mirror of
https://github.com/hanzoai/engine.git
synced 2026-11-20 23:45:22 +00:00
1) Drop stale QMatMul::VulkanQuantBank refs (8 sites in hanzo-quant + hanzo-engine): the ml's unified-quant refactor folded the resident-MoE-bank into VulkanQuant, but the engine still matched the removed variant -> vulkan lib failed to compile. 2) Route Vulkan RoPE to rope_positions_via_gather (in apply_rotary_positions_qk + qk_rms_norm_rope_positions): Vulkan has no fused positions kernel in hanzo-quant, so it fell through to cpu_apply_rotary_qk which bails 'RoPE positions must be CPU u32'. Mirrors the existing ROCm handling. Validated: native Windows build (msvc + vulkan, precompiled .spv) runs Qwen3-0.6B on vulkan[0] (AMD 8060S), 28/28 layers on GPU, correct output, 22.3 tok/s decode (WSL2 GPU is env-blocked, ~1 tok/s CPU fallback).