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

tweaks

Browse Source
This commit is contained in:
mrq 2024-11-11 12:40:41 -06:00
parent 48490757da
commit a748e223ce
2 changed files with 6 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
vall_e/config.py
View File

@ -948,6 +948,7 @@ class Config(BaseConfig):
if not isinstance( model, dict ): if not isinstance( model, dict ):
continue continue
# to-do: prune unused keys in here too automatically
if "experimental" not in model or not model["experimental"]: if "experimental" not in model or not model["experimental"]:
model["experimental"] = {} model["experimental"] = {}
@ -962,6 +963,9 @@ class Config(BaseConfig):
if "p_rvq_levels" in model["experimental"]: if "p_rvq_levels" in model["experimental"]:
model["experimental"]["rvq_levels_p"] = model["experimental"]["p_rvq_levels"] model["experimental"]["rvq_levels_p"] = model["experimental"]["p_rvq_levels"]
del model["experimental"]["p_rvq_levels"] del model["experimental"]["p_rvq_levels"]
if "p_len_train" in model["experimental"]:
del model["experimental"]["p_len_train"]
self.models = [ Model(**model) if isinstance(model, dict) else model for model in self.models ] self.models = [ Model(**model) if isinstance(model, dict) else model for model in self.models ]
self.loras = [ LoRA(**lora) if isinstance(lora, dict) else lora for lora in self.loras ] self.loras = [ LoRA(**lora) if isinstance(lora, dict) else lora for lora in self.loras ]

4
vall_e/models/ar_nar.py
View File

@ -275,9 +275,9 @@ class AR_NAR(Base):
_super = super() _super = super()
# to-do: allow for batch processing (it should probably work batched anyways) # to-do: allow for batch processing (it should probably work batched anyways)
def demask_sampling( batch_index, seq_len ): def demask_sampling( batch_index, seq_len ):
# overrides # overrides, to be user-controllable soonTM
max_steps = 10 max_steps = 10
temperature = 0.3 temperature = 1.0
cfg_strength = 1.0 cfg_strength = 1.0
sampling_repetition_penalty = 1.0 # force rep pen off, because this caused false positives due to how rep pen was being naively applied...... sampling_repetition_penalty = 1.0 # force rep pen off, because this caused false positives due to how rep pen was being naively applied......
sampling_top_p = 0.9 # a lot of demasking samplers use a top-k of seq_len * 0.9 sampling_top_p = 0.9 # a lot of demasking samplers use a top-k of seq_len * 0.9