24 Pull requests merged by 14 people

• Lint fixes torchao/profiler and torchao/testing

  #1368 merged Dec 2, 2024

• Fix find_multiple import in GPTQ.py

  #1367 merged Dec 2, 2024

• Lint fixes torchao files

  #1366 merged Dec 2, 2024

• bump main version to 0.8

  #1364 merged Dec 2, 2024

• check scale.ndim before applying t/transpose

  #1339 merged Dec 2, 2024

• add option to use SAC in float8 training profiling script

  #1354 merged Dec 2, 2024

• Lint fixes test sparsity

  #1360 merged Dec 2, 2024

• Lint fixes test/quantization

  #1359 merged Dec 2, 2024

• Update hardware check conditions

  #1356 merged Dec 1, 2024

• Update README.md: Fix bibtex and sglang links

  #1361 merged Nov 30, 2024

• Benchmark intel xpu

  #1259 merged Nov 29, 2024

• Add support for quantize_() with Float8Linear module

  #1344 merged Nov 28, 2024

• Reduce startup time for SAM2 AMG by using torch.export

  #1358 merged Nov 27, 2024

• Add Int4CPULayout and update int4 woq

  #1278 merged Nov 27, 2024

• Add floating point options for autoquant and add accuracy measurement

  #1355 merged Nov 27, 2024

• Benchamarking

  #1353 merged Nov 27, 2024

• Enable 8-bit

  #1254 merged Nov 26, 2024

• Reduce SAM2 AMG cli startup by using deploy

  #1350 merged Nov 26, 2024

• [NF4] .to() fixes

  #1312 merged Nov 26, 2024

• [low-bit optim] Fix edge cases for FSDP2 integration

  #1269 merged Nov 26, 2024

• SAM2 AMG cli.py on modal

  #1349 merged Nov 26, 2024

• Enable CPU Offload for Intel GPU

  #1324 merged Nov 26, 2024

• Fixed invalid url in citation section

  #1348 merged Nov 26, 2024

• Remove lm_eval warning

  #1347 merged Nov 26, 2024

5 Pull requests opened by 4 people

• [wip] float8 training: invert the meaning of scale

  #1351 opened Nov 26, 2024

• zero dim support for tensorwise float8 training

  #1352 opened Nov 26, 2024

• testing build linux wheels

  #1357 opened Nov 27, 2024

• Fix bfloat16/float16/float32 options

  #1369 opened Dec 2, 2024

• Move profiler -> prototype

  #1370 opened Dec 3, 2024

2 Issues closed by 2 people

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

  #1310 closed Nov 26, 2024

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

  #1338 closed Nov 26, 2024

2 Issues opened by 2 people

• GPTQ.py import path bug?: find_multiple

  #1365 opened Dec 2, 2024

• Can't use AdamW?

  #1363 opened Dec 2, 2024

18 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Update SAM AMG README with more descriptive install instructions

  #1337 commented on Nov 28, 2024 • 7 new comments

• [float8] Allow specifying arbitrary dtype for each tensor

  #1326 commented on Nov 26, 2024 • 5 new comments

• [float8] Re-enable slow-accum in the bwd of axis-wise scaling schemes

  #1325 commented on Nov 26, 2024 • 1 new comment

• Test meta CDN

  #1081 commented on Nov 29, 2024 • 1 new comment

• Self-Compression QAT and Linear

  #1342 commented on Nov 26, 2024 • 0 new comments

• metal lowbit kernels: executorch ops

  #1322 commented on Nov 27, 2024 • 0 new comments

• [WIP] Codebook quantization flow

  #1299 commented on Dec 2, 2024 • 0 new comments

• Fixes observer attachment to model based on config for wanda sparsifier

  #1265 commented on Nov 27, 2024 • 0 new comments

• [low-bit optim] Add coat for float8 optimizer

  #1231 commented on Nov 26, 2024 • 0 new comments

• Add TTFT benchmarks + update sparsity benchmarks

  #1140 commented on Dec 2, 2024 • 0 new comments

• gemlite integration in torchao

  #1034 commented on Nov 26, 2024 • 0 new comments

• Enable ROCM in CI

  #999 commented on Dec 3, 2024 • 0 new comments

• W4A8 based on CUTLASS

  #880 commented on Nov 29, 2024 • 0 new comments

• [Tracker] autoquant tracker

  #1215 commented on Dec 2, 2024 • 0 new comments

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

  #1308 commented on Dec 2, 2024 • 0 new comments

• Sporadic Bad Alloc Failures on CI

  #1229 commented on Dec 2, 2024 • 0 new comments

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

  #1320 commented on Nov 27, 2024 • 0 new comments

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

  #1309 commented on Nov 26, 2024 • 0 new comments