv_heads, head_dim]
+ const paddle::optional<paddle::Tensor>& smooth_weight, // [num_kv_heads, head_dim]
+ const paddle::Tensor& seq_lens_q,
+ const paddle::Tensor& seq_lens_kv,
+ const paddle::Tensor& seq_lens_encoder,
+ const paddle::Tensor& padding_offsets,
+ const paddle::Tensor& cum_offsets,
+ const paddle::Tensor& block_table,
+ const paddle::Tensor& batch_ids,
+ const paddle::Tensor& tile_ids_per_batch,
+ const std::string& cache_quant_type_str,
+ const int num_blocks,
+ const int block_shape_q,
+ const int max_seq_len,
+ const int max_dec_len,
+ const int num_heads,
+ const int kv_num_heads,
+ const int head_dim,
+ const float in_scale,
+ const int max_partition_size,
+ const int encoder_max_partition_size,
+ const int speculate_max_draft_token_num,
+ const bool causal,
+ const bool is_decoder,
+ const bool enable_prefill,
+ cudaStream_t& stream,
+ paddle::Tensor* out);

csrc/gpu/append_attn/append_attention_bfloat16_int8_kernel.cu

@@ -0,0 +1,51 @@
+// Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+#include "append_attention_kernel.h"
+
+template void CascadeAppendAttentionKernel<paddle::bfloat16, int8_t>(
+ const paddle::Tensor& qkv, // [token_num, num_heads, head_dim]
+ const paddle::Tensor& cache_k, // [max_block_num, num_heads, block_size, head_dim]
+ const paddle::Tensor& cache_v, // [max_block_num, num_heads, head_dim, block_size]
+ const paddle::optional<paddle::Tensor>& attn_mask,
+ const paddle::optional<paddle::Tensor>& cache_k_scale, // [num_kv_heads, head_dim]
+ const paddle::optional<paddle::Tensor>& cache_v_scale, // [num_kv_heads, head_dim]
+ const paddle::optional<paddle::Tensor>& cache_k_zp, // [num_kv_heads, head_dim]
+ const paddle::optional<paddle::Tensor>& cache_v_zp, // [num_kv_heads, head_dim]
+ const paddle::optional<paddle::Tensor>& shift_bias, // [num_kv_heads, head_dim]
+ const paddle::optional<paddle::Tensor>& smooth_weight, // [num_kv_heads, head_dim]
+ const paddle::Tensor& seq_lens_q,
+ const paddle::Tensor& seq_lens_kv,
+ const paddle::Tensor& seq_lens_encoder,
+ const paddle::Tensor& padding_offsets,
+ const paddle::Tensor& cum_offsets,
+ const paddle::Tensor& block_table,
+ const paddle::Tensor& batch_ids,
+ const paddle::Tensor& tile_ids_per_batch,
+ const std::string& cache_quant_type_str,
+ const int num_blocks,
+ const int block_shape_q,
+ const int max_seq_len,
+ const int max_dec_len,
+ const int num_heads,
+ const int kv_num_heads,
+ const int head_dim,
+ const float in_scale,
+ const int max_partition_size,
+ const int encoder_max_partition_size,
+ const int speculate_max_draft_token_num,
+ const bool causal,
+ const bool is_decoder,
+ const bool enable_prefill,
+ cudaStream_t& stream,
+ paddle::Tensor* out);