diff --git a/vllm/transformers_utils/configs/speculators/algos.py b/vllm/transformers_utils/configs/speculators/algos.py index 8ec978d735..9f45183328 100644 --- a/vllm/transformers_utils/configs/speculators/algos.py +++ b/vllm/transformers_utils/configs/speculators/algos.py @@ -130,6 +130,30 @@ def update_dflash(config_dict: dict, pre_trained_config: dict) -> None: ) +@register_speculator("dflash2") +def update_dflash2(config_dict: dict, pre_trained_config: dict) -> None: + """ + Apply DFlash2 specific configuration transformations to the `dict` used + to construct the Transformers PreTrainedConfig. + + DFlash2 extends DFlash with a grouped convolution layer and a + low-rank candidate selector head. It reuses the same DFlash runtime + (method="dflash") but has its own architecture (DFlash2DraftModel) + and additional config fields for the convolution and selector. + """ + update_dflash(config_dict, pre_trained_config) + pre_trained_config["architectures"] = ["DFlash2DraftModel"] + + for key in ( + "conv_kernel_size", + "conv_group_size", + "selector_rank", + "selector_top_k", + ): + if config_dict.get(key) is not None: + pre_trained_config["dflash_config"][key] = config_dict[key] + + @register_speculator("dspark") def update_dspark(config_dict: dict, pre_trained_config: dict) -> None: """ diff --git a/vllm/transformers_utils/configs/speculators/base.py b/vllm/transformers_utils/configs/speculators/base.py index 08368d346f..1cbaf6e801 100644 --- a/vllm/transformers_utils/configs/speculators/base.py +++ b/vllm/transformers_utils/configs/speculators/base.py @@ -134,4 +134,6 @@ class SpeculatorsConfig(PretrainedConfig): } if result["method"] == "peagle": result.update({"method": "eagle3", "parallel_drafting": True}) + elif result["method"] == "dflash2": + result["method"] = "dflash" return result