Misc. bug: Unsupported op "CPY" / SIGABRT on Apple CPU

[stephen-ebenezar]

Name and Version

version: 5171 (2cca6c0)
built with Apple clang version 16.0.0 (clang-1600.0.26.6) for arm64-apple-darwin24.4.0

Operating systems

Mac (M4 Pro / 48 GB)

Which llama.cpp modules do you know to be affected?

llama-cli

Command line

bin/llama-cli --model ./DeepSeek-R1-GGUF/DeepSeek-R1-UD-IQ1_S/DeepSeek-R1-UD-IQ1_S-00001-of-00003.gguf --cache-type-k q4_0 --threads 1 -no-cnv --prio 2 --temp 0.6 --ctx-size 8 --seed 3407 --prompt "<｜User｜>The first six words of a Tale of Two Cities<｜Assistant｜>"

Problem description & steps to reproduce

Please find the full log here:
deepseek_llama_cpp_sigabrt.log

Possibly related issues/PRs:
Closed issue #12715
Closed PR #11987

First Bad Commit

Unknown. Takes a while to run on CPU.

Relevant log output

Last few lines of uploaded log:

The first six words of a Tale of Two Cities<think>
Okay, so the user is asking about the first six words of "A Tale of/Users/stephenebenezar/dev/llama.cpp/ggml/src/ggml-blas/ggml-blas.cpp:250: ggml_backend_blas_graph_compute: unsupported op CPY
Process finished with exit code 134 (interrupted by signal 6:SIGABRT)

[slaren]

This happens during the K shift because the operation is being forced to run on the BLAS backend with ggml_backend_sched_set_tensor_backend:

llama.cpp/src/llama-context.cpp

Lines 496 to 502 in 558a764

	for (const auto & backend : backends) {
	// Figure out which backend KV cache belongs to
	if (ggml_backend_supports_buft(backend.get(), ggml_backend_buffer_get_type(bbuf))) {
	ggml_backend_sched_set_tensor_backend(sched.get(), tmp, backend.get());
	break;
	}
	}

@ggerganov do you remember why was this necessary?

[ggerganov]

@ggerganov do you remember why was this necessary?

Maybe this is related: #5653 (comment)