Commits

Branch selector

 master

User selector

All users

Datepicker

All time

Commit History

Commits on Jul 9, 2023

• Add CUDA device_guard to 2nd part of fused attn

  turboderp

  committed
  e61d4d3

  Copy full SHA for e61d4d3

• Fix potential Safetensors memory leak and limit system RAM usage while loading model

  turboderp

  committed
  62fa472

  Copy full SHA for 62fa472

Commits on Jul 8, 2023

• Fix potential Safetensors memory leak and limit system RAM usage while loading model

  turboderp

  committed
  617fd2a

  Copy full SHA for 617fd2a

Commits on Jul 5, 2023

• Fix install steps order in README (turboderp#133)

  VldmrB

  authored
  d769533

  Copy full SHA for d769533

Commits on Jul 3, 2023

• Fixed: batch memory corruption due to overrunning temp_state buffer. (turboderp#128)

  Show description for 63ebf2e
  aljungberg

  authored
  63ebf2e

  Copy full SHA for 63ebf2e

Commits on Jul 2, 2023

• Fix logit concatenation when chunking

  turboderp

  committed
  693ef40

  Copy full SHA for 693ef40

• More models validated, benchmark for 3b.

  turboderp

  committed
  6bf0046

  Copy full SHA for 6bf0046

• One more fix for odd tensor shapes in OpenLlama-3b

  turboderp

  committed
  4dab525

  Copy full SHA for 4dab525

Commits on Jul 1, 2023

• Make CUDA kernels tolerate head_dim other than 128

  turboderp

  committed
  b56a9c2

  Copy full SHA for b56a9c2

• Merge remote-tracking branch 'origin/master'

  turboderp

  committed
  36b49f1

  Copy full SHA for 36b49f1

• Fix broken directive when compiling with HIP (turboderp#124)

  Show description for b5e7cf7
  ardfork

  authored
  b5e7cf7

  Copy full SHA for b5e7cf7

• Fix for pascal generation (turboderp#123)

  Show description for 088e9d3
  Ph0rk0z

  authored
  088e9d3

  Copy full SHA for 088e9d3

• (Experimental) Add support to NTK RoPE scaling (turboderp#118)

  Show description for 8229d55
  

  Panchovix

  authored
  8229d55

  Copy full SHA for 8229d55

Commits on Jun 28, 2023

• Forced sync to accurately measure prompt speed

  turboderp

  committed
  f4848e6

  Copy full SHA for f4848e6

• Moved input ID chunking to model, added logic to limit size of attention for very long contexts

  turboderp

  committed
  93d50d1

  Copy full SHA for 93d50d1

Commits on Jun 24, 2023

• Add batch support to generate_simple(), also example

  turboderp

  committed
  a01b25c

  Copy full SHA for a01b25c

• Add batched encode/decode to tokenizer

  turboderp

  committed
  cdb6f54

  Copy full SHA for cdb6f54

• Merge remote-tracking branch 'origin/master'

  Show description for b854e64
  turboderp

  committed
  b854e64

  Copy full SHA for b854e64

• Add optional input mask to forward pass, masking for batch test

  turboderp

  committed
  696c2bb

  Copy full SHA for 696c2bb

• Fixed: batching lead to faulty results, crashes and men wielding bananas. (turboderp#101)

  Show description for 4f1684b
  aljungberg

  authored
  4f1684b

  Copy full SHA for 4f1684b

Commits on Jun 23, 2023

• Minor clean-up.

  aljungberg

  committed
  20feb55

  Copy full SHA for 20feb55

• Fixed: batching lead to faulty results and crashes.

  Show description for f24617b
  aljungberg

  committed
  f24617b

  Copy full SHA for f24617b

• Add LoRA option to CLI chatbot example

  turboderp

  committed
  248f59f

  Copy full SHA for 248f59f

• Limit attention size during generation

  turboderp

  committed
  7f31d5e

  Copy full SHA for 7f31d5e

Commits on Jun 22, 2023

• Web UI: forward max_seq_len to client

  turboderp

  committed
  9a686ca

  Copy full SHA for 9a686ca

• Merge remote-tracking branch 'origin/master'

  turboderp

  committed
  32c028f

  Copy full SHA for 32c028f

• Rework repetition penalty

  turboderp

  committed
  47ee2fd

  Copy full SHA for 47ee2fd

• Fix compiling in venv on Windows (turboderp#89)

  Show description for 4ee693d
  EyeDeck

  authored
  4ee693d

  Copy full SHA for 4ee693d

• Confirm LoRA bias is zero before ignoring

  turboderp

  committed
  4ba6527

  Copy full SHA for 4ba6527

• Merge remote-tracking branch 'origin/master'

  turboderp

  committed
  91c7e55

  Copy full SHA for 91c7e55

• Windows doesn't use utf-8 encoding by default. (turboderp#97)

  allenbenz

  authored
  21e5630

  Copy full SHA for 21e5630

• Fix AttributeError: 'torch.device' object has no attribute 'startswith' (turboderp#96)

  

  Panchovix

  authored
  fc96ae8

  Copy full SHA for fc96ae8

• Consistent naming for command line args

  turboderp

  committed
  7f11b1e

  Copy full SHA for 7f11b1e

• Add option to compress positional embeddings

  turboderp

  committed
  d05f5d0

  Copy full SHA for d05f5d0

• Fix perplexity test for chunk_size > 2048

  turboderp

  committed
  2c7bca4

  Copy full SHA for 2c7bca4

Pagination

Previous

Next