philipjohnbasile/glm52-idx-extract

GLM-5.2 — DSA indexer weights

Explore the model guide · All public work

Release at a glance

	This artifact
Purpose	Sparse-attention indexer tensors extracted from GLM 5.2 for runtime and conversion work.
Runtime	A component for engines that implement the GLM DSA path. It is not a complete language model and cannot generate by itself.
Status	Component only; see the evidence and limits below.
Tensor download	0.20 GB (0.18 GiB) of root .safetensors files, including any root sidecars. This is a file-size total, not peak RAM.
Read first	The root tensor files total about 197 MB in this revision. The model config identifies the source architecture, not standalone completeness.

The DSA (sparse-attention) indexer weights extracted from zai-org/GLM-5.2 (743B total / 39B active, MIT), split out as a component artifact (197.2 MB of root tensor files in this revision).

GLM-5.2's attention uses a learned indexer to select a top-k subset of past tokens once the context exceeds index_topk, instead of attending densely over everything. Those indexer tensors are a tiny fraction of the model and are used by engines implementing GLM-5.2's sparse-attention path — and pulling them out of the full checkpoint otherwise means downloading large source shards to keep the small indexer payload.

Contents

• out-idx-*.safetensors — the extracted indexer tensors
• config.json / generation_config.json — GLM-5.2 config (GlmMoeDsaForCausalLM)
• _meta/ — tokenizer files

Use with iliria

iliria runs GLM-5.2 off SSD by streaming experts from disk. Its converter takes --keep-idx to produce exactly this artifact; this repo saves you that step. Point the engine at a container that includes these tensors and the DSA path activates once context passes index_topk.

Provenance

Derived from zai-org/GLM-5.2 (MIT). Extraction only — no retraining, no quantization of the indexer tensors themselves. The weights are unmodified values lifted from the source checkpoint.