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

Update vision.py

Browse source
This commit is contained in:
Daniel Han 2025-03-14 05:40:56 -07:00
commit 26045d8efe
1 changed files with 27 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

29
unsloth/models/vision.py
View file

@ -75,6 +75,8 @@ FORCE_EAGER_ATTENTION = [
"pixtral", # Pixtral SDPA not implemented "pixtral", # Pixtral SDPA not implemented
] ]
global NUM_LOGITS_TO_KEEP
NUM_LOGITS_TO_KEEP = dict()
def unsloth_base_fast_generate( def unsloth_base_fast_generate(
self, self,
@ -85,18 +87,39 @@ def unsloth_base_fast_generate(
dtype = _get_dtype(self.config.torch_dtype) dtype = _get_dtype(self.config.torch_dtype)
# Check if VLM # Check if VLM
is_vlm = ( is_vlm = any(
x.endswith(("ForConditionalGeneration", "ForVisionText2Text")) x.endswith(("ForConditionalGeneration", "ForVisionText2Text"))
for x in self.config.architectures for x in self.config.architectures
) )
is_vlm = is_vlm or hasattr(self.config, "vision_config") is_vlm = is_vlm or hasattr(self.config, "vision_config")
arch = self.config.architectures[0]
# Remove token_type_ids # Remove token_type_ids
kwargs.pop("token_type_ids", None) kwargs.pop("token_type_ids", None)
# VLMs do not allow logits_to_keep # VLMs do not allow logits_to_keep
if not is_vlm: if not is_vlm:
kwargs["logits_to_keep"] = 1 global NUM_LOGITS_TO_KEEP
if arch not in NUM_LOGITS_TO_KEEP:
m = self
while hasattr(m, "model"):
if hasattr(m, "forward"):
keys = inspect.signature(m.forward).parameters.keys()
if "num_logits_to_keep" in keys:
NUM_LOGITS_TO_KEEP[arch] = "num_logits_to_keep"
break
elif "logits_to_keep" in keys:
NUM_LOGITS_TO_KEEP[arch] = "logits_to_keep"
break
m = m.model
pass
if arch not in NUM_LOGITS_TO_KEEP:
NUM_LOGITS_TO_KEEP[arch] = None
pass
pass
key = NUM_LOGITS_TO_KEEP[arch]
if key is not None:
kwargs[key] = 1
else: else:
kwargs.pop("logits_to_keep", None) kwargs.pop("logits_to_keep", None)
kwargs.pop("num_logits_to_keep", None) kwargs.pop("num_logits_to_keep", None)
@ -112,6 +135,8 @@ def unsloth_base_fast_generate(
try: kwargs["pixel_values"] = kwargs["pixel_values"].to(dtype) try: kwargs["pixel_values"] = kwargs["pixel_values"].to(dtype)
except: pass except: pass
print(kwargs)
# Mixed precision autocast # Mixed precision autocast
if os.environ.get("UNSLOTH_FORCE_FLOAT32", "0") == "1": dtype = torch.float32 if os.environ.get("UNSLOTH_FORCE_FLOAT32", "0") == "1": dtype = torch.float32
with torch.inference_mode(), torch.autocast(device_type = "cuda", dtype = dtype): with torch.inference_mode(), torch.autocast(device_type = "cuda", dtype = dtype):