pytorch/ao

View on GitHub

PyTorch native quantization for training and inference

Stars: 3k
Forks: 627
Open beginner issues: 13
Indexed issues: 301
Avg merge: 4d 14h
Merged PRs (30d): 34

Dominant language: Python
License: No license data
Last GitHub push: Sep 18, 2026
Latest indexed: Sep 19, 2026
Contributing guide: Contributing guide
Code of conduct: Code of conduct
Beginner labels: good first issue

Beginner friendly Everything

Open All

301 open issues indexed

[Bug] AdamW8bit picks a different state precision after sharding the same DTensor parameter Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 66/100

pytorch/ao#4912 · Opened Sep 18, 2026
pt2e observers silently discard the imaginary part of complex tensors Open

bug

Difficulty 3/5 1-2 days Newbie friendliness 74/100

pytorch/ao#4910 · 1 comment · Opened Sep 17, 2026
aarch64 regression tests (macos-14): ExecuTorch shared_kernels build fails — 'executorch/extension/threadpool/threadpool.h' file not found Open

ci-failure

Difficulty 4/5 3-5 days Newbie friendliness 52/100

pytorch/ao#4904 · 6 comments · Opened Sep 16, 2026
aarch64 CI: ExecuTorch shared_kernels build fails — 'executorch/extension/threadpool/threadpool.h' file not found (unpinned pip install executorch) Open

ci-failure

Difficulty 3/5 1-2 days Newbie friendliness 65/100

pytorch/ao#4903 · 1 comment · Opened Sep 16, 2026
Support input-channel MX quantization for nn.Conv2d weights Open

Difficulty 4/5 3-5 days Newbie friendliness 52/100

pytorch/ao#4866 · 1 comment · Opened Sep 4, 2026
Support nn.Conv2d in the observer-based Static FP8 quantization flow Open

Difficulty 4/5 3-5 days Newbie friendliness 63/100

pytorch/ao#4865 · 1 comment · Opened Sep 4, 2026
test_dtensor.py (float8 TP compile): aten.abs.default NotImplemented on Float8TrainingTensor after removing allow-in-graph autograd wrappers Open

ci-failure

Difficulty 4/5 3-5 days Newbie friendliness 45/100

pytorch/ao#4864 · 14 comments · Opened Sep 2, 2026
Avoid re-registering native Enum pytree constants Open

Difficulty 2/5 1-3 hours Newbie friendliness 78/100

pytorch/ao#4848 · Opened Aug 29, 2026
[MXFP8 training] Weight wrapper drops torch.nn.functional.linear bias Open

Difficulty 3/5 1-2 days Newbie friendliness 76/100

pytorch/ao#4846 · Opened Aug 29, 2026
Int8Tensor/Float8Tensor missing aten.view/abs/arithmetic op registrations in 0.18 (quantize fails silently, unquantized) Open

Difficulty 4/5 3-5 days Newbie friendliness 52/100

pytorch/ao#4845 · 5 comments · Opened Aug 28, 2026
Stale MXFP4 training entry in workflow documentation Open

Difficulty 1/5 Under an hour Newbie friendliness 92/100

pytorch/ao#4841 · Opened Aug 28, 2026
[WIP] MXTensor inference weights cannot be sharded with FSDP2 Open

Difficulty 4/5 3-5 days Newbie friendliness 48/100

pytorch/ao#4837 · Opened Aug 27, 2026
MXFP8 conversion unnecessarily requires PyTorch nightly for non-DTensor paths Open

Difficulty 3/5 1-2 days Newbie friendliness 78/100

pytorch/ao#4830 · Opened Aug 26, 2026
generate_permute_indices: the GPU path writes past the end of its max_len output buffer Open

Difficulty 4/5 3-5 days Newbie friendliness 68/100

pytorch/ao#4825 · Opened Aug 25, 2026
IndexError in _fqn_to_config_handler when multiple parameter FQNs are set to None Open

bug

Difficulty 2/5 1-3 hours Newbie friendliness 84/100

pytorch/ao#4818 · Opened Aug 24, 2026
Bitpacking default pack and unpack dimensions do not round-trip multidimensional tensors Open

Difficulty 2/5 1-3 hours Newbie friendliness 85/100

pytorch/ao#4815 · Opened Aug 23, 2026
Track deletion of low-usage TorchAO APIs Open

module: core topic: for developers tracker

Difficulty 4/5 3-5 days Newbie friendliness 20/100

pytorch/ao#4797 · Opened Aug 19, 2026
Float8Tensor zero inputs quantize to NaN Open

Difficulty 2/5 1-3 hours Newbie friendliness 68/100

pytorch/ao#4713 · Opened Aug 6, 2026
test_fp32_fp8_multi_module_parity (FSDP2 float8): NaN/0.0 loss parity mismatch on 4xH100 (torch 2.13.0) Open

ci-failure

Difficulty 4/5 3-5 days Newbie friendliness 35/100

pytorch/ao#4708 · 2 comments · Opened Aug 5, 2026
ConvRot — Group-wise Regular Hadamard Rotation for W4A4/W8A8 Diffusion Transformer Quantization Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 38/100

pytorch/ao#4695 · 5 comments · Opened Aug 5, 2026
ZeroPointDomain (class) vs zero_point_domain (variable) typo Open

bug

Difficulty 1/5 Under an hour Newbie friendliness 88/100

pytorch/ao#4693 · Opened Aug 5, 2026
convrot for int8 and fp4 Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 25/100

pytorch/ao#4685 · Opened Aug 3, 2026
ROCm Nightly (gfx950): fp8 hipSPARSELt sparse tests fail "hipSPARSELt not supported" plus gfx950 mxfp8/GPTQ numeric mismatches (6 failed) Open

ci-failure module: rocm

Difficulty 4/5 3-5 days Newbie friendliness 43/100

pytorch/ao#4663 · 49 comments · Opened Jul 30, 2026
Blockwise FP8 training GEMMs read past the end of both scale tensors Open

bug

Difficulty 2/5 1-3 hours Newbie friendliness 74/100

pytorch/ao#4634 · Opened Jul 29, 2026
`int_scaled_matmul` reads out of bounds and returns wrong results when K is even but not a multiple of `BLOCK_K Open

bug

Difficulty 2/5 1-3 hours Newbie friendliness 78/100

pytorch/ao#4633 · 1 comment · Opened Jul 29, 2026
RFC: 2:4 sparsity for Linear and Conv2d Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 38/100

pytorch/ao#4626 · Opened Jul 28, 2026
MXFP8 + compile + TP2 error: mat2 must be col_major Open

Difficulty 4/5 3-5 days Newbie friendliness 48/100

pytorch/ao#4624 · 1 comment · Opened Jul 27, 2026
[Bug] DISABLED backward path crashes with dtype mismatch under autocast Open

bug

Difficulty 2/5 1-3 hours Newbie friendliness 68/100

pytorch/ao#4616 · 1 comment · Opened Jul 27, 2026
Inverted bias flag in replace_conv2d_1x1 inside swap_conv2d_1x1_to_linear Open

bug

Difficulty 1/5 Under an hour Newbie friendliness 88/100

pytorch/ao#4615 · Opened Jul 27, 2026
`UIntxWeightOnlyConfig` with `bit_width=4` broken when model contains `bfloat16` weights Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 58/100

pytorch/ao#4614 · Opened Jul 26, 2026
[RFC] Decouple `int_scaled_matmul` / `int_matmul` op schemas from triton availability Open

enhancement

Difficulty 4/5 3-5 days Newbie friendliness 55/100

pytorch/ao#4602 · Opened Jul 24, 2026
[Bug] Hardcoded CUDA autocast queries cause silent fp32 fallback on non-CUDA backends Open

bug

Difficulty 3/5 1-2 days Newbie friendliness 74/100

pytorch/ao#4601 · 1 comment · Opened Jul 24, 2026
GPTQ hardcodes CUDA device — blocks all non-CUDA backends Open

bug

Difficulty 2/5 1-3 hours Newbie friendliness 78/100

pytorch/ao#4600 · Opened Jul 24, 2026
RFC：Extract FP8/INT8 inference compute kernels into `torch.library` ops for third-party backend extensibility Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 45/100

pytorch/ao#4599 · 3 comments · Opened Jul 24, 2026
float8nocompile `_to_fp8_row_major_t`: transposed-store bounds mask has its axes swapped Open

bug

Difficulty 3/5 1-2 days Newbie friendliness 76/100

pytorch/ao#4584 · Opened Jul 21, 2026
Rowwise fp8 (Float8DynamicActivationFloat8Weight, PerRow) fails with CUBLAS_STATUS_NOT_SUPPORTED for large-K + small-M GEMMs on H100 Open

Difficulty 4/5 3-5 days Newbie friendliness 58/100

pytorch/ao#4582 · Opened Jul 20, 2026
torchao 0.17.0+xpu: MPS module raises RuntimeError, escaping pkgutil's ImportError handler Open

bug

Difficulty 3/5 1-2 days Newbie friendliness 68/100

pytorch/ao#4577 · Opened Jul 16, 2026
Bounds-check `block_size` against `scale_and_zero` in the int4 tile-packed load path (avoids an OOB read in `_weight_int4pack_mm`) Open

Difficulty 2/5 1-3 hours Newbie friendliness 72/100

pytorch/ao#4572 · Opened Jul 14, 2026
X86InductorQuantizer + torch.export/aoti_compile_and_package never fuses quantized Linear into onednn::qlinear_pointwise (silent fp32 fallback) Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 45/100

pytorch/ao#4566 · 3 comments · Opened Jul 12, 2026
RFC: Weight Clustering for torchao Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 35/100

pytorch/ao#4557 · 1 reaction · Opened Jul 6, 2026
broken CI for float8 training due to upstream triton pin update Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 32/100

pytorch/ao#4556 · 1 comment · Opened Jul 6, 2026
`CantSplit` in inductor codegen for a TorchAO int8 matmul whose input is a `cat` over two dynamic sequence dims Open

Difficulty 4/5 3-5 days Newbie friendliness 52/100

pytorch/ao#4546 · 2 comments · 1 reaction · Opened Jul 1, 2026
Backport workflow-hardening fix (excessive-permissions) to 5 release branches Open

Difficulty 2/5 1-3 hours Newbie friendliness 68/100

pytorch/ao#4543 · Opened Jun 29, 2026
Module-wise type hinting and docstring coverage audit for PT2E Open

Difficulty 5/5 Over a week Newbie friendliness 42/100

pytorch/ao#4541 · 1 comment · Opened Jun 28, 2026
Add scale and zero-point dimension validation to low-level quantization primitives Open

Difficulty 4/5 3-5 days Newbie friendliness 68/100

pytorch/ao#4538 · 1 comment · Opened Jun 27, 2026
Question: Lowering a custom TorchAO QAT model to XNNPACK backend Open

pytorch/ao#4529 · 1 comment · 1 assignee · Opened Jun 23, 2026
Linear+BatchNorm1d not folded during QAT, unlike Conv+BN Open

enhancement

Difficulty 3/5 1-2 days Newbie friendliness 74/100

pytorch/ao#4463 · 3 comments · Opened Jun 4, 2026
convert_pt2e crashes in DuplicateDQPass — GraphModule rebuild drops submodules only referenced by other submodules' internal graphs Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 52/100

pytorch/ao#4455 · 2 comments · Opened May 29, 2026
`Float8DynamicActivationFloat8WeightConfig` + `KernelPreference.MSLK` + `torch.compile` fails on torchao 0.17.0 (missing `import mslk.gemm` in `float8_tensor.py`) Open

Difficulty 2/5 1-3 hours Newbie friendliness 45/100

pytorch/ao#4438 · 1 comment · Opened May 25, 2026
constant_fold deletes shared get_attr target while another get_attr for the same target is still live Open

bug module: pt2e_quant

Difficulty 3/5 1-2 days Newbie friendliness 70/100

pytorch/ao#4420 · Opened May 18, 2026

pytorch/ao

Get new issues in your inbox