Issues

• Error: backward is not supported for conv-transpose1d

  huggingface/candle#4064 · Filter by author LunaticWyrm467 opened on Oct 24, 2026

• ConvTranspose1D: unsigned subtraction underflow for valid output shapes

  huggingface/candle#4063 · Filter by author LunaticWyrm467 opened on Oct 24, 2026

• Current version BROKEN due to ZIP being yanked

  huggingface/candle#4058 · Filter by author Goldziher opened on Oct 22, 2026

• The fused CPU RMSNorm rounds the scale to the storage dtype, and the two implementations in ops.rs disagree with each other

  huggingface/candle#4052 · Filter by author lu-zero opened on Oct 20, 2026
  1

• candle-onnx: Shape op ignores ONNX semantics for start and end

  huggingface/candle#4049 · Filter by author RaphaelFakhri opened on Oct 20, 2026

• Implement backward pass for flash attention

  huggingface/candle#4047 · Filter by author Enderlook opened on Oct 19, 2026

• wasm32 simd128: Q8_0 vec_dot_2 / vec_dot_4

  huggingface/candle#4043 · Filter by author mgoldio opened on Oct 16, 2026

• Release zip 9 support: zip 8.6.0 is yanked

  huggingface/candle#4042 · Filter by author jbg opened on Oct 16, 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
  1

• 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
  1

• 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

• 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