bot-lab-21 / GLM-5.3-Int4-Int8Mix-Pollard-GPTQ-MTP-Draft

The model-author multi-token-prediction layer of zai-org/GLM-5.3 (layer 78: full MoE decoder block + ehproj/enorm/hnorm + shared-head norm), packaged as a standalone speculative draft for vLLM method: mtp: 13 GB, compressed-tensors pack-quantized, with the embeddings and lmhea…