Function kv_cache_init

ds4.c:8484–8524

Allocate per-layer KV state: a raw sliding window for all layers, plus * compressed attention/indexer caches for layers whose ratio is nonzero. */

    matvec_iq2_xxs_accum_ctx ctx = {
        .out = out,
        .in_dim = in_dim0,
        .n_expert = n_expert,
    };
    for (int i = 0; i < n_expert; i++) {
        ctx.base[i] = base[i];
        ctx.row_bytes[i] = row_bytes[i];
        ctx.xq[i] = xq + (uint64_t)i * n_blocks;
    }

    ds4_parallel_for(out_dim0, matvec_iq2_xxs_accum_worker, &ctx);
}

typedef struct {
    uint32_t token;
    uint32_t slot;
} ds4_expert_pair;

typedef struct {
    float *mid;
    const uint8_t *gate_base[DS4_MAX_EXPERT];
    const uint8_t *up_base[DS4_MAX_EXPERT];
    const block_q8_K *xq;
    const ds4_expert_pair *pairs;
    const uint32_t *pair_ids;
    const uint32_t *expert_offset;
    const uint32_t *active_expert;
    const float *pair_weight;
    float clamp;
    uint64_t in_dim;
    uint64_t out_dim;
    uint64_t gate_row_bytes[DS4_MAX_EXPERT];
    uint64_t up_row_bytes[DS4_MAX_EXPERT];
    uint64_t xq_blocks;
} matvec_q2_k_batch_mid_ctx;

static void matvec_q2_k_batch_mid_worker(void *vctx, uint64_t task0, uint64_t task1) {
    matvec_q2_k_batch_mid_ctx *ctx = vctx;

    for (uint64_t task = task0; task < task1; task++) {
        const uint32_t active_idx = (uint32_t)(task / ctx->out_dim);
        const uint64_t row = task - (uint64_t)active_idx * ctx->out_dim;
        const uint32_t expert = ctx->active_expert[active_idx];
        const uint32_t begin = ctx->expert_offset[expert];
        const uint32_t end = ctx->expert_offset[expert + 1];