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

trust_remote_code

Browse source
This commit is contained in:
Daniel Han-Chen 2024-03-18 04:07:54 +11:00
commit e6a758d2c3
3 changed files with 16 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

10
unsloth/models/llama.py
View file

@ -758,6 +758,7 @@ def CausalLM_fast_forward(fast_forward_inference):
else:
logits = self.lm_head(hidden_states)
pass
logits = logits.to(self.config.torch_dtype)
loss = None
if labels is not None:
@ -928,6 +929,7 @@ class FastLlamaModel:
fix_tokenizer = True,
model_patcher = None,
tokenizer_name = None,
trust_remote_code = False,
**kwargs,
):
if model_patcher is None: model_patcher = FastLlamaModel
@ -988,6 +990,7 @@ class FastLlamaModel:
token = token,
rope_scaling = rope_scaling,
max_position_embeddings = max_position_embeddings,
trust_remote_code = trust_remote_code,
**kwargs,
)
@ -995,9 +998,10 @@ class FastLlamaModel:
tokenizer_name = model_name if tokenizer_name is None else tokenizer_name
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name,
model_max_length = max_position_embeddings,
padding_side = "right",
token = token,
model_max_length = max_position_embeddings,
padding_side = "right",
token = token,
trust_remote_code = trust_remote_code,
)
model, tokenizer = patch_tokenizer(model, tokenizer)

2
unsloth/models/loader.py
View file

@ -74,6 +74,7 @@ class FastLanguageModel(FastLlamaModel):
device_map = "sequential",
rope_scaling = None,
fix_tokenizer = True,
trust_remote_code = False,
use_gradient_checkpointing = True,
*args, **kwargs,
):
@ -139,6 +140,7 @@ class FastLanguageModel(FastLlamaModel):
fix_tokenizer = fix_tokenizer,
model_patcher = dispatch_model,
tokenizer_name = tokenizer_name,
trust_remote_code = trust_remote_code,
*args, **kwargs,
)

10
unsloth/models/mistral.py
View file

@ -230,6 +230,7 @@ def MistralForCausalLM_fast_forward(
else:
logits = self.lm_head(hidden_states)
pass
logits = logits.to(self.config.torch_dtype)
loss = None
if labels is not None:
@ -295,6 +296,7 @@ class FastMistralModel(FastLlamaModel):
fix_tokenizer = True,
model_patcher = None,
tokenizer_name = None,
trust_remote_code = False,
**kwargs,
):
if model_patcher is None: model_patcher = FastMistralModel
@ -353,6 +355,7 @@ class FastMistralModel(FastLlamaModel):
quantization_config = bnb_config,
token = token,
# rope_scaling = rope_scaling,
trust_remote_code = trust_remote_code,
**kwargs,
)
@ -360,9 +363,10 @@ class FastMistralModel(FastLlamaModel):
tokenizer_name = model_name if tokenizer_name is None else tokenizer_name
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name,
model_max_length = max_position_embeddings,
padding_side = "right",
token = token,
model_max_length = max_position_embeddings,
padding_side = "right",
token = token,
trust_remote_code = trust_remote_code,
)
model, tokenizer = patch_tokenizer(model, tokenizer)