Spaces:
Running on Zero
Running on Zero
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,15 +1,4 @@
|
|
| 1 |
-
"""P-H3-Image Studio — gr.Server studio + gr.Workflow canvas.
|
| 2 |
-
|
| 3 |
-
The bespoke studio at / replaces the classic form UI; the backend is unchanged from
|
| 4 |
-
PrunaAI/P-H3-Image: the pipeline is loaded once at startup, one component at a time (the host has
|
| 5 |
-
~97 GB), and slimmed to fit the slice with p_h3_image_fast.py (kernels only, no torch.compile):
|
| 6 |
-
AdaLN tables instead of the transformer's AdaLN projections, FP8 linears everywhere, fused Triton
|
| 7 |
-
DiT blocks, batched VAE tiles, and for the prompt upsampler a weight-only FP8 GEMV, decode
|
| 8 |
-
attention and speculative decoding with its MTP head in one CUDA graph per step. The studio's
|
| 9 |
-
Generate/Edit calls and the workflow canvas at /workflow all run through the same _render.
|
| 10 |
-
|
| 11 |
-
Server + workflow structure adapted from https://huggingface.co/spaces/briaai/Fibo-2.0-editing.
|
| 12 |
-
"""
|
| 13 |
|
| 14 |
import base64
|
| 15 |
import json
|
|
|
|
| 1 |
+
"""P-H3-Image Studio — gr.Server studio + gr.Workflow canvas."""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
|
| 3 |
import base64
|
| 4 |
import json
|