TensorRT-LLMs/cpp/tensorrt_llm/kernels/beamSearchKernels/beamSearchKernelsTemplate.h
2024-04-16 19:40:08 +08:00

910 lines
38 KiB
C++

/*
* Copyright (c) 2020-2023, 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.
*/
#ifndef CUDART_VERSION
#error CUDART_VERSION Undefined!
#elif (CUDART_VERSION >= 11050)
#include <cub/cub.cuh>
#else
#include "3rdparty/cub/cub.cuh"
#endif
#include "tensorrt_llm/common/assert.h"
#include "tensorrt_llm/common/cudaUtils.h"
#include "tensorrt_llm/common/reduceKernelUtils.cuh"
#include "tensorrt_llm/common/stringUtils.h"
#include "tensorrt_llm/kernels/beamSearchKernels.h"
#include "tensorrt_llm/kernels/decodingCommon.h"
using namespace tensorrt_llm::common;
namespace tensorrt_llm
{
namespace kernels
{
#define DO_SPLIT_SMALL_TOP_K_SOFTMAX
#define TOPK_FP16_STORAGE 0
#pragma nv_diag_suppress static_var_with_dynamic_init
template <typename T, int MAX_K2, int THREADBLOCK_SIZE>
__launch_bounds__(THREADBLOCK_SIZE) __global__
void batchBeamKernel(int const* __restrict topk_id_buffer, T const* __restrict topk_val_buffer, BeamHypotheses bh)
{
int const tid = threadIdx.x;
int const bid = blockIdx.x;
int const gbid{bh.ite * bh.local_batch_size + bid}; // global batch index
int const K{bh.beam_width};
int const V{bh.vocab_size};
int const nCandidate{K * K * 2};
T const MAX_T_VAL = (std::is_same<T, half>::value) ? HALF_FLT_MAX : FLT_MAX;
float const diversity_rate{bh.diversity_rates[gbid]};
float const length_penalty{bh.length_penalties[gbid]};
int const early_stopping{bh.early_stoppings[gbid]};
__shared__ int nBeamForNextStep;
__shared__ float smem_cum_log_probs[MAX_K2 / 2];
if (tid == 0)
{
nBeamForNextStep = 0;
}
if (tid < K)
{
smem_cum_log_probs[tid] = bh.cum_log_probs[bid * K + tid];
}
__syncthreads();
if (bh.num_beams != nullptr)
{
// Beam search is enabled
if (bh.num_beams[gbid] == 0 && tid == 0)
{
// Initialize worst_score in the first time
bh.min_normed_scores[gbid] = FLT_MAX;
}
else if (early_stopping == 1 && bh.num_beams[gbid] == K
|| early_stopping != 1 && bh.finished[bid * K].isFinished())
{
// New but false condition:
// else if (early_stopping == 1 && bh.num_beams[gbid] == K || early_stopping != 1 && bh.is_done[bid])
// Condition of early return:
// 1. In EarlyStopping mode, and we have got enough beams
// 2. In NonEarlyStopping mode, and this batch has been marked as done
return;
}
}
// Get top 2K tokens from candidates
topk_id_buffer += bid * nCandidate;
topk_val_buffer += bid * nCandidate;
using cub_kvp = cub::KeyValuePair<int, T>;
cub_kvp partial_topk{nCandidate - 1, -MAX_T_VAL};
cub::ArgMax arg_max;
extern __shared__ char smem[];
T* smem_topk = reinterpret_cast<T*>(smem);
for (int id = tid; id < nCandidate; id += THREADBLOCK_SIZE)
{
int const index = bh.num_beams == nullptr ? id % K : id / 2 / K;
T val = topk_val_buffer[id] + static_cast<T>(diversity_rate * index);
cub_kvp new_elem{id, val};
partial_topk = arg_max(partial_topk, new_elem);
smem_topk[id] = val;
}
__syncthreads();
using BlockReduce = cub::BlockReduce<cub_kvp, THREADBLOCK_SIZE>;
__shared__ typename BlockReduce::TempStorage reduce_buffer;
__shared__ cub_kvp cta_topk[MAX_K2];
__shared__ int thread_requiring_update;
for (int i = 0; i < 2 * K; ++i)
{
cub_kvp total_topk = BlockReduce(reduce_buffer).Reduce(partial_topk, arg_max);
if (tid == 0)
{
cta_topk[i] = total_topk;
smem_topk[total_topk.key] = -MAX_T_VAL;
thread_requiring_update = total_topk.key % THREADBLOCK_SIZE;
}
__syncthreads();
// Only one thread needs to update the old partial before the next block reduce.
// No need to do this in the last iteration.
if (tid == thread_requiring_update && i < (2 * K - 1))
{
partial_topk.key = nCandidate - 1;
partial_topk.value = -MAX_T_VAL;
for (int index = tid; index < nCandidate; index += THREADBLOCK_SIZE)
{
cub_kvp new_elem{index, smem_topk[index]};
partial_topk = arg_max(partial_topk, new_elem);
}
}
}
if (tid == 0)
{
// Adjust beams or select completed beams sequentially
// Reference (might be changed along HF in the future):
// https://github.com/huggingface/transformers/blob/main/src/transformers/generation/beam_search.py#L272
for (int i = 0; i < 2 * K; ++i)
{
int const current_key = cta_topk[i].key;
T const current_value = cta_topk[i].value;
bool const is_end_token = topk_id_buffer[current_key] % V == bh.end_ids[bid];
if (i < K && bh.num_beams != nullptr && is_end_token)
{
// Condition of this branch
// In Beam search mode, this token is end_token and belongs to top K range in Beam search mode
int const seq_len = bh.seq_len[bid * K + i] + 1 - bh.input_lengths[gbid * K + i];
float const normed_score = applyLengthPenalty(current_value, seq_len, length_penalty);
int beam_idx = bh.num_beams[gbid];
if (beam_idx == K)
{
// There are already K beams
if (normed_score < bh.min_normed_scores[gbid])
{
// Current score is worse than the worst one in candidate beams
if (early_stopping)
{
// Stop since we have got enough beams
break;
}
else
{
// Continue since there might be longer but better beams
continue;
}
}
else
{
// Current score is better than the worst one in candidate beams
// Find the candidate beam index with the worst score and erase it
for (int j = 0; j < K; j++)
{
if (bh.normed_scores_cba[gbid * (K * 2) + j] == bh.min_normed_scores[gbid])
{
beam_idx = j;
bh.num_beams[gbid]--;
bh.min_normed_scores[gbid] = FLT_MAX;
bh.normed_scores_cba[gbid * (K * 2) + j] = normed_score;
for (int l = 0; l < K; l++)
{
bh.min_normed_scores[gbid]
= min(bh.min_normed_scores[gbid], bh.normed_scores_cba[gbid * (K * 2) + l]);
}
break;
}
}
}
}
int prev_id = (topk_id_buffer[current_key] / V) % K;
int const current_step = bh.seq_len[bid * K + prev_id];
int const tgt_id_offset = ((bid + bh.ite * bh.local_batch_size) * (K * 2) + beam_idx) * bh.max_seq_len;
bh.output_ids_cba[tgt_id_offset + current_step] = bh.end_ids[bid];
if (bh.log_probs_cba != nullptr)
{
bh.log_probs_cba[tgt_id_offset + current_step] = (float) topk_val_buffer[current_key]
- smem_cum_log_probs[(topk_id_buffer[current_key] / V) % K];
}
// Write finished beam from work tree to CBA
for (int j = current_step - 1; j >= 0; j--)
{
bh.output_ids_cba[tgt_id_offset + j] = bh.output_ids_ptr[bid][prev_id * bh.max_seq_len + j];
prev_id = bh.parent_ids_ptr[bid][prev_id * bh.max_seq_len + j];
}
if (bh.log_probs_cba != nullptr && bh.log_probs != nullptr)
{
prev_id = (topk_id_buffer[current_key] / V) % K;
for (int j = current_step - 1; j >= 0; j--)
{
int const index = j * bh.batch_size * K + bh.ite * bh.local_batch_size * K + bid * K + prev_id;
bh.log_probs_cba[tgt_id_offset + j] = bh.log_probs[index];
prev_id = bh.parent_ids_ptr[bid][prev_id * bh.max_seq_len + j];
}
}
int const tgt_beam_idx = gbid * (K * 2) + beam_idx;
bh.seq_len_cba[tgt_beam_idx] = current_step;
bh.normed_scores_cba[tgt_beam_idx] = normed_score;
bh.min_normed_scores[gbid] = min(bh.min_normed_scores[gbid], bh.normed_scores_cba[tgt_beam_idx]);
bh.num_beams[gbid]++;
bh.cum_log_probs_cba[tgt_beam_idx] = (float) topk_val_buffer[current_key];
}
else if (i < K || bh.num_beams != nullptr && !is_end_token)
{
// Condition of this branch
// 1. bh.num_beams == nullptr && i < K, i.e., beam search is disable
// 2. bh.num_beams != nullptr && i < K && is_end_token == false, i.e., add token at the end
// 3. bh.num_beams != nullptr && i >= K && is_end_token == false, i.e., add token at the end
int const current_step = bh.seq_len[bid * K + nBeamForNextStep];
// Write the selected token to work tree
bh.output_ids_ptr[bid][nBeamForNextStep * bh.max_seq_len + current_step] = topk_id_buffer[current_key];
if (bh.log_probs != nullptr)
{
bh.log_probs[current_step * bh.batch_size * K + bid * K + nBeamForNextStep]
= (float) topk_val_buffer[current_key]
- smem_cum_log_probs[(topk_id_buffer[current_key] / V) % K];
}
bh.cum_log_probs[bid * K + nBeamForNextStep] = (float) topk_val_buffer[current_key];
nBeamForNextStep++;
}
else
{
// Condition of this branch, which we do nothing for it
// 1. bh.num_beams == nullptr && i >= K, i.e., beam search is disable
// 2. bh.num_beams != nullptr && i >= K && is_end_token == true, i.e., ignore the worse beams
}
// if (early_stopping == 1 && bh.num_beams[gbid] >= K || nBeamForNextStep >= K)
if (nBeamForNextStep >= K)
{
// Condition of this branch:
// 1. In EarlyStopping mode, and get enough candidate beams
// 2. In EarlyStopping mode, and get enough tokens for the next generation step
// 3. In NonEarlyStopping mode, and get enough tokens for the next generation step
break;
}
}
}
// Update bh.is_done
if (tid == 0 && bh.num_beams != nullptr)
{
if (bh.num_beams[bid] < K)
{
// no enough beams
bh.is_done[bid] = false;
}
else if (early_stopping == 1)
{
// enough candidate beams in EarlyStopping mode
bh.is_done[bid] = true;
}
else
{
// enough beams in NonEarlyStopping mode
int seq_len = bh.seq_len[bid * K] + 1 - bh.input_lengths[gbid * K];
float const best_sum_logprobs = cta_topk[0].value;
// According to semantics of HF, cta_topk[0].value is used as best_sum_logprobs
// But maybe bh.cum_log_probs[bid * K + i] is more suitable?
