v1.3

ansible/roles/llama_cpp/tasks/main.yml

@@ -12,6 +12,22 @@
   failed_when: false
   when: ansible_system == "Linux"
 
+- name: Check for CUDA compiler in standard install locations
+  shell: |
+    for candidate in /usr/local/cuda/bin/nvcc /usr/local/cuda-*/bin/nvcc; do
+      if [ -x "$candidate" ]; then
+        printf '%s\n' "$candidate"
+        exit 0
+      fi
+    done
+    exit 1
+  args:
+    executable: /bin/bash
+  register: courseware_llama_nvcc_fallback
+  changed_when: false
+  failed_when: false
+  when: ansible_system == "Linux"
+
 - name: Check for CUDA runtime header on Linux
   stat:
     path: "{{ item }}"
@@ -48,24 +64,63 @@
         ls /usr/local/cuda/include/cuda_runtime.h
   when:
     - ansible_system == "Linux"
-    - courseware_llama_nvcc.rc != 0 or (courseware_llama_cuda_headers.results | selectattr('stat.exists', 'equalto', true) | list | length == 0)
+    - courseware_llama_nvcc.rc != 0
+    - courseware_llama_nvcc_fallback.rc != 0
+    - (courseware_llama_cuda_headers.results | selectattr('stat.exists', 'equalto', true) | list | length == 0)
+
+- name: Set CUDA compiler path for llama.cpp builds
+  set_fact:
+    courseware_llama_cuda_compiler: "{{ (courseware_llama_nvcc.stdout | trim) if courseware_llama_nvcc.rc == 0 else (courseware_llama_nvcc_fallback.stdout | trim) }}"
+    courseware_llama_cuda_bin_dir: "{{ ((courseware_llama_nvcc.stdout | trim) if courseware_llama_nvcc.rc == 0 else (courseware_llama_nvcc_fallback.stdout | trim)) | dirname }}"
+  when:
+    - ansible_system == "Linux"
+    - courseware_llama_nvcc.rc == 0 or courseware_llama_nvcc_fallback.rc == 0
 
 - name: Set llama.cpp backend flag
   set_fact:
     courseware_llama_backend_flag: "{{ '-DGGML_METAL=ON' if ansible_system == 'Darwin' else '-DGGML_CUDA=ON' }}"
 
+- name: Set llama.cpp build parallelism
+  set_fact:
+    courseware_llama_build_jobs: >-
+      {{
+        [
+          1,
+          [
+            ansible_processor_vcpus | default(ansible_processor_nproc | default(1)) | int,
+            ((ansible_memtotal_mb | int) // 8192) | int,
+            4
+          ] | min
+        ] | max
+      }}
+
+- name: Build llama.cpp CMake configure arguments
+  set_fact:
+    courseware_llama_cmake_configure_argv: >-
+      {{
+        [
+          'cmake',
+          '-S',
+          courseware_repos_dir ~ '/llama.cpp',
+          '-B',
+          courseware_repos_dir ~ '/llama.cpp/build',
+          '-DCMAKE_BUILD_TYPE=Release',
+          courseware_llama_backend_flag
+        ]
+        +
+        (
+          ['-DCMAKE_CUDA_COMPILER=' ~ courseware_llama_cuda_compiler]
+          if ansible_system == 'Linux' and (courseware_llama_cuda_compiler | default('') | length > 0)
+          else []
+        )
+      }}
+
 - name: Configure llama.cpp
   command:
-    argv:
-      - cmake
-      - -S
-      - "{{ courseware_repos_dir }}/llama.cpp"
-      - -B
-      - "{{ courseware_repos_dir }}/llama.cpp/build"
-      - -DCMAKE_BUILD_TYPE=Release
-      - "{{ courseware_llama_backend_flag }}"
-  args:
-    creates: "{{ courseware_repos_dir }}/llama.cpp/build/CMakeCache.txt"
+    argv: "{{ courseware_llama_cmake_configure_argv }}"
+  environment:
+    PATH: "{{ (courseware_llama_cuda_bin_dir ~ ':' ~ ansible_env.PATH) if ansible_system == 'Linux' and (courseware_llama_cuda_bin_dir | default('') | length > 0) else ansible_env.PATH }}"
+    CUDACXX: "{{ courseware_llama_cuda_compiler | default(omit) }}"
 
 - name: Build llama.cpp tools
   command:
@@ -73,12 +128,16 @@
       - cmake
       - --build
       - "{{ courseware_repos_dir }}/llama.cpp/build"
+      - --parallel
+      - "{{ courseware_llama_build_jobs }}"
       - --target
       - llama-cli
       - llama-quantize
       - llama-perplexity
       - llama-server
-      - -j
+  environment:
+    PATH: "{{ (courseware_llama_cuda_bin_dir ~ ':' ~ ansible_env.PATH) if ansible_system == 'Linux' and (courseware_llama_cuda_bin_dir | default('') | length > 0) else ansible_env.PATH }}"
+    CUDACXX: "{{ courseware_llama_cuda_compiler | default(omit) }}"
 
 - name: Check system PATH slots for llama.cpp tools
   stat: