Function metal_graph_debug_dump_tensor

ds4.c:10837–10864

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 xq_blocks;
    uint64_t gate_row_bytes[DS4_MAX_EXPERT];
    uint64_t up_row_bytes[DS4_MAX_EXPERT];
} matvec_q8_k_batch_mid_ctx;

static void matvec_iq2_xxs_batch_mid_worker(void *vctx, uint64_t task0, uint64_t task1) {
    matvec_iq2_xxs_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];

        const block_iq2_xxs *gate_row = (const block_iq2_xxs *)(ctx->gate_base[expert] + row * ctx->gate_row_bytes[expert]);
        const block_iq2_xxs *up_row = (const block_iq2_xxs *)(ctx->up_base[expert] + row * ctx->up_row_bytes[expert]);

        for (uint32_t i = begin; i < end; i++) {