vllm.models.qwen4_exp.amd

Modules:

• hyperconnection –

  HyperConnection (Gated Residual) utilities for the AMD model variant.

• indexer_qsa –

  Qwen4Exp weight-free QSA indexer.

• low_latency_gemm –

  Qwen4Exp low-latency GEMM hook for AMD ROCm.

• model –

  Inference-only Qwen4Exp model.

• model_state –

  Model-runner state for Qwen4Exp PLE inputs.

• mtp –

  Inference-only Qwen4Exp MTP (Multi-Token Predictor) model.

• ops –

  AMD ROCm Qwen4Exp kernels; import leaf modules directly.

• ple_layer –

  GPU-resident Qwen4Exp position-learning enhancement layers.

• qsa –

  AMD ROCm QSA owner with Triton kernels.