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

Update gemma.py

Browse source
This commit is contained in:
Daniel Han-Chen 2024-02-24 20:11:23 +11:00
commit 6d94cf88a7
1 changed files with 29 additions and 27 deletions
Download patch file Download diff file Expand all files Collapse all files

52
unsloth/models/gemma.py
View file

@ -205,8 +205,8 @@ 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 = 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,
@ -222,8 +222,8 @@ 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 = 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
@ -385,7 +385,9 @@ def GemmaModel_fast_forward(
if output_attentions:
all_self_attns += (layer_outputs[1],)
hidden_states = self.norm(hidden_states)
# hidden_states = self.norm(hidden_states)
hidden_states = fast_rms_layernorm_inference(self.norm, hidden_states)
# add hidden states from the last decoder layer
if output_hidden_states:
@ -549,22 +551,22 @@ class FastGemmaModel(FastLlamaModel):
# Also patch all dtypes - BnB seems to not allocate the correct type?
# BnB default dtype seems to be float16!
# correct_dtype = lm_head.weight.dtype
correct_dtype = lm_head.weight.dtype
# for name, module in model.named_modules():
# if isinstance(module, (Bnb_Linear4bit, Peft_Linear4bit)):
# weight = module.weight
# quant_state = weight.quant_state
for name, module in model.named_modules():
if isinstance(module, (Bnb_Linear4bit, Peft_Linear4bit)):
weight = module.weight
quant_state = weight.quant_state
# if type(quant_state) is list:
# # BnB seems to have float16 as default!
# module.weight.quant_state[2] = correct_dtype # Cast to correct dtype
# else:
# # https://github.com/TimDettmers/bitsandbytes/pull/763/files
# quant_state.dtype = correct_dtype
# pass
# pass
# pass
if type(quant_state) is list:
# BnB seems to have float16 as default!
module.weight.quant_state[2] = correct_dtype # Cast to correct dtype
else:
# https://github.com/TimDettmers/bitsandbytes/pull/763/files
quant_state.dtype = correct_dtype
pass
pass
pass
# Add 1 to weight
# return output * (1 + self.weight)
@ -581,12 +583,12 @@ 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