Two NVIDIA DGX Sparks, one model. DeepSeek-V4-Flash, 284B params, full 1M-token context, across two DGX Sparks linked by QSFP/RoCE, driven entirely from my Mac over SSH.
The 30-megabyte mistake that froze both nodes, the off-by-1 that wasn't, and every script I wrote so it never happens twice.
Two NVIDIA DGX Sparks, one model. DeepSeek-V4-Flash, 284B params, full 1M-token context, across two DGX Sparks linked by QSFP/RoCE, driven entirely from my Mac over SSH. The 30-megabyte mistake that froze both nodes, the off-by-1 that wasn't, and every script I wrote so it never happens twice.
0 vue
Publié le 19/09/2026 03:33
0:55
Two NVIDIA DGX Sparks, one model. DeepSeek-V4-Flash, 284B params, full 1M-token context, across two DGX Sparks linked by QSFP/RoCE, driven entirely from my Mac over SSH.
The 30-megabyte mistake that froze both nodes, the off-by-1 that wasn't, and every script I wrote so it never happens twice.
0 commentaire
?