Commits

Breadcrumbs

History for

onmaster

User selector

All users

Datepicker

All time

Commit History

Commits on Jun 21, 2024

• Move conversion script into exllamav2 package

  turboderp

  committed
  a0ea2b0

  Copy full SHA for a0ea2b0

Commits on Jun 8, 2024

• Option to resume conversion job with no other args

  turboderp

  committed
  6030517

  Copy full SHA for 6030517

Commits on May 22, 2024

• add layer GPU offloading for hidden/target states

  kallewoof

  committed
  0ece2f3

  Copy full SHA for 0ece2f3

Commits on May 19, 2024

• Add machine-parseable output to convert script

  turboderp

  committed
  83baa98

  Copy full SHA for 83baa98

Commits on May 9, 2024

• Allow quantizing models with max_seq_len < 2048

  turboderp

  committed
  a847f48

  Copy full SHA for a847f48

Commits on Apr 18, 2024

• Fix scale rounding during quant

  turboderp

  committed
  893e73c

  Copy full SHA for 893e73c

Commits on Apr 5, 2024

• Convert: Limit max_output_len to save VRAM while measuring

  turboderp

  committed
  f52612c

  Copy full SHA for f52612c

Commits on Jan 31, 2024

• Adding graceful exit signal handling and status box for estimating time remaining in quantization process

  bgorlick

  committed
  56a0d6d

  Copy full SHA for 56a0d6d

Commits on Dec 28, 2023

• Fix rope_scale display in convert.py

  turboderp

  committed
  970af13

  Copy full SHA for 970af13

• Read RoPE linear scale from model config

  turboderp

  committed
  47df040

  Copy full SHA for 47df040

Commits on Dec 14, 2023

• Oopsie

  turboderp

  committed
  5044e38

  Copy full SHA for 5044e38

Commits on Dec 13, 2023

• Rework quantization and optimization

  turboderp

  committed
  0d63d64

  Copy full SHA for 0d63d64

Commits on Dec 10, 2023

• Fix quant resume

  turboderp

  committed
  1f36c4a

  Copy full SHA for 1f36c4a

• Revert some changes, calibrate to q state again (fixes 70B low bitrate)

  turboderp

  committed
  3c43bad

  Copy full SHA for 3c43bad

• Faster and more stable quant

  turboderp

  committed
  a89d85a

  Copy full SHA for a89d85a

Commits on Dec 8, 2023

• New quant optimization procedure

  turboderp

  committed
  2e91239

  Copy full SHA for 2e91239

Commits on Nov 30, 2023

• Tree: Force utf8 when opening files

  Show description for 6bfcefe
  kingbri1

  committed
  6bfcefe

  Copy full SHA for 6bfcefe

Commits on Nov 21, 2023

• Add RoPE arguments to quantizer script

  turboderp

  committed
  09b981f

  Copy full SHA for 09b981f

Commits on Oct 5, 2023

• Add QKV embeddings

  turboderp

  committed
  d4e9020

  Copy full SHA for d4e9020

Commits on Sep 21, 2023

• Bump to 0.0.3

  turboderp

  committed
  21c44b7

  Copy full SHA for 21c44b7

Commits on Sep 20, 2023

• Make sure temp buffers are allocated for length of calibration data

  turboderp

  committed
  09227a7

  Copy full SHA for 09227a7

Commits on Sep 18, 2023

• More options for converter to facilitate scripting

  turboderp

  committed
  6fd006b

  Copy full SHA for 6fd006b

Commits on Sep 17, 2023

• Fix arg type for shard_size

  turboderp

  committed
  fae6fb2

  Copy full SHA for fae6fb2

Commits on Sep 16, 2023

• More memory tweaks, made swapping states to CPU the default to accommodate quanting 70B on 24GB GPUs

  turboderp

  committed
  5c247f9

  Copy full SHA for 5c247f9

• Add more quant options

  turboderp

  committed
  2f72437

  Copy full SHA for 2f72437

• Conversion: ability to save sharded models (addresses OoM when compiling output file)

  turboderp

  committed
  af1398f

  Copy full SHA for af1398f

Commits on Sep 9, 2023

• Update defaults for convert script

  turboderp

  committed
  952c67c

  Copy full SHA for 952c67c

Commits on Aug 30, 2023

• Initial commit

  turboderp

  committed
  bb83469

  Copy full SHA for bb83469

Loading