checkpoint

Author: LeaveNhA

ggml/src/ggml-metal/ggml-metal.m

@@ -5967,12 +5967,17 @@ static enum ggml_status ggml_backend_metal_split_buffer_init_tensor(ggml_backend
         return GGML_STATUS_ALLOC_FAILED;
     }
 
-    // For a dynamic array, we need to manually manage the array
-    ctx->tensor_extras = realloc(ctx->tensor_extras, (ctx->tensor_extras_size + 1) * sizeof(struct ggml_tensor_extra_metal *));
-    if (ctx->tensor_extras == NULL) {
-        GGML_LOG_ERROR("%s: failed to reallocate tensor_extras array\n", __func__);
-        free(extra);
-        return GGML_STATUS_ALLOC_FAILED;
+    // For a dynamic array, we need to manually manage the array with proper capacity tracking
+    if (ctx->tensor_extras_size >= ctx->tensor_extras_capacity) {
+        size_t new_capacity = ctx->tensor_extras_capacity == 0 ? 16 : ctx->tensor_extras_capacity * 2;
+        struct ggml_tensor_extra_metal ** new_tensor_extras = realloc(ctx->tensor_extras, new_capacity * sizeof(struct ggml_tensor_extra_metal *));
+        if (new_tensor_extras == NULL) {
+            GGML_LOG_ERROR("%s: failed to reallocate tensor_extras array\n", __func__);
+            free(extra);
+            return GGML_STATUS_ALLOC_FAILED;
+        }
+        ctx->tensor_extras = new_tensor_extras;
+        ctx->tensor_extras_capacity = new_capacity;
     }
     ctx->tensor_extras[ctx->tensor_extras_size] = extra;
     ctx->tensor_extras_size++;
@@ -6023,15 +6028,17 @@ static enum ggml_status ggml_backend_metal_split_buffer_init_tensor(ggml_backend
     GGML_LOG_DEBUG("%s: tensor '%s' initializing buffer with zeros\n", __func__, tensor->name);
     void * bufferContents = [extra->data_device[id] contents];
     if (bufferContents == NULL) {
-        GGML_LOG_ERROR("%s: Metal buffer contents is NULL for tensor '%s'\n", __func__, tensor->name);
-        [extra->data_device[id] release];
-        free(extra);
-        return GGML_STATUS_ALLOC_FAILED;
+        // If we can't access the buffer contents directly, we'll skip initialization
+        // Buffers with StorageModePrivate are typically zero-initialized by Metal
+        // or will be initialized when first used
+        GGML_LOG_DEBUG("%s: Metal buffer contents is NULL for tensor '%s', skipping zero initialization\n", __func__, tensor->name);
+    } else {
+        // We can access the buffer contents directly, so initialize with zeros
+        memset(bufferContents, 0, size);
     }
-    memset(bufferContents, 0, size);
 
     tensor->extra = extra;
-    GGML_LOG_DEBUG("%s: tensor '%s' initialization completed\n", __func__, tensor->name);
+    GGML_LOG_DEBUG("%s: tensor '%s' initialization completed successfully\n", __func__, tensor->name);
     return GGML_STATUS_SUCCESS;
 }