Function edgetam_repvit_block_forward

sam3.cpp:4732–4765

RepViT block forward: token_mixer DW → optional SE → channel_mixer + residual

// RepViT block forward: token_mixer DW → optional SE → channel_mixer + residual
static struct ggml_tensor* edgetam_repvit_block_forward(struct ggml_context* ctx,
                                                         struct ggml_tensor* x,
                                                         const edgetam_repvit_block& blk) {
    // x layout: [W, H, C, 1]
    // Token mixer: DW 3×3 conv, stride=1, pad=1 (RepVGG-reparameterized, no residual)
    // Use ggml_conv_2d_dw_direct which requires F32 kernel
    auto* tm_w_f32 = (blk.tm_w->type == GGML_TYPE_F32) ? blk.tm_w
                     : ggml_cast(ctx, blk.tm_w, GGML_TYPE_F32);
    x = ggml_conv_2d_dw_direct(ctx, tm_w_f32, x, 1, 1, 1, 1, 1, 1);
    x = edgetam_conv2d_bias(ctx, x, blk.tm_b);

    // Squeeze-and-excitation (optional, on even-indexed blocks)
    if (blk.has_se) {
        int W = (int)x->ne[0];
        int H = (int)x->ne[1];
        // Global average pool: [W, H, C, 1] → [1, 1, C, 1]
        auto* pooled = ggml_pool_2d(ctx, x, GGML_OP_POOL_AVG, W, H, W, H, 0, 0);
        // SE operates on tiny [1,1,C,1] — keep using ggml_conv_2d (overhead negligible)
        auto* se = ggml_conv_2d(ctx, blk.se_fc1_w, pooled, 1, 1, 0, 0, 1, 1);
        se = edgetam_conv2d_bias(ctx, se, blk.se_fc1_b);
        se = ggml_relu(ctx, se);
        se = ggml_conv_2d(ctx, blk.se_fc2_w, se, 1, 1, 0, 0, 1, 1);
        se = edgetam_conv2d_bias(ctx, se, blk.se_fc2_b);
        se = ggml_sigmoid(ctx, se);
        x = ggml_mul(ctx, x, se);
    }

    // Channel mixer: 1×1 expand → GELU → 1×1 project (use mul_mat path)
    auto* identity = x;
    auto* cm = edgetam_conv1x1_mulmat(ctx, x, blk.cm_conv1_w, blk.cm_conv1_b);
    cm = ggml_gelu(ctx, cm);
    cm = edgetam_conv1x1_mulmat(ctx, cm, blk.cm_conv2_w, blk.cm_conv2_b);
    return ggml_add(ctx, cm, identity);
}

// Downsample: pre_block → spatial DW s=2 → channel expand → FFN + residual
static struct ggml_tensor* edgetam_repvit_downsample_forward(struct ggml_context* ctx,