deepseek-ai

/

DeepSeek-V3.2

like 741

Follow

DeepSeek 109k

I am using the DSV32 checkpoint and can successfully host it with 8 B200 GPUs. However, when I ask the model to program in C++/Cuda, I consistently get syntax errors, not limited to following examples:

Macro definition (extra semicolons)

#define IDX(i, j) ((i + m_it * WGMMA_M) * N + (j + n_it * WGMMA_N));

missing or extra brackets

tokenB = cuda::device::barrier_arrive_tx(barB, 1, sizeof(sB)));

I have to fallback to V31-terminus for C++/Cuda generation.