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

resolve deprecation warning that doesn't show on my old training rig but does on my new one

Browse Source
This commit is contained in:
mrq 2024-05-09 23:25:44 -05:00
parent 1547de5020
commit 2109712e5b
2 changed files with 11 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
vall_e/models/ar_nar.py
View File

@ -365,7 +365,7 @@ def example_usage():
'n_tokens': 1024, 'n_tokens': 1024,
'd_model': 1024, # 256, # 1024, # 1536 'd_model': 1024, # 256, # 1024, # 1536
'n_heads': 16, # 4, # 16, # 24 'n_heads': 16, # 4, # 16, # 24
'n_layers': 4, # 32 'n_layers': 12, # 32
'n_experts': 1, 'n_experts': 1,
'l_padding': 8 if cfg.optimizations.fp8 else 0, 'l_padding': 8 if cfg.optimizations.fp8 else 0,

15
vall_e/models/base.py
View File

@ -481,7 +481,7 @@ class Base(nn.Module):
is_encoder_decoder=False, is_encoder_decoder=False,
is_decoder=True, is_decoder=True,
attn_implementation=attention, attn_implementation=attention,
gradient_checkpointing=self.activation_checkpointing, #gradient_checkpointing=self.activation_checkpointing,
)) ))
else: else:
self.model = MixtralModel(MixtralConfig( self.model = MixtralModel(MixtralConfig(
@ -501,8 +501,14 @@ class Base(nn.Module):
num_local_experts=n_experts, num_local_experts=n_experts,
num_experts_per_tok=min(2, n_experts), num_experts_per_tok=min(2, n_experts),
attn_implementation=attention, attn_implementation=attention,
gradient_checkpointing=self.activation_checkpointing, #gradient_checkpointing=self.activation_checkpointing,
)) ))
if self.activation_checkpointing and not self.model.gradient_checkpointing:
self.model.gradient_checkpointing_enable()
if training:
self.model.training = True
elif self.arch_type == "llama": elif self.arch_type == "llama":
if n_experts <= 1: if n_experts <= 1:
self.model = LlamaModel(LlamaConfig( self.model = LlamaModel(LlamaConfig(
@ -519,7 +525,7 @@ class Base(nn.Module):
is_encoder_decoder=False, is_encoder_decoder=False,
is_decoder=True, is_decoder=True,
attn_implementation=attention, attn_implementation=attention,
gradient_checkpointing=self.activation_checkpointing, #gradient_checkpointing=self.activation_checkpointing,
)) ))
else: else:
self.model = MixtralModel(MixtralConfig( self.model = MixtralModel(MixtralConfig(
@ -539,10 +545,9 @@ class Base(nn.Module):
num_local_experts=n_experts, num_local_experts=n_experts,
num_experts_per_tok=min(2, n_experts), num_experts_per_tok=min(2, n_experts),
attn_implementation=attention, attn_implementation=attention,
gradient_checkpointing=self.activation_checkpointing, #gradient_checkpointing=self.activation_checkpointing,
)) ))
print("Checkpointing:", self.activation_checkpointing, self.model.gradient_checkpointing)
if self.activation_checkpointing and not self.model.gradient_checkpointing: if self.activation_checkpointing and not self.model.gradient_checkpointing:
self.model.gradient_checkpointing_enable() self.model.gradient_checkpointing_enable()