Commits

Breadcrumbs

History for

onmaster

User selector

All users

Datepicker

All time

Commit History

Commits on Jul 21, 2023

• Experimental flash-attn-2 support. Probably don't use yet.

  turboderp

  committed
  ce687cd

  Copy full SHA for ce687cd

Commits on Jul 19, 2023

• Support for GQA and Llama2-70b

  turboderp

  committed
  b3aea52

  Copy full SHA for b3aea52

Commits on Jun 24, 2023

• Add batched encode/decode to tokenizer

  turboderp

  committed
  cdb6f54

  Copy full SHA for cdb6f54

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

  turboderp

  committed
  696c2bb

  Copy full SHA for 696c2bb

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

Commits on Jun 22, 2023

• Confirm LoRA bias is zero before ignoring

  turboderp

  committed
  4ba6527

  Copy full SHA for 4ba6527

• Refactor perplexity test, add options

  turboderp

  committed
  2747e00

  Copy full SHA for 2747e00

• Allow LoRA to ignore (empty) bias

  turboderp

  committed
  58e9e69

  Copy full SHA for 58e9e69

Commits on Jun 18, 2023

• Add LoRA support

  turboderp

  committed
  3160ae2

  Copy full SHA for 3160ae2

Commits on Jun 17, 2023

• Fix -v mode

  turboderp

  committed
  e6f2beb

  Copy full SHA for e6f2beb

• Allow forcing single-token code paths in perplexity tests, for debugging fused kernels

  turboderp

  committed
  127e603

  Copy full SHA for 127e603

Commits on Jun 16, 2023

• More versatile perplexity tests

  Show description for 998fd46
  turboderp

  committed
  998fd46

  Copy full SHA for 998fd46

Commits on Jun 12, 2023

• Refactor of perplexity into its own module for easier extensibility. Changed command options -ppl [method] specifies how we calculate perplexity, and -ppl-ds lets us load either JSONL or raw text d…

  Show description for ecb414a
  lhl

  authored
  ecb414a

  Copy full SHA for ecb414a

Commits on Jun 11, 2023

• Benchmark >2048 token sequence prompts in batches

  turboderp

  committed
  896da5d

  Copy full SHA for 896da5d

Commits on Jun 10, 2023

• Removed debug mode, forced no_grad() in forward pass

  turboderp

  committed
  fecebe2

  Copy full SHA for fecebe2

Commits on Jun 9, 2023

• (Partial) fused attn

  turboderp

  committed
  91c9c8c

  Copy full SHA for 91c9c8c

Commits on Jun 5, 2023

• Add no-half2 option to matmul and MLP

  turboderp

  committed
  cddb6ee

  Copy full SHA for cddb6ee

Commits on Jun 4, 2023

• Refactoring

  turboderp

  committed
  ba8666c

  Copy full SHA for ba8666c

• Refactoring

  turboderp

  committed
  1c25134

  Copy full SHA for 1c25134

Commits on Jun 2, 2023

• Remove layer streaming

  turboderp

  committed
  0cbf29b

  Copy full SHA for 0cbf29b

Commits on May 29, 2023

• Warmup pass before benchmark

  turboderp

  committed
  a88ded2

  Copy full SHA for a88ded2

Commits on May 27, 2023

• Maybe fix for the disappearing hidden states?

  turboderp

  committed
  ca4436b

  Copy full SHA for ca4436b

• Maybe fix for the disappearing hidden states?

  turboderp

  committed
  b11d80f

  Copy full SHA for b11d80f

Commits on May 26, 2023

• Add debug mode

  turboderp

  committed
  6d7d128

  Copy full SHA for 6d7d128

• Test

  turboderp

  committed
  37d7c40

  Copy full SHA for 37d7c40

• Some tidying up, a few fixes and improvements

  turboderp

  committed
  3dc7cd9

  Copy full SHA for 3dc7cd9

Commits on May 25, 2023

• CUDA RoPE implementation, other little tweaks, best-case benchmark results

  turboderp

  committed
  4dbd304

  Copy full SHA for 4dbd304

Commits on May 24, 2023

• Tested some models

  turboderp

  committed
  4a0b461

  Copy full SHA for 4a0b461

Commits on May 23, 2023

• Finish dequant stuff

  Show description for a0b6668
  turboderp

  committed
  a0b6668

  Copy full SHA for a0b6668

Commits on May 22, 2023

• Option to dequantize at load-time

  turboderp

  committed
  64f32bc

  Copy full SHA for 64f32bc

• Added option to auto-split model based on per-GPU allocation of VRAM

  Show description for b585395
  turboderp

  committed
  b585395

  Copy full SHA for b585395

Commits on May 19, 2023

• Testing out layer streaming

  Show description for 11fbd4f
  turboderp

  committed
  11fbd4f

  Copy full SHA for 11fbd4f

Commits on May 18, 2023

• Started work on fused layers

  Show description for bc4c7e7
  turboderp

  committed
  bc4c7e7

  Copy full SHA for bc4c7e7

• Cleaned up CUDA code

  Show description for 8d9b755
  turboderp

  committed
  8d9b755

  Copy full SHA for 8d9b755

Pagination

Previous

Next