Update app.py
This commit is contained in:
107
app.py
107
app.py
@@ -254,113 +254,6 @@ pipe.fuse_lora()
|
|||||||
print("lightning lora fused.")
|
print("lightning lora fused.")
|
||||||
|
|
||||||
|
|
||||||
# --- 2. manual surgery for lokr (snofs) ---
|
|
||||||
print("attempting manual lokr injection for snofs...")
|
|
||||||
|
|
||||||
try:
|
|
||||||
from safetensors.torch import load_file
|
|
||||||
from huggingface_hub import hf_hub_download
|
|
||||||
import torch.nn as nn
|
|
||||||
|
|
||||||
# download
|
|
||||||
lora_path = hf_hub_download(repo_id="headlesssetton/kjfakjf", filename="Qwen_Snofs_1_2.safetensors")
|
|
||||||
state_dict = load_file(lora_path)
|
|
||||||
|
|
||||||
lokr_scale = 1.0
|
|
||||||
|
|
||||||
def get_target_module(root, path_str):
|
|
||||||
"""
|
|
||||||
recursively resolves a path like 'transformer_blocks_25_attn_add_v_proj'
|
|
||||||
against the actual model structure.
|
|
||||||
"""
|
|
||||||
if not path_str:
|
|
||||||
return root
|
|
||||||
|
|
||||||
# try to find the longest attribute/index match at the start of path_str
|
|
||||||
# we split by underscores, but we have to be careful about names that contain underscores
|
|
||||||
parts = path_str.split('_')
|
|
||||||
|
|
||||||
# greedy search: try to consume as many parts as possible to form a valid attribute name
|
|
||||||
current_attr = ""
|
|
||||||
for i in range(len(parts), 0, -1):
|
|
||||||
candidate = "_".join(parts[:i])
|
|
||||||
|
|
||||||
# check if candidate is an attribute
|
|
||||||
if hasattr(root, candidate):
|
|
||||||
next_root = getattr(root, candidate)
|
|
||||||
remaining = "_".join(parts[i:])
|
|
||||||
return get_target_module(next_root, remaining)
|
|
||||||
|
|
||||||
# check if candidate is an index (for ModuleList/Sequential)
|
|
||||||
if candidate.isdigit() and isinstance(root, (nn.Sequential, nn.ModuleList)):
|
|
||||||
idx = int(candidate)
|
|
||||||
if idx < len(root):
|
|
||||||
next_root = root[idx]
|
|
||||||
remaining = "_".join(parts[i:])
|
|
||||||
return get_target_module(next_root, remaining)
|
|
||||||
|
|
||||||
# if we are here, we failed to match any attribute.
|
|
||||||
# heuristic: maybe the model uses 'proj' but lora says 'linear'? (unlikely here)
|
|
||||||
return None
|
|
||||||
|
|
||||||
updates = 0
|
|
||||||
with torch.no_grad():
|
|
||||||
prefixes = set()
|
|
||||||
for key in state_dict.keys():
|
|
||||||
if "lokr_w1" in key:
|
|
||||||
prefixes.add(key.replace(".lokr_w1", ""))
|
|
||||||
|
|
||||||
for prefix in prefixes:
|
|
||||||
# clean prefix: remove 'lora_unet_'
|
|
||||||
search_path = prefix.replace("lora_unet_", "")
|
|
||||||
|
|
||||||
# traverse
|
|
||||||
target_module = get_target_module(pipe.transformer, search_path)
|
|
||||||
|
|
||||||
if target_module and hasattr(target_module, "weight"):
|
|
||||||
# load weights
|
|
||||||
w1 = state_dict[f"{prefix}.lokr_w1"].to(device, dtype=dtype)
|
|
||||||
w2 = state_dict[f"{prefix}.lokr_w2"].to(device, dtype=dtype)
|
|
||||||
alpha = state_dict.get(f"{prefix}.alpha", None)
|
|
||||||
|
|
||||||
# scale math
|
|
||||||
current_scale = lokr_scale
|
|
||||||
if alpha is not None:
|
|
||||||
if isinstance(alpha, torch.Tensor):
|
|
||||||
alpha = alpha.to(device, dtype=dtype)
|
|
||||||
current_scale *= (alpha / w1.shape[0])
|
|
||||||
|
|
||||||
# kron
|
|
||||||
delta = torch.kron(w1, w2) * current_scale
|
|
||||||
|
|
||||||
# device align
|
|
||||||
if delta.device != target_module.weight.device:
|
|
||||||
delta = delta.to(target_module.weight.device)
|
|
||||||
|
|
||||||
# inject
|
|
||||||
if target_module.weight.shape == delta.shape:
|
|
||||||
target_module.weight.add_(delta)
|
|
||||||
updates += 1
|
|
||||||
elif target_module.weight.shape == delta.T.shape:
|
|
||||||
target_module.weight.add_(delta.T)
|
|
||||||
updates += 1
|
|
||||||
else:
|
|
||||||
print(f"shape mismatch {search_path}: {target_module.weight.shape} vs {delta.shape}")
|
|
||||||
else:
|
|
||||||
# keep this silent unless debugging, 120 success is usually enough for visual impact
|
|
||||||
pass
|
|
||||||
|
|
||||||
print(f"recursive solver successfully injected {updates} lokr layers.")
|
|
||||||
|
|
||||||
except Exception as e:
|
|
||||||
import traceback
|
|
||||||
traceback.print_exc()
|
|
||||||
print(f"lokr injection failed: {e}")
|
|
||||||
|
|
||||||
# --- end of surgery ---
|
|
||||||
|
|
||||||
# --- end of surgery ---
|
|
||||||
|
|
||||||
# # Apply the same optimizations from the first version
|
# # Apply the same optimizations from the first version
|
||||||
# pipe.transformer.__class__ = QwenImageTransformer2DModel
|
# pipe.transformer.__class__ = QwenImageTransformer2DModel
|
||||||
# pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
|
# pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
|
||||||
|
|||||||
Reference in New Issue
Block a user