We introduce AMDKernelVault, an open corpus and training framework for optimizing HIP and Triton kernels on recent AMD CDNA GPUs. Existing LLM-based kernel agents are predominantly designed for CUDA and NVIDIA hardware and often rely on repeated calls to frontier models for generation, reflection, and optimization. To address this gap, we develop HIPKernelGen and TritonKernelGen, agent-driven pipelines that translate PyTorch references into GPU kernels, compile and validate candidates under ROCm, and profile their latency on AMD hardware. AMDKernelVault contains 62,153 execution-verified HIP kernel samples, 2,377 production-grounded ROCm Libraries QA entries, and 39,893 Triton kernels. To demonstrate the corpus’s utility, we specialize Qwen3-8B through supervised fine-tuning and execution-aware reinforcement learning. Under fixed evaluation budgets and identical agent settings, the resulting model achieves the highest correctness among the evaluated models on PyTorch-to-HIP (34.0% Pass@1), TritonBench-G (33.2% Corr@3), and ROCmBench (41.94% Corr@3), while running locally on AMD hardware. These advantages are specific to the reported correctness metrics and do not extend uniformly to compilation accuracy or execution speed. The corpus and documentation are publicly available at https://huggingface.co/datasets/amd/AIG-Datasets.

Large Language Models (LLMs) inherently use autoregressive decoding, which lacks parallelism in inference and results in significantly slow inference speed. While methods such as Medusa constructs parallelized heads, they lack adequate information interaction across different prediction positions. To overcome this limitation, we introduce Amphista, an enhanced speculative decoding framework that builds upon Medusa. Specifically, Amphista models an Auto-embedding Block capable of parallel inference, incorporating bi-directional attention to enable interaction between different drafting heads. Additionally, Amphista integrates Staged Adaptation Layers, which ensure a seamless transition of semantic information from the target model’s autoregressive inference to the drafting heads’ non-autoregressive inference, effectively achieving paradigm shift and feature fusion. Experimental results on Vicuna models using MT-Bench and Spec-Bench demonstrate that Amphista achieves substantial acceleration while maintaining generation quality. On MT-Bench, Amphista delivers up to 2.75× speedup over vanilla autoregressive decoding and 1.40× over Medusa on Vicuna 33B in wall-clock time.