Modernize: pin CPU torch, download weights from model repo

Preventive fix for the same failure mode hitting other spaces in this
org: unpinned ultralytics/torch pulls full CUDA wheels on cpu-basic
hardware. Pin CPU-only torch/torchvision and a stable ultralytics
version, redirect framework caches to /tmp, and load the model from
the FathomNet/megalodon-2023-yolov8 model repo via huggingface_hub
instead of bundling the weight file. Bump gradio to 6.9.0.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

README.md

@@ -4,7 +4,7 @@ emoji: 🦈
 colorFrom: green
 colorTo: blue
 sdk: gradio
-sdk_version: 4.36.1
+sdk_version: 6.9.0
 app_file: app.py
 pinned: true
 license: cc-by-4.0

app.py

@@ -1,9 +1,23 @@
+import os
+
+# Keep framework caches on ephemeral scratch space rather than the Space's
+# persistent disk, to avoid exceeding the storage quota:
+# https://discuss.huggingface.co/t/how-to-fix-workload-evicted-storage-limit-exceeded-50g-error-in-huggingface-spaces/169258
+os.environ.setdefault("TORCH_HOME", "/tmp/torch_cache")
+os.environ.setdefault("HUB_DIR", "/tmp/torch_hub")
+os.environ.setdefault("HF_HOME", "/tmp/hf_home")
+os.environ.setdefault("TMPDIR", "/tmp")
+
 import glob
 import gradio as gr
+from huggingface_hub import hf_hub_download
 from ultralytics import YOLO
 
-model_path = "best.pt"
-model = YOLO(model_path)
+MODEL_PATH = hf_hub_download(
+    repo_id="FathomNet/megalodon-2023-yolov8",
+    filename="mbari-megalodon-yolov8x.pt",
+)
+model = YOLO(MODEL_PATH)
 
 
 PREDICT_KWARGS = {

best.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:24fe6d9808fdd44b64d2fbfc23bbe6455520eb7b392d3664c92cbcdf88a09196
-size 136714110

requirements.txt

@@ -1 +1,5 @@
-ultralytics==8.0.158
+--extra-index-url https://download.pytorch.org/whl/cpu
+torch==2.14.0
+torchvision==0.29.0
+ultralytics==8.3.108
+huggingface_hub