Skip to content

fix: some tensors are in bf16 #82

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Nov 20, 2023
Merged

fix: some tensors are in bf16 #82

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9e14d55
fix: some tensors are in bf16
Green-Sky Nov 19, 2023
45c339d
did it in preprocess_lora
leejet Nov 20, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 11 additions & 5 deletions models/convert.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -101,7 +101,7 @@ def quantize_q5_1(x):
def quantize_q8_0(x):
assert x.shape[-1] % QK8_0 == 0 and x.shape[-1] > QK8_0
x = x.reshape(-1, QK8_0)
amax = np.max(np.abs(x), axis=-1, keepdims=True)
amax = np.max(np.abs(x), axis=-1, keepdims=True)
d = amax / ((1 << 7) - 1)
qs = (x / d).round().clip(min=-128, max=127).astype(np.int8)
d = d.astype(np.float16).view(np.int8)
Expand Down Expand Up @@ -178,7 +178,7 @@ def preprocess(state_dict):
print("no alphas_cumprod in file, generate new one")
alphas_cumprod = get_alpha_comprod()
state_dict["alphas_cumprod"] = alphas_cumprod

new_state_dict = {}
for name, w in state_dict.items():
# ignore unused tensors
Expand All @@ -192,7 +192,7 @@ def preprocess(state_dict):
if skip:
continue

# # convert BF16 to FP16
# convert BF16 to FP16
if w.dtype == torch.bfloat16:
w = w.to(torch.float16)

Expand Down Expand Up @@ -251,7 +251,7 @@ def preprocess(state_dict):
new_state_dict[new_name] = w
print(f"preprocess {name} => {new_name}")
continue

# convert unet transformer linear to conv2d 1x1
if name.startswith("model.diffusion_model.") and (name.endswith("proj_in.weight") or name.endswith("proj_out.weight")):
if len(w.shape) == 2:
Expand Down Expand Up @@ -342,6 +342,11 @@ def preprocess_lora(state_dict):
for name, w in state_dict.items():
if not isinstance(w, torch.Tensor):
continue

# convert BF16 to FP16
if w.dtype == torch.bfloat16:
w = w.to(torch.float16)

name_without_network_parts, network_part = name.split(".", 1)
new_name_without_network_parts = convert_diffusers_name_to_compvis(name_without_network_parts)
if new_name_without_network_parts == None:
Expand Down Expand Up @@ -421,6 +426,7 @@ def convert(model_path, out_type = None, out_file=None, lora=False):
continue
if name in unused_tensors:
continue

data = state_dict[name].numpy()

n_dims = len(data.shape)
Expand Down Expand Up @@ -452,7 +458,7 @@ def convert(model_path, out_type = None, out_file=None, lora=False):
else:
data = data.astype(np.float32)
ttype = "f32"

print("Processing tensor: {} with shape {}, {} -> {}".format(name, data.shape, old_type, ttype))

# header
Expand Down