Commit Graph

8 Commits

Author SHA1 Message Date
Aaron Miller
47fbc0e309
non-llama: explicitly greedy sampling for temp<=0 (#901)
copied directly from llama.cpp - without this temp=0.0 will just
scale all the logits to infinity and give bad output
2023-06-08 11:08:30 -07:00
Aaron Miller
b14953e136
sampling: remove incorrect offset for n_vocab (#900)
no effect, but avoids a *potential* bug later if we use
actualVocabSize - which is for when a model has a larger
embedding tensor/# of output logits than actually trained token
to allow room for adding extras in finetuning - presently all of our
models have had "placeholder" tokens in the vocab so this hasn't broken
anything, but if the sizes did differ we want the equivalent of
`logits[actualVocabSize:]` (the start point is unchanged), not
`logits[-actualVocabSize:]` (this.)
2023-06-08 11:08:10 -07:00
AT
48275d0dcc
Dlopen backend 5 (#779)
Major change to the backend that allows for pluggable versions of llama.cpp/ggml. This was squashed merged from dlopen_backend_5 where the history is preserved.
2023-05-31 17:04:01 -04:00
Adam Treat
7f9f91ad94 Revert "New tokenizer implementation for MPT and GPT-J"
This reverts commit bbcee1ced5.
2023-05-30 12:59:00 -04:00
Aaron Miller
bbcee1ced5 New tokenizer implementation for MPT and GPT-J
Improves output quality by making these tokenizers more closely
match the behavior of the huggingface `tokenizers` based BPE
tokenizers these models were trained with.

Featuring:
 * Fixed unicode handling (via ICU)
 * Fixed BPE token merge handling
 * Complete added vocabulary handling
2023-05-30 12:05:57 -04:00
Aaron Miller
d14936bfd6 backend: dedupe tokenizing code in mpt/gptj 2023-05-16 10:30:19 -04:00
Aaron Miller
6182026c70 backend: dedupe tokenizing code in gptj/mpt 2023-05-16 10:30:19 -04:00
Adam Treat
d918b02c29 Move the llmodel C API to new top-level directory and version it. 2023-05-10 11:46:40 -04:00