mirror of
https://github.com/huggingface/accelerate.git
synced 2025-10-20 10:03:46 +08:00
847ae58c74
* single-gpu tests passing * install deepspeed in fp8 container * revert mixed_precision check
19 lines
797 B
YAML
19 lines
797 B
YAML
# This config template simply setups up the TransformersEngine config (and a config for a single GPU),
|
|
# this can interop with the other configs in this folder
|
|
distributed_type: "NO"
|
|
mixed_precision: "fp8"
|
|
# Then we specify the fp8 configuration:
|
|
fp8_config:
|
|
backend: TE # Can be TE | MS-AMP
|
|
# The following are TE specific arguments.
|
|
# See https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/api/common.html#common-api for more details
|
|
amax_history_len: 1024
|
|
fp8_format: E4M3
|
|
interval: 1
|
|
margin: 0
|
|
override_linear_precision: [false, false, false]
|
|
# Generally this should always be set to `false` to have the most realistic fp8 eval performance
|
|
use_autocast_during_eval: false
|
|
# If using MS-AMP, we ignore all of the prior and set a opt_level
|
|
#opt_level: O1
|