报错：Cannot copy out of meta tensor; no data!

[zhaojigang]

File "D:\ai\self_use_package\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1362, in convert
raise NotImplementedError(
NotImplementedError: Cannot copy out of meta tensor; no data! Please use torch.nn.Module.to_empty() instead of torch.nn.Module.to() when moving module from meta to a different device.

[idleeyan]

Same error here (ComfyUI 0.34.5 / torch 2.9.1+cu130 / RTX 2060 SUPER, INT8 + Q4_K_M GGUF).

Root cause: load_llada2_gguf_encoder builds the text encoder with init_empty_weights(include_buffers=True), so every buffer starts on the meta device. The GGUF only provides parameters, and the rotary embedding's inv_freq buffer is never materialized — the first forward pass then dies in rotary_emb at .to(x.device) with "Cannot copy out of meta tensor; no data!".

Local fix (verified): after all GGUF weights are bound, materialize any meta buffer — recompute inv_freq with the module's own rope init function:

# in load_llada2_gguf_encoder, after the "Load all remaining tensors" loop, before model.eval()
for _m in model.modules():
    for _name, _buf in list(_m.named_buffers(recurse=False)):
        if _buf is None or not _buf.is_meta:
            continue
        if _name == "inv_freq" and hasattr(_m, "rope_init_fn") and hasattr(_m, "config"):
            _inv, _scale = _m.rope_init_fn(_m.config, "cpu")
            _m.register_buffer(_name, _inv, persistent=False)
            if hasattr(_m, "attention_scaling"):
                _m.attention_scaling = _scale
        else:
            _m.register_buffer(_name, torch.zeros(_buf.shape, dtype=dtype), persistent=False)

After the patch: 0 meta buffers remain on load and _encode_text runs clean (embeds (1, 277, 2560) bf16). The other init_empty_weights sites (INT8/BF16 transformer, VAE) contain no buffers, so they are unaffected.