perf(ds4): overlap attention across heterogeneous GPUs

server/deps/llama.cpp/ggml/include/ggml-alloc.h

@@ -46,11 +46,10 @@ GGML_API enum ggml_status    ggml_tallocr_alloc(struct ggml_tallocr * talloc, st
 typedef struct ggml_gallocr * ggml_gallocr_t;
 
 GGML_API ggml_gallocr_t ggml_gallocr_new(ggml_backend_buffer_type_t buft);
-// Uses max_chunk_size as the preferred backing-allocation limit while
-// preserving a single logical graph allocator. Individual tensors are never
-// split and may exceed the limit. This is useful on devices without virtual
-// memory support, where a large contiguous allocation can fail despite
-// sufficient aggregate free memory.
+// Limits each backing allocation while preserving a single logical graph
+// allocator. Tensors are never split across chunks. This is useful on devices
+// without virtual memory support, where a large contiguous allocation can fail
+// despite sufficient aggregate free memory.
 GGML_API ggml_gallocr_t ggml_gallocr_new_with_max_chunk_size(
     ggml_backend_buffer_type_t buft,
     size_t max_chunk_size);

server/deps/llama.cpp/ggml/include/ggml-backend.h

@@ -385,6 +385,14 @@ extern "C" {
                                             ggml_backend_sched_t sched,
                                             bool enabled);
 
+    // In a non-pipelined scheduler, protect reusable split-input buffers with
+    // stream events instead of blocking the host on every destination stream.
+    // The source waits until the destination has finished the prior generation;
+    // the existing async-copy contract then publishes the new generation.
+    GGML_API void                 ggml_backend_sched_set_single_copy_event_fences(
+                                            ggml_backend_sched_t sched,
+                                            bool enabled);
+
     //
     // Meta backend
     //

server/deps/llama.cpp/ggml/include/ggml.h

@@ -677,6 +677,10 @@ extern "C" {
         GGML_TENSOR_FLAG_PARAM   =  4, // ...contains trainable parameters
         GGML_TENSOR_FLAG_LOSS    =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
         GGML_TENSOR_FLAG_COMPUTE = 16, // ...must be computed
+        // ...is a persistent fork payload whose cross-device copy may run on
+        // the destination stream. Transient allocator storage must never set
+        // this flag because the producer may otherwise reuse it too early.
+        GGML_TENSOR_FLAG_DST_STREAM_COPY = 32,
     };
 
     enum ggml_tri_type {