Upload folder using huggingface_hub

AGENTS.md

@@ -21,8 +21,7 @@ For tasks of **reboot** from a new codex session:
 1. Read doc/v0/PRD.md, doc/v0/Plan.md, doc/v0/Progress.md for baseline implementation
 2. Read doc/v1/PRD.md, doc/v1/Plan.md, doc/v1/Progress.md for edge refinement (v1)
 3. Read doc/v4/PRD.md, doc/v4/Plan.md, doc/v4/Progress.md for SAM 2.1 integration (card + hand)
-4. Read doc/v5/PRD.md, doc/v5/Plan.md, doc/v5/Progress.md for the Gradio/ZeroGPU deployment port
-5. Assume this is a continuation of an existing project.
+4. Assume this is a continuation of an existing project.
 5. Summarize your understanding of the current state and propose the next concrete step without writing code yet.
 
 ## Project Overview

CLAUDE.md

@@ -21,8 +21,7 @@ For tasks of **reboot** from a new codex session:
 1. Read doc/v0/PRD.md, doc/v0/Plan.md, doc/v0/Progress.md for baseline implementation
 2. Read doc/v1/PRD.md, doc/v1/Plan.md, doc/v1/Progress.md for edge refinement (v1)
 3. Read doc/v4/PRD.md, doc/v4/Plan.md, doc/v4/Progress.md for SAM 2.1 integration (card + hand)
-4. Read doc/v5/PRD.md, doc/v5/Plan.md, doc/v5/Progress.md for the Gradio/ZeroGPU deployment port
-5. Assume this is a continuation of an existing project.
+4. Assume this is a continuation of an existing project.
 5. Summarize your understanding of the current state and propose the next concrete step without writing code yet.
 
 ## Project Overview

README.md

@@ -3,10 +3,8 @@ title: Ring Sizer
 emoji: "\U0001F48D"
 colorFrom: blue
 colorTo: purple
-sdk: gradio
-sdk_version: 5.49.0
-app_file: app.py
-python_version: "3.10"
+sdk: docker
+app_port: 7860
 ---
 
 # Ring Sizer

requirements.txt

@@ -7,22 +7,8 @@ flask>=3.0.0
 gunicorn>=21.2.0
 openai>=1.0.0
 supabase>=2.0.0
-# SAM 2.1 via HuggingFace transformers (card segmentation).
-# transformers pinned to the 4.x track for stability (SAM integration was
-# validated there). 4.x caps huggingface_hub at <1.0, which is fine for
-# everything downstream.
+# SAM 2.1 via HuggingFace transformers (card segmentation)
 torch>=2.4.0
 torchvision>=0.19.0
-transformers>=4.47.0,<5.0
-huggingface_hub>=0.33.5,<1.0
+transformers>=4.47.0
 pillow>=10.0.0
-# v5: HF ZeroGPU requires Gradio SDK; `spaces` provides @spaces.GPU (no-op off ZeroGPU).
-# Pinned to Gradio 5.x because:
-#   - Gradio 4.x imports `HfFolder` from huggingface_hub; 5.x dropped it.
-#   - Gradio 4.44 passes a dict as the template name to Starlette's
-#     `TemplateResponse(...)`, which breaks against the newer
-#     `starlette>=0.40` that Gradio's own deps pull in — crashes with
-#     `TypeError: unhashable type: 'dict'` on every page load.
-#   - Gradio 5.49 fixes both above.
-gradio>=5.49.0
-spaces>=0.30.0

src/sam_backend.py

@@ -23,35 +23,6 @@ INFERENCE_MAX_SIDE = 1024
 
 _model = None
 _processor = None
-_device: str = "cpu"
-
-
-def _select_device() -> str:
-    """Pick a torch device for SAM inference.
-
-    Returns ``"cuda"`` when a GPU is visible (HF ZeroGPU exposes CUDA even
-    at module import time via an emulation shim, so this picks the right
-    path both at startup and inside ``@spaces.GPU`` functions), otherwise
-    ``"cpu"``. Import of torch is local so CLI users without it still see
-    a clean error from the caller.
-    """
-    try:
-        import torch
-        if torch.cuda.is_available():
-            return "cuda"
-    except Exception:
-        pass
-    return "cpu"
-
-
-def get_sam2_device() -> str:
-    """Return the device the SAM singleton was loaded on.
-
-    Callers use this to move their ``processor(..., return_tensors="pt")``
-    outputs onto the same device as the model before the forward pass.
-    Returns ``"cpu"`` before ``get_sam2()`` has been called.
-    """
-    return _device
 
