NVIDIA-only (cutedsl/cutlass) kernels for DeepSeek V4.
These modules import cutlass/cutedsl at module top level, so they must not be imported on non-CUDA platforms. Callers should gate on vllm.utils.import_utils.has_cutedsl() before importing from here.
Modules:
| Name | Description |
|---|---|
attention | DeepseekV4 MLA Attention Layer |
fused_indexer_q_cutedsl | |
prepare_megamoe | Triton input-staging kernel for DeepSeek V4 MegaMoE. |