"""Dump the token-embedding matrix of a GGUF model as raw float32 (n_vocab x n_embd), dequantized exactly as llama.cpp
reads it, so that feeding E[t] as an input embedding is the same as feeding token t.
usage: dump_embd.py gguf/lfm3-orig-q5km.gguf gguf/lfm3-q5km.embd.f32
"""
import sys

import numpy as np

sys.path.insert(0, "C:/Users/user/llama-src/gguf-py")
from gguf import GGUFReader  # noqa: E402
from gguf.quants import dequantize  # noqa: E402

r = GGUFReader(sys.argv[1])
t = next(t for t in r.tensors if t.name == "token_embd.weight")
print(t.name, t.tensor_type.name, t.shape)
E = dequantize(t.data, t.tensor_type).astype(np.float32)
E = E.reshape(-1, E.shape[-1])
print("dequantized", E.shape, float(np.abs(E).mean()))
E.tofile(sys.argv[2])