Environment Variable Usage

This section describes the definition of the DashInfer environment variables and their function.

Memory Mangament

Environment Var: Memory

EnvVar Name

Describe

Type

Default

Options

BFC_ALLOCATOR

Use BFC Allocator or raw cudaMalloc API

for management of CUDA Device memory.

bool

ON

ON - Enable BFC Allocator

OFF - Disable BFC Allocator

BFC_MEM_RATIO

The max ratio of device memory that will be managemented by BFC Allocator.

float

0.9

float value between (0.0,1.0]

BFC_LEFTOVER_MB

The amount of GPU memory that cannot be used by the BFC allocator,

typically the sum of GPU memory occupied by PyTorch, CUDA driver,

default context, etc.

int

350

Formula for the actual GPU memory allocated

by the BFC Allocator on each GPU:

(Total Physical Memory - BFC_LEFTOVER_MB) * BFC_MEM_RATIO

CPU_CACHE_RATIO

DashInfer will set CPU_CACHE_RATIO * 100% of the current remaining CPU memory

for kv cache storage, and when CPU_CACHE_RATIO=0, no CPU memory is used to

store kv cache.

float

0.0

float value between [0.0, 1.0]

Logging

Environment Var: Logging

EnvVar Name

Describe

Type

Default

Options

ALLSPARK_TIME_LOG

Whether logging the generation and context step detailed time in different phase.

int

0

0 - not print; 1 - print log.

ALLSPARK_DUMP_OUTPUT_TOKEN

Whether print output token in log.

int

0

0 - not print; 1 - print log.

HIE_LOG_SATAUS_INTERVAL

The threshold control for printing statistical log when text generation.

int

5

In second, should be greater than 0.

Engine Behavior

Environment Var: Engine Behavior

EnvVar Name

Describe

Type

Default

Options

ALLSPARK_USE_TORCH_SAMPLE

Use the same sampler as vllm and PyTorch.

The generation speed may decrease by 5%-10%.

int

1

0 - use torch sampler

1 - use DashInfer native sampler,

which provides the same distribution,

but not exactly the same value.

AS_FLASH_THRESH

Threshold for enable Flash Attention do context attention calculation.

Flash Attention will be used if context length is greater than this threshold.

int

1024

int value between (0, int64_max).

ALLSPARK_DISABLE_WARMUP

Disable warm up step when model is start up.

int

0

1: disable warm up

0: not disable warm up