@sudoingX Ran Ling 3.1 Flash on Four Sparks for 12 Hours on One Prompt
7.5 tok/s over llama.cpp RPC, 325k tokens out, and 12 of 12 tests passing at the end.
Slow, and it finished.
Ling 3.1 Flash does not fit on @sudoingX’s two DGX Sparks, so @niklaslenz_ai lent him four. With no official FP4 build for vLLM yet, he ran AtomicChat’s AD-Q4_K_M GGUF over llama.cpp RPC, with Hermes Agent driving. The prompt was his floating tree 3D scene.
| Measure | Result |
|---|---|
| Wall time | 12 hours 26 minutes |
| Requests and output | 86 requests, 325k tokens |
| Decode | 7.5 tok/s, q4 over RPC |
| Peak context | 143k of 262k |
| Tests | 12 of 12 pass, smoke green |
| Code | 2,166 lines across 12 modules |
It built the island, the tree with a full crown, two houses, the pond and the dusk sky. With no vision on this setup, it checked its renders with pixel stats and frame diffs, and caught and fixed a flaky test.
Four of its turns took about 9 hours and 77% of all its tokens and wrote very little; one ran 2h52m and left a 10 line file. Next, the same prompt goes to the full model on the API, to show whether the long turns come from the model or his setup.
“This was my first time running 4x dgx sparks in my entire life,” he writes. He ran the previous Ling 3 on a single Spark and can see how fast Ling is improving.
Sources
- @sudoingX on X · Oct 22, 6:59 PM PT
- @sudoingX on X · Oct 23, 6:27 AM PT
- @sudoingX on X · Oct 23, 6:37 AM PT