 
 def get_sam2() -> Tuple[object, object]:
@@ -61,26 +32,19 @@ def get_sam2() -> Tuple[object, object]:
     the HEAD-request retry storm that happens when huggingface.co is slow or
     unreachable but the weights are already on disk. On a true cache miss we
     fall through to a normal online load.
-
-    The model is placed on the device returned by ``_select_device()``.
-    HF ZeroGPU docs require CUDA placements to happen at module-level
-    startup for best performance — callers in ZeroGPU Spaces should invoke
-    ``get_sam2()`` once at import time so this runs before the first
-    ``@spaces.GPU``-wrapped request.
     """
-    global _model, _processor, _device
+    global _model, _processor
     if _model is None or _processor is None:
         from transformers import Sam2Model, Sam2Processor
-        _device = _select_device()
         t0 = time.time()
-        print(f"  Loading SAM 2.1 ({SAM2_MODEL_ID}) on {_device}...")
+        print(f"  Loading SAM 2.1 ({SAM2_MODEL_ID})...")
         try:
             _processor = Sam2Processor.from_pretrained(SAM2_MODEL_ID, local_files_only=True)
-            _model = Sam2Model.from_pretrained(SAM2_MODEL_ID, local_files_only=True).to(_device).eval()
-            print(f"  SAM 2.1 loaded (offline cache, {_device}) in {time.time() - t0:.1f}s")
+            _model = Sam2Model.from_pretrained(SAM2_MODEL_ID, local_files_only=True).to("cpu").eval()
+            print(f"  SAM 2.1 loaded (offline cache) in {time.time() - t0:.1f}s")
         except (OSError, ValueError):
             # Cache miss — fall back to online download.
             _processor = Sam2Processor.from_pretrained(SAM2_MODEL_ID)
-            _model = Sam2Model.from_pretrained(SAM2_MODEL_ID).to(_device).eval()
-            print(f"  SAM 2.1 loaded (online, {_device}) in {time.time() - t0:.1f}s")
+            _model = Sam2Model.from_pretrained(SAM2_MODEL_ID).to("cpu").eval()
+            print(f"  SAM 2.1 loaded (online) in {time.time() - t0:.1f}s")
     return _model, _processor

src/sam_card_detection.py

@@ -26,7 +26,7 @@ from .card_detection import (
     get_quad_dimensions,
     order_corners,
 )
-from .sam_backend import INFERENCE_MAX_SIDE as PROMPT_INFERENCE_MAX_SIDE, get_sam2, get_sam2_device
+from .sam_backend import INFERENCE_MAX_SIDE as PROMPT_INFERENCE_MAX_SIDE, get_sam2
 
 # HF Hub model id — tiny, small, base-plus, large
 SAM2_MODEL_ID = "facebook/sam2.1-hiera-small"
@@ -531,13 +531,6 @@ def detect_credit_card_sam_prompt(
         input_labels=input_labels,
         return_tensors="pt",
     )
-    # `original_sizes` is used after the forward pass for mask post-processing
-    # and scale calculations. Pull it to CPU before moving `inputs` to the
-    # model device so downstream code never has to chase device placement.
-    original_sizes_cpu = inputs["original_sizes"].cpu() if hasattr(inputs["original_sizes"], "cpu") else inputs["original_sizes"]
-    device = get_sam2_device()
-    if device != "cpu":
-        inputs = inputs.to(device)
     with torch.inference_mode():
         # multimask_output=True gives 3 masks per seed (small / medium / large
         # disambiguation of the prompt). Empirically this matters for card
@@ -549,13 +542,13 @@ def detect_credit_card_sam_prompt(
 
     # Score masks in the scaled 1024-space. Only the single winner is
     # upscaled to full resolution afterward, which avoids O(N) 12 MP resizes.
-    scaled_h = int(original_sizes_cpu[0][0].item())
-    scaled_w = int(original_sizes_cpu[0][1].item())
+    scaled_h = inputs["original_sizes"][0][0].item()
+    scaled_w = inputs["original_sizes"][0][1].item()
     scaled_area = float(scaled_h * scaled_w)
 
     masks_list = processor.post_process_masks(
         outputs.pred_masks.cpu(),
-        original_sizes_cpu,
+        inputs["original_sizes"],
         mask_threshold=0.0,
     )
     masks_tensor = masks_list[0]  # (num_prompts, num_candidates, H_s, W_s)

src/sam_hand_segmentation.py

@@ -22,7 +22,7 @@ from typing import List, Optional, Tuple
 import cv2
 import numpy as np
 
-from .sam_backend import INFERENCE_MAX_SIDE, get_sam2, get_sam2_device
+from .sam_backend import INFERENCE_MAX_SIDE, get_sam2
 
 
 def _downscale(image_bgr: np.ndarray) -> Tuple[np.ndarray, float]:
@@ -87,9 +87,6 @@ def segment_hand_sam(
         input_labels=[[prompt_labels]],
         return_tensors="pt",
     )
-    device = get_sam2_device()
-    if device != "cpu":
-        inputs = inputs.to(device)
     with torch.inference_mode():
         outputs = model(**inputs, multimask_output=True)