vllm.models.deepseek_v41.nvidia.specialized_decode.runtime
¶
Modules:
-
decode–The specialized decode executor.
-
draft–The specialized decode of the DSpark draft model.
-
metadata_graphs–CUDA graphs of the attention metadata builds of FULL decode steps.
-
model_state–The model states of the specialized decode and of SM80 and SM90.
-
warmup–Tuning and compiling the specialized decode's kernels.