Spaces:
Sleeping
Sleeping
Trim preload to 1 model + LoRA (150GB cap)
Browse files
README.md
CHANGED
|
@@ -13,17 +13,15 @@ startup_duration_timeout: "30m"
|
|
| 13 |
preload_from_hub:
|
| 14 |
- techfreakworm/wan-lightning-loras
|
| 15 |
- techfreakworm/wan2.2-t2v-a14b
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
#
|
| 21 |
-
#
|
| 22 |
-
# and load via cached repo IDs above. Mounts are still created by
|
| 23 |
-
# scripts/create_space.py for the LoRA mirror (works correctly there).
|
| 24 |
# ZeroGPU hardware is set programmatically by scripts/create_space.py
|
| 25 |
-
# (SpaceHardware.ZERO_A10G — empirically the live Blackwell ZeroGPU V2
|
| 26 |
-
# as of May 2026).
|
| 27 |
---
|
| 28 |
|
| 29 |
# Wan Studio
|
|
|
|
| 13 |
preload_from_hub:
|
| 14 |
- techfreakworm/wan-lightning-loras
|
| 15 |
- techfreakworm/wan2.2-t2v-a14b
|
| 16 |
+
# Preload trimmed to the single most-used model (~58 GB) + LoRA mirror
|
| 17 |
+
# (~5 GB) to stay under ZeroGPU's 150 GB ephemeral storage cap. Other
|
| 18 |
+
# checkpoints (Wan 2.1 T2V, Wan 2.x I2V variants) download on first
|
| 19 |
+
# generate of that mode. HF Volume mounts truncate small JSON files
|
| 20 |
+
# (transformer/config.json comes back as 290B instead of 495B), so we
|
| 21 |
+
# bypass mounts in pipelines/handle.py and load via cached repo IDs.
|
|
|
|
|
|
|
| 22 |
# ZeroGPU hardware is set programmatically by scripts/create_space.py
|
| 23 |
+
# (SpaceHardware.ZERO_A10G — empirically the live Blackwell ZeroGPU V2
|
| 24 |
+
# pool as of May 2026).
|
| 25 |
---
|
| 26 |
|
| 27 |
# Wan Studio
|