// https://github.com/huggingface/transformers/blob/main/src/transformers/generation/beam_search.py#L307
if (early_stopping != 0 && length_penalty > 0.0f)
{
// Specialization for early_stopping == "never" and length_penalty > 0 in HF
seq_len = bh.max_seq_len - bh.input_lengths[gbid * K];
}
float const highest_attainable_score = applyLengthPenalty(best_sum_logprobs, seq_len, length_penalty);
bh.is_done[bid] = bh.min_normed_scores[gbid] >= highest_attainable_score;
}
}
__syncthreads();
// Update sequence_lengths, parent_ids, output_ids and finished
__shared__ int s_sequence_lengths[MAX_K2 / 2];
if (tid < K)
{
s_sequence_lengths[tid] = bh.seq_len[bid * K + tid];
}
__syncthreads();
if (tid < K)
{
int const bb_index = bid * K + tid;
int const current_step = s_sequence_lengths[tid];
if (!bh.finished[bb_index].isFinished())
{
s_sequence_lengths[tid]++;
}
int const new_id = bh.output_ids_ptr[bid][tid * bh.max_seq_len + current_step];
int const new_beam_id = (new_id / V) % K;
int const new_word_id = new_id % V;
bh.seq_len[bb_index] = s_sequence_lengths[new_beam_id];
if (new_word_id == bh.end_ids[bid])
{
bh.finished[bb_index].setFinishedEOS();
}
bh.parent_ids_ptr[bid][tid * bh.max_seq_len + current_step] = new_beam_id;
bh.output_ids_ptr[bid][tid * bh.max_seq_len + current_step] = new_word_id;
if ((early_stopping == 1) && (bh.num_beams != nullptr && bh.num_beams[gbid] == K)
|| (early_stopping != 1) && bh.is_done[bid])
{
bh.is_done[bid] = true;
bh.finished[bb_index].setFinished();
}
}
}
struct __align__(8) MD
{
float m;
float d;
};
__device__ __forceinline__ MD reduce_md_op(MD a, MD b)
{
bool const is_a_bigger = a.m > b.m;
MD const bigger = is_a_bigger ? a : b;
MD const smaller = is_a_bigger ? b : a;
MD res{bigger.m, bigger.d + smaller.d * __expf(smaller.m - bigger.m)};
return res;
}
template <typename T, int MAX_K>
struct TopKMD
{
MD md;
TopK<T, MAX_K> topk;
};
template <typename T, int MAX_K>
__device__ __forceinline__ TopKMD<T, MAX_K> reduce_topk_md_op(TopKMD<T, MAX_K> const& a, TopKMD<T, MAX_K> const& b)
{
TopKMD<T, MAX_K> res;
res.md = reduce_md_op(a.md, b.md);
res.topk = reduce_topk_op(a.topk, b.topk);
return res;
}
template <typename T, int ITEMS_PER_THREAD, int MAX_K, int THREADBLOCK_SIZE>
__launch_bounds__(THREADBLOCK_SIZE) __global__ void beamKernel(T const* __restrict logits, T const* __restrict bias,
float const* __restrict cum_log_probs, FinishedState const* __restrict finished, int* __restrict topk_id_buffer,
T* __restrict topk_val_buffer, int V, int K, int const* __restrict end_ids)
{
int const tid = threadIdx.x;
int const bid = blockIdx.x;
T const MAX_T_VAL = (std::is_same<T, half>::value) ? HALF_FLT_MAX : FLT_MAX;
TopKMD<float, MAX_K> partial;
partial.md.m = -MAX_T_VAL;
partial.md.d = 0.0F;
partial.topk.init();
if (finished[bid].isFinished())
{
for (int id = tid; id < V; id += THREADBLOCK_SIZE)
{
float const val = id == end_ids[bid / K] ? MAX_T_VAL : -MAX_T_VAL;
MD new_elem{val, 1.0F};
partial.md = reduce_md_op(partial.md, new_elem);
partial.topk.insert(val, id);
}
}
else
{
T const* local_logits = logits + bid * V;
for (int id = tid; id < V; id += THREADBLOCK_SIZE)
{
float const val = local_logits[id] + bias[id];
MD new_elem{val, 1.0F};
partial.md = reduce_md_op(partial.md, new_elem);
partial.topk.insert(val, id);
}
}
typedef cub::BlockReduce<TopKMD<float, MAX_K>, THREADBLOCK_SIZE> BlockReduce;
__shared__ typename BlockReduce::TempStorage reduce_buffer;
TopKMD<float, MAX_K> total = BlockReduce(reduce_buffer).Reduce(partial, reduce_topk_md_op<float, MAX_K>);
if (tid == 0)
{
int* local_topk_id = topk_id_buffer + bid * K;
T const* local_topk_val = topk_val_buffer + bid * K;
float const total_m = total.md.m;
float const total_d = logf(total.md.d);
float local_cum_log_probs = cum_log_probs[bid];
