Support tie embedding for chatglm models

[piDack]

Make sure to read the contributing guidelines before submitting a PR

The GLM 1.5B model uses tied embeddings, but the chatglm architecture does not take this into account, resulting in a ‘missing tensor ‘output.weight’’ error.this pr is fixed this issue.

[CISC]

BTW, I checked both the safetensors and GGUFs, and even though it has tied word embeddings enabled the safetensors have both sets of (most likely identical) weights, and thus the GGUFs also.

Which GGUF file gave you this error?

To make such a GGUF you would have to do something similar to this (which could make sense to incorporate into this PR, but I'd like to know where and why such a GGUF already exists first):

llama.cpp/convert_hf_to_gguf.py

Lines 4309 to 4313 in 764b856

	# assuming token_embd.weight is seen before output.weight
	if self._tok_embd is not None and new_name == output_name:
	if torch.equal(self._tok_embd, data_torch):
	logger.debug(f"{output_name} is equivalent to {tok_embd_name}, omitting")
	return []

[piDack]

BTW, I checked both the safetensors and GGUFs, and even though it has tied word embeddings enabled the safetensors have both sets of (most likely identical) weights, and thus the GGUFs also.

Which GGUF file gave you this error?

To make such a GGUF you would have to do something similar to this (which could make sense to incorporate into this PR, but I'd like to know where and why such a GGUF already exists first):

llama.cpp/convert_hf_to_gguf.py

Lines 4309 to 4313 in 764b856

	# assuming token_embd.weight is seen before output.weight
	if self._tok_embd is not None and new_name == output_name:
	if torch.equal(self._tok_embd, data_torch):
	logger.debug(f"{output_name} is equivalent to {tok_embd_name}, omitting")
	return []

I’m sorry, this weight does not come from Hugging Face, but is for my own testing purposes. Indeed, there is no output.weight in the safetensor.
huggingface weight output

myself weight output

I think adding relevant checks is the right thing to do; at the very least, it can prevent some problems.right？

[engineer1109]

Good Job