Upload modeling.py

modeling.py

@@ -28,6 +28,111 @@ except ImportError:
     )
 
 
+_LAYER_PATTERN = (["linear_attention"] * 3 + ["full_attention"]) * 10
+
+LOCKED_TEXT_FIELDS = {
+    "hidden_size": 2048,
+    "num_hidden_layers": 40,
+    "num_attention_heads": 16,
+    "num_key_value_heads": 2,
+    "vocab_size": 248320,
+    "num_experts": 256,
+    "num_experts_per_tok": 8,
+    "moe_intermediate_size": 512,
+}
+
+
+class Qwen35MoeTextConfig:
+    """Nested language-model config (``text_config`` in the JSON).
+
+    This is kept as a standalone dataclass-like container for callers
+    that need to introspect text-level defaults without instantiating
+    the full ``PretrainedConfig`` chain.
+    """
+
+    model_type = "qwen3_5_moe_text"
+
+    def __init__(
+        self,
+        hidden_size: int = 2048,
+        num_hidden_layers: int = 40,
+        num_attention_heads: int = 16,
+        num_key_value_heads: int = 2,
+        vocab_size: int = 248320,
+        head_dim: int = 256,
+        hidden_act: str = "silu",
+        intermediate_size: int = 2048,
+        moe_intermediate_size: int = 512,
+        shared_expert_intermediate_size: int = 512,
+        num_experts: int = 256,
+        num_experts_per_tok: int = 8,
+        max_position_embeddings: int = 262144,
+        rms_norm_eps: float = 1e-6,
+        rope_theta: float = 1000000.0,
+        partial_rotary_factor: float = 0.25,
+        attention_bias: bool = False,
+        attention_dropout: float = 0.0,
+        attn_output_gate: bool = True,
+        full_attention_interval: int = 4,
+        layer_types: Optional[List[str]] = None,
+        linear_conv_kernel_dim: int = 4,
+        linear_key_head_dim: int = 128,
+        linear_num_key_heads: int = 16,
+        linear_num_value_heads: int = 32,
+        linear_value_head_dim: int = 128,
+        mtp_num_hidden_layers: int = 1,
+        mtp_use_dedicated_embeddings: bool = False,
+        tie_word_embeddings: bool = False,
+        initializer_range: float = 0.02,
+        use_cache: bool = True,
+        router_aux_loss_coef: float = 0.001,
+        output_router_logits: bool = False,
+        bos_token_id: int = 248044,
+        eos_token_id: int = 248044,
+        dtype: str = "bfloat16",
+        **kwargs,
+    ):
+        self.hidden_size = hidden_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.vocab_size = vocab_size
+        self.head_dim = head_dim
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.moe_intermediate_size = moe_intermediate_size
+        self.shared_expert_intermediate_size = shared_expert_intermediate_size
+        self.num_experts = num_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.max_position_embeddings = max_position_embeddings
+        self.rms_norm_eps = rms_norm_eps
+        self.rope_theta = rope_theta
+        self.partial_rotary_factor = partial_rotary_factor
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.attn_output_gate = attn_output_gate
+        self.full_attention_interval = full_attention_interval
+        self.layer_types = layer_types or list(_LAYER_PATTERN)
+        self.linear_conv_kernel_dim = linear_conv_kernel_dim
+        self.linear_key_head_dim = linear_key_head_dim
+        self.linear_num_key_heads = linear_num_key_heads
+        self.linear_num_value_heads = linear_num_value_heads
+        self.linear_value_head_dim = linear_value_head_dim
+        self.mtp_num_hidden_layers = mtp_num_hidden_layers
+        self.mtp_use_dedicated_embeddings = mtp_use_dedicated_embeddings
+        self.tie_word_embeddings = tie_word_embeddings
+        self.initializer_range = initializer_range
+        self.use_cache = use_cache
+        self.router_aux_loss_coef = router_aux_loss_coef
+        self.output_router_logits = output_router_logits
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.dtype = dtype
+
+    def to_dict(self) -> Dict[str, Any]:
+        return {k: v for k, v in self.__dict__.items() if not k.startswith("_")}
+
+
 class Qwen35MoeConfig(_BaseConfig):
     """Universal config shim.  Delegates everything to PretrainedConfig
     which reads all fields from config.json via ``**kwargs``."""