apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-12-24 13:28:13 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

【Hackathon 9th No.86】autogen MultiQueryDecoderAttention template_instantiation -part (#4383)

Browse Source 
* split MultiQueryDecoderAttention template_instantiation

* update comment

* CI
This commit is contained in:
Zhenghai Zhang
2025-10-16 17:08:19 +08:00
committed by GitHub
parent f72be7a2c8
commit 6adfbe07ad
27 changed files with 3975 additions and 3836 deletions
Download Patch File Download Diff File Expand all files Collapse all files

92
custom_ops/gpu_ops/append_attn/append_attention_kernel.h
View File

@@ -15,97 +15,9 @@
#include "helper.h"
#include "utils.cuh"
#include "append_attention_c16_impl.cuh"
#include "append_attention_c8_impl.cuh"
template <typename T, typename OutT>
void CascadeAppendAttentionC16Kernel(
const AppendAttnMetaData& meta_data,
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& batch_id_per_token,
const paddle::Tensor& cu_seqlens_q,
const paddle::Tensor& block_table,
const paddle::Tensor& batch_ids,
const paddle::Tensor& tile_ids_per_batch,
const int num_blocks,
const int block_shape_q,
const int max_seq_len,
const int max_dec_len,
const float quant_max_bound,
const float quant_min_bound,
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);
template <typename T, typename OutT>
void CascadeAppendAttentionC4Kernel(
const AppendAttnMetaData& meta_data,
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& batch_id_per_token,
const paddle::Tensor& cu_seqlens_q,
const paddle::Tensor& block_table,
const paddle::Tensor& batch_ids,
const paddle::Tensor& tile_ids_per_batch,
const int num_blocks,
const int block_shape_q,
const int max_seq_len,
const int max_dec_len,
const float quant_max_bound,
const float quant_min_bound,
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);
#include "append_attention_c4_impl.cuh"
template <typename T, typename OutT>
void CascadeAppendAttentionKernel(