EAGLE-3 Korean draft for gemma-4-31B-it (seq-len 8192)

config.json

@@ -0,0 +1,66 @@
+{
+  "architectures": [
+    "Eagle3DraftModel"
+  ],
+  "auto_map": {
+    "": "config.Eagle3SpeculatorConfig"
+  },
+  "draft_vocab_size": 32000,
+  "dtype": "bfloat16",
+  "eagle_aux_hidden_state_layer_ids": [
+    2,
+    30,
+    57
+  ],
+  "embed_requires_grad": false,
+  "norm_before_fc": false,
+  "norm_before_residual": true,
+  "speculators_config": {
+    "algorithm": "eagle3",
+    "default_proposal_method": "greedy",
+    "proposal_methods": [
+      {
+        "accept_tolerance": 0.0,
+        "proposal_type": "greedy",
+        "speculative_tokens": 3,
+        "verifier_accept_k": 1
+      }
+    ],
+    "verifier": {
+      "architectures": [],
+      "name_or_path": "google/gemma-4-31B-it"
+    }
+  },
+  "speculators_model_type": "eagle3",
+  "speculators_version": "0.5.0.dev27",
+  "target_hidden_size": null,
+  "tie_word_embeddings": false,
+  "transformer_layer_config": {
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "bos_token_id": 1,
+    "eos_token_id": 2,
+    "head_dim": 256,
+    "hidden_act": "silu",
+    "hidden_size": 5376,
+    "initializer_range": 0.02,
+    "intermediate_size": 21504,
+    "max_position_embeddings": 262144,
+    "mlp_bias": false,
+    "model_type": "llama",
+    "num_attention_heads": 32,
+    "num_hidden_layers": 1,
+    "num_key_value_heads": 16,
+    "pad_token_id": null,
+    "pretraining_tp": 1,
+    "rms_norm_eps": 1e-06,
+    "rope_parameters": {
+      "rope_theta": 10000.0,
+      "rope_type": "default"
+    },
+    "tie_word_embeddings": false,
+    "use_cache": true,
+    "vocab_size": 262144
+  },
+  "transformers_version": "5.12.1"
+}

config.py

@@ -0,0 +1,93 @@
+from typing import Any, Literal
+
+from pydantic import Field, field_serializer, field_validator
+from transformers import AutoConfig, PretrainedConfig
+from transformers.models.llama.configuration_llama import LlamaConfig
+
+from speculators import SpeculatorModelConfig
+
+__all__ = [
+    "Eagle3SpeculatorConfig",
+]
+
+
+@SpeculatorModelConfig.register("eagle3")
+class Eagle3SpeculatorConfig(SpeculatorModelConfig):
+    """
+    Configuration for EAGLE-3 speculator with vocabulary mapping.
+
+    EAGLE-3 features vocabulary mapping between draft (32K) and target (128K)
+    vocabularies, enabling cross-tokenizer speculation.
+
+    :param transformer_layer_config: Configuration for the transformer decoder layer
+    :param draft_vocab_size: Size of draft model vocabulary for speculation
+    :param norm_before_residual: Apply hidden_norm before storing residual
+    """
+
+    speculators_model_type: Literal["eagle3"] = "eagle3"
+    architectures: list[str] = Field(
+        default_factory=lambda: ["Eagle3Speculator"],
+        description="Model architectures that can load these weights",
+    )
+
+    transformer_layer_config: PretrainedConfig = Field(
+        default_factory=LlamaConfig,
+        description="Configuration for the transformer decoder layer",
+    )
+
+    draft_vocab_size: int = Field(
+        default=32000,
+        description="Size of draft model vocabulary for speculation",
+    )
+
+    norm_before_residual: bool = Field(
+        default=False,
+        description="Apply hidden_norm before storing residual",
+    )
+
+    target_hidden_size: int | None = Field(
+        default=None,
+        description="Hidden size of the target model (if different from draft model)",
+    )
+
+    eagle_aux_hidden_state_layer_ids: list[int] | None = Field(
+        default=None,
+        description="Layer IDs of the Eagle auxiliary hidden state layers",
+    )
+
+    norm_before_fc: bool = Field(
+        default=False,
+        description=(
+            "If True, vLLM will add and apply RMSNorm before the fc layer when loading "
+            "this draft model (e.g. for gpt-oss draft checkpoints). Set in config when "
+            "converting or saving gpt-oss draft models."
+        ),
+    )
+
+    embed_requires_grad: bool = Field(
+        default=False,
+        description="Whether embedding layer weights require gradients during training",
+    )
+
+    @field_serializer("transformer_layer_config")
+    def serialize_transformer_config(self, value: PretrainedConfig) -> dict:
+        """Serialize transformer config to dict."""
+        return value.to_diff_dict()
+
+    @field_validator("transformer_layer_config", mode="before")
+    @classmethod
+    def validate_transformer_config(cls, value: Any) -> PretrainedConfig:
+        """Validate and convert transformer config."""
+        if isinstance(value, dict):
+            config_class: type[PretrainedConfig] = LlamaConfig
+            if "model_type" in value:
+                config_class = AutoConfig.for_model(
+                    model_type=value["model_type"]
+                ).__class__
+            return config_class(**value)
+        return value
+
+    @property
+    def target_vocab_size(self) -> int:
+        """Get target vocabulary size from transformer config."""
+        return self.transformer_layer_config.vocab_size

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c3067b281a6612a41f6f195ccfff574a386121f99fc817ff61e0a9bdfb16c6c
+size 4470642280

optimizer_state_dict.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e9c412c819e6fed56ffacc244252f40754b86c298d905ef98fbfe81f5780f892
+size 2614984665

scheduler_state_dict.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e4e931e23da4f9772178ac65b35ad8b532e12ae61cd24396c58918328ad97e2
+size 1531

val_metrics.json

@@ -0,0 +1 @@
+{"loss_0_epoch": 1.9602829618689581, "full_acc_0_epoch": 0.6379972797349239, "cond_acc_0_epoch": 0.6379972797349239, "loss_1_epoch": 3.1759868421052633, "full_acc_1_epoch": 0.3797289321761222, "cond_acc_1_epoch": 0.5951889517991245, "loss_2_epoch": 4.022086466165414, "full_acc_2_epoch": 0.23484306047796374, "cond_acc_2_epoch": 0.6184231131679151, "loss_epoch": 9.158356270139635}