for (int i = 0; i < K; ++i)
{
local_topk_id[i] = total.topk.p[i] + bid * V;
local_topk_val[i] = total.topk.u[i] - total_m - total_d + local_cum_log_probs;
}
}
}
template <typename T, int ITEMS_PER_THREAD, int MAX_K2, int THREADBLOCK_SIZE>
__launch_bounds__(THREADBLOCK_SIZE, 1) __global__ void beamStage1BaseKernel(T const* __restrict logits,
T const* __restrict bias, FinishedState const* __restrict finished, float* __restrict temp_buffer, int V, int K,
int const* __restrict end_ids)
{
// Compare to beamStage1FastKernel, here is no share memory for storage of logits,
// and each ThreadBlock is responsible for `V / voc_parts` elements
constexpr int PACKED_TOP_KMD_SIZE = 2 * MAX_K2 + 2;
int const tid = threadIdx.x;
int const bid = blockIdx.x;
int const V_local = (V + gridDim.y - 1) / gridDim.y;
int const section_start = V_local * blockIdx.y;
int const section_end = std::min(section_start + V_local, V);
T const MAX_T_VAL = (std::is_same<T, half>::value) ? HALF_FLT_MAX : FLT_MAX;
// Load element from logits to do reduce_md and arg_max meanwhile
#if TOPK_FP16_STORAGE == 1
TopKMD<__half, MAX_K2> partial;
#else
TopKMD<T, MAX_K2> partial;
#endif
partial.md.m = -MAX_T_VAL;
partial.md.d = 0.0F;
partial.topk.init();
if (finished[bid].isFinished())
{
#pragma unroll 1
for (int id = section_start + tid; id < section_end; id += THREADBLOCK_SIZE)
{
float const val = (id == end_ids[bid / K]) ? MAX_T_VAL : -MAX_T_VAL;
MD const new_elem_md{val, 1.0F};
partial.md = reduce_md_op(partial.md, new_elem_md);
partial.topk.insert(val, id);
}
}
else
{
T const* local_logits = logits + bid * V;
#pragma unroll 1
for (int id = section_start + tid; id < section_end; id += THREADBLOCK_SIZE)
{
T const b = bias == nullptr ? (T) 0.0f : bias[id];
T const val = local_logits[id] + b;
MD new_elem_md{val, 1.0F};
partial.md = reduce_md_op(partial.md, new_elem_md);
partial.topk.insert(val, id);
}
}
// Search the top 2K elements among `V` elements and write into smem_output
#if TOPK_FP16_STORAGE == 1
typedef cub::BlockReduce<TopKMD<__half, MAX_K2>, THREADBLOCK_SIZE> BlockReduce;
__shared__ typename BlockReduce::TempStorage reduce_buffer;
TopKMD<__half, MAX_K2> total = BlockReduce(reduce_buffer).Reduce(partial, reduce_topk_md_op<__half, MAX_K2>);
#else
typedef cub::BlockReduce<TopKMD<T, MAX_K2>, THREADBLOCK_SIZE> BlockReduce;
__shared__ typename BlockReduce::TempStorage reduce_buffer;
TopKMD<T, MAX_K2> total = BlockReduce(reduce_buffer).Reduce(partial, reduce_topk_md_op<T, MAX_K2>);
#endif
__shared__ float smem_output[PACKED_TOP_KMD_SIZE];
if (tid == 0)
{
for (int i = 0; i < 2 * K; i++)
{
int const index = bid * V + total.topk.p[i];
reinterpret_cast<int*>(smem_output)[i] = index;
smem_output[MAX_K2 + i] = total.topk.u[i];
}
smem_output[2 * MAX_K2] = total.md.d;
smem_output[2 * MAX_K2 + 1] = total.md.m;
}
__syncthreads();
// Write the smem_output into temp_buffer
float* local_temp_buffer = temp_buffer + bid * PACKED_TOP_KMD_SIZE * gridDim.y + blockIdx.y * PACKED_TOP_KMD_SIZE;
#pragma unroll
for (int id = tid; id < PACKED_TOP_KMD_SIZE; id += THREADBLOCK_SIZE)
{
local_temp_buffer[id] = smem_output[id];
}
}
template <typename T, int ITEMS_PER_THREAD, int MAX_K2, int THREADBLOCK_SIZE>
__launch_bounds__(THREADBLOCK_SIZE, 1) __global__ void beamStage1FastKernel(T const* __restrict logits,
T const* __restrict bias, FinishedState const* __restrict finished, float* __restrict temp_buffer, int V, int K,
int const* __restrict end_ids, int const V_local)
{
constexpr int PACKED_TOP_KMD_SIZE = 2 * MAX_K2 + 2;
int const tid = threadIdx.x;
int const bid = blockIdx.x;
int const section_start = V_local * blockIdx.y;
int const section_end = std::min(section_start + V_local, V);
int const valid_smem_length = section_end - section_start;
