mrq/vall-e
5
9
Fork 0
Code Issues 8 Pull Requests Packages Projects Releases Wiki Activity

the learning rate scheduler pill is a tough pill to swallow

Browse Source
This commit is contained in:
mrq 2025-02-28 22:07:57 -06:00
parent b97faa8173
commit 94861677d3
6 changed files with 112 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
vall_e/config.py
View File

@ -504,7 +504,6 @@ class Hyperparameters:
warmup_steps: int = 0 # number of steps to warm up the optimizer before performing updates, I think, this is just passed to deepspeed
scheduler: str = "" # scheduler to use, currently don't ever use one so this doesn't really matter
scheduler_type: str = "" # deprecated
scheduler_params: dict = field(default_factory=lambda: {}) # to pass through deepspeed config
autotune: bool = False # to do deepspeed's autotuning
@ -1063,10 +1062,6 @@ class Config(BaseConfig):
if model.training:
model.teacher = False
if self.hyperparameters.scheduler_type and not self.hyperparameters.scheduler:
self.hyperparameters.scheduler = self.hyperparameters.scheduler_type
self.hyperparameters.scheduler_type = ""
# do not combine the two
if self.hyperparameters.scheduler