vllm.models.hy_v4.nvidia.dfly
¶
Classes:
-
HYV4DFlyFeedForward–Keep the module identity while waiting for the streamed FFN width.
-
HYV4DFlyQuantizationConfig–Match draft linears against their checkpoint paths, without rewriting regexes.
HYV4DFlyFeedForward
¶
Bases: HYV4FeedForward
Keep the module identity while waiting for the streamed FFN width.
Source code in vllm/models/hy_v4/nvidia/dfly.py
HYV4DFlyQuantizationConfig
¶
Bases: CompressedTensorsConfig
Match draft linears against their checkpoint paths, without rewriting regexes.