Turn @config into a dataclass_transform (#31541)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2026-02-03 17:40:59 +00:00
committed by GitHub
parent b1bb18de8d
commit 61e632aea1
32 changed files with 153 additions and 191 deletions
@@ -13,9 +13,10 @@ def register_speculator(name):
@register_speculator("eagle3")
def update_eagle3(config_dict: dict, vllm_config: dict) -> None:
def update_eagle3(config_dict: dict, pre_trained_config: dict) -> None:
"""
Apply Eagle-3 specific configuration transformations.
Apply Eagle-3 specific configuration transformations to the `dict` used to
construct the Transformers PreTrainedConfig.
Eagle-3 specific fields:
- draft_vocab_size: Size of the draft model's vocabulary
@@ -27,12 +28,14 @@ def update_eagle3(config_dict: dict, vllm_config: dict) -> None:
predictions. This is the standard field used in Eagle3 checkpoints.
"""
vllm_config["draft_vocab_size"] = config_dict.get("draft_vocab_size")
pre_trained_config["draft_vocab_size"] = config_dict.get("draft_vocab_size")
if config_dict.get("target_hidden_size") is not None:
vllm_config["target_hidden_size"] = config_dict["target_hidden_size"]
vllm_config["norm_before_residual"] = config_dict.get("norm_before_residual", True)
vllm_config["architectures"] = ["Eagle3LlamaForCausalLM"]
pre_trained_config["target_hidden_size"] = config_dict["target_hidden_size"]
pre_trained_config["norm_before_residual"] = config_dict.get(
"norm_before_residual", True
)
pre_trained_config["architectures"] = ["Eagle3LlamaForCausalLM"]
if config_dict.get("eagle_aux_hidden_state_layer_ids"):
vllm_config["eagle_aux_hidden_state_layer_ids"] = config_dict[
pre_trained_config["eagle_aux_hidden_state_layer_ids"] = config_dict[
"eagle_aux_hidden_state_layer_ids"
]
@@ -24,13 +24,16 @@ class SpeculatorsConfig(PretrainedConfig):
"""Load speculators Eagle config and convert to vLLM format."""
config_dict, _ = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
vllm_config = cls.extract_vllm_speculative_config(config_dict)
vllm_config = cls.extract_transformers_pre_trained_config(config_dict)
return cls(**vllm_config)
@classmethod
def extract_vllm_speculative_config(
def extract_transformers_pre_trained_config(
cls, config_dict: dict[str, Any]
) -> dict[str, Any]:
"""
Extract standard Transformers PreTrainedConfig config from speculators config.
"""
speculators_model_type = config_dict.get("speculators_model_type")
if speculators_model_type not in SUPPORTED_SPECULATORS_TYPES:
raise ValueError(
@@ -38,15 +41,23 @@ class SpeculatorsConfig(PretrainedConfig):
"Please ensure you're loading a speculators-format model."
)
# Start with transformer layer configuration if present
pre_trained_config = config_dict.get("transformer_layer_config", {})
# Apply anything specific to the supported algorithm
algo_updater = SUPPORTED_SPECULATORS_TYPES[speculators_model_type]
algo_updater(config_dict=config_dict, pre_trained_config=pre_trained_config)
return pre_trained_config
@classmethod
def extract_vllm_speculative_config(
cls, config_dict: dict[str, Any]
) -> dict[str, Any]:
"""Extract vLLM speculative config from speculators config."""
# validate fields
# TODO: @dsikka - use speculators pydantic model to validate
cls.validate_speculators_config(config_dict=config_dict)
# Convert from speculators config -> format that can be ingested by vLLM
vllm_config = cls.build_vllm_speculative_config(config_dict=config_dict)
# Apply anything specific to the supported algorithm
algo_updater = SUPPORTED_SPECULATORS_TYPES[speculators_model_type]
algo_updater(config_dict=config_dict, vllm_config=vllm_config)
return vllm_config
return cls.build_vllm_speculative_config(config_dict=config_dict)
@classmethod
def validate_speculators_config(cls, config_dict: dict[str, Any]) -> None:
@@ -101,14 +112,7 @@ class SpeculatorsConfig(PretrainedConfig):
)
# Build base vLLM speculative configuration
vllm_config = {
return {
"method": config_dict.get("speculators_model_type"),
"num_speculative_tokens": num_speculative_tokens,
"target_model": spec_config.get("verifier")["name_or_path"],
}
# Merge transformer layer configuration if present
transformer_config = config_dict.get("transformer_layer_config", {})
vllm_config.update(transformer_config)
return vllm_config