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

maybe not

Browse Source
This commit is contained in:
mrq 2024-08-09 11:38:08 -05:00
parent c658a7b440
commit ed373957e2
2 changed files with 3 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
vall_e/config.py
View File

@ -614,7 +614,7 @@ class Trainer:
amp: bool = False # automatic mixed precision
ddp: bool = False # torch's internal DDP, automatically set if local backend is used and multiple GPUs are requested
scale_loss: bool = False # whether to perform loss scaling (for FP16 training) (it actually seems more harmful than not for this specific workload)
#scale_loss: bool = False # whether to perform loss scaling (for FP16 training) (it actually seems more harmful than not for this specific workload)
load_webui: bool = False # not working, but loads the web UI to allow inferencing during training
no_logger: bool = False # deprecated, but reroutes some logger calls to normal print statements for when logger broke because of BitNet
@ -634,14 +634,12 @@ class Trainer:
return torch.float8_e4m3fn
return torch.float32
"""
@cached_property
def scale_loss(self):
# currently cannot feasibly apply loss scaling with DeepSpeed backend (it can handle it itself anyways)
if self.backend != "local":
return False
return self.dtype == torch.float16
"""
"""
@dataclass()

2
vall_e/models/ar_nar.py
View File

@ -316,7 +316,7 @@ class AR_NAR(Base):
def example_usage():
cfg.trainer.backend = "local"
# cfg.trainer.backend = "local"
cfg.hyperparameters.gradient_accumulation_steps = 1
if cfg.audio_backend == "dac":
cfg.sample_rate = 44_100