huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-19 19:58:35 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add flash attention (#241)

Browse Source 
* Add some flash-attn kernel, import the code for flash-attn v2 from Dao-AILab.

* More flash attn.

* Set up the flash attn parameters.

* Get things to compile locally.

* Move the flash attention files in a different directory.

* Build the static C library with nvcc.

* Add more flash attention.

* Update the build part.

* Better caching.

* Exclude flash attention from the default workspace.

* Put flash-attn behind a feature gate.

* Get the flash attn kernel to run.

* Move the flags to a more appropriate place.

* Enable flash attention in llama.

* Use flash attention in llama.
This commit is contained in:
Laurent Mazare
2023-07-26 07:48:10 +01:00
committed by GitHub
parent c97d51243c
commit d9f9c859af
22 changed files with 2699 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

66
candle-flash-attn/kernels/static_switch.h Normal file
View File

@ -0,0 +1,66 @@
// Inspired by
// https://github.com/NVIDIA/DALI/blob/main/include/dali/core/static_switch.h
// and https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/Dispatch.h
#pragma once
/// @param COND - a boolean expression to switch by
/// @param CONST_NAME - a name given for the constexpr bool variable.
/// @param ... - code to execute for true and false
///
/// Usage:
/// ```
/// BOOL_SWITCH(flag, BoolConst, [&] {
/// some_function<BoolConst>(...);
/// });
/// ```
#define BOOL_SWITCH(COND, CONST_NAME, ...) \
[&] { \
if (COND) { \
constexpr static bool CONST_NAME = true; \
return __VA_ARGS__(); \
} else { \
constexpr static bool CONST_NAME = false; \
return __VA_ARGS__(); \
} \
}()
#define FP16_SWITCH(COND, ...) \
[&] { \
if (COND) { \
using elem_type = cutlass::half_t; \
return __VA_ARGS__(); \
} else { \
using elem_type = cutlass::bfloat16_t; \
return __VA_ARGS__(); \
} \
}()
#define FWD_HEADDIM_SWITCH(HEADDIM, ...) \
[&] { \
if (HEADDIM <= 32) { \
constexpr static int kHeadDim = 32; \
return __VA_ARGS__(); \
} else if (HEADDIM <= 64) { \
constexpr static int kHeadDim = 64; \
return __VA_ARGS__(); \
} else if (HEADDIM <= 96) { \
constexpr static int kHeadDim = 96; \
return __VA_ARGS__(); \
} else if (HEADDIM <= 128) { \
constexpr static int kHeadDim = 128; \
return __VA_ARGS__(); \
} else if (HEADDIM <= 160) { \
constexpr static int kHeadDim = 160; \
return __VA_ARGS__(); \
} else if (HEADDIM <= 192) { \
constexpr static int kHeadDim = 192; \
return __VA_ARGS__(); \
} else if (HEADDIM <= 224) { \
constexpr static int kHeadDim = 224; \
return __VA_ARGS__(); \
} else if (HEADDIM <= 256) { \
constexpr static int kHeadDim = 256; \
return __VA_ARGS__(); \
} \
}()