Compare commits

...

3 Commits

Author SHA1 Message Date
comfyanonymous
483004dd1d Support newer glora format. 2024-09-03 17:02:19 -04:00
comfyanonymous
00a5d08103 Lower fp8 lora memory usage. 2024-09-03 01:25:05 -04:00
comfyanonymous
d043997d30 Flux onetrainer lora. 2024-09-02 08:22:15 -04:00
2 changed files with 34 additions and 9 deletions

View File

@@ -41,9 +41,8 @@ def manual_stochastic_round_to_float8(x, dtype, generator=None):
(2.0 ** (exponent - EXPONENT_BIAS)) * (1.0 + abs_x),
(2.0 ** (-EXPONENT_BIAS + 1)) * abs_x
)
del abs_x
return sign.to(dtype=dtype)
return sign
@@ -57,6 +56,11 @@ def stochastic_rounding(value, dtype, seed=0):
if dtype == torch.float8_e4m3fn or dtype == torch.float8_e5m2:
generator = torch.Generator(device=value.device)
generator.manual_seed(seed)
return manual_stochastic_round_to_float8(value, dtype, generator=generator)
output = torch.empty_like(value, dtype=dtype)
num_slices = max(1, (value.numel() / (4096 * 4096)))
slice_size = max(1, round(value.shape[0] / num_slices))
for i in range(0, value.shape[0], slice_size):
output[i:i+slice_size].copy_(manual_stochastic_round_to_float8(value[i:i+slice_size], dtype, generator=generator))
return output
return value.to(dtype=dtype)

View File

@@ -324,6 +324,7 @@ def model_lora_keys_unet(model, key_map={}):
to = diffusers_keys[k]
key_map["transformer.{}".format(k[:-len(".weight")])] = to #simpletrainer and probably regular diffusers flux lora format
key_map["lycoris_{}".format(k[:-len(".weight")].replace(".", "_"))] = to #simpletrainer lycoris
key_map["lora_transformer_{}".format(k[:-len(".weight")].replace(".", "_"))] = to #onetrainer
return key_map
@@ -527,20 +528,40 @@ def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32):
except Exception as e:
logging.error("ERROR {} {} {}".format(patch_type, key, e))
elif patch_type == "glora":
if v[4] is not None:
alpha = v[4] / v[0].shape[0]
else:
alpha = 1.0
dora_scale = v[5]
old_glora = False
if v[3].shape[1] == v[2].shape[0] == v[0].shape[0] == v[1].shape[1]:
rank = v[0].shape[0]
old_glora = True
if v[3].shape[0] == v[2].shape[1] == v[0].shape[1] == v[1].shape[0]:
if old_glora and v[1].shape[0] == weight.shape[0] and weight.shape[0] == weight.shape[1]:
pass
else:
old_glora = False
rank = v[1].shape[0]
a1 = comfy.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, intermediate_dtype)
a2 = comfy.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, intermediate_dtype)
b1 = comfy.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, intermediate_dtype)
b2 = comfy.model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, intermediate_dtype)
if v[4] is not None:
alpha = v[4] / rank
else:
alpha = 1.0
try:
lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1).to(dtype=intermediate_dtype), a2), a1)).reshape(weight.shape)
if old_glora:
lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1).to(dtype=intermediate_dtype), a2), a1)).reshape(weight.shape) #old lycoris glora
else:
if weight.dim() > 2:
lora_diff = torch.einsum("o i ..., i j -> o j ...", torch.einsum("o i ..., i j -> o j ...", weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape)
else:
lora_diff = torch.mm(torch.mm(weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape)
lora_diff += torch.mm(b1, b2).reshape(weight.shape)
if dora_scale is not None:
weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype))
else: