vllm.models.deepseek_v41.nvidia.ops
¶
Modules:
-
cute_dsl–CuTe DSL kernels for DSV4.1 mHC.
-
fused_wo_a–DSV4.1 small-batch MXFP8 WO-A with inverse RoPE and quantization.
-
mega_mhc– -
mhc–Dispatch DSV4.1 mHC operations and overlap coefficient generation.
-
o_proj–DSV4.1 output projection with a small-batch SM100/SM103 fusion.