vllm.model_executor.kernels.linear.mixed_precision.rdna3_w4a16 ¶

W4A16 GPTQ kernel for AMD RDNA3 (gfx1100) — fp16 + bf16.

Drop-in replacement for ExllamaLinearKernel on RDNA3 that adds native bf16 support. The HIP kernel lives in csrc/rocm/q_gemm_rdna3.cu and is exposed via torch.ops._rocm_C.gptq_gemm_rdna3.

Registered ahead of TritonW4A16LinearKernel for the ROCm-RDNA3 path; falls through to the Triton kernel on non-RDNA3 ROCm devices (e.g. CDNA/MI300).