ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

norm

Browse source
This commit is contained in:
Daniel Han-Chen 2024-02-24 14:02:26 +11:00
commit 4c6e122caa
2 changed files with 9 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

16
unsloth/models/gemma.py
View file

@ -86,7 +86,7 @@ def GemmaDecoderLayer_fast_forward(
hidden_states += residual
else:
residual = hidden_states
hidden_states = fast_rms_layernorm(self.input_layernorm, hidden_states)
hidden_states = self.input_layernorm(hidden_states)
hidden_states, self_attn_weights, present_key_value = self.self_attn(
hidden_states=hidden_states,
causal_mask=causal_mask,
@ -101,7 +101,7 @@ def GemmaDecoderLayer_fast_forward(
# Fully Connected
residual = hidden_states
hidden_states = fast_rms_layernorm(self.post_attention_layernorm, hidden_states)
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states = self.mlp(hidden_states)
hidden_states = residual + hidden_states
pass
@ -328,12 +328,12 @@ class FastGemmaModel(FastLlamaModel):
pass
print("Unsloth: Patching Gemma RMS Layernorm + 1")
for name, module in model.named_modules():
if isinstance(module, GemmaRMSNorm):
module.weight += 1.0 # return output * (1 + self.weight)
if not hasattr(module, "variance_epsilon"):
module.variance_epsilon = module.eps # Gemma doesn't use variance_epsilon
pass
# for name, module in model.named_modules():
# if isinstance(module, GemmaRMSNorm):
# module.weight += 1.0 # return output * (1 + self.weight)
# if not hasattr(module, "variance_epsilon"):
# module.variance_epsilon = module.eps # Gemma doesn't use variance_epsilon
# pass
# Clear deleted GPU items
import gc

2
unsloth/models/llama.py
View file

@ -609,7 +609,7 @@ def LlamaModel_fast_forward(
all_self_attns += (layer_outputs[1],)
pass
hidden_states = fast_rms_layernorm(self.norm, hidden_states)
hidden_states = self.norm(hidden_states)
# add hidden states from the last decoder layer
if output_hidden_states: