ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-02 20:51:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

Implement fp8 quant for layernorm and rmsnorm (#1814)

Browse Source
This commit is contained in:
ruanjm
2025-01-24 16:40:43 +08:00
committed by GitHub
parent 5b9b083dbc
commit 64d5c4d6cb
9 changed files with 67 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
example/ck_tile/02_layernorm2d/generate.py
View File

@@ -39,7 +39,8 @@ FUSED_FUSED_SWEEP_STR_MAP = [
DATA_TYPE_MAP = {'fp32' : 'float',
'fp16' : 'ck_tile::fp16_t',
'bf16' : 'ck_tile::bf16_t',
'int8' : 'ck_tile::int8_t'}
'int8' : 'ck_tile::int8_t',
'fp8' : 'ck_tile::fp8_t'}
def BOOL_MAP(b_) -> str:
if b_:
@@ -504,12 +505,13 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
h_traits = layernorm_fwd_codegen.h_traits
h_instance = layernorm_fwd_codegen.h_instance
dynamic_quant_out_dtype = ['int8']
dynamic_quant_out_dtype = ['int8', 'fp8']
# some predefined support range
# (prec_i,prec_o) for simplicity this string will be used as key for dict
scale_list = [('fp32,fp32')]
dtype_list = [('fp16,fp16'), ('bf16,bf16'),
('fp16,int8'), ('bf16,int8')] # NOTE: only fused-dynamic-quant use int8 out
('fp16,int8'), ('bf16,int8'),
('fp16,fp8'), ('bf16,fp8')] # NOTE: only fused-dynamic-quant use int8 or fp8 out
types_8bit = ('int8', 'fp8')
types_16bit = ('int16', 'fp16', 'bf16')
#fused_add_list = [0, 1, 2]