Spaces:
Running on Zero
Running on Zero
Update app.py
Browse files
app.py
CHANGED
|
@@ -253,6 +253,7 @@ pipe.load_lora_weights(
|
|
| 253 |
pipe.fuse_lora()
|
| 254 |
print("lightning lora fused.")
|
| 255 |
|
|
|
|
| 256 |
# --- 2. manual surgery for lokr (snofs) ---
|
| 257 |
print("attempting manual lokr injection for snofs...")
|
| 258 |
|
|
@@ -274,32 +275,38 @@ try:
|
|
| 274 |
prefixes.add(key.replace(".lokr_w1", ""))
|
| 275 |
|
| 276 |
for prefix in prefixes:
|
| 277 |
-
# extract weights
|
| 278 |
w1 = state_dict[f"{prefix}.lokr_w1"].to(device, dtype=dtype)
|
| 279 |
w2 = state_dict[f"{prefix}.lokr_w2"].to(device, dtype=dtype)
|
| 280 |
alpha = state_dict.get(f"{prefix}.alpha", None)
|
| 281 |
|
| 282 |
-
#
|
| 283 |
-
|
| 284 |
-
# if alpha is present, scale = alpha / w1.shape[0] (or similar convention)
|
| 285 |
-
# here we will assume simple multiplication or alpha scaling if provided
|
| 286 |
-
scale = lokr_scale
|
| 287 |
if alpha is not None:
|
| 288 |
-
|
|
|
|
|
|
|
|
|
|
| 289 |
|
| 290 |
# compute delta: kronecker product
|
| 291 |
# w1: (a, b), w2: (c, d) -> result: (a*c, b*d)
|
| 292 |
-
|
| 293 |
-
delta = torch.kron(w1, w2) * scale
|
| 294 |
|
| 295 |
# find target layer in model
|
| 296 |
-
# prefix example: "transformer_blocks.0.attn.add_k_proj"
|
| 297 |
-
# pipe.transformer matches this structure directly
|
| 298 |
path_parts = prefix.split('.')
|
| 299 |
target = pipe.transformer
|
|
|
|
|
|
|
| 300 |
try:
|
| 301 |
for part in path_parts:
|
| 302 |
target = getattr(target, part)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 303 |
|
| 304 |
# check shapes
|
| 305 |
if target.weight.shape == delta.shape:
|
|
@@ -307,12 +314,14 @@ try:
|
|
| 307 |
updates += 1
|
| 308 |
else:
|
| 309 |
print(f"shape mismatch for {prefix}: model {target.weight.shape} vs lora {delta.shape}")
|
| 310 |
-
|
| 311 |
print(f"layer not found: {prefix}")
|
| 312 |
|
| 313 |
print(f"successfully injected {updates} lokr layers manually.")
|
| 314 |
|
| 315 |
except Exception as e:
|
|
|
|
|
|
|
| 316 |
print(f"lokr injection failed: {e}")
|
| 317 |
print("running with lightning lora only.")
|
| 318 |
|
|
|
|
| 253 |
pipe.fuse_lora()
|
| 254 |
print("lightning lora fused.")
|
| 255 |
|
| 256 |
+
|
| 257 |
# --- 2. manual surgery for lokr (snofs) ---
|
| 258 |
print("attempting manual lokr injection for snofs...")
|
| 259 |
|
|
|
|
| 275 |
prefixes.add(key.replace(".lokr_w1", ""))
|
| 276 |
|
| 277 |
for prefix in prefixes:
|
| 278 |
+
# extract weights and FORCE TO DEVICE
|
| 279 |
w1 = state_dict[f"{prefix}.lokr_w1"].to(device, dtype=dtype)
|
| 280 |
w2 = state_dict[f"{prefix}.lokr_w2"].to(device, dtype=dtype)
|
| 281 |
alpha = state_dict.get(f"{prefix}.alpha", None)
|
| 282 |
|
| 283 |
+
# handle scale/alpha math carefully
|
| 284 |
+
current_scale = lokr_scale
|
|
|
|
|
|
|
|
|
|
| 285 |
if alpha is not None:
|
| 286 |
+
# alpha is a tensor, move it to gpu
|
| 287 |
+
if isinstance(alpha, torch.Tensor):
|
| 288 |
+
alpha = alpha.to(device, dtype=dtype)
|
| 289 |
+
current_scale *= (alpha / w1.shape[0])
|
| 290 |
|
| 291 |
# compute delta: kronecker product
|
| 292 |
# w1: (a, b), w2: (c, d) -> result: (a*c, b*d)
|
| 293 |
+
delta = torch.kron(w1, w2) * current_scale
|
|
|
|
| 294 |
|
| 295 |
# find target layer in model
|
|
|
|
|
|
|
| 296 |
path_parts = prefix.split('.')
|
| 297 |
target = pipe.transformer
|
| 298 |
+
|
| 299 |
+
layer_found = True
|
| 300 |
try:
|
| 301 |
for part in path_parts:
|
| 302 |
target = getattr(target, part)
|
| 303 |
+
except AttributeError:
|
| 304 |
+
layer_found = False
|
| 305 |
+
|
| 306 |
+
if layer_found:
|
| 307 |
+
# double check devices before adding
|
| 308 |
+
if target.weight.device != delta.device:
|
| 309 |
+
delta = delta.to(target.weight.device)
|
| 310 |
|
| 311 |
# check shapes
|
| 312 |
if target.weight.shape == delta.shape:
|
|
|
|
| 314 |
updates += 1
|
| 315 |
else:
|
| 316 |
print(f"shape mismatch for {prefix}: model {target.weight.shape} vs lora {delta.shape}")
|
| 317 |
+
else:
|
| 318 |
print(f"layer not found: {prefix}")
|
| 319 |
|
| 320 |
print(f"successfully injected {updates} lokr layers manually.")
|
| 321 |
|
| 322 |
except Exception as e:
|
| 323 |
+
import traceback
|
| 324 |
+
traceback.print_exc()
|
| 325 |
print(f"lokr injection failed: {e}")
|
| 326 |
print("running with lightning lora only.")
|
| 327 |
|