candle · Issues· 908 open

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

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

Regression: `layer_norm()`/`rms_norm()` always fails to find weight/bias tensors [main branch]

Gemma 4 models cannot use

mimi uses the interleaved RoPE convention but the checkpoints are stored in the transformers layout

CUDA `copy_strided_src` copies past the end of an offset source view, silently corrupting `slice_scatter` (CPU disagrees)

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

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

Design check: Intel Arc (SYCL) backend

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

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

Quantization Error unsupported safetensor dtype F8_E4M3

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

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

Candle leaves ~ model size of memory unused on Spark