Issues

• ONNX Transpose with omitted perm swaps only the last two axes

  huggingface/candle#4034 · Filter by author vera-rubin opened on Oct 12, 2026

• conv1d/encoder gives different results for strided (narrow) vs contiguous input with identical values

  huggingface/candle#4031 · Filter by author Trisert opened on Oct 10, 2026

• x86 repacked matmul: add a Q5K tile

  huggingface/candle#4029 · Filter by author bhaskargurram-ai opened on Oct 9, 2026

• Metal: Tensor::copy() shares the source's buffer, so in-place updates (Var::set, zero_set) show through the "copy"

  huggingface/candle#4027 · Filter by author skilesare opened on Oct 8, 2026
  1

• ONNX ReduceMax/ReduceMin treat explicitly empty axes as a no-op by default

  huggingface/candle#4023 · Filter by author hulkbig opened on Oct 5, 2026

• Metal SDPA: support head dim 48 (d_model 384 / 8 heads)

  huggingface/candle#4021 · Filter by author OmniaFlux opened on Oct 3, 2026

• Metal SDPA silently returns wrong values for non-contiguous q/k/v

  huggingface/candle#4020 · Filter by author OmniaFlux opened on Oct 3, 2026
  1

• Apertus model support

  huggingface/candle#4018 · Filter by author saona-raimundo opened on Sep 30, 2026
  1

• wasm32 simd128 backend doesn't define CurrentCpuF16/CurrentCpuBF16, breaking any build with target-feature=+simd128

  huggingface/candle#4014 · Filter by author willarrington opened on Sep 29, 2026
  1

• quantized_phi3: partial rotary embeddings fail with "inconsistent last dim size in rope"

  huggingface/candle#4004 · Filter by author czeti opened on Sep 27, 2026

• Unsound APIs (Layout::new, Layout::contiguous_with_offset) — unchecked start_offset leads to heap-buffer-overflow

  huggingface/candle#3986 · Filter by author ksj1230 opened on Sep 17, 2026

• Vulkan accelerator backend (Device::Vulkan): implementation available, seeking to upstream

  huggingface/candle#3985 · Filter by author rexlunae opened on Sep 16, 2026
  2

• Gemma 4 models cannot use

  huggingface/candle#3967 · Filter by author Zhang-JK opened on Sep 9, 2026

• Design check: Intel Arc (SYCL) backend

  huggingface/candle#3921 · Filter by author Hahihula opened on Aug 28, 2026

• qwen3_vl: no way to reset the KV cache, so a model instance is single-use

  huggingface/candle#3901 · Filter by author sidd-27 opened on Aug 17, 2026
  1

• CUDA is bound at link time, so a cuda-enabled binary cannot start without CUDA installed

  huggingface/candle#3900 · Filter by author nabbisen opened on Aug 17, 2026
  4

• F8E4M3 CUDA kernels are unreachable: DType::as_str() yields f8e4m3, the kernels are named f8_e4m3

  huggingface/candle#3886 · Filter by author astorise opened on Aug 14, 2026
  1

• get_or_load_func drops the kernel name from CUDA symbol-lookup errors

  huggingface/candle#3884 · Filter by author astorise opened on Aug 14, 2026
  2

• potential deadlock risk from multiple read locks over storage in tensor operations

  huggingface/candle#3878 · Filter by author scaspin opened on Aug 12, 2026
  3

• qwen3_5::Config deserializes without checking its fields agree, so an incoherent one fails deep instead of at load

  huggingface/candle#3868 · Filter by author astorise opened on Aug 9, 2026

• ModelOpt checkpoints: the weight source stops at the projection; detection, shards and config are still on every consumer

  huggingface/candle#3866 · Filter by author astorise opened on Aug 8, 2026

• Qwen3.5 quantized (GGUF) MoE checkpoints are not supported (qwen35moe)

  huggingface/candle#3865 · Filter by author astorise opened on Aug 8, 2026

• QTensor::quantize on a narrowed view quantizes the parent (silently in release, debug_assert in debug)

  huggingface/candle#3864 · Filter by author oetiker opened on Aug 8, 2026
  1

• Metal: K-quant mul_mv kernels store past ne01; Q2K over-dispatches 2x on top of it

  huggingface/candle#3863 · Filter by author oetiker opened on Aug 8, 2026
  1

• qwen3_5: the sparse-MoE layers and partial rotary the released checkpoints need

  huggingface/candle#3837 · Filter by author astorise opened on Aug 4, 2026