vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

metal : print more verbose device info + handle errors

Browse source
This commit is contained in:
Georgi Gerganov 2023-06-16 18:53:18 +03:00
parent 12156fb4a8
commit 71443d757e
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
1 changed files with 31 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

34
ggml-metal.m
View file

@ -177,6 +177,14 @@ struct ggml_metal_context * ggml_metal_init(void) {
#undef GGML_METAL_ADD_KERNEL #undef GGML_METAL_ADD_KERNEL
} }
fprintf(stderr, "%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
fprintf(stderr, "%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
if (ctx->device.maxTransferRate != 0) {
fprintf(stderr, "%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
} else {
fprintf(stderr, "%s: maxTransferRate = built-in GPU\n", __func__);
}
return ctx; return ctx;
} }
@ -250,11 +258,11 @@ bool ggml_metal_add_buffer(
ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil]; ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
if (ctx->buffers[ctx->n_buffers].metal == nil) { if (ctx->buffers[ctx->n_buffers].metal == nil) {
fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0); fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
return false; return false;
} }
fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0); fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
++ctx->n_buffers; ++ctx->n_buffers;
} else { } else {
@ -281,11 +289,21 @@ bool ggml_metal_add_buffer(
return false; return false;
} }
fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld\n", __func__, name, size_step_aligned / 1024.0 / 1024.0, i); fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
++ctx->n_buffers; ++ctx->n_buffers;
} }
} }
fprintf(stderr, ", (%8.2f / %8.2f)",
ctx->device.currentAllocatedSize / 1024.0 / 1024.0,
ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) {
fprintf(stderr, ", warning: current allocated size is greater than the recommended max working set size\n");
} else {
fprintf(stderr, "\n");
}
} }
return true; return true;
@ -862,4 +880,14 @@ void ggml_metal_graph_compute(
dispatch_barrier_sync(queue, ^{}); dispatch_barrier_sync(queue, ^{});
[command_buffers[n_cb - 1] waitUntilCompleted]; [command_buffers[n_cb - 1] waitUntilCompleted];
// check status of command buffers
// needed to detect if the device ran out-of-memory for example (#1881)
for (int i = 0; i < n_cb; i++) {
MTLCommandBufferStatus status = (MTLCommandBufferStatus) [command_buffers[i] status];
if (status != MTLCommandBufferStatusCompleted) {
fprintf(stderr, "%s: command buffer %d failed with status %lu\n", __func__, i, status);
GGML_ASSERT(false);
}
}
} }