mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
116 lines
3.9 KiB
C++
116 lines
3.9 KiB
C++
/*
|
|
* Copyright (c) 2024, 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/assert.h"
|
|
#include "tensorrt_llm/common/cudaUtils.h"
|
|
#include "tensorrt_llm/runtime/iTensor.h"
|
|
#include <vector>
|
|
|
|
namespace tensorrt_llm::runtime
|
|
{
|
|
|
|
class MedusaModule
|
|
{
|
|
public:
|
|
using TensorPtr = ITensor::SharedPtr;
|
|
using MedusaChoices = std::vector<std::vector<SizeType>>;
|
|
|
|
explicit MedusaModule(SizeType medusaHeads, SizeType maxMedusaTokens) noexcept
|
|
: mMedusaHeads(medusaHeads)
|
|
, mMaxMedusaTokens(maxMedusaTokens)
|
|
, mTokensPerStep(mMaxMedusaTokens + 1)
|
|
{
|
|
mNumPackedMasks = tensorrt_llm::common::divUp(mTokensPerStep, 32);
|
|
}
|
|
|
|
explicit MedusaModule() noexcept
|
|
: MedusaModule(0, 0)
|
|
{
|
|
}
|
|
|
|
MedusaModule(MedusaModule const& o) = default;
|
|
MedusaModule& operator=(MedusaModule const& o) = default;
|
|
|
|
[[nodiscard]] SizeType medusaHeads() const noexcept
|
|
{
|
|
return mMedusaHeads;
|
|
}
|
|
|
|
[[nodiscard]] SizeType maxMedusaTokens() const noexcept
|
|
{
|
|
return mMaxMedusaTokens;
|
|
}
|
|
|
|
[[nodiscard]] SizeType tokensPerStep() const noexcept
|
|
{
|
|
return mTokensPerStep;
|
|
}
|
|
|
|
[[nodiscard]] SizeType numPackedMasks() const noexcept
|
|
{
|
|
return mNumPackedMasks;
|
|
}
|
|
|
|
[[nodiscard]] MedusaChoices const& getMedusaChoices() const noexcept
|
|
{
|
|
return mDefaultMedusaChoices;
|
|
}
|
|
|
|
void initMedusaTensorsFromChoices(MedusaChoices const& choices, std::vector<SizeType>& topKs,
|
|
TensorPtr& positionOffsets, TensorPtr& treeIds, TensorPtr& paths, TensorPtr& packedMask,
|
|
SizeType& totalPaths) const noexcept;
|
|
|
|
private:
|
|
using Prefix = uint64_t;
|
|
static SizeType constexpr PREFIX_CHUNK_SIZE_BITS = 4;
|
|
static SizeType constexpr PREFIX_MAX_VALUE = 16;
|
|
|
|
struct MedusaTreeNode
|
|
{
|
|
SizeType nodeId;
|
|
SizeType depth;
|
|
SizeType parentLinearIdx;
|
|
SizeType linearIdx;
|
|
std::vector<SizeType> childLinearIndices;
|
|
};
|
|
|
|
SizeType computePathsAndMask(
|
|
std::vector<MedusaTreeNode> const& tree, TensorPtr& packedMask, TensorPtr& paths) const;
|
|
|
|
void copyPackedMask(TensorPtr& mask, SizeType srcIdx, SizeType dstIdx) const;
|
|
void setOnePackedMask(TensorPtr& mask, SizeType row, SizeType col) const;
|
|
Prefix computePrefix(std::vector<SizeType> const& vec, SizeType len) const;
|
|
|
|
void dumpChoices(MedusaChoices const& choices, std::vector<SizeType> const& indices) const;
|
|
|
|
private:
|
|
SizeType mMedusaHeads;
|
|
SizeType mMaxMedusaTokens;
|
|
SizeType mTokensPerStep;
|
|
SizeType mNumPackedMasks;
|
|
// FIXME(nkorobov): this should come from outside to setup or per request
|
|
// mc_sim_7b_63
|
|
MedusaChoices mDefaultMedusaChoices = {{0}, {0, 0}, {1}, {0, 1}, {2}, {0, 0, 0}, {1, 0}, {0, 2}, {3}, {0, 3}, {4},
|
|
{0, 4}, {2, 0}, {0, 5}, {0, 0, 1}, {5}, {0, 6}, {6}, {0, 7}, {0, 1, 0}, {1, 1}, {7}, {0, 8}, {0, 0, 2}, {3, 0},
|
|
{0, 9}, {8}, {9}, {1, 0, 0}, {0, 2, 0}, {1, 2}, {0, 0, 3}, {4, 0}, {2, 1}, {0, 0, 4}, {0, 0, 5}, {0, 0, 0, 0},
|
|
{0, 1, 1}, {0, 0, 6}, {0, 3, 0}, {5, 0}, {1, 3}, {0, 0, 7}, {0, 0, 8}, {0, 0, 9}, {6, 0}, {0, 4, 0}, {1, 4},
|
|
{7, 0}, {0, 1, 2}, {2, 0, 0}, {3, 1}, {2, 2}, {8, 0}, {0, 5, 0}, {1, 5}, {1, 0, 1}, {0, 2, 1}, {9, 0},
|
|
{0, 6, 0}, {0, 0, 0, 1}, {1, 6}, {0, 7, 0}};
|
|
};
|
|
} // namespace tensorrt_llm::runtime
|