| """Configuration class for the MiniQwen3 standalone model. |
| |
| Mirrors HF's `Qwen3Config` but with every architectural knob exposed |
| as a constructor kwarg, so this file is the single source of truth for |
| ablation experiments. Keep `MiniQwen3Config.model_type` unique so the |
| HF Auto* registry never silently falls back to the upstream Qwen3 |
| implementation. |
| """ |
|
|
| from __future__ import annotations |
|
|
| from transformers import PretrainedConfig |
|
|
|
|
| class MiniQwen3Config(PretrainedConfig): |
| """Standalone, easy-to-edit Qwen3-style transformer config. |
| |
| Defaults target ~0.5B parameters (24L x d=1024, 16 attn heads, |
| GQA 2:1, SwiGLU MLP 3072, RoPE theta=1e6, tied embeddings, QK-Norm). |
| """ |
|
|
| model_type = "mini_qwen3" |
| keys_to_ignore_at_inference = ["past_key_values"] |
|
|
| base_model_tp_plan = None |
| base_model_pp_plan = None |
|
|
| def __init__( |
| self, |
| vocab_size: int = 151936, |
| hidden_size: int = 1024, |
| intermediate_size: int = 3072, |
| num_hidden_layers: int = 24, |
| num_attention_heads: int = 16, |
| num_key_value_heads: int = 8, |
| head_dim: int = 64, |
| hidden_act: str = "silu", |
| max_position_embeddings: int = 4096, |
| initializer_range: float = 0.02, |
| rms_norm_eps: float = 1e-6, |
| use_cache: bool = True, |
| rope_theta: float = 1_000_000.0, |
| attention_bias: bool = False, |
| mlp_bias: bool = False, |
| attention_dropout: float = 0.0, |
| tie_word_embeddings: bool = True, |
| use_qk_norm: bool = True, |
| pad_token_id: int | None = None, |
| bos_token_id: int | None = 151643, |
| eos_token_id: int | None = 151643, |
| **kwargs, |
| ) -> None: |
| if num_attention_heads % num_key_value_heads != 0: |
| raise ValueError( |
| f"num_attention_heads ({num_attention_heads}) must be divisible " |
| f"by num_key_value_heads ({num_key_value_heads}) for GQA." |
| ) |
|
|
| self.vocab_size = vocab_size |
| self.hidden_size = hidden_size |
| self.intermediate_size = intermediate_size |
| self.num_hidden_layers = num_hidden_layers |
| self.num_attention_heads = num_attention_heads |
| self.num_key_value_heads = num_key_value_heads |
| self.head_dim = head_dim |
| self.hidden_act = hidden_act |
| self.max_position_embeddings = max_position_embeddings |
| self.initializer_range = initializer_range |
| self.rms_norm_eps = rms_norm_eps |
| self.use_cache = use_cache |
| self.rope_theta = rope_theta |
| self.attention_bias = attention_bias |
| self.mlp_bias = mlp_bias |
| self.attention_dropout = attention_dropout |
| self.use_qk_norm = use_qk_norm |
|
|
| super().__init__( |
| pad_token_id=pad_token_id, |
| bos_token_id=bos_token_id, |
| eos_token_id=eos_token_id, |
| tie_word_embeddings=tie_word_embeddings, |
| **kwargs, |
| ) |
|
|
