Spaces:
Running on Zero
Running on Zero
Pin torch 2.9.1 (tested v2 version), hide UI-only event from API, document ZeroGPU size
Browse files- README.md +1 -1
- app.py +1 -1
- requirements.txt +2 -2
README.md
CHANGED
|
@@ -54,7 +54,7 @@ How this Space runs it:
|
|
| 54 |
|
| 55 |
Measured end to end on an RTX 5090 (GPU phase, warm, 89 frames): 540p 19s, 720p 28s, 1080p 48-54s, and 720p with 267 frames 68s. The first call after the model is loaded takes longer.
|
| 56 |
The GPU time requested from ZeroGPU is estimated from those numbers, so a clip that does not fit gets an error naming the limit instead of being cut off.
|
| 57 |
-
The status line reports the GPU that actually ran.
|
| 58 |
|
| 59 |
**Temporal context**: Auto picks a window of 80 frames up to 720p and 24 frames at 1080p. An 80-frame window costs about the same as 8 frames up to 720p, but at 1080p it is about 2.5x slower, which is why Auto narrows it there. Peak VRAM depends on resolution and window size, not clip length (about 6 GB at 720p with 8 frames, 13 GB with 80, 25 GB at 1080p with 80).
|
| 60 |
|
|
|
|
| 54 |
|
| 55 |
Measured end to end on an RTX 5090 (GPU phase, warm, 89 frames): 540p 19s, 720p 28s, 1080p 48-54s, and 720p with 267 frames 68s. The first call after the model is loaded takes longer.
|
| 56 |
The GPU time requested from ZeroGPU is estimated from those numbers, so a clip that does not fit gets an error naming the limit instead of being cut off.
|
| 57 |
+
The status line reports the GPU that actually ran. ZeroGPU's default size is half of an NVIDIA RTX PRO 6000 Blackwell (48 GB), and v2 peaks at 6-25 GB, so the default size is enough (the 96 GB `xlarge` size costs 2x quota and is not needed). PyTorch is pinned to 2.9.1, the version v2 was tested with: a newer PyTorch changes autocast dtypes and breaks its attention code.
|
| 58 |
|
| 59 |
**Temporal context**: Auto picks a window of 80 frames up to 720p and 24 frames at 1080p. An 80-frame window costs about the same as 8 frames up to 720p, but at 1080p it is about 2.5x slower, which is why Auto narrows it there. Peak VRAM depends on resolution and window size, not clip length (about 6 GB at 720p with 8 frames, 13 GB with 80, 25 GB at 1080p with 80).
|
| 60 |
|
app.py
CHANGED
|
@@ -1169,7 +1169,7 @@ with gr.Blocks(title="CorridorKey") as demo:
|
|
| 1169 |
version.change(
|
| 1170 |
_on_version, inputs=[version],
|
| 1171 |
outputs=[v2_controls, v1_controls, process_btn_v2, process_btn, v2_examples, v1_examples],
|
| 1172 |
-
|
| 1173 |
)
|
| 1174 |
|
| 1175 |
process_btn.click(
|
|
|
|
| 1169 |
version.change(
|
| 1170 |
_on_version, inputs=[version],
|
| 1171 |
outputs=[v2_controls, v1_controls, process_btn_v2, process_btn, v2_examples, v1_examples],
|
| 1172 |
+
api_visibility="private",
|
| 1173 |
)
|
| 1174 |
|
| 1175 |
process_btn.click(
|
requirements.txt
CHANGED
|
@@ -6,8 +6,8 @@ onnx
|
|
| 6 |
onnxconverter-common
|
| 7 |
spaces
|
| 8 |
gradio[mcp]
|
| 9 |
-
torch
|
| 10 |
-
torchvision
|
| 11 |
timm
|
| 12 |
https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.9.0/flash_attn-2.8.3+cu126torch2.9-cp310-cp310-linux_x86_64.whl
|
| 13 |
transformers>=4.57,<5
|
|
|
|
| 6 |
onnxconverter-common
|
| 7 |
spaces
|
| 8 |
gradio[mcp]
|
| 9 |
+
torch==2.9.1
|
| 10 |
+
torchvision==0.24.1
|
| 11 |
timm
|
| 12 |
https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.9.0/flash_attn-2.8.3+cu126torch2.9-cp310-cp310-linux_x86_64.whl
|
| 13 |
transformers>=4.57,<5
|