mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
138 lines
3.7 KiB
C++
138 lines
3.7 KiB
C++
/*
|
|
* SPDX-FileCopyrightText: Copyright (c) 2022-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
* SPDX-License-Identifier: Apache-2.0
|
|
*
|
|
* 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 "envUtils.h"
|
|
#include "tensorrt_llm/common/logger.h"
|
|
#include <cstdlib>
|
|
|
|
namespace tensorrt_llm::common
|
|
{
|
|
|
|
static std::optional<int32_t> getIntEnv(char const* name)
|
|
{
|
|
char const* const env = std::getenv(name);
|
|
if (env == nullptr)
|
|
{
|
|
return std::nullopt;
|
|
}
|
|
int32_t const val = std::stoi(env);
|
|
if (val <= 0)
|
|
{
|
|
return std::nullopt;
|
|
}
|
|
return {val};
|
|
};
|
|
|
|
// XQA kernels (optimized kernels for generation phase).
|
|
bool forceXQAKernels()
|
|
{
|
|
static bool const forceXQA = (getIntEnv("TRTLLM_FORCE_XQA").value_or(0) != 0);
|
|
return forceXQA;
|
|
}
|
|
|
|
int32_t xqaMaxNbCtaPerKVHeadFactor()
|
|
{
|
|
return envXqaNbCtaPerKVHead().value_or(8);
|
|
}
|
|
|
|
std::optional<int32_t> envXqaNbCtaPerKVHead()
|
|
{
|
|
static std::optional<int32_t> const ret = getIntEnv("TRTLLM_XQA_BLOCKS_PER_SEQUENCE");
|
|
return ret;
|
|
}
|
|
|
|
bool getEnvDisableXQAJIT()
|
|
{
|
|
static bool init = false;
|
|
static bool disableXQAJIT = false;
|
|
if (!init)
|
|
{
|
|
init = true;
|
|
char const* disable_xqa_jit_var = std::getenv("TRTLLM_DISABLE_XQA_JIT");
|
|
if (disable_xqa_jit_var)
|
|
{
|
|
if (disable_xqa_jit_var[0] == '1' && disable_xqa_jit_var[1] == '\0')
|
|
{
|
|
disableXQAJIT = true;
|
|
}
|
|
}
|
|
}
|
|
return disableXQAJIT;
|
|
}
|
|
|
|
// Tune the number of blocks per sequence for accuracy/performance purpose.
|
|
bool getEnvMmhaMultiblockDebug()
|
|
{
|
|
static bool init = false;
|
|
static bool forceMmhaMaxSeqLenTile = false;
|
|
if (!init)
|
|
{
|
|
init = true;
|
|
char const* enable_mmha_debug_var = std::getenv("TRTLLM_ENABLE_MMHA_MULTI_BLOCK_DEBUG");
|
|
if (enable_mmha_debug_var)
|
|
{
|
|
if (enable_mmha_debug_var[0] == '1' && enable_mmha_debug_var[1] == '\0')
|
|
{
|
|
forceMmhaMaxSeqLenTile = true;
|
|
}
|
|
}
|
|
}
|
|
return forceMmhaMaxSeqLenTile;
|
|
}
|
|
|
|
int getEnvMmhaBlocksPerSequence()
|
|
{
|
|
static bool init = false;
|
|
static int mmhaBlocksPerSequence = 0;
|
|
if (!init)
|
|
{
|
|
init = true;
|
|
char const* mmhaBlocksPerSequenceEnv = std::getenv("TRTLLM_MMHA_BLOCKS_PER_SEQUENCE");
|
|
if (mmhaBlocksPerSequenceEnv)
|
|
{
|
|
mmhaBlocksPerSequence = std::atoi(mmhaBlocksPerSequenceEnv);
|
|
if (mmhaBlocksPerSequence <= 0)
|
|
{
|
|
TLLM_LOG_WARNING("Invalid value for TRTLLM_MMHA_BLOCKS_PER_SEQUENCE. Will use default values instead!");
|
|
}
|
|
}
|
|
}
|
|
return mmhaBlocksPerSequence;
|
|
}
|
|
|
|
int getEnvMmhaKernelBlockSize()
|
|
{
|
|
static bool init = false;
|
|
static int mmhaKernelBlockSize = 0;
|
|
if (!init)
|
|
{
|
|
init = true;
|
|
char const* mmhaKernelBlockSizeEnv = std::getenv("TRTLLM_MMHA_KERNEL_BLOCK_SIZE");
|
|
if (mmhaKernelBlockSizeEnv)
|
|
{
|
|
mmhaKernelBlockSize = std::atoi(mmhaKernelBlockSizeEnv);
|
|
if (mmhaKernelBlockSize <= 0)
|
|
{
|
|
TLLM_LOG_WARNING("Invalid value for TRTLLM_MMHA_KERNEL_BLOCK_SIZE. Will use default values instead!");
|
|
}
|
|
}
|
|
}
|
|
return mmhaKernelBlockSize;
|
|
}
|
|
|
|
} // namespace tensorrt_llm::common
|