qwen3_vl_text¶
Qwen3-VL's language model on vLLM (vllm.model_executor.models.qwen3_vl), text-only prompts.
vLLM runs the whole Qwen3VLForConditionalGeneration, whose language model
is Qwen3LLMForCausalLM at language_model: the stack at
language_model.model.{embed_tokens, layers, norm}, built from vLLM's
Qwen3 classes, so the fused block ((hidden_states, residual)) and the
attention (queries and keys normed, then rotated) are qwen3's under that
prefix. The positions are interleaved M-RoPE's [3, tokens]; on a
text-only prompt the three rows are the same 1-D positions, so the rotation
is plain Qwen3's.
DeepStack. The model adds a deepstack feature to hidden_states after
each of the first blocks, outside the block, only when it is called with
inputs_embeds, which is vLLM's multimodal mode. As a language model there
is none, and layers[k+1].layer_input == layers[k].layer_output on every
block; deepstack_output is not mapped.
Text-only. The vision tower is not mapped and images are out of scope here;
run the engine as a language model (language_model_only=True, or every
limit_mm_per_prompt at 0). In its multimodal mode vLLM embeds every
prompt outside the model's forward and calls it with inputs_embeds and
no ids, so input_ids and token_embeddings are not there to read.
module-attribute
¶
RENAME = {'language_model.model.embed_tokens': 'embed_tokens', 'language_model.model.layers': 'layers', 'language_model.model.norm': 'norm', 'language_model.lm_head': 'lm_head', 'language_model.logits_processor': 'logits_processor'}
module-attribute
¶
ENVOYS = {Qwen3DecoderLayer: Layer, Qwen3Attention: Attention, Qwen3MLP: Mlp}
Bases: FusedLayer
Qwen3's decoder block: forward(positions, hidden_states, residual) -> (hidden_states, residual), with M-RoPE positions.