pytorch/ao

Issues

• Docs Revamp

  #181 opened a month ago by msaroufim
  0

• Saving autoquant quantization plan

  #320 opened 2 days ago by RobinKa
  1

• ARM builds in CI

  #335 opened 2 days ago by msaroufim
  0

• Improvement ideas for `hf_eval.py`

  #332 opened 2 days ago by gau-nernst
  0

• FloatQuantization subclass

  #228 opened a month ago by msaroufim
  3

• FP6 dtype!

  #208 opened a month ago by NicolasMejiaPetit
  16

• [feature request] np.packbits / np.unpackbits, general BitTensors (maybe can be just tensors with dtype torch.bits8 or have a new dtype torch.bits introduced) and bit packed tensors utilities for saving memory / accesses, support for BitTensors wherever BoolTensors are used

  #292 opened 4 years ago by vadimkantorov
  82

• [Tracker] WIP features for torchao 0.3

  #252 opened 22 days ago by supriyar
  3

• Error when using to_nf4 function, inside NF4Tensor Class

  #268 opened 16 days ago by FabioDataGeek
  2

• Bitnet 1.58 prework, POC, and staging

  #281 opened 13 days ago by CoffeeVampir3
  2

• `dequantize_affine` modified the `input` in-place

  #289 opened 5 days ago by yiliu30
  7

• apply_dynamic_quant for vit_b_16

  #63 opened 5 days ago by cpuhrsch
  2

• torch.iinfo() support for sub byte dtypes

  #308 opened 5 days ago by vayuda
  0

• custom cuda extensions make installing ao hard

  #288 opened 11 days ago by msaroufim
  4

• Numerics checks between NF4 and bnb nf4

  #295 opened 10 days ago by msaroufim
  0

• Generic packing algorithms from size N to M

  #284 opened 13 days ago by vayuda
  4

• torchao.utils.benchmark_model support cpu and mps benchmarking

  #287 opened 11 days ago by jerryzh168
  0

• [NF4][FSDP2] avoid peaking GPU memory when constructing NF4 tensors

  #204 opened a month ago by weifengpy
  1

• torchao init: ImportError: libcudart.so.12: cannot open shared object file: No such file or directory

  #260 opened 19 days ago by mikekgfb
  1

• HQQ Tracker

  #255 opened 22 days ago by HDCharles
  1

• [Tracker] WIP Features for torchao v0.2

  #132 opened 22 days ago by supriyar
  1

• [Question] MBU in automated CI?

  #237 opened a month ago by cadedaniel
  2

• Building torchao from source installs unnecessary torch and nvidia packages every time

  #231 opened a month ago by gau-nernst
  1

• Doc build failing on main

  #221 opened a month ago by msaroufim
  0

• [BUG] No module named 'expecttest' when import `torchao`

  #222 opened a month ago by gau-nernst
  0

• [RFC] Plans for sparsity

  #143 opened 2 months ago by jcaip
  2

• 2:4 sparsity + PTQ(int8) model's inference

  #134 opened 2 months ago by RanchiZhao
  7

• [RFC] More general affine quantization primitives

  #160 opened 2 months ago by jerryzh168
  1

• Run semi-structured spare benchmarks on consumer hardware

  #174 opened a month ago by jcaip
  12

• [NF4][FSDP2]: enable multi-gpu CI

  #202 opened a month ago by weifengpy
  0

• NF4Tensor uses 8 bits of memory

  #209 opened a month ago by cuichenx
  7

• Sparsity OSS colab tracker

  #113 opened 2 months ago by jcaip
  0

• [RFC] Plans for torchao

  #47 opened 3 months ago by supriyar
  21

• [New Feature] CUTLASS kernels for w4a8 quantization

  #64 opened 3 months ago by supriyar
  4

• [RFC] Plans for LLM QAT

  #86 opened 2 months ago by andrewor14
  0

• [Tracker] General feature requests for torchao

  #65 opened 3 months ago by supriyar
  1

• 1 bit LLM implementation

  #67 opened 3 months ago by msaroufim
  0

• [NF4][FSDP2] DTensor + fused adam on cpu

  #205 opened a month ago by weifengpy
  0

• Semi-Structured Sparsity unsupported for Windows

  #191 opened a month ago by philipbutler
  1

• Custom CUDA extensions

  #137 opened a month ago by msaroufim
  0

• [Tracker] Outstanding Issues and WIP Features for version 0.1

  #20 opened 2 months ago by HDCharles
  2

• how do you implement int8 mm on cuda?

  #34 opened 4 months ago by ThisisBillhe
  1

• Project implicitly depends on torch nightly

  #29 opened 5 months ago by ELanning
  4

• Handle NonDynamicallyQuantizableLinear in smoothquant module

  #30 opened 5 months ago by ELanning
  2

• Reconsider using class method for Int8DynamicallyQuantizedLinearWeight.from_float

  #21 opened 6 months ago by HDCharles
  1

• SDPA: scaled_dot_product_attention to use this attetino donot need to retain ?

  #14 opened 6 months ago by ganliqiang
  3