-
Notifications
You must be signed in to change notification settings - Fork 9.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
convert : fix vocab size when not defined in hparams #3421
Conversation
7d50d3c
to
9c621af
Compare
# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py | ||
tokenizer = AutoTokenizer.from_pretrained(dir_model) | ||
|
||
# The number of tokens in tokenizer.json can differ from the expected vocab size. | ||
# This causes downstream issues with mismatched tensor sizes when running the inference | ||
vocab_size = hparams.get("vocab_size", len(tokenizer.vocab)) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Using len(tokenizer.vocab) is not ideal, we really should use vocab_size from AutoConfig, which has an architecture-specific default. Although, this whole thing is a hack to work around downstream issues, right? I would much rather just store the vocab size in the GGUF and avoid this whole padding tokens mess.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I guess the change is fine, although I don't fully understand all intricacies with the vocab size. As long as this does not break conversion of the standard Falcon and Starcoder models, it should be OK to merge
I wasn't able to use the Falcon convert script as-is, because it was renamed from RWForCausalLM to FalconForCausalLM. I'll make a PR for that. |
…example * 'master' of github.com:ggerganov/llama.cpp: (24 commits) convert : fix Baichuan2 models by using vocab size in config.json (ggerganov#3299) readme : add project status link ggml : fix build after ggerganov#3329 llm : add Refact model (ggerganov#3329) sync : ggml (conv 1d + 2d updates, UB fixes) (ggerganov#3468) finetune : readme fix typo (ggerganov#3465) ggml : add RISC-V Vector Support for K-Quants and improved the existing intrinsics (ggerganov#3453) main : consistent prefix/suffix coloring (ggerganov#3425) llama : fix session saving/loading (ggerganov#3400) llama : expose model's rope_freq_scale in the API (ggerganov#3418) metal : alibi for arbitrary number of heads (ggerganov#3426) cmake : make LLAMA_NATIVE flag actually use the instructions supported by the processor (ggerganov#3273) Work on the BPE tokenizer (ggerganov#3252) convert : fix vocab size when not defined in hparams (ggerganov#3421) cmake : increase minimum version for add_link_options (ggerganov#3444) CLBlast: Add broadcast support for matrix multiplication (ggerganov#3402) gguf : add BERT, MPT, and GPT-J arch info (ggerganov#3408) gguf : general usability improvements (ggerganov#3409) cmake : make CUDA flags more similar to the Makefile (ggerganov#3420) finetune : fix ggerganov#3404 (ggerganov#3437) ...
If vocab_size is somehow missing from config.json, or in the case of the previous GPT-NeoX script, it is ignored entirely, we can end up in a case where vocab_size is less than
len(reverse_vocab)
, even though the purpose of vocab_size is to enlarge the vocabulary with padding tokensUse
len(tokenizer.vocab)
instead of attempting to interpret JSON directly, to account for added tokens. Also, add the missing hparams check to the GPT-NeoX script.With this change, GPT-NeoX is now attempting to use added tokens, though it is failing due to reasons described in PR #3405. Before this change, it wasn't even trying.
cc @goerch (yes, I know this conflicts with your PR)