T const MAX_T_VAL = (std::is_same<T, half>::value) ? HALF_FLT_MAX : FLT_MAX;
// Load element from logits to smem_logprobs, doing reduce_md and arg_max meanwhile
// Each thread is responsible for `V_local / THREADBLOCK_SIZE` elements
extern __shared__ char smem_[];
T* smem_logprobs = reinterpret_cast<T*>(smem_);
MD partial_md{-MAX_T_VAL, 0.0f};
#if TOPK_FP16_STORAGE == 1
using cub_kvp = cub::KeyValuePair<int, __half>;
#else
using cub_kvp = cub::KeyValuePair<int, T>;
#endif
cub_kvp partial_topk{V - 1, -MAX_T_VAL};
cub::ArgMax arg_max;
if (finished[bid].isFinished())
{
#pragma unroll 1
for (int id = section_start + tid; id < section_end; id += THREADBLOCK_SIZE)
{
float const val = (id == end_ids[bid / K]) ? MAX_T_VAL : -MAX_T_VAL;
int const smem_index = id - section_start;
smem_logprobs[smem_index] = val;
MD const new_elem_md{val, 1.0F};
partial_md = reduce_md_op(partial_md, new_elem_md);
cub_kvp const new_elem_topk{smem_index, val};
partial_topk = arg_max(partial_topk, new_elem_topk);
}
}
else
{
T const* local_logits = logits + bid * V;
#pragma unroll 1
for (int id = section_start + tid; id < section_end; id += THREADBLOCK_SIZE)
{
T const b = bias == nullptr ? (T) 0.0f : bias[id];
T const val = local_logits[id] + b;
int const smem_index = id - section_start;
smem_logprobs[smem_index] = val;
MD new_elem_md{val, 1.0F};
partial_md = reduce_md_op(partial_md, new_elem_md);
cub_kvp new_elem_topk{smem_index, val};
partial_topk = arg_max(partial_topk, new_elem_topk);
}
}
__syncthreads();
// Search the top 2K elements among `V_local` elements of this ThreadBlock and write into smem_output
__shared__ float smem_output[PACKED_TOP_KMD_SIZE];
__shared__ int thread_requiring_update;
using BlockReduceMD = cub::BlockReduce<MD, THREADBLOCK_SIZE>;
using BlockReduceTopK = cub::BlockReduce<cub_kvp, THREADBLOCK_SIZE>;
__shared__ union
{
typename BlockReduceMD::TempStorage md_smem;
typename BlockReduceTopK::TempStorage topk_smem;
} reduce_buffer;
for (int i = 0; i < 2 * K; ++i)
{
// Pop the element with largest value to "smem_output" per iteration
cub_kvp total_topk = BlockReduceTopK(reduce_buffer.topk_smem).Reduce(partial_topk, arg_max);
if (tid == 0)
{
int const index = bid * V + section_start + total_topk.key;
reinterpret_cast<int*>(smem_output)[i] = index;
smem_output[MAX_K2 + i] = total_topk.value;
smem_logprobs[total_topk.key] = -MAX_T_VAL; // pollute the value of the popped element
thread_requiring_update = total_topk.key % THREADBLOCK_SIZE;
}
__syncthreads();
if (tid == thread_requiring_update && i < 2 * K - 1)
{
// The thread popped the element need to update its partial_topk
// No need to do this in the last iteration
partial_topk.key = V - 1;
partial_topk.value = -MAX_T_VAL;
for (int index = tid; index < valid_smem_length; index += THREADBLOCK_SIZE)
{
cub_kvp new_elem{index, smem_logprobs[index]};
partial_topk = arg_max(partial_topk, new_elem);
}
}
}
// Do reduce_md among the top 2K elements in the smem_output and write into tail of smem_output
auto reduce_md_func = [](const MD& a, const MD& b) { return reduce_md_op(a, b); };
MD total_md = BlockReduceMD(reduce_buffer.md_smem).Reduce(partial_md, reduce_md_func);
if (tid == 0)
{
smem_output[2 * MAX_K2] = total_md.d;
smem_output[2 * MAX_K2 + 1] = total_md.m;
}
__syncthreads();
// Write the smem_output into temp_buffer
float* local_temp_buffer = temp_buffer + bid * PACKED_TOP_KMD_SIZE * gridDim.y + blockIdx.y * PACKED_TOP_KMD_SIZE;
#pragma unroll
for (int id = tid; id < PACKED_TOP_KMD_SIZE; id += THREADBLOCK_SIZE)
{
local_temp_buffer[id] = smem_output[id];
}
}
template <typename T, int MAX_K2, int THREADBLOCK_SIZE>
__launch_bounds__(THREADBLOCK_SIZE) __global__
void beamStage2Kernel(float const* __restrict temp_buffer, float const* __restrict cum_log_probs,
int* __restrict topk_id_buffer, T* __restrict topk_val_buffer, int const K, int const voc_parts, int const V)
{
constexpr int PACKED_TOP_KMD_SIZE = 2 * MAX_K2 + 2;
int const bid = blockIdx.x;
int const tid = threadIdx.x;
T const MAX_T_VAL = (std::is_same<T, half>::value) ? HALF_FLT_MAX : FLT_MAX;
using cub_kvp = cub::KeyValuePair<int, T>;
using BlockReduceTopK = cub::BlockReduce<cub_kvp, THREADBLOCK_SIZE>;
using BlockReduceMD = cub::BlockReduce<MD, THREADBLOCK_SIZE>;
extern __shared__ char smem[];
float* smem_topk = reinterpret_cast<float*>(smem);
__shared__ cub_kvp buf_smem_kv[MAX_K2];
__shared__ union
{
typename BlockReduceTopK::TempStorage topk_smem;
typename BlockReduceMD::TempStorage md_smem;
} shared_temp_storage;
cub::ArgMax arg_max;
MD partial_md{-MAX_T_VAL, 0.0f};
cub_kvp total_topk{V - 1, -MAX_T_VAL};
auto reduce_md_func = [](const MD& a, const MD& b) { return reduce_md_op(a, b); };
// Load and unpack into registers through smem
float const* local_temp_storage = temp_buffer + PACKED_TOP_KMD_SIZE * bid * voc_parts;
for (int idx = tid; idx < PACKED_TOP_KMD_SIZE * voc_parts; idx += THREADBLOCK_SIZE)
{
smem_topk[idx] = local_temp_storage[idx];
}
__syncthreads();
// Find the argmax within each voc_parts
// Find the topK across all voc_parts
for (int k = 0; k < 2 * K; ++k)
{
cub_kvp partial_topk{V - 1, -MAX_T_VAL};
// Only threads responsible for a chunk will do the computation
if (tid < voc_parts)
{
for (int i = 0; i < 2 * K; ++i)
{
int const current_index = tid * PACKED_TOP_KMD_SIZE + i;
T current_value = smem_topk[current_index + MAX_K2];
cub_kvp new_elem = {current_index, current_value};
partial_topk = arg_max(partial_topk, new_elem);
}
}
cub_kvp total_topk = BlockReduceTopK(shared_temp_storage.topk_smem).Reduce(partial_topk, arg_max);
__syncthreads();
if (tid == 0)
{
// Store kv pairs in shared mem buffer
int temp_offset = total_topk.key;
int global_offset = reinterpret_cast<int*>(smem_topk)[temp_offset];
total_topk.key = global_offset;
buf_smem_kv[k] = total_topk;
// Invalidate the maximum value within the chunk
reinterpret_cast<int*>(smem_topk)[temp_offset] = V - 1; // id in share memory
smem_topk[temp_offset + MAX_K2] = -MAX_T_VAL; // value in share memory
}
__syncthreads();
}
// Extract and reduce MD values across the chunks
if (tid < voc_parts)
{
partial_md.d = smem_topk[tid * PACKED_TOP_KMD_SIZE + 2 * MAX_K2];
partial_md.m = smem_topk[tid * PACKED_TOP_KMD_SIZE + 2 * MAX_K2 + 1];
}
__syncthreads();
MD total_md = BlockReduceMD(shared_temp_storage.md_smem).Reduce(partial_md, reduce_md_func);
if (tid == 0)
{
float d_total_log = logf(total_md.d);
for (int i = 0; i < MAX_K2; ++i)
{
float val = (float) buf_smem_kv[i].value - total_md.m - d_total_log;
if (i < 2 * K)
{
topk_id_buffer[bid * 2 * K + i] = buf_smem_kv[i].key;
topk_val_buffer[bid * 2 * K + i] = val + cum_log_probs[bid];
}
}
}
}
template <typename T, int MAX_K2>
void beamStage2KernelLauncher(float const* temp_buffer, float const* cum_log_probs, int* topk_id_buffer,
T* topk_val_buffer, int const batch_size, int const beam_width, int const voc_parts, int const V,
cudaStream_t stream)
{
// TODO: rewrite kernel to remove dependence of constant block size to reduce compilation time
size_t const smem_size = sizeof(float) * voc_parts * (2 * MAX_K2 + 2);
if (voc_parts <= 32)
{
beamStage2Kernel<T, MAX_K2, 32><<<batch_size * beam_width, 32, smem_size, stream>>>(
temp_buffer, cum_log_probs, topk_id_buffer, topk_val_buffer, beam_width, voc_parts, V);
return;
}
if (voc_parts <= 64)
{
beamStage2Kernel<T, MAX_K2, 64><<<batch_size * beam_width, 64, smem_size, stream>>>(
temp_buffer, cum_log_probs, topk_id_buffer, topk_val_buffer, beam_width, voc_parts, V);
return;
}
if (voc_parts <= 128)
{
beamStage2Kernel<T, MAX_K2, 128><<<batch_size * beam_width, 128, smem_size, stream>>>(
temp_buffer, cum_log_probs, topk_id_buffer, topk_val_buffer, beam_width, voc_parts, V);
return;
}
assert(0);
}
template <typename T, int MAX_K>
void topK_softMax_kernelLauncher(
T const* logits, T const* bias, void* workspace, BeamHypotheses& bh, cudaStream_t stream)
{
// Workflow of this function (reference: https://github.com/NVIDIA/online-softmax)
// Using batch_size (BS) = 2, beam_width (BM) = 5, vocab_size (V) = 32000 as an example:
// nPaddedBeamWidth (pBM) = 8 = 2 ^ ceil(log(BM)), nSmallTopKMaxVocParts (nVP) = 128 (Constant)
// MAX_K = 8 = pBM, MAX_K2 = 16 = 2 * pBM
// logits.shape = [BS, BM, V]
// blockSize = 128, voc_parts = 13, voc_size_chunk = 2462 = ceil(32000/13)
// The content of workspace (length aligned to 4):
// | allocated size | used size | data type |
// ┏━━━━━━━━━━━━━━━━━┓ ---------------------------------------------------------------------------
// ┃ topk_id_buffer ┃ BS * pBM * pBM * 2 | | int |
// ┣━━━━━━━━━━━━━━━━━┫ -------------------------------------- Change "pBM" into "BM" -------------
// ┃ topk_val_buffer ┃ BS * pBM * pBM * 2 | | float |
// ┣━━━━━━━━━━━━━━━━━┫ -------------------------------------- in the left formulas -------------
// ┃ temp_buffer ┃ BS * pBM * nVP * (2 * (pBM * 2) + 2) | | float |
// ┗━━━━━━━━━━━━━━━━━┛ ---------------------------------------------------------------------------
// Stage1: gridDim(BS*BM,voc_parts,1), blockDim(blockSize,1,1)
// Each ThreadBlock takes `voc_size_chunk` contiguous elements in logits to do TopK and reduce_md,
// then writes output into temp_buffer.
// At end of this kernel, each ThreadBlock holds the indexes and values of the top 2*K elements,
// as well as the m(x) and l(x) of those elements (see paper of Flash Attention, arXiv:2205.14135)
// temp_buffer.shape = [BS*BM, voc_parts, 2*MAX_K2+2]
// The content of the last dimension of temp_buffer (updated by each ThreadBlock, we call it "Tile"):
// ┏━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━┓
// ┃ topk_id ┃ topk_val ┃ md ┃
// ┗━━━━━━━━━┻━━━━━━━━━━┻━━━━━━━┛
// | allocated size | MAX_K2 | MAX_K2 | 2 |
// | used size | 2*BM | 2*BM | 2 |
// | data type | int | float | float |
// Stage2: gridDim(BS*BM,1,1), blockDim(32/64/128,1,1)
// Each TheadBlock takes `voc_parts` contiguous Tiles in temp_buffer to do reduce_topk and reduce_md,
// writes output topk_id into in topk_id_buffer, writes topk_value + cum_log_probs into topk_val_buffer.
// batchBeamKernel: gridDim(BS,1,1), blockDim(128,1,1)
// Each TheadBlock is responsible for one batch, doing work below:
// + moves one beam into candidate-beam-array if it is finished (gemerated end_id in this step).
// + selects BM elements for the next generation step if not.
// + maintains related score array, min_normed_score / is_done / finished, etc..
constexpr int items_per_thread = 1;
constexpr int blockSize = (MAX_K < 16) ? ((MAX_K < 8) ? nSmallTopKBlockSize : 128) : 64;
int const batch_size{bh.local_batch_size};
int const beam_width{bh.beam_width};
int const V{bh.vocab_size};
int const* end_ids{bh.end_ids};
float* cum_log_probs{bh.cum_log_probs};
FinishedState const* finished{bh.finished};
int const offset = roundUp(batch_size * beam_width * beam_width * 2, 4);
int* topk_id_buffer = reinterpret_cast<int*>(workspace);
T* topk_val_buffer = reinterpret_cast<T*>(topk_id_buffer + offset);
float* temp_buffer = reinterpret_cast<float*>(topk_val_buffer + offset);
#ifdef DO_SPLIT_SMALL_TOP_K_SOFTMAX
// Upper limit count of ThreadBlock, gotten by using no share memory
int max_active_blocks = -1;
TLLM_CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(
&max_active_blocks, beamStage1FastKernel<T, items_per_thread, 2 * MAX_K, blockSize>, blockSize, 0));
// Find the max smem on the device and use that to determine the vocab parts in the best case.
int max_smem_per_sm = -1;
int max_smem_per_block = -1;
int const device = tensorrt_llm::common::getDevice();
TLLM_CUDA_CHECK(cudaDeviceGetAttribute(&max_smem_per_sm, cudaDevAttrMaxSharedMemoryPerMultiprocessor, device));
TLLM_CUDA_CHECK(cudaDeviceGetAttribute(&max_smem_per_block, cudaDevAttrMaxSharedMemoryPerBlockOptin, device));
cudaFuncAttributes attr;
TLLM_CUDA_CHECK(cudaFuncGetAttributes(&attr, beamStage1FastKernel<T, items_per_thread, 2 * MAX_K, blockSize>));
// One ThreadBlock must at least have share memory of `sizeof(T) * V / nSmallTopKMaxVocParts` bytes
int const static_smem = attr.sharedSizeBytes;
int const max_dyn_smem_per_block = max_smem_per_block - static_smem;
TLLM_CHECK_WITH_INFO(sizeof(T) * V <= max_dyn_smem_per_block * nSmallTopKMaxVocParts,
"Vocab size is too large for split-k TopK beam search fast path.");
// Find the maximum of ThreadBlock (maximum of voc_parts, minimum of smem),
// satisfying voc_parts <= nSmallTopKMaxVocParts && dyn_smem_size * voc_parts >= sizeof(T) * V
int const driver_smem_per_block = max_smem_per_sm - max_smem_per_block;
int const extra_smem = driver_smem_per_block + static_smem;
int voc_parts = nSmallTopKMaxVocParts + 1;
for (int n_block = max_active_blocks - 1; n_block > 0 && voc_parts > nSmallTopKMaxVocParts; --n_block)
{
int smem_per_block = max_smem_per_sm / n_block;
int dyn_smem_size = smem_per_block - extra_smem;
dyn_smem_size -= dyn_smem_size % sizeof(T);
voc_parts = (sizeof(T) * V + dyn_smem_size - 1) / dyn_smem_size;
}
if (voc_parts <= nSmallTopKMaxVocParts)
{
// Use stage 1 fast kernel
int const voc_size_chunk = (V + voc_parts - 1) / voc_parts;
int const dyn_smem_size = sizeof(T) * voc_size_chunk;
if (dyn_smem_size >= (48 << 10))
{
TLLM_CUDA_CHECK(cudaFuncSetAttribute(beamStage1FastKernel<T, items_per_thread, 2 * MAX_K, blockSize>,
cudaFuncAttributeMaxDynamicSharedMemorySize, dyn_smem_size));
}
dim3 gridSize(batch_size * beam_width, voc_parts);
beamStage1FastKernel<T, items_per_thread, 2 * MAX_K, blockSize><<<gridSize, blockSize, dyn_smem_size, stream>>>(
logits, bias, finished, temp_buffer, V, beam_width, end_ids, voc_size_chunk);
}
else
{
// use stage 1 base kernel
int voc_parts = 4;
if (batch_size * beam_width < 256)
{
// TODO: add heuristics for base stage 1 kernel
// Volta has 80 SMs, so we aim for three waves
voc_parts = (240 + batch_size * beam_width - 1) / (batch_size * beam_width);
voc_parts = std::min(128, voc_parts); // we implement up to 128
}
cudaFuncSetAttribute(beamStage1BaseKernel<T, items_per_thread, 2 * MAX_K, blockSize>,
cudaFuncAttributePreferredSharedMemoryCarveout, cudaSharedmemCarveoutMaxL1);
dim3 gridSize(batch_size * beam_width, voc_parts);
beamStage1BaseKernel<T, items_per_thread, 2 * MAX_K, blockSize>
<<<gridSize, blockSize, 0, stream>>>(logits, bias, finished, temp_buffer, V, beam_width, end_ids);
}
sync_check_cuda_error();
beamStage2KernelLauncher<T, 2 * MAX_K>(
temp_buffer, cum_log_probs, topk_id_buffer, topk_val_buffer, batch_size, beam_width, voc_parts, V, stream);
#else
beamKernel<T, items_per_thread, MAX_K, blockSize><<<batch_size * beam_width, blockSize, 0, stream>>>(
logits, bias, cum_log_probs, finished, topk_id_buffer, topk_val_buffer, V, beam_width, end_ids);
#endif
sync_check_cuda_error();
// Keep 2 * beam_width candidates in case of k candidates finishes in one iteration
size_t const smem_size = sizeof(T) * beam_width * beam_width * 2;
if (smem_size >= (48 << 10))
{
TLLM_CUDA_CHECK(cudaFuncSetAttribute(
batchBeamKernel<T, MAX_K * 2, 32>, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size));
}
batchBeamKernel<T, MAX_K * 2, 32><<<batch_size, 32, smem_size, stream>>>(topk_id_buffer, topk_val_buffer, bh);
sync_check_cuda_error();
}
#define INSTANTIATE_BEAMSEARCH_K(T, MAX_K) \
template void topK_softMax_kernelLauncher<T, MAX_K>( \
T const* logits, T const* bias, void* workspace, BeamHypotheses& bh, cudaStream_t stream);
} // namespace kernels
} // namespace tensorrt_llm