Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
// src/prefill/moe_f16_kq.hpp - gfx11: the K-quant prompt experts on the F16 matrix cores (moe_f16_kq.cu; the switch,
// STRATA_PF_F16=1 with STRATA_PF_FUSED_KQ=1, is moe_f16_common.cuh's).
// Internal to moe_fused_iq.cu, which owns the native experts' tables and launches.
#pragma once
#include <cstdint>
#include <cuda_runtime.h>
#include "strata/prefill/moe_fused.hpp"
#include "strata/prefill/moe_fused_iq.hpp"
namespace strata::prefill::fused::f16kq {
/// The gate/up type has a kernel here (and this build has gfx11 device code).
bool covers(int gu_type);
/// Gate/up, SwiGLU and H for experts [b.e0, b.e1): the same tiles, activations and H as native_w11_kernel<WT, true>.
/// `ts`, `tiles`, `off` are fused::group's tables (at most `work_bound` tiles); `n` the layer's routed rows over
/// `n_expert` experts; `act` the int8 activations (quantize_act_native); `out` H.
void gate_up(const Batch& b, const NativeGeom& g, const int32_t* ts, const int2* tiles, const int32_t* off, int64_t n,
int n_expert, int64_t work_bound, const uint8_t* act, const int32_t* src, uint8_t* out, cudaStream_t s);
/// The down type has a kernel here (Q5_1).
bool covers_down(int d_type);
/// Down for the same experts and tiles: H (int8, `h`) to the routed rows' FP32 outputs in `dm`.
void down(const Batch& b, const NativeGeom& g, const int32_t* ts, const int2* tiles, const int32_t* off, int64_t n, int n_expert,
int64_t work_bound, const uint8_t* h, float* dm, cudaStream_t s);
} // namespace strata::prefill::fused::f16kq