fix(vae): order Sol attention and FFN writes to reduce VRAM

src/model/vae/minimax_h3_sol_attention.hpp

@@ -172,6 +172,7 @@ namespace MiniMaxH3VAE {
                                                                heads, nullptr, false, runner_ctx->flash_attn_enabled);
                         auto indices  = ggml_reshape_1d(ctx, batch.output_rows, ggml_nelements(batch.output_rows));
                         output        = ggml_set_rows(ctx, output, ggml_reshape_2d(ctx, attended, dim, indices->ne[0]), indices);
+                        runner_ctx->expand_graph(output);
                     }
                     // Later windows can only read this block and its immediate successor.
                     previous = current;

src/model/vae/minimax_h3_vae.hpp

@@ -430,6 +430,7 @@ namespace MiniMaxH3VAE {
                 auto input   = ggml_ext_slice(ctx, x, 1, row, std::min(row + rows_per_chunk, x->ne[1]), false);
                 auto indices = ggml_ext_slice(ctx, sol->row_ids, 0, row, std::min(row + rows_per_chunk, x->ne[1]), false);
                 output       = ggml_set_rows(ctx->ggml_ctx, output, finish(input), indices);
+                ctx->expand_graph(output);
             }
             return output;
         }