Repository navigation
metal : use more command buffers on iOS - #427
Merged
Merged
Conversation
iOS can stop a GPU command buffer that runs for more than a few seconds. With n_cb = 1, about 90% of each graph goes into a single command buffer, so a long prompt batch on a slow device can hit that limit. The buffer is then discarded with kIOGPUCommandBufferCallbackErrorInnocentVictim and llama_decode returns -3. Default to 8 command buffers on iOS. macOS is unchanged.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Default the Metal backend to 8 command buffers on iOS instead of 1. macOS is unchanged.
Why
iOS can stop a GPU command buffer that runs for more than a few seconds. With
n_cb = 1, the graph is split into a small first buffer (about 10% of the nodes) and one buffer with everything else. On a long prompt withn_ubatch = 512, that second buffer can run for 4 to 6 seconds on an A17 Pro. When iOS stops it, the buffer is discarded withkIOGPUCommandBufferCallbackErrorInnocentVictim(status 5), the backend goes into its error state, andllama_decodereturns -3. Loweringn_ubatchavoids it, at a large cost in prefill speed.How
ggml_backend_metal_set_n_cbis now called withGGML_METAL_DEFAULT_N_CB, which is 8 whenTARGET_OS_IPHONEand 1 otherwise. 8 is the existingGGML_METAL_MAX_COMMAND_BUFFERScap. At startup on iOS you will see the existing "n_cb > 2 is not recommended" warning; it is expected here.Testing
iPhone 15 Pro Max, a small quantized model, 16,553-token prompt, n_ctx 33,280, n_batch and n_ubatch 512, flash attention on, F16 K/V, all layers on GPU. Same upstream base for all runs; per-buffer GPU time logged from
GPUStartTime/GPUEndTime.These are one or two runs per setup, and the phone's thermal state varied between them, so the prefill numbers are noisy. Over the same token range of the first pair, n_cb = 8 was not slower (139 vs 134 tok/s).
The device runs used upstream llama.cpp from 2026-10-15 with this same change applied, not this branch. On this branch,
ggml-metal.cpppasses a syntax-only compile for iOS and macOS, and the define resolves to 8 and 1 respectively. Not built or run end to end on this branch.