Upload folder using huggingface_hub

sansi/hub.py

@@ -66,10 +66,26 @@ def load(repo_or_dir, device="cuda", revision=None, backbone=None):
     lm.config.use_cache = False
     hidden_size = lm.lm_head.in_features
     letter_ids, _ = option_token_ids(tok)
-    lm = PeftModel.from_pretrained(lm, path)
+    # On ZeroGPU there is no GPU at import time: peft's infer_device() would route the
+    # adapter safetensors onto CUDA and fail. Load the adapter on CPU; model.to(device)
+    # below moves everything at once.
+    import peft as _peft
+    _infer = _peft.peft_model.infer_device
+    _peft.peft_model.infer_device = lambda: "cpu"
+    try:
+        lm = PeftModel.from_pretrained(lm, path)
+    finally:
+        _peft.peft_model.infer_device = _infer
+    # Same for the readout: load_file must not touch CUDA before a GPU is attached.
+    from safetensors.torch import load_file as _load_file
+    from safetensors import safe_open
+    _readout = {}
+    with safe_open(os.path.join(path, "readout.safetensors"), framework="pt", device="cpu") as f:
+        for k in f.keys():
+            _readout[k] = f.get_tensor(k)
     model = OuroSanSi(lm, letter_ids, hidden_size, n_loop_params=cfg["n_loop_params"],
                       readout_rank=cfg["readout_rank"])
-    model.load_readout({"readout": load_file(os.path.join(path, "readout.safetensors"))})
+    model.load_readout({"readout": _readout})
     model.loops = cfg["loops"]
     return model.to(device).eval(), tok