vLLM corrected three weight loading and shuffling bugs in the AITER MXFP4 MoE backend that previously caused garbled outputs for DeepSeek-V4 on AMD gfx950 GPUs, restoring numerical correctness and unlocking significant latency and throughput gains.
Corrects tensor-parallel weight offset calculations, removes an incorrect de-interleave step for standard checkpoints, and aligns weight shuffling functions with AITER FlyDSL kernel expectations, enabling stable MXFP4 quantized MoE inference.
Operators running DeepSeek-V4 on AMD MI300-series GPUs can now deploy MXFP4-quantized models without encountering garbled text, reducing time-to-first-token by 46% and increasing request throughput by 34%. The fix resolves tensor-parallel offset miscalculations, checkpoint layout mismatches, and incorrect kernel shuffling routines that previously broke the AITER MoE backend. Teams should monitor GSM8K accuracy stability across different tensor-parallel sizes and watch for potential regression when switching between triton_unfused and aiter backends on other MoE architectures.