U-Cast param-count Modal job

ucast_params.py

@@ -0,0 +1,35 @@
+"""Non-blocked U-Cast claim: count parameters from the RELEASED checkpoint
+(salv47/u-cast/ucast.ckpt) — weights confirmed present on HF, no ERA5 needed."""
+import modal
+
+image = (modal.Image.debian_slim(python_version="3.11")
+         .pip_install("torch", "huggingface_hub", "safetensors"))
+app = modal.App("ucast-params", image=image)
+
+
+@app.function(timeout=1200)
+def count():
+    import torch, collections
+    from huggingface_hub import hf_hub_download
+    p = hf_hub_download("salv47/u-cast", "ucast.ckpt")
+    ckpt = torch.load(p, map_location="cpu", weights_only=False)
+    sd = ckpt.get("state_dict", ckpt)
+    # model weights only (exclude optimizer/EMA duplicates if any obvious prefix)
+    total = sum(v.numel() for v in sd.values() if hasattr(v, "numel"))
+    # try to separate EMA / model if both present
+    groups = collections.Counter()
+    for k in sd:
+        groups[k.split(".")[0]] += sd[k].numel() if hasattr(sd[k], "numel") else 0
+    return {"total_state_dict_params": int(total),
+            "millions": round(total / 1e6, 1),
+            "top_level_groups": dict(groups.most_common(8)),
+            "num_tensors": len(sd)}
+
+
+@app.local_entrypoint()
+def main():
+    import json
+    r = count.remote()
+    print(json.dumps(r, indent=2))
+    with open("ucast_params.json", "w") as f:
+        json.dump(r, f, indent=2)