Add cache-busting to refresh button

app.py

@@ -5,7 +5,7 @@ import torch.nn.functional as F
 import math
 import time
 from transformers import AutoTokenizer
-from huggingface_hub import hf_hub_download, list_repo_files
+from huggingface_hub import hf_hub_download, HfApi
 
 class TuneableAttentionMHA(nn.Module):
     def __init__(self, d: int, h: int, r: int):
@@ -85,7 +85,8 @@ class ARHead(nn.Module):
 MODEL_REPO = "OpenTransformer/AGILLM-3-large"
 
 def get_latest_checkpoint():
-    files = list_repo_files(MODEL_REPO)
+    api = HfApi()
+    files = api.list_repo_files(MODEL_REPO, revision="main")
     ckpts = [f for f in files if f.startswith("checkpoints/") and f.endswith(".pt")]
     if not ckpts:
         raise ValueError("No checkpoints found in repo")
@@ -106,14 +107,14 @@ model_state = {
     "vocab": 0,
 }
 
-def load_model(ckpt_name=None):
+def load_model(ckpt_name=None, force_download=False):
     global model_state
     
     if ckpt_name is None:
         ckpt_name = get_latest_checkpoint()
     
     print(f"Loading checkpoint: {ckpt_name}")
-    ckpt_path = hf_hub_download(MODEL_REPO, ckpt_name)
+    ckpt_path = hf_hub_download(MODEL_REPO, ckpt_name, force_download=force_download)
     ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)
     cfg = ckpt["cfg"]
     
@@ -163,7 +164,8 @@ def check_for_updates():
     current_step = model_state["step"]
     
     if latest_step > current_step:
-        new_step, new_name = load_model(latest)
+        # Force fresh download, bypass cache
+        new_step, new_name = load_model(latest, force_download=True)
         return f"✅ Updated! Step {current_step:,} → {new_step:,}"
     else:
         return f"Already on latest (step {current_step:,})"