From bb3676d2c8c869bb4c903b43b086e639185bd692 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 3 Jun 2026 03:16:06 -0700
Subject: [PATCH] llama.cpp CUDA detection: handle dlopen-ed backend (split
 build layout)

Current llama.cpp ships the CUDA backend as a dynamically-loaded plugin
(libggml-cuda.so* next to the binary), NOT a load-time dependency, so
ldd llama-server | grep libggml-cuda is a false negative: it reports no
CUDA on a perfectly good CUDA build. That made both is_cuda_server()
(provision_llama_cuda.sh) and _have_cuda_llama_server() (setup.sh) force a
needless full rebuild every run.

Fix both: keep the ldd check (old monolithic builds) and additionally treat
the presence of libggml-cuda.so* beside the binary as the CUDA signal. A
CPU-only build has no such backend, so this stays correct for the CPU case.

Verified on an N1X/sm_121 WSL build: llama-server --list-devices shows
CUDA0 JMJWOA-Generic-GPU and serves on the GPU, while ldd lists no
libggml-cuda; the new check correctly returns CUDA-present.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---
 studio/scripts/provision_llama_cuda.sh | 14 +++++++++++++-
 studio/setup.sh                        | 10 +++++++++-
 2 files changed, 22 insertions(+), 2 deletions(-)

diff --git a/studio/scripts/provision_llama_cuda.sh b/studio/scripts/provision_llama_cuda.sh
index 694979ab46..3f20527288 100644
--- a/studio/scripts/provision_llama_cuda.sh
+++ b/studio/scripts/provision_llama_cuda.sh
@@ -21,7 +21,19 @@ LLAMA_DIR="${UNSLOTH_LLAMA_CPP_PATH:-$HOME/.unsloth/llama.cpp}"
 SERVER="$LLAMA_DIR/build/bin/llama-server"
 log() { printf '  - %s\n' "$*"; }
 
-is_cuda_server() { [ -x "$1" ] && ldd "$1" 2>/dev/null | grep -qi 'libggml-cuda'; }
+# A llama-server is CUDA-capable in either of two build layouts:
+#   * old monolithic build -> libggml-cuda is a direct load-time dependency (ldd shows it)
+#   * current split build   -> CUDA ships as a dlopen-ed backend plugin, libggml-cuda.so*,
+#                              sitting next to the binary; ldd will NOT list it
+# Checking only ldd (the old behaviour) is a false negative on current llama.cpp and would
+# force a pointless full rebuild every run. A CPU-only build has no libggml-cuda.so at all,
+# so the presence of that backend beside the binary is the reliable CUDA signal.
+is_cuda_server() {
+    [ -x "$1" ] || return 1
+    ldd "$1" 2>/dev/null | grep -qi 'libggml-cuda' && return 0
+    for _so in "$(dirname "$1")"/libggml-cuda.so*; do [ -e "$_so" ] && return 0; done
+    return 1
+}
 
 # 0. Already provisioned?
 if is_cuda_server "$SERVER"; then
diff --git a/studio/setup.sh b/studio/setup.sh
index 5aa7f0497a..3e17444244 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -1410,8 +1410,16 @@ fi  # end _SKIP_GGUF_BUILD check
 # already built a CUDA server are byte-for-byte unaffected. Opt out with
 # UNSLOTH_NO_LLAMA_CUDA=1. Best-effort: never aborts setup (provision script always
 # exits 0; failures leave the prior CPU/degraded state for the fallback below).
+# CUDA-capable in either build layout: old monolithic (libggml-cuda is a direct
+# ldd dependency) or current split build (CUDA is a dlopen-ed backend, libggml-cuda.so*,
+# beside the binary -- ldd will NOT list it). Checking only ldd is a false negative on
+# current llama.cpp and would force a needless rebuild; a CPU-only build has no
+# libggml-cuda.so at all, so its presence beside the binary is the reliable signal.
 _have_cuda_llama_server() {
-    [ -x "$LLAMA_SERVER_BIN" ] && ldd "$LLAMA_SERVER_BIN" 2>/dev/null | grep -qi 'libggml-cuda'
+    [ -x "$LLAMA_SERVER_BIN" ] || return 1
+    ldd "$LLAMA_SERVER_BIN" 2>/dev/null | grep -qi 'libggml-cuda' && return 0
+    for _so in "$(dirname "$LLAMA_SERVER_BIN")"/libggml-cuda.so*; do [ -e "$_so" ] && return 0; done
+    return 1
 }
 if [ "$_HOST_SYSTEM" = "Linux" ] \
         && { [ "$_HOST_MACHINE" = "aarch64" ] || [ "$_HOST_MACHINE" = "arm64" ]; } \