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

Update llama.py

Browse source
This commit is contained in:
Daniel Han-Chen 2024-03-11 03:07:44 +11:00
commit aba595de9c
1 changed files with 18 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

28
unsloth/models/llama.py
View file

@ -511,21 +511,29 @@ def LlamaModel_fast_forward(
# Mormalized from Gemma # Mormalized from Gemma
IS_GEMMA = self.config.model_type == "gemma" IS_GEMMA = self.config.model_type == "gemma"
if IS_GEMMA: if IS_GEMMA:
inputs_requires_grad = inputs_embeds.requires_grad
if not inputs_embeds.is_leaf:
inputs_embeds = inputs_embeds.detach()
inputs_requires_grad = True
elif inputs_requires_grad:
inputs_embeds.requires_grad_(False)
pass
# Match Gemma exactly by casting to bfloat16 / float16 # Match Gemma exactly by casting to bfloat16 / float16
# inputs_embeds *= math_sqrt(self.config.hidden_size) # inputs_embeds *= math_sqrt(self.config.hidden_size)
# Ie 3072**0.5 = 55.5000 in bfloat16, whilst 55.4256 in float32 # Ie 3072**0.5 = 55.5000 in bfloat16, whilst 55.4256 in float32
# & 2048**0.5 = 45.2500 in bfloat16, whilst 45.2548 in float32 # & 2048**0.5 = 45.2500 in bfloat16, whilst 45.2548 in float32
inputs_embeds *= torch.tensor(math_sqrt(self.config.hidden_size), dtype = inputs_embeds.dtype) normalizer = torch.tensor(math_sqrt(self.config.hidden_size), dtype = inputs_embeds.dtype)
# inputs_embeds *= math_sqrt(self.config.hidden_size)
if inputs_requires_grad: inputs_embeds.requires_grad_(True) if self.embed_tokens.weight.requires_grad:
# Careful we must not do an inplace op!
inputs_embeds = inputs_embeds * normalizer
else:
inputs_requires_grad = inputs_embeds.requires_grad
if not inputs_embeds.is_leaf:
inputs_embeds = inputs_embeds.detach()
inputs_requires_grad = True
elif inputs_requires_grad:
inputs_embeds.requires_grad_(False)
pass
inputs_embeds *= scalar
# inputs_embeds *= math_sqrt(self.config.hidden_size)
if inputs_requires_grad: inputs_embeds.requires_grad_(True)
pass
pass pass
# Fix up attention mask by setting elements to 0 # Fix up attention mask by setting elements to 0