TensorRT-LLMs/cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/xqaParams.h
Pengbo Wang c0e25e5418
[TRTLLM-10022][feat] Add hopper xqa decode support for skip softmax attention (#10264)
Signed-off-by: Pengbo Wang <221450789+pengbowang-nv@users.noreply.github.com>
2026-01-11 19:26:10 -05:00

225 lines
11 KiB
C++

/*
* Copyright (c) 2020-2025, NVIDIA CORPORATION. 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.
*/
#pragma once
#include "tensorrt_llm/common/config.h"
#include "tensorrt_llm/common/quantization.h"
#include "tensorrt_llm/kernels/gptKernels.h"
#include "tensorrt_llm/kernels/multiHeadAttentionCommon.h"
#include "tensorrt_llm/kernels/sparseAttentionKernels.h"
TRTLLM_NAMESPACE_BEGIN
namespace kernels
{
using XQADataType = Data_type;
struct XQAParams
{
XQADataType data_type = DATA_TYPE_FP16;
XQADataType kv_cache_data_type = DATA_TYPE_FP16;
XQADataType output_data_type = DATA_TYPE_FP16;
void* output = nullptr;
void* output_sf = nullptr;
void const* qkv = nullptr;
int32_t const* cache_indir = nullptr;
float const* attention_sinks = nullptr;
float const* kv_scale_orig_quant = nullptr;
float const* kv_scale_quant_orig = nullptr;
int32_t const* host_past_key_value_lengths = nullptr;
int32_t const* host_context_lengths = nullptr;
int32_t* semaphores = nullptr;
void* workspaces = nullptr;
uint32_t batch_size = 0;
int32_t beam_width = 0;
int32_t chunked_attention_size = INT_MAX;
int32_t max_attention_window_size = 0;
int32_t cyclic_attention_window_size = 0;
int32_t sink_token_length = 0;
int max_past_kv_length = 0;
void const* qkv_bias;
int32_t const* sequence_lengths; //
int32_t const* context_lengths; // maybe not used now
void const* alibi_slopes; // maybe not used now
float const* rotary_embedding_inv_freq_cache; // precomputed rotary inv freq
int32_t const* spec_decoding_packed_mask;
int const* spec_decoding_position_offsets; // for position embedding.
int const* spec_decoding_generation_lengths; // variable input lengths.
bool spec_decoding_is_generation_length_variable; // whether the generation lengths actually vary
int32_t spec_decoding_max_generation_length; // max possible input length
int64_t* spec_decoding_bl_tree_mask_offset; // for blackwell spec-dec tree mask offset
uint32_t* spec_decoding_bl_tree_mask; // for blackwell spec-dec tree mask
int32_t* spec_bl_tree_first_sparse_mask_offset_kv; // for blackwell spec-dec tree first sparse mask offset kv
int32_t const* mrope_position_deltas = nullptr;
// almost copy from GPTAttentionPluginCommon.
// maybe use one struct for parameters in GPTAttentionPluginCommon and share the same here.
int32_t generation_input_length;
int32_t num_q_heads = 0;
int32_t num_kv_heads = 0;
int32_t head_size = 0;
int unidirectional;
float q_scaling = 0;
int32_t rotary_embedding_dim = 0;
float rotary_embedding_base = 0.0f;
tensorrt_llm::kernels::RotaryScalingType rotary_embedding_scale_type;
float rotary_embedding_scale;
int rotary_embedding_max_positions;
int rotary_vision_start;
int rotary_vision_length;
float2 const* rotary_cos_sin;
tensorrt_llm::kernels::PositionEmbeddingType position_embedding_type;
bool position_shift_enabled = false;
bool remove_padding = false;
tensorrt_llm::kernels::AttentionMaskType mask_type;
// Paged KV cache parameters.
bool paged_kv_cache;
int tokens_per_block;
int max_blocks_per_sequence;
tensorrt_llm::common::QuantMode kv_cache_quant_mode;
int tp_size = 1;
int tp_rank = 0;
bool qkv_bias_enabled;
bool cross_attention;
int max_distance = 0;
bool multi_block_mode;
bool multi_query_tokens = false;
bool is_spec_dec_tree
= true; // by default, XQA spec-dec expect tree-based draft token, only affective when multi_query_tokens = true
float const* logn_scaling_ptr = nullptr; // for logn scaling in XQA
int32_t total_num_input_tokens; // total number of input tokens. may differ from batch_size due to medusa.
bool is_fp8_output;
float const* fp8_out_scale = nullptr; // fp8 output scale in case we need post-processing to convert output to fp8.
// nullptr means no conversion.
float const* fp4_out_sf_scale = nullptr; // SF scale for FP4 output.
int32_t start_token_idx_sf = 0; // The start token index in SF tensor.
void* quant_q_buffer_ptr = nullptr;
// for cross attention
int32_t const* encoder_input_lengths = nullptr;
// sparse attention parameters
SparseAttentionParams sparse_params;
bool use_sparse_attention = false;
// Skip softmax threshold.
float skip_softmax_threshold_scale_factor = 0;
#ifdef SKIP_SOFTMAX_STAT
uint32_t* skip_softmax_total_blocks = nullptr;
uint32_t* skip_softmax_skipped_blocks = nullptr;
#endif
cudaStream_t stream = 0;
// layer index
int32_t layer_idx = 0;
std::string toString() const
{
std::stringstream ss;
ss << "XQAParams ====================" << std::endl
<< "data_type: " << static_cast<int>(data_type) << std::endl
<< "kv_cache_data_type: " << static_cast<int>(kv_cache_data_type) << std::endl
<< "output: " << output << std::endl
<< "qkv: " << qkv << std::endl
<< "cache_indir: " << cache_indir << std::endl
<< "kv_scale_orig_quant: " << kv_scale_orig_quant << std::endl
<< "kv_scale_quant_orig: " << kv_scale_quant_orig << std::endl
<< "host_past_key_value_lengths: " << host_past_key_value_lengths << std::endl
<< "host_context_lengths: " << host_context_lengths << std::endl
<< "semaphores: " << semaphores << std::endl
<< "workspaces: " << workspaces << std::endl
<< "batch_size: " << batch_size << std::endl
<< "beam_width: " << beam_width << std::endl
<< "max_attention_window_size: " << max_attention_window_size << std::endl
<< "cyclic_attention_window_size: " << cyclic_attention_window_size << std::endl
<< "sink_token_length: " << sink_token_length << std::endl
<< "max_past_kv_length: " << max_past_kv_length << std::endl
<< "qkv_bias: " << qkv_bias << std::endl
<< "sequence_lengths: " << sequence_lengths << std::endl
<< "context_lengths: " << context_lengths << std::endl
<< "alibi_slopes: " << alibi_slopes << std::endl
<< "rotary_embedding_inv_freq_cache: " << rotary_embedding_inv_freq_cache << std::endl
<< "spec_decoding_packed_mask: " << spec_decoding_packed_mask << std::endl
<< "spec_decoding_position_offsets: " << spec_decoding_position_offsets << std::endl
<< "spec_decoding_generation_lengths: " << spec_decoding_generation_lengths << std::endl
<< "spec_decoding_is_generation_length_variable: "
<< (spec_decoding_is_generation_length_variable ? "true" : "false") << std::endl
<< "spec_decoding_max_generation_length: " << spec_decoding_max_generation_length << std::endl
<< "spec_decoding_bl_tree_mask_offset: " << spec_decoding_bl_tree_mask_offset << std::endl
<< "spec_decoding_bl_tree_mask: " << spec_decoding_bl_tree_mask << std::endl
<< "spec_bl_tree_first_sparse_mask_offset_kv: " << spec_bl_tree_first_sparse_mask_offset_kv << std::endl
<< "mrope_position_deltas: " << mrope_position_deltas << std::endl
<< "generation_input_length: " << generation_input_length << std::endl
<< "num_q_heads: " << num_q_heads << std::endl
<< "num_kv_heads: " << num_kv_heads << std::endl
<< "head_size: " << head_size << std::endl
<< "unidirectional: " << unidirectional << std::endl
<< "q_scaling: " << q_scaling << std::endl
<< "rotary_embedding_dim: " << rotary_embedding_dim << std::endl
<< "rotary_embedding_base: " << rotary_embedding_base << std::endl
<< "rotary_embedding_scale_type: " << static_cast<int>(rotary_embedding_scale_type) << " (enum value)"
<< std::endl
<< "rotary_embedding_scale: " << rotary_embedding_scale << std::endl
<< "rotary_embedding_max_positions: " << rotary_embedding_max_positions << std::endl
<< "rotary_vision_start: " << rotary_vision_start << std::endl
<< "rotary_vision_length: " << rotary_vision_length << std::endl
<< "rotary_cos_sin: " << rotary_cos_sin << std::endl
<< "position_embedding_type: " << static_cast<int>(position_embedding_type) << " (enum value)" << std::endl
<< "position_shift_enabled: " << (position_shift_enabled ? "true" : "false") << std::endl
<< "remove_padding: " << (remove_padding ? "true" : "false") << std::endl
<< "mask_type: " << static_cast<int>(mask_type) << " (enum value)" << std::endl
<< "paged_kv_cache: " << (paged_kv_cache ? "true" : "false") << std::endl
<< "tokens_per_block: " << tokens_per_block << std::endl
<< "max_blocks_per_sequence: " << max_blocks_per_sequence << std::endl
<< "tp_size: " << tp_size << std::endl
<< "tp_rank: " << tp_rank << std::endl
<< "qkv_bias_enabled: " << (qkv_bias_enabled ? "true" : "false") << std::endl
<< "cross_attention: " << (cross_attention ? "true" : "false") << std::endl
<< "max_distance: " << max_distance << std::endl
<< "multi_block_mode: " << (multi_block_mode ? "true" : "false") << std::endl
<< "multi_query_tokens: " << (multi_query_tokens ? "true" : "false") << std::endl
<< "logn_scaling_ptr :" << logn_scaling_ptr << std ::endl
<< "total_num_input_tokens :" << total_num_input_tokens << std ::endl
<< "is_fp8_output :" << (is_fp8_output ? "true" : "false") << std ::endl
<< "fp8_out_scale :" << fp8_out_scale << std ::endl
<< "encoder_input_lengths: " << encoder_input_lengths << std::endl
<< "sparse_params: " << sparse_params.toString() << std::endl
<< "use_sparse_attention :" << (use_sparse_attention ? "true" : "false") << std ::endl
<< "skip_softmax_threshold_scale_factor :" << skip_softmax_threshold_scale_factor << std ::endl
#ifdef SKIP_SOFTMAX_STAT
<< "skip_softmax_total_blocks :" << skip_softmax_total_blocks << std ::endl
<< "skip_softmax_skipped_blocks :" << skip_softmax_skipped_blocks << std ::endl
#endif
<< "stream :" << stream;
return ss.str();
}
bool isMLA() const
{
return head_size == 576 && num_q_heads == 128 && num_kv_heads == 1;
}
};
} // namespace kernels
TRTLLM_NAMESPACE_END