qwen3

qwen3

Qwen3 on vLLM (vllm.model_executor.models.qwen3).

Llama's names and Llama's fused block. The attention is its own class (it norms the queries and keys); the model and the MLP are Qwen2's, which vLLM reuses.

RENAME module-attribute

RENAME = {'model.embed_tokens': 'embed_tokens', 'model.layers': 'layers', 'model.norm': 'norm'}

ENVOYS module-attribute

ENVOYS = {Qwen3DecoderLayer: Layer, Qwen3Attention: Attention, Qwen3MLP: Mlp}

Layer

Layer(*args: Any, **kwargs: Any)

Bases: FusedLayer

Qwen3's decoder block: forward(positions, hidden_states, residual) -> (hidden_states, residual).