chatglm

Model type	chatglm
Class	ChatGLMCausalLMModel
Task	text-generation
Source	models/chatglm.py

Description

ChatGLM model with partial rotary and fused projection imports.

GLM-4’s original ChatGLM checkpoint layout nests decoder weights below transformer.encoder and stores QKV and gate/up projections fused. Canonicalize those names before generic GPTQ preprocessing, then split the converted QKV and gate/up tensors along their output dimension. Splitting the repacked gate/up tensors is lossless and avoids a runtime activation Split between captured CUDA graph segments.

Usage

mobius build --model <MODEL_ID> output_dir/

from mobius import build

model = build("<MODEL_ID>")