pytorch/ao

Issues

• [Feature Request] Support of `int8_dynamic_activation_int8_weight` with asymmetrically quantized weights

  #1320 opened a month ago by sanchitintel
  3

• [NF4] Various bugs in how NF4 handles `.to()` to move to a different device

  #1310 opened a month ago by gau-nernst
  0

• [AQT] Failed to move compiled module with AQT to a different device

  #1309 opened a month ago by gau-nernst
  2

• [CI] CUDA nightly regression test is failing due to bnb + `triton.ops`

  #1338 opened a month ago by gau-nernst
  0

• [FLOAT8] Add Hardware Compatibility Check for FP8 Quantization

  #1188 opened a month ago by drisspg
  1

• `attempting to run aten.abs.default, this is not supported` with latest torchtitan + torchao

  #1313 opened a month ago by lchu-ibm
  3

• Very large discrepancy in the quantized model's output compared to the original model when quantizing on CPU

  #1335 opened a month ago by JohnnyRacer
  4

• [Tracker] autoquant v2 tracker

  #1215 opened 2 months ago by jerryzh168
  0

• Use unittest instead of pytest everywhere

  #1321 opened a month ago by jerryzh168
  0

• `int8_dynamic_activation_int8_weight` uses zero-points for weight when activation is asymmetrically quantized

  #1317 opened a month ago by sanchitintel
  2

• pip install torchao cannot get latest versions (only 0.1 and 2 other version in the same level)

  #1300 opened a month ago by moreAImore
  8

• fp8dq requires both dimensions to be divisible by 16

  #1268 opened a month ago by piotr-bazan-nv
  7

• How to trigger torchao unit tests?

  #1315 opened a month ago by goldhuang
  3

• Why my QAT's convert doesn't work,still float32

  #1290 opened a month ago by Xxxgrey
  5

• [float8] DDP GPT1.5B Torch.compile dynamo error

  #1308 opened a month ago by OrenLeung
  1

• torch.compile(sync_float8_amax_and_scale_history) not working with triton latest main

  #1311 opened a month ago by goldhuang
  2

• [float8] FP8 GPT 1.5B Delayed Scaling 2x Slower than BF16

  #1297 opened a month ago by OrenLeung
  5

• [Sparsity] When sparsifying using Wanda on only Linear layers, PerChannelNormObserver() being added to embedding layers, leading to RuntimeError: linalg.vector_norm: Expected a floating point or complex tensor as input. Got Long

  #1133 opened 2 months ago by agrawal-aka
  6

• CPUoffloadOptimizer issues

  #1209 opened 2 months ago by felipemello1
  4

• Can you export a quantized model to Torchscript?

  #1293 opened a month ago by will-rice
  2

• SEO not helping ao

  #1298 opened a month ago by msaroufim
  0

• Add codebook (look up table based) quantization flow in torchao

  #1195 opened 2 months ago by jerryzh168
  2

• ZeroPointDomain as an arguments

  #1264 opened a month ago by airMeng
  7

• AMD integration tracker

  #1260 opened a month ago by jcaip
  1

• Int4wo quantization raises exception for FP32/FP16 i.e. works only for BF16

  #1267 opened a month ago by piotr-bazan-nv
  3

• Sporadic Bad Alloc Failures on CI

  #1229 opened a month ago by drisspg
  2

• How to skip decomposition of dequantize_affine and quantize_affine custom ops in inductor?

  #1230 opened a month ago by Nullkooland
  3

• Low bit shampoo

  #1257 opened a month ago by msaroufim
  0

• Cannot run FSDP2 with low bit optim from AO

  #1189 opened a month ago by nighting0le01
  16

• aarch64 jobs failing because conda is missing

  #1246 opened a month ago by msaroufim
  1

• Unskip `test_jit_trace` in `test/prototype/test_parametrization.py`

  #1220 opened a month ago by jerryzh168
  0

• torchao.float8 not working on PyTorch 2.4.1 and how does torchao handle FP8 autocast?

  #1159 opened 2 months ago by zigzagcai
  15

• [QAT] Low-bit FSDP all-gather for QAT

  #1224 opened 2 months ago by gau-nernst
  1

• [RFC] Sparsity Future Plans

  #1136 opened 2 months ago by jcaip
  1

• [low-bit optim] Add COAT optimizer

  #1190 opened 2 months ago by gau-nernst
  5

• Low Bit Optim Instability.

  #1218 opened 2 months ago by nighting0le01
  4

• [feat request] PagedAdamw8bit for FSDP2

  #1213 opened 2 months ago by felipemello1
  4

• [New method] VPTQ Vector Post-Training Quantization Support

  #1204 opened 2 months ago by YangWang92
  2

• Why swapping nn.Linear with FP8Linear brings throughput decrease for 7B LLAMA2-like model?

  #1199 opened 2 months ago by zigzagcai
  3

• Does Float8Linear support Tensor Parallelism and Sequence Parallelism?

  #1198 opened 2 months ago by zigzagcai
  1

• torchao.float8 + torch.compile does not work on HuggingFace's Mixtral model

  #1200 opened 2 months ago by vkuzo
  7

• Unable to save checkpoints when Use low bit optimizers with FSDP1 or FSDP2

  #1185 opened 2 months ago by nighting0le01
  3

• # Add API for allowing quant_api's to register their own filter_fns

  #1193 opened 2 months ago by drisspg
  0

• FP6 Speed on A100 80g

  #1181 opened 2 months ago by shihaobai
  6

• [sparse] Add GPTQ support for sparse-marlin

  #1134 opened 2 months ago by jcaip
  1

• [RFC] Follow Up for torchao developer experience discussion

  #1184 opened 2 months ago by jerryzh168
  0

• 2 sparse tests failing on main nightly job

  #1158 opened 2 months ago by msaroufim
  1

• cannot import name 'quantize_' from 'torchao.quantization' (C:\ai\Meissonic\venv\lib\site-packages\torchao\quantization\__init__.py)

  #1146 opened 2 months ago by nitinmukesh
  3

• Nightly wheels don't have correct version string for Windows and MacOs

  #1149 opened 2 months ago by gau-nernst
  0

• Fail to reproduce benchmark results

  #1135 opened 2 months ago by ThisisBillhe
  3