GLM-5.3-Flash-0.1B-A0.1B-MTP

Tiny version of zai-org/GLM-5.3-Flash with MTP (Multi-Token Prediction) layer included, intended for testing MTP quantization pipelines in llm-compressor.

Base tiny model: inference-optimization/GLM-5.3-Flash-0.1B-A0.1B

What's different from the base tiny model

• model_mtp.safetensors is added with 55 synthetic MTP tensors at model.language_model.layers.5.*
• model.safetensors.index.json is created mapping layers 0–4 → model.safetensors, layer 5 (MTP) → model_mtp.safetensors
• MTP tensors are zero-initialized bfloat16, scaled to match the tiny model's dimensions
• MTP layer structure: deepseek_sparse_attention (mirrors layer 3) + 4 GLM5-specific tensors (eh_proj, enorm, hnorm, shared_head.norm)

Config

Field	Value
model_type	glm5_next
hidden_size	256
num_hidden_layers	5
num_nextn_predict_layers	1
n_routed_experts	8

Notes

This model is not intended for inference. It is a synthetic test fixture for validating MTP quantization code paths in llm-compressor. GLM stores MTP tensors as the last N layers of model.language_model.layers.*, unlike Qwen/Nemotron which use an mtp.* prefix.

The base tiny model omits MTP weights because the glm5_next model type is not yet in transformers and _keys_to_ignore_on_load_unexpected drops them at load time. This variant restores them for pipeline testing.