Spaces:
Paused
Paused
fix OOM: use mmap loading + assign (no memory copy)
Browse files
app.py
CHANGED
|
@@ -48,7 +48,12 @@ def _load_generator():
|
|
| 48 |
print("[init] Loading pre-quantized INT8 DiT...")
|
| 49 |
t0 = time.time()
|
| 50 |
|
| 51 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 52 |
step1x_params = Step1XParams(
|
| 53 |
in_channels=64, out_channels=64, vec_in_dim=768, context_in_dim=4096,
|
| 54 |
hidden_size=3072, mlp_ratio=4.0, num_heads=24, depth=19,
|
|
@@ -56,7 +61,9 @@ def _load_generator():
|
|
| 56 |
)
|
| 57 |
dit = Step1XEdit(step1x_params).float()
|
| 58 |
dit = torch.quantization.quantize_dynamic(dit, {torch.nn.Linear}, dtype=torch.qint8)
|
| 59 |
-
dit.load_state_dict(
|
|
|
|
|
|
|
| 60 |
print(f"[init] DiT loaded in {time.time()-t0:.0f}s")
|
| 61 |
|
| 62 |
# Load VAE
|
|
|
|
| 48 |
print("[init] Loading pre-quantized INT8 DiT...")
|
| 49 |
t0 = time.time()
|
| 50 |
|
| 51 |
+
# Use mmap to avoid doubling memory during load
|
| 52 |
+
print("[init] Loading with mmap (memory-mapped, no copy)...")
|
| 53 |
+
sd = torch.load(dit_path, map_location="cpu", weights_only=False, mmap=True)
|
| 54 |
+
print(f"[init] State dict mmap'd ({len(sd)} keys)")
|
| 55 |
+
|
| 56 |
+
# Build quantized model structure and load weights via assign (no copy)
|
| 57 |
step1x_params = Step1XParams(
|
| 58 |
in_channels=64, out_channels=64, vec_in_dim=768, context_in_dim=4096,
|
| 59 |
hidden_size=3072, mlp_ratio=4.0, num_heads=24, depth=19,
|
|
|
|
| 61 |
)
|
| 62 |
dit = Step1XEdit(step1x_params).float()
|
| 63 |
dit = torch.quantization.quantize_dynamic(dit, {torch.nn.Linear}, dtype=torch.qint8)
|
| 64 |
+
dit.load_state_dict(sd, strict=False, assign=True)
|
| 65 |
+
del sd
|
| 66 |
+
gc.collect()
|
| 67 |
print(f"[init] DiT loaded in {time.time()-t0:.0f}s")
|
| 68 |
|
| 69 |
# Load VAE
|