Repository Issue Activity (beta)

#3065 [Bug]: autowrap_forward fails with NameError for models defined in __main__ under cProfile/runpy

#3083 [Bug]: requires_compute_capability crashes at collection time on Apple Silicon

#3057 [Bug] sequential pipeline cannot trace models whose forward is decorated without functools.wraps (Qwen3.8 / Qwen3_5GatedDeltaNet)

#3037 Qwen3.5-MoE misses ARCH_TO_2D_MAPPINGS, forcing a 2D→3D→2D expert round-trip that doubles expert memory at load

#2296 [AWQ] add option to take smooth layer quantization into accout

#3054 [Bug] AutoRoundModifier: calibration inputs retained on GPU exhaust VRAM before optimization

#2939 GPTQ silently corrupts fused-expert MoE (Qwen3.6-A3B) on 0.12.0 — fixed on main by #2847/#2848, please cut a release

#3069 [Bug]: Cannot dispatch MOE correctly if using naming `w1,w3` or `block_sparse_moe`

#3071 [Bug]: Qwen3-VL-2B (dense) AWQ/GPTQ calibration fails, vision tower not FX-traceable (rot_pos_emb CUDA driver error), even with visual.* in ignore

#2984 [Bug]: Loading models with DDP leads to race condition on HF `cached_files`

#2936 How to achieve high accuracy on NVFP4？

#3064 [Bug]: tools/collect_env.py crashes on Apple Silicon (torch.mps has no get_device_name)

#2949 [Feature] Zero-copy DDP weight sharing for memory-constrained large MoE quantization

#3059 [Bug]: Noisy gpt_oss test

#2741 [Feature] Native intra-calibration resume in `SequentialPipeline`

#3040 [Bug] Fix ignore list for linear_attn with Qwen3.8

#2919 Q3 Roadmap

#2522 [Bug]: Evaluate AWQ and GPTQ for Gemma

#2690 [RFC] HIGGS Integration into llm-compressor

#3032 [Feature] Multimodal calibration dataset support in oneshot

#3023 [Bug] AutoRoundModifier: input_capture_hook accumulates GPU memory during optimization when gradient_accumulate_steps > 1

#2693 [Bug]: `compressed-tensors` W4A8 INT (int4 weights + int8 activations) fails to find any compatible kernel on H100

#2975 [ModelFreePTQ] Memory issues with multi-gpu

#1939 HOW To: Quantization: Qwen/Qwen3-VL-30B-A3B-Instruct AWQ

#2979 [Kimi-K3] Integrate KimiSparseMoeBlock with LinearExperts2D for REAP Support