Repository Issue Activity (beta)

#2378 Qwen3.6-35B-A3B (Qwen3_5MoeForConditionalGeneration): every completion request fails with 'moe experts forward / PendingIsqLayer is in an invalid transitional state'

#1714 Qwen 3 VL support for GGUF models?

#2262 Cargo Crate out of date

#2125 GGUF loading fails for Qwen3.5 models with `Unknown GGUF architecture qwen35`

#2367 Don't force images and other media to come first in context for VLMs

#2361 Support qwen35 architecture

#2363 Chat template output differs from transformers on tool-calling prompts

#1951 Support realtime API for audio streaming

#156 Model Wishlist

#1114 cuda+cpu mode panic at unwrap

#1948 Docker: Missing mistralrs binary and libc dependencies in cpu-0.7 image

#2340 Unsupported group_size: 6 on CUDA

#2360 Unknown GGUF architecture `gemma4`

#2024 paged_attention: O(N²) memory allocator thrashing and continuous batching failure

#2343 Concurrent text-serving throughput does not scale above serial despite PagedAttention engaged (v0.9.0, H100, Llama-3.2-1B bf16)

#2323 Batched multimodal prefill crashes with rank mismatch; mixed-length image batches silently corrupt output (Gemma 4)

#2336 Web UI: sending an image returns HTTP 500 `invalid image source` (v1 API works)

#2335 Llama-3.2-11B-Vision (mllama): errors with PagedAttention by default, and image inference crashes on CUDA with `--paged-attn off`

#2329 sampler: greedy paths report log of raw logit instead of log of probability

#2328 Re-export pinned candle versions (or at least candle-core::Device)

#2322 MemoryUsage::query runs sysinfo System::new_all() in the attention hot path (CPU + integrated CUDA)

#2321 PagedAttentionScheduler::schedule livelocks with concurrent multimodal requests

#2128 Add architecture support for `paddleocr_vl` (PaddleOCR-VL-1.5)

#2160 GGUF Qwen3 MoE on Metal loads successfully but fails on first inference via CUDA-only indexed_moe_forward path

#2272 All looks good on GB10 with 0.8.21 but doesn't work due to CUDA mismatch (?)