执行失败

[yeyueluke]

CLI=build/examples/cli/da3-cli
M=models/depth-anything-base-f32.gguf
$CLI depth --model $M --input mountains.jpg --pfm depth.pfm --png depth.png
非法指令 (核心已转储)

[yeyueluke]

#0 0x000064da7abc38c6 in da::winograd_conv3x3(ggml_context*, ggml_tensor*, ggml_tensor*, int) ()
(gdb) bt
#0 0x000064da7abc38c6 in da::winograd_conv3x3(ggml_context*, ggml_tensor*, ggml_tensor*, int) ()
#1 0x000064da7abc31b0 in da::conv2d(ggml_context*, ggml_tensor*, ggml_tensor*, ggml_tensor*, int, int) ()
#2 0x000064da7aba8199 in da::DptHead::build_depth_graph(ggml_context*, ggml_tensor* const*, int, int, da::GraphInputPool&, std::vector<float, std::allocator >, std::vector<float, std::allocator >, std::vector<float, std::allocator >, std::vector<float, std::allocator >) ()
#3 0x000064da7ab7a8be in std::_Function_handler<ggml_tensor* (ggml_context*), da::Engine::depth_native_fused(da::Image const&, std::vector<float, std::allocator >&, std::vector<float, std::allocator >&, int&, int&)::{lambda(ggml_context*)#1}>::_M_invoke(std::_Any_data const&, ggml_context*&&) ()
#4 0x000064da7ab34292 in da::Backend::compute(std::function<ggml_tensor* (ggml_context*)> const&, std::vector<float, std::allocator >&, unsigned long) ()
#5 0x000064da7ab7f5cd in da::Engine::depth_native_fused(da::Image const&, std::vector<float, std::allocator >&, std::vector<float, std::allocator >&, int&, int&) ()
#6 0x000064da7ab7fc5d in da::Engine::depth_native(std::__cxx11::basic_string<char, std::char_traits, std::allocator > const&, std::vector<float, std::allocator >&, std::vector<float, std::allocator >&, int&, int&) ()
#7 0x000064da7ab2d214 in main ()

[localai-org-maint-bot]

The backtrace points to a CPU instruction-set portability problem worth checking first. On current master (14f7461d1f704761a038ac9f50dbde8fdb7275e2), CMakeLists.txt:91-108 enables AVX-512 for all of src/winograd.cpp whenever the compiler accepts the flag. It does not check the executing CPU or perform runtime dispatch. Compiler support alone does not establish CPU support, and the compiler can emit these instructions in winograd_conv3x3 itself, before the compute callback runs. GGML_NATIVE=OFF does not disable this separate per-source flag.

Could you provide your CPU model and flags (lscpu), compiler/build command, commit (git rev-parse HEAD), and x/i $pc at the SIGILL in gdb? Those will distinguish an unsupported instruction from another cause. I have not reproduced the crash locally.

No open PR currently covers this path; #4 concerns Metal tensor residency. A portable fix needs to retain a baseline implementation and enable the AVX-512 implementation only when the CPU supports its required features (or make that build an explicit opt-in). The reported crash should be verified on a CPU without AVX-512 before calling it resolved.