🔌
New: connect Claude & other AIs to GenAIList over MCP — research the catalog and contribute to the shared knowledge base. Learn how →
// RUN — FITS
Can the NVIDIA DGX Spark run DEQ-Transformer (Post-LN) + Jacobian Regularisation?
Yes — here's how fast and up to what context, with real measured numbers.
DEQ-Transformer (Post-LN) + Jacobian Regularisation on the NVIDIA DGX Spark
| Quant | Fits? | Max context | Speed | Beyond context |
|---|---|---|---|---|
| FP16 | yes | 1732k | 1016.8-1211.8 t/s | ~37.2-44.4 (slow) |
| Q8 | yes | 1733k | 2033.6-2423.6 t/s | ~74.5-88.8 (slow) |
| Q4 | yes | 1734k | 3631.4-4327.8 t/s | ~133-158.5 (slow) |
Estimate (memory-bound). Beyond "max context" the KV cache spills to SSD swap and speed drops to the "beyond" figure.
→ All models on the NVIDIA DGX Spark
→ DEQ-Transformer (Post-LN) + Jacobian Regularisation specs
→ Other hardware
FAQ
Can the NVIDIA DGX Spark run DEQ-Transformer (Post-LN) + Jacobian Regularisation?▶
Yes. At Q4 it generates ~3631.4-4327.8 tok/s and fits up to 1734k context; beyond that it spills to SSD swap and slows to ~133-158.5 tok/s.
How much context fits?▶
At Q4, up to about 1734k tokens stay in fast memory; longer context spills to SSD swap and throughput collapses.