Repository Issue Activity (beta)

#3752 [Invalid — misfiled downstream issue] LayerNorm fast path silently severs autograd graph (apply_op3_no_bwd)

#3750 DeBERTa-v2 models fail to run at F16

#3739 Track ROCm follow-ups for PR #3424 (RDNA4 / ROCm 7.1)

#3745 Gemma models use the wrong activation function

#3744 CPU matmul silently returns incorrect values for stride-zero broadcast batch

#3430 SafeTensors loading produces corrupted weights on big-endian architectures (s390x)

#3737 regression in support for nvidia rtx 20xx GPUs with newer nvcc toolchains

#3464 Metal buffer pool grows beyond physical RAM on small unified-memory devices

#3735 CPU conv2d backward produces incorrect gradients for non-contiguous kernels

#3651 [feature] CUDA Graph capture for autoregressive decode + optional FlashInfer backend

#3727 Softmax on f16/bf16 is numerically broken on long axes (rows sum to ~2.0, not 1.0)

#3653 [feature] Add DeepSeek-V3, Llama 4 and Qwen3-MoE model architectures

#3654 [feature] Add min_p sampling and frequency/presence penalties to generation

#3732 PagedKvCache::write_new_kv's host readback breaks CudaGraph capture for the decode step

#3698 phi3: long-prompt output degrades into word-salad — BF16 RoPE table + ignored `sliding_window` (rope dtype also affects qwen2/gemma/gemma2)

#3728 Add a graph-replay-safe decode position seam (`Cache::set_decode_position`) for CUDA graph decode capture

#3725 Add an additive use_flashinfer_attention seam to candle-transformers::models::llama for external decode-attention wiring

#3687 candle-datasets ssl

#3720 Add an additive Cache::Paged seam to candle-transformers::models::llama for external paged-attention wiring

#3708 No lazy/mmap loading for GGUF weights (MmapedSafetensors equivalent)

#3707 CPU quantized matmul has no batch path: prefill runs at token-at-a-time speed

#3705 Metal: deterministic corruption in quantized generation once the KV cache reaches depth 8,192 (evidence points at a missed synchronization)

#3701 [feature] Opt-in block metadata for long-context KV pruning experiments

#3696 Add first-class LoRA adapter injection to candle-transformers model loaders

#3619 Deterministic panic (DoS) via `Long1` opcode in candle-core's pickle reader