CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-12 10:47:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

metal : retain src and dst buffers during async ops (#17101)

Browse Source
This commit is contained in:
Georgi Gerganov
2025-11-09 08:28:51 +02:00
committed by GitHub
parent aa3b7a90b4
commit 0750a59903
1 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
ggml/src/ggml-metal/ggml-metal-context.m
View File

@@ -289,7 +289,7 @@ void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor,
// queue the copy operation into the queue of the Metal context
// this will be queued at the end, after any currently ongoing GPU operations
id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBufferWithUnretainedReferences];
id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBuffer];
id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
[encoder copyFromBuffer:buf_src
@@ -300,6 +300,7 @@ void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor,
[encoder endEncoding];
[cmd_buf commit];
[buf_src release];
// do not wait here for completion
//[cmd_buf waitUntilCompleted];
@@ -330,7 +331,7 @@ void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * te
// queue the copy operation into the queue of the Metal context
// this will be queued at the end, after any currently ongoing GPU operations
id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBufferWithUnretainedReferences];
id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBuffer];
id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
[encoder copyFromBuffer:bid_src.metal
@@ -341,6 +342,7 @@ void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * te
[encoder endEncoding];
[cmd_buf commit];
[buf_dst release];
// do not wait here for completion
//[cmd_buf waitUntilCompleted];