Fix RWKV v6 model conversion

[MollySophia]

Make sure to read the contributing guidelines before submitting a PR

It seems that some rwkv tensors are made FP16 rather than FP32 after specific commits. However, ggml_cuda_op_bin_bcast requires src1->type == FP32. As a result, newly converted RWKV models cannot run with cuda, while existing files aren't affected. This PR fixes the issue above.

Also add LLAMA_EXAMPLE_PERPLEXITY in the examples list of parameter --no-context-shift so that models without context shift support can do llama-perplexity again.

[ggerganov]

This is likely caused by this change where I removed the squeeze() during conversion:

llama.cpp/convert_hf_to_gguf.py

Lines 298 to 301 in 0bf2d10

	for new_name, data_torch in (self.modify_tensors(data_torch, name, bid)):
	# TODO: why do we squeeze here?
	# data = data_torch.squeeze().numpy()
	data = data_torch.numpy()

[MollySophia]

This is likely caused by this change where I removed the squeeze() during conversion:

llama.cpp/convert_hf_to_gguf.py

Lines 298 to 301 in 0bf2d10

	for new_name, data_torch in (self.modify_tensors(data_torch, name, bid)):
	# TODO: why do we squeeze here?
	# data = data_torch.squeeze().numpy()
	data = data_torch.numpy()

I see. So there's another way to fix this: squeeze them in rwkv6's modify_tensors(), rather than adding them to the F32 list?
Should both solve the issue. I wonder which way is better

[ggerganov]

Don't think there is any significant advantage one way or the other. Maybe squeezing in modify_tensors is a bit more localized.

[MollySophia]

Don't think there is any significant advantage one way or the other. Maybe squeezing in modify_tensors is a bit more localized.

Yeah that's what I meant. Let me change to use the more localized way then.

[ggerganov]

This is good to merge?

[MollySophia]

This is good to merge?

Yes. Thanks a lot for your time!