vllm.model_executor.models.qwen2 ¶

def qwen_2_model_invariants(
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    intermediate_tensors: IntermediateTensors | None = None,
    inputs_embeds: torch.Tensor | None = None,
):
    """Shape invariants for Qwen2Model Model, those are translated to
    runtime assertions for unbacked dynamic shapes and are compiled away for
    backed"""
    # All these should be equal.
    # input_ids.size()[0]
    # positions.size()[-1]
    # intermediate_tensors["hidden_states"].size()[0]
    # inputs_embeds.size()[0]
    torch._check(input_ids.size()[0] == positions.size()[-1])
    if intermediate_tensors is not None:
        torch._check(
            input_ids.size()[0] == intermediate_tensors["hidden_states"].size()[0]
        )

    if inputs_embeds is not None:
        torch._check(input_ids.size()[0] == inputs_embeds.size()[0])

    # Hidden dimensions should match (hidden_size)
    # intermediate_tensors["hidden_states"].size()[1]
    # inputs_embeds.size()[1]
    if inputs_embeds is not None and intermediate_tensors is not None:
        torch._check(
            inputs_embeds.size()[1] == intermediate_tensors["hidden_states"].size()[1]
        )