Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
#version 450
// Gated DeltaNet's causal depthwise convolution (vk_chain.c, VKC_P_DNCONV): one thread
// per channel c walks the S rows in order with the channel's ring of the last ck-1
// inputs (which lives on the device across steps):
// acc = sum_k w[c][k] * ring[k] + w[c][ck-1] * in[s][c] (order 0: the taps first)
// acc = w[c][ck-1] * in[s][c] + sum_k w[c][k] * ring[k] (order 1: the input first)
// out[s][c] = silu(acc); the ring drops its oldest input and takes in[s][c].
// The two orders are the two engines' (qwen36.c, qwen38_core.h), so the sum rounds as
// on the CPU. snap_row >= 0: the ring as it stands after that row is copied to snap
// (the state an MTP verify's first row leaves, for the rollback of its draft).
layout(local_size_x = 64) in;
layout(std430, binding = 0) readonly buffer IN { float xin[]; };
layout(std430, binding = 1) readonly buffer WB { float w[]; };
layout(std430, binding = 2) buffer RB { float ring[]; };
layout(std430, binding = 3) writeonly buffer OB { float o[]; };
layout(std430, binding = 4) writeonly buffer SB { float snap[]; };
layout(push_constant) uniform P {
int S, CD, CK;
int in_off, in_row, out_off, out_row;
int snap_row, order;
int w_off, ring_off, snap_off;
} p;
void main() {
int c = int(gl_GlobalInvocationID.x + gl_GlobalInvocationID.y * gl_NumWorkGroups.x * 64u);
if (c >= p.CD) return;
int nh = p.CK - 1;
float hist[8];
for (int k = 0; k < 8; k++) hist[k] = k < nh ? ring[p.ring_off + c * nh + k] : 0.0;
int wb = p.w_off + c * p.CK;
float wl = w[wb + p.CK - 1];
for (int s = 0; s < p.S; s++) {
float cur = xin[p.in_off + s * p.in_row + c];
float acc;
/* constant trip counts: the ring stays in registers */
if (p.order == 0) {
acc = 0.0;
for (int k = 0; k < 8; k++) if (k < nh) acc += w[wb + k] * hist[k];
acc += wl * cur;
} else {
acc = wl * cur;
for (int k = 0; k < 8; k++) if (k < nh) acc += w[wb + k] * hist[k];
}
o[p.out_off + s * p.out_row + c] = acc / (1.0 + exp(-acc));
for (int k = 0; k < 7; k++) if (k + 1 < nh) hist[k] = hist[k + 1];
for (int k = 0; k < 8; k++) if (k == nh - 1) hist[k] = cur;
if (s == p.snap_row)
for (int k = 0; k < 8; k++) if (k < nh) snap[p.snap_off + c * nh + k] = hist[k];
}
for (int k = 0; k < 8; k++) if (k < nh) ring[p.ring_off + c * nh + k] = hist[k];
}