AlpinDale
|
313e6e1ec7
feat: add typical acceptance sampling
|
преди 7 месеца |
AlpinDale
|
4d1e613804
chore: minor simplifications
|
преди 7 месеца |
AlpinDale
|
e0886ee929
feat: add `ProposerWorkerBase` abstract class
|
преди 7 месеца |
AlpinDale
|
344ddaac5a
properly disable speculative decoding
|
преди 7 месеца |
AlpinDale
|
5b0c11d190
support pipeline parallel pynccl groups
|
преди 7 месеца |
AlpinDale
|
de62ceb18c
refactor: eliminate parallel worker per-step task scheduling overhead
|
преди 7 месеца |
AlpinDale
|
236be273e5
feat: tensor parallel speculative decoding (#554)
|
преди 7 месеца |
AlpinDale
|
197a6d2c16
auto disable speculative decoding by the running queue size
|
преди 7 месеца |
AlpinDale
|
438f5bdce9
fix ngrams
|
преди 7 месеца |
AlpinDale
|
ef733aee43
implement ExecuteModelData to reduce executor complexity
|
преди 8 месеца |
AlpinDale
|
79901b76de
logprobs for target model (spec decoding)
|
преди 8 месеца |
AlpinDale
|
723c6acb84
re-add ngram speculative decoding
|
преди 8 месеца |
AlpinDale
|
7bcf4c3fc9
centralize gpu worker construction
|
преди 8 месеца |
AlpinDale
|
9d81716bfd
[v0.5.3] Release Candidate (#388)
|
преди 10 месеца |
AlpinDale
|
f8dfac6372
chore: attention refactor and upstream sync apr01 (#365)
|
преди 11 месеца |