Issues

Issues

• test_codebook_quant.py collection fails: module 'torch.jit' has no attribute 'script' on torch 2.16 nightly (ROCm), aborts whole pytest run

  pytorch/ao#5098 · opened on Nov 11, 2026
  2

• Int4WeightOnlyConfig silently skips quantization when in_features is not divisible by group_size

  pytorch/ao#5096 · opened on Nov 11, 2026

• CPU Nightly CI fails: torch.jit.script was removed in torch nightly

  pytorch/ao#5093 · opened on Nov 10, 2026

• Importing torchao.prototype.mx_formats initializes CUDA (device capability checked at module level)

  pytorch/ao#5086 · opened on Nov 9, 2026
  3

• CPU kernel C++ tests and benchmarks don't compile with TORCHAO_ENABLE_ARM_I8MM=ON

  pytorch/ao#5083 · opened on Nov 9, 2026

• CPU kernel C++ tests no longer build on aarch64 (fallback test_linear link error, GCC errors on Linux)

  pytorch/ao#5082 · opened on Nov 9, 2026

• BlockwiseQuantLinear.forward always fails: self.dtype is never set from the dtype argument

  pytorch/ao#5070 · opened on Nov 6, 2026

• [PT2E] Fix dynamic FP16 Linear fusion for mm and bmm

  pytorch/ao#5060 · opened on Nov 5, 2026

• [MXFP8] Native token padding corrupts float32 inputs

  pytorch/ao#5055 · opened on Nov 4, 2026

• [NVFP4] Padded grouped GEMM fails with the native CUDA extension

  pytorch/ao#5054 · opened on Nov 4, 2026

• [PT2E] prepare_pt2e folds every BatchNorm into a conv whose output has multiple users → wrong FP32 graph

  pytorch/ao#5048 · opened on Nov 2, 2026
  1

• prepare_pt2e observes a tensor before an in-place write through a view, so the quantized model ignores its input

  pytorch/ao#5040 · opened on Oct 30, 2026
  1

• [RFC] Unify Converter Carriers: Module Swap / Tensor Wrap / In-Place Patch

  pytorch/ao#5039 · opened on Oct 30, 2026
  2

• [Issue] MXFakeQuantizedLinear does not support weight-only quantization

  pytorch/ao#5038 · opened on Oct 30, 2026
  2

• [RFC] Unify Workflow Configs Under AOBaseConfig: Fix "Orphan" Configs

  pytorch/ao#5037 · opened on Oct 30, 2026
  3

• [pt2e] Should convert_pt2e insert the quant primitive ops for all granularities?

  pytorch/ao#5029 · opened on Oct 27, 2026
  1

• Int8Tensor ASYMMETRIC activations: zero-point correction is subtracted after casting to bf16/fp16 (NaN/inf in fp16, bf16 error up to 16.6x the rounding floor)

  pytorch/ao#5026 · opened on Oct 26, 2026

• config_from_dict cannot deserialize torchao's own QAT configs (torchao.quantization.qat.api missing from ALLOWED_AO_MODULES)

  pytorch/ao#5022 · opened on Oct 23, 2026
  1

• [INT4][WOQ] Non-contiguous activation to int4 weight-only Linear fails on XPU and CUDA

  pytorch/ao#5017 · opened on Oct 22, 2026

• [CI] Linux nightly wheel builds fail due to Python 3.10 pin

  pytorch/ao#5007 · opened on Oct 16, 2026
  1

• Quick survey: your experience as a PyTorch Ecosystem maintainer

  pytorch/ao#5006 · opened on Oct 16, 2026

• Build Linux Wheels (x86): rocm10.1 leg fails — "No matching distribution found for torch" (nightly ROCm 10.1 wheels missing from index)

  pytorch/ao#5003 · opened on Oct 15, 2026
  34

• [RFC] Extend KernelPreference: add a vendor-native kernel member ASCEND (case study: Ascend NPU / torch_npu)

  pytorch/ao#4999 · opened on Oct 15, 2026
  4

• Build Docs upload job: git push -f to gh-pages rejected (cannot lock ref, concurrent update)

  pytorch/ao#4964 · opened on Sep 30, 2026
  1

• test_triton_mxfp8_dim1_randn[FLOOR-256-128]: exact e8m0 scale mismatch (12/1024) on ROCm MI350

  pytorch/ao#4953 · opened on Sep 25, 2026
  1