mirror of
https://github.com/ollama/ollama.git
synced 2025-11-11 15:06:49 +01:00
d5a0d8d904
This changes the memory allocation strategy from upfront estimation to tracking actual allocations done by the engine and reacting to that. The goal is avoid issues caused by both under-estimation (crashing) and over-estimation (low performance due to under-utilized GPUs). It is currently opt-in and can be enabled for models running on the Ollama engine by setting OLLAMA_NEW_ESTIMATES=1. Behavior in other cases is unchanged and will continue to use the existing estimates.
28 lines
1003 B
Diff
28 lines
1003 B
Diff
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
|
|
From: Daniel Hiltgen <daniel@ollama.com>
|
|
Date: Wed, 30 Jul 2025 08:43:46 -0700
|
|
Subject: [PATCH] BF16 macos version guard
|
|
|
|
Only enable BF16 on supported MacOS versions (v14+)
|
|
---
|
|
ggml/src/ggml-metal/ggml-metal.m | 6 +++++-
|
|
1 file changed, 5 insertions(+), 1 deletion(-)
|
|
|
|
diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
|
|
index fe7b2f0a..e4c31268 100644
|
|
--- a/ggml/src/ggml-metal/ggml-metal.m
|
|
+++ b/ggml/src/ggml-metal/ggml-metal.m
|
|
@@ -106,7 +106,11 @@ static id<MTLDevice> ggml_backend_metal_device_acq(struct ggml_backend_metal_dev
|
|
ctx->has_bfloat |= [ctx->mtl_device supportsFamily:MTLGPUFamilyApple6];
|
|
|
|
#if defined(GGML_METAL_USE_BF16)
|
|
- ctx->use_bfloat = ctx->has_bfloat;
|
|
+ if (@available(macOS 14.0, *)) {
|
|
+ ctx->use_bfloat = ctx->has_bfloat;
|
|
+ } else {
|
|
+ ctx->use_bfloat = false;
|
|
+ }
|
|
#else
|
|
ctx->use_bfloat = false;
|
|
#endif
|