python/src/converter.py:222

• With the hardlinking removed, tie_word_embeddings=True + embedding_found=False now silently results in no output projection being written and no warning being emitted. Consider restoring an explicit warning/error in this branch when embeddings weren’t exported (since the runtime will not have an output weight to use).

    if tie_word_embeddings:
        if embedding_found:
            for name in OUTPUT_NAMES:
                if name in state_dict:
                    saved_tensor_full_names.add(name)

python/src/converter.py:223

• When tie_word_embeddings is enabled, the converter no longer overwrites/removes any pre-existing output_weight.weights in the output directory. If a previous (partial) conversion left that file behind, it will remain and can make the output directory inconsistent. Consider deleting output_weight.weights when tie_word_embeddings is true (or ensuring the output dir is cleaned before conversion).

    if tie_word_embeddings:
        if embedding_found:
            for name in OUTPUT_NAMES:
                if name in state_dict:
                    saved_tensor_full_names.add(name)
    else: