howtospark
models / zai-org/GLM-5.2-FP8
Z.ai logo

753B params~51B activefp8MoE 256×8MTP ×11.0M ctxmit
Sparks
~887 GB
~510 GB
~350 GB
~343 GB
~329 GB
110 usable
REAP dialno pruning · 256 experts
Spec decode
Context33K / 1.0MKV
On the Sparks — 1× · 2-bit experts + NVFP4 dense · 33K ctx · FP16 KV~0 ctx fits usable
spark
KV cache125.6* GBFP16
Dense weights14.3* GB4-bit
Expert planes181.2* GB2-bit
110 usable
+219 over
329* / 110 GB
299% of usable
Expert planesDense weightsKV cacheActivations + graphsOS reserve

* segment sizes marked with an asterisk are estimates pending a measured run

Estimated per-node footprint, weights + KV sharded across the Spark (tensor parallel). Overhead ~8 GB/node; KV at FP16.

layer map
78 layersdense FFN×3MoE FFN×75
in
out

one MoE layer · 256 experts, 8+1 fire per token

each expert ≈ 38M params

resident: 75 MoE layers × 256 experts × 38M 724.8B

active/token: 75 × 8 × 38M 22.6B experts

+ 28.6B always-on dense (attention, embeddings, shared stack) = 51.2B active

Recipes (all recipes)

Worked deployments that run this model on DGX Spark — each shows how it fills each node's unified memory.

Optimize
Verdict

753B params, ~753 GB at already fp8-quantized and still ~377 GB at 4-bit — one Spark can't hold it; you need surgery or more Sparks.

Make it fit

Get the weights (and headroom for KV) inside the 128 GB pool — the hard gate.

Prune experts (then quantize)

With 256 routed experts and only 8 firing per token, REAP-style pruning cuts total params without touching active compute.

pruning to half the experts ≈ halves the routed-expert weights, ~207 GB at 4-bit

Split or stream it

Even 4-bit is ~377 GB against a 110 GB usable budget; shard across ≥4 Sparks or stream cold weights from NVMe.

Go below 4 bits

Codebook/outlier methods reach 2–3 bpw when 4-bit still overflows. — for MoE, the proven shape is a 2-bit sign-symmetric codebook on the routed experts with the dense stack kept at FP8 (vLLM-Moet).

~210 GB with 2-bit experts + FP8 dense (est.)

Make it fast

Once it fits: fewer bytes per token and fewer decode steps against 273 GB/s.

Use its MTP heads

Ships 1 multi-token-prediction layer(s) — self-speculative decode with no separate draft model; check vLLM/SGLang support for this arch.

Tame the KV cache

At 1048576 tokens the KV cache alone is ~4020 GB at FP16; quantize or evict.

Make it yours

Change what the model does — adapt, edit, or steer it — independent of size.

Fine-tune it

LoRA/QLoRA adapts behavior on one Spark without touching the base weights.