Repository Issue Activity (beta)

#2715 [Bug]: # Internvl quantization for language part: KeyError: 'vision_tower.embeddings.cls_token'

#2879 [Bug]: mistral quantization w4a6 and w8a8 int8 schemes slower than bf16

#2883 Transformers 5.13.0 Failure for GraniteMoeParallelExperts

#2730 [Bug]: Mix-precision vLLM loading failure (MXFP4+MXFP8)

#2560 [Bug]: Cannot properly compress remaining layers of already compressed model (gpt-oss-20b)

#2706 [AutoRound] Q2 2026 Roadmap

#2339 [Bug]: Granite 4.0 h small FP8 quantization fails (model can't load)

#2853 KV-cache calibration silently produces zero scales for models that read the cache structurally (e.g. DiffusionGemma)

#2480 [Feature] Enable GPU offloading for reduced weight movement

#2512 Has anyone performed inference on a GLM-5 model quantized with AWQ+INT4 using VLLM?

#2563 [Bug]: Encountered a type error related to offload.device.type when running kimi_k2_example.py

#2069 [Performance]: Quantized Qwen3 MoE

#2494 [Feature Request/Question]fused QKV linear layers

#2568 Getting "KeyError: language_model.language_model.layers.0.mlp.experts.w2_weight_packed" while trying to infer AWQ quantized Qwen3.5-122B model using vllm

#2880 [Bug]: compression fails silently if all model files except for tokenizer are missing

#2862 GLM-5.2 (glm_moe_dsa) GPTQ: whole-MoE-layer sequential target OOMs, and sequential_targets=Linear can't fx-trace GlmMoeDsaIndexer (data-dependent control flow)

#2479 AWQ various enhancements and research

#2624 Q2 2026 Roadmap

#2872 [RFC] Agent skill for hardware-aware quantization + example generation

#2843 [Bug]: GPTQ modifier error in llm-compressor

#1442 Add Additional Model Mappings for `AWQ` and `SmoothQuant`

#1927 [Python 3.10+] Modernize codebase with Python 3.10+ idioms and features

#2268 LLMCompressor version compatibility with vllm

#2397 [Bug]: compatibility issue between torch fxtrace, llmcompressor and qwen architecture

#2680 AWQ calibration for Qwen3.5 thinking models: non-thinking leakage of <think> and failure to emit </think> after quantization