Got GLM-5.2 + MTP speculative decode running on 4× DGX Spark (GB10) — and the build piece the public recipe is missing
Summary
The author successfully ran GLM-5.2 with MTP speculative decoding on a 4× DGX Spark (GB10) setup, revealing a missing component in the public build recipe.
Similar Articles
Follow-up: GLM-5.2 NVFP4 on four DGX Sparks — the MTP mystery is solved, and it's now ~24 tok/s at 128K context
A bug in vLLM's speculative decoding configuration for GLM-5.2 NVFP4 on four DGX Sparks was fixed, resolving a performance tradeoff and achieving ~24 tok/s at 128K context with MTP4.
GLM-5.2-Int4-Int8 on 8× GB10: ~1,200 t/s prefill, 33–54 t/s avg decode
Describes deployment and benchmarking of the quantized GLM-5.2-Int4-Int8Mix model on an 8-node DGX Spark (GB10) cluster using a custom vLLM fork, achieving ~1,200 t/s prefill and ~35 t/s decode with MTP tool calling.
@Tech2Wild: Running GLM-5.2 at home the FULL 744B, all 256 experts, UNPRUNED across 4× NVIDIA DGX Spark (GB10). 200K context · MTP …
A detailed recipe for running the unpruned GLM-5.2 model (744B parameters, 256 experts) across 4 NVIDIA DGX Spark nodes with 200K context, achieving up to 60.5 tok/s aggregate. Includes performance benchmarks, credits, and patches.
4-bit GLM-5.2 (753B MoE) on 4× DGX Spark: 70.8% on Terminal-Bench 2.1 vs 81.0% for the full model
Running a 4-bit quantized version of GLM-5.2 (753B MoE) on 4 DGX Spark machines achieves 70.8% on Terminal-Bench 2.1, compared to 81.0% from the full model.
@Ex0byt: Update: the road to GLM-5.2: we're getting there, folks! non-quantized, non-pruned DeepSeek-v4-Flash. 11tok/s on a sing…
Update on running a non-quantized DeepSeek-v4-Flash model at 11 tok/s on a single DGX Spark using sglang inference and a custom mega-kernel, progressing towards GLM-5.2.