Add llama-run

container-images/asahi/Containerfile

@@ -1,11 +1,11 @@
 FROM fedora:41
 
-ARG LLAMA_CPP_SHA=1329c0a75e6a7defc5c380eaf80d8e0f66d7da78
+ARG LLAMA_CPP_SHA=0827b2c1da299805288abbd556d869318f2b121e
 # renovate: datasource=git-refs depName=ggerganov/whisper.cpp packageName=https://github.com/ggerganov/whisper.cpp gitRef=master versioning=loose type=digest
-ARG WHISPER_CPP_SHA=6266a9f9e56a5b925e9892acf650f3eb1245814d
+ARG WHISPER_CPP_SHA=3de9deead5759eb038966990e3cb5d83984ae467
 
 COