Deploy MODUS 3-tab any-to-any demo (ZeroGPU)

README.md

@@ -4,7 +4,7 @@ emoji: 🎨
 colorFrom: indigo
 colorTo: purple
 sdk: gradio
-sdk_version: 4.44.1
+sdk_version: 5.12.0
 app_file: app.py
 pinned: false
 short_description: MODUS any-to-any multimodal demo (16 aligned modalities)

demo_modus.py

@@ -557,7 +557,7 @@ def build_ui() -> gr.Blocks:
                     if ex:
                         gr.Examples(ex, inputs=t1_image, label="Example images")
                 with gr.Column():
-                    t1_gallery = gr.Gallery(label="Results", columns=[4])
+                    t1_gallery = gr.Gallery(label="Results", columns=4)
                     t1_status = gr.Markdown()
             t1_cond.change(_toggle_condition_input, t1_cond, [t1_image, t1_text])
             t1_btn.click(

requirements.txt

@@ -1,22 +1,14 @@
 # MODUS demo — HF ZeroGPU Space requirements.
 #
-# NOTE: the versions that passed the cluster dry-run are ARM / nvidia-container
-# builds (torch 2.6 nv24.12, pycocotools nv, flash-attn 2.4.2 for GH200). HF
-# ZeroGPU is x86 + H200, so torch/flash-attn are pinned to standard x86 wheels
-# below and MAY need adjustment after reading the first Space build log.
-
-# ZeroGPU runtime helper (provides @spaces.GPU). Not installed on the cluster.
+# gradio >= 5.12 is REQUIRED for ZeroGPU to attribute a logged-in / PRO user's
+# quota from the browser (older gradio does not forward the user's ZeroGPU token,
+# so every browser request falls back to the tiny anonymous tier). Let gradio pull
+# its own matching gradio-client / fastapi / starlette / jinja2.
 spaces
-
-# --- web stack (verified working with the demo; keep these pins) ---
-gradio==4.44.1
-gradio-client==1.3.0
-fastapi==0.112.2
-starlette==0.38.6
-jinja2==3.1.4
+gradio==5.12.0
 
 # --- HF / model stack ---
-huggingface-hub>=0.26,<1.0   # >=0.26 for `spaces`; <1.0 for gradio 4.44 compat
+huggingface-hub>=0.26,<1.0
 transformers==4.46.2
 accelerate==1.1.1
 safetensors==0.4.5
@@ -34,13 +26,6 @@ scikit-learn==1.7.2
 pyyaml==6.0.3
 tqdm==4.67.1
 
-# --- torch/torchvision: the ZeroGPU base image ships a CUSTOM torch (2.11.0) with
-#     NO matching torchvision wheel; installing torchvision downgrades torch to a
-#     vanilla build and breaks ZeroGPU. So we do NOT install either — torch comes
-#     from the base image and data/transforms.py was patched to be torchvision-free. ---
-
-# --- flash-attn: HARD top-level import in modeling/bagel/{qwen2_navit,siglip_navit}.py.
-#     Must match the Space torch/CUDA/python. Use a prebuilt wheel; building from
-#     source on the Space has no nvcc and will fail. Adjust the version/URL after
-#     the first build log tells you the Space's torch+cu+cp tags. Example:
-# flash-attn==2.6.3
+# --- torch/torchvision: provided by the ZeroGPU base image (custom torch 2.11);
+#     do NOT install (a pip torchvision would downgrade torch and break ZeroGPU).
+#     data/transforms.py is torchvision-free. No flash-attn (SDPA fallback). ---