Bases: DraftModelSpeculatorConfig
DFly behavior shared by model-specific configuration subclasses.
Source code in vllm/v1/worker/gpu/spec_decode/dfly/config.py
| class DFlySpeculatorConfig(DraftModelSpeculatorConfig):
"""DFly behavior shared by model-specific configuration subclasses."""
base_method = "dfly"
uses_target_hidden_states = True
shares_target_positions = False
parallel_drafting = True
@classmethod
def default_model(cls, spec: "SpeculativeConfig") -> str:
if spec.target_model_config is None:
raise ValueError("DFly requires target_model_config")
return spec.target_model_config.model
@classmethod
def draft_hf_overrides(
cls, spec: "SpeculativeConfig"
) -> Callable[["PreTrainedConfig"], "PreTrainedConfig"]:
target = spec.target_model_config
if target is None:
raise ValueError("DFly requires target_model_config")
raise ValueError(
"No DFly configuration is registered for target model type "
f"{target.hf_text_config.model_type!r}"
)
@classmethod
def resolve_num_speculative_tokens(cls, spec: "SpeculativeConfig") -> None:
n = spec.num_speculative_tokens
if n is None:
block_size = getattr(
spec.target_model_config.hf_config, "dfly_block_size", None
)
if type(block_size) is not int or not 2 <= block_size <= 17:
raise ValueError("DFly requires dfly_block_size in [2, 17]")
# The checkpoint block includes the separate anchor query.
n = block_size - 1
if type(n) is not int or not 1 <= n <= 16:
raise ValueError("DFly requires num_speculative_tokens in [1, 16]")
spec.num_speculative_tokens = n
@classmethod
def apply_options(cls, spec: "SpeculativeConfig") -> None:
if spec.enable_adaptive_verification:
raise ValueError("DFly does not support adaptive verification")
if spec.rejection_sample_method != "standard":
raise ValueError("DFly requires standard target verification")
@property
def num_lookahead_tokens(self) -> int:
return self.num_speculative_tokens + 1
def max_num_draft_seqs(self, max_num_batched_tokens: int) -> int:
return max_num_batched_tokens // self.num_lookahead_tokens
|