metal: encode window over the paged (boundary-scheduled) compute path

ggml/include/ggml-backend.h

@@ -328,13 +328,6 @@ extern "C" {
     GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
     GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
 
-    // Get the scheduler's current compute graph (the last one split/reserved).
-    // Its nodes are in execution order; each node's assigned backend is
-    // queryable via ggml_backend_sched_get_tensor_backend. Valid after
-    // ggml_backend_sched_reserve (which llama runs at context construction);
-    // before any graph is split the returned graph has n_nodes == 0. The
-    // scheduler owns the graph — do not free or mutate it. Never NULL for a
-    // valid sched. (titanium-chicken: memory/scheduler observability.)
     GGML_API struct ggml_cgraph * ggml_backend_sched_get_graph(ggml_backend_sched_t sched);
 
     GGML_API ggml_backend_buffer_type_t ggml_backend_sched_get_buffer_type(ggml_backend_sched_t sched, ggml_backend_t backend);

ggml/include/ggml-metal.h

@@ -56,6 +56,34 @@ GGML_BACKEND_API void ggml_backend_metal_capture_next_compute(ggml_backend_t bac
 
 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_metal_reg(void);
 
+typedef struct ggml_metal_event * ggml_metal_event_t;
+
+GGML_BACKEND_API ggml_metal_event_t ggml_backend_metal_event_init(ggml_backend_t backend);
+GGML_BACKEND_API void               ggml_backend_metal_event_free(ggml_backend_t backend, ggml_metal_event_t ev);
+
+GGML_BACKEND_API void     ggml_backend_metal_event_host_signal(ggml_metal_event_t ev, uint64_t value);
+GGML_BACKEND_API bool     ggml_backend_metal_event_host_wait  (ggml_metal_event_t ev, uint64_t value, uint64_t timeout_ms);
+GGML_BACKEND_API uint64_t ggml_backend_metal_event_host_value (ggml_metal_event_t ev);
+
+GGML_BACKEND_API void ggml_backend_metal_set_boundary_schedule(
+        ggml_backend_t backend,
+        int n_cuts,  struct ggml_tensor * const * cut_nodes,  ggml_metal_event_t * sig_ev,  const uint64_t * sig_val,
+        int n_waits, struct ggml_tensor * const * wait_nodes, ggml_metal_event_t * wait_ev, const uint64_t * wait_val);
+
+// Restrict the boundary-scheduled (paged) compute path to the segments
+// intersecting the inclusive node range [first_node, last_node], with an
+// optional device-side blit of a boundary activation into the window before
+// its first segment (in_src -> in_dst) and out of it after its last
+// (out_src -> out_dst). Persistent and pointer-keyed like the boundary
+// schedule; a graph containing neither endpoint encodes in full. Only
+// meaningful while a boundary schedule is set.
+GGML_BACKEND_API void ggml_backend_metal_set_encode_window(
+        ggml_backend_t backend,
+        struct ggml_tensor * first_node,  struct ggml_tensor * last_node,
+        struct ggml_tensor * blit_in_src, struct ggml_tensor * blit_in_dst,
+        struct ggml_tensor * blit_out_src, struct ggml_tensor * blit_out_dst);
+GGML_BACKEND_API void ggml_backend_metal_clear_encode_window(ggml_backend_t backend);
+
 #ifdef __cplusplus
 }
 #endif

ggml/src/ggml-metal/ggml-metal-context.h

@@ -31,6 +31,20 @@ void ggml_metal_event_wait  (ggml_metal_t ctx, ggml_metal_event_t ev);
 
 ggml_metal_event_t ggml_metal_get_ev_cpy(ggml_metal_t ctx);
 
+ggml_metal_device_t ggml_metal_get_device(ggml_metal_t ctx);
+
+void ggml_metal_set_boundary_schedule(
+        ggml_metal_t ctx,
+        int n_cuts,  struct ggml_tensor * const * cut_nodes,  ggml_metal_event_t * sig_ev, const uint64_t * sig_val,
+        int n_waits, struct ggml_tensor * const * wait_nodes, ggml_metal_event_t * wait_ev, const uint64_t * wait_val);
+
+void ggml_metal_set_encode_window(
+        ggml_metal_t ctx,
+        struct ggml_tensor * first_node,  struct ggml_tensor * last_node,
+        struct ggml_tensor * blit_in_src, struct ggml_tensor * blit_in_dst,
+        struct ggml_tensor * blit_out_src, struct ggml_tensor * blit_out_dst);
+void ggml_metal_clear_encode_window(ggml_metal_t ctx);
+
 void ggml_metal_set_n_cb            (ggml_metal_t ctx, int n_cb);
 void ggml_metal_set_abort_callback  (ggml_metal_t ctx, ggml_abort_callback abort_callback, void * user_data);
 bool ggml_metal_supports_family     (ggml_metal_t ctx, int family);