Hugging Face's logo

Spaces:
Ruurd
/
llada-8b-instruct
Running on Zero

App Files Community
llada-8b-instruct / src

Commit History

Yesterday

  1. Add optional temperature-scaled Gumbel randomized Viterbi, retaining fast exact MAP decoding

    Ruurd
    Verified signature
    67a9c62

  2. Run small exact trigram MAP chains on CPU to reduce GPU launch overhead

    Ruurd
    Verified signature
    aee543d

  3. Speed up exact trigram scoring and avoid GPU scalar traceback synchronization

    Ruurd
    Verified signature
    945f648

  4. Add optional expanded-corpus bigram and exact trigram guidance

    Ruurd
    Verified signature
    b31a17a

  5. Add optional expanded-corpus bigram and exact trigram guidance

    Ruurd
    Verified signature
    40f551e

  6. Guarantee a minimum token commitment quota each APD step

    Ruurd
    Verified signature
    942fb38

  7. Add adjustable APD commitment aggressiveness; preserve default behavior

    Ruurd
    Verified signature
    24666c5

  8. Fade repetition penalties over early block iterations; retain constant-strength option

    Ruurd
    Verified signature
    4742018

  9. Add optional joint bigram repetition-risk penalty

    Ruurd
    Verified signature
    82e83e1

  10. Decouple PMI weight from sequence sampling temperature

    Ruurd
    Verified signature
    35914ff

  11. Add optional bigram-guided sequence sampling and candidate-occurrence repetition counts

    Ruurd
    Verified signature
    a9be260

2 days ago

  1. Add optional adaptive parallel decoding; preserve APD-off behavior

    Ruurd
    Verified signature
    1473488

  2. Add optional PMI Viterbi transition repetition penalty; preserve existing defaults

    Ruurd
    Verified signature
    6e41363

3 days ago

  1. Deploy native LLaDA-Instruct with shared diffusion and PMI Viterbi controls

    Ruurd
    Verified signature
    153c673