Fix core dump issue for binary built for old GPUs (cm<89)

[jasonchen31]

Fixing #227

PR: Gate SageAttention2 for pre-sm_89 CUDA builds and expose --cuda-arch in the Linux build script

Problem

Two bugs broke CUDA builds targeting old GPUs (compute capability 6.1, e.g. Pascal GTX 10xx):

1. SageAttention2 is unconditionally compiled and can't build for old GPUs.
   The vendored ggml CUDA backend globs sage-attn2.cu into every build. Its
   headers include <cuda/barrier> / <cuda/pipeline> unconditionally, which
   CUDA hard-errors on sub-sm_70 targets:

   #error "CUDA synchronization primitives are only supported for sm_70 and up."
   

2. scripts/build_linux.sh --backend cuda never sets the target GPU architecture.
   It silently inherits CMAKE_CUDA_ARCHITECTURES=native (the build machine's
   GPU). Building on any modern GPU produced a single-arch image with no sm_61
   code, and left the Sage gate "ON" (arch >= 89), so kernels with SM89+
   instructions were packaged into a binary that core-dumped when run on a 6.1 GPU.

A direct cmake -S . -B build -DENGINE_ENABLE_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=61
worked because it pins the arch (and 61 < 89 disables Sage at compile time).

Changes

external/ggml/src/ggml-cuda/CMakeLists.txt (build gate)

Add a fail-closed build-time gate for SageAttention2:

• Default ON, but turned OFF if any CMAKE_CUDA_ARCHITECTURES entry
  parses to < 89 (or is unparseable, e.g. native).
• When OFF, sage-attn2.cu is removed from the source list and a
  GGML_CUDA_SAGE_ATTN2_ENABLED compile definition is not added; a
  message(STATUS ...) reports CUDA SageAttention2 disabled.

external/ggml/src/ggml-cuda/sage-attn2.cuh (stub definitions)

Because ggml-cuda.cu unconditionally references the four Sage entry points
(ggml_cuda_sage_attn2, _i8, _supported, _i8_supported), the header now:

• keeps the real declarations when GGML_CUDA_SAGE_ATTN2_ENABLED is defined
  (arch >= 89), and
• provides static inline stubs when it is not: supported() returns false
  (op never attached to a graph) and the compute entry points GGML_ABORT.
  Since ggml-cuda.cu already includes this header, no dispatcher edits are
  needed.

scripts/build_linux.sh (target architecture control)

• New --cuda-arch <spec> flag that forwards -DCMAKE_CUDA_ARCHITECTURES=<spec>
  (with -UCMAKE_CUDA_ARCHITECTURES to avoid the sticky cache entry, mirroring
  the existing HIP --gpu-targets handling).
• Status output prints the effective CUDA architectures (or
  <auto: machine-native ...> when unset).

Docs

• docs/build/linux.md: document --cuda-arch, including the
  old-GPU example and the equivalence with the working direct CMake invocation.
  Noted that ENGINE_ENABLE_NATIVE_CPU=OFF is not required, since
  GGML_NATIVE affects CUDA arch only as a fallback when
  CMAKE_CUDA_ARCHITECTURES is undefined.

Why this is safe for correctness

• Disabling Sage does not leave MiniMax-H3 (the only Sage consumer) without
  an attention path: src/community_models/minimax_h3/dit_denoiser.cpp already
  falls back to ggml_flash_attn_ext when the Sage op is unsupported.
• Flash attention has a real Pascal (cc 610) kernel path
  (BEST_FATTN_KERNEL_TILE in external/ggml/src/ggml-cuda/fattn.cu), which is
  compiled into the sm_61 build.
• Sage physically cannot run below SM89 anyway; the runtime supported()
  guard already returns false for cc < 890, so a 6.1 GPU would have used flash
  attention regardless. The real fix is compiling the whole backend for sm_61.

Usage

# Old GPU (compute capability 6.1)
./scripts/build_linux.sh --backend cuda --cuda-arch 61 --target audiocpp_cli --target audiocpp_server

Validation status

• git diff for CMakeLists.txt is a pure additive change (22 insertions /
  0 deletions); CRLF line endings of the vendored file are preserved.
• Verified locally via a cmake -P-equivalent simulation that the gate resolves
  correctly: 61/52/mixed-low/native -> Sage OFF; 89-real/120a-real -> ON.

[jasonchen31]

I compiled again while fixing the SensVoice ASR Webui missing issue, Sage Attention was somehow included and the build failed again. Still working on it.

[jasonchen31]

It should be fixed this time. clean build success, and GPU inference correctly.