Search results

Ouro llm support for candle

Status: Open.

#3659 In huggingface/candle;

· silicology1 opened on Jun 26, 2026

[feature] CPU/Metal (device-agnostic) PagedAttention to complement the CUDA kernels from #3655

Status: Open.

#3656 In huggingface/candle;

· astorise opened on Jun 25, 2026

[feature] Add min_p sampling and frequency/presence penalties to generation

Status: Open.

#3654 In huggingface/candle;

· astorise opened on Jun 25, 2026

[feature] Add DeepSeek-V3, Llama 4 and Qwen3-MoE model architectures

Status: Open.

#3653 In huggingface/candle;

· astorise opened on Jun 25, 2026

[feature] First-class LoRA adapters in candle-nn (load + merge + multi-adapter)

Status: Open.

#3652 In huggingface/candle;

· astorise opened on Jun 25, 2026

[feature] CUDA Graph capture for autoregressive decode + optional FlashInfer backend

Status: Open.

#3651 In huggingface/candle;

· astorise opened on Jun 25, 2026

[feature] Weight quantization kernels: GPTQ/Marlin, AWQ, block-wise FP8

Status: Open.

#3650 In huggingface/candle;

· astorise opened on Jun 25, 2026

Deterministic panic (DoS) via Long1 opcode in candle-core's pickle reader

Status: Open.

#3619 In huggingface/candle;

· professor-moody opened on Jun 14, 2026

Pickle memo bomb — CPU + memory exhaustion (algorithmic-complexity DoS) in candle-core pickle reader

Status: Open.

#3620 In huggingface/candle;

· professor-moody opened on Jun 14, 2026

Unbounded pickle-VM working set in candle-core/src/pickle.rs (DoS via crafted .pth)

Status: Open.

#3617 In huggingface/candle;

· PCfVW opened on Jun 13, 2026

candle-flash-attn-v3 fails to build with CUDA 13.x (pinned CUTLASS uses unversioned PFN_cuTensorMapEncodeTiled)

Status: Open.

#3605 In huggingface/candle;

· Xueying-VirtueAI opened on Jun 11, 2026

candle-core CUDA backend: LaunchConfig::for_num_elems(el as u32) silently truncates; ops fail on tensors with >= 2^32 elements

Status: Open.

#3604 In huggingface/candle;

· Xueying-VirtueAI opened on Jun 11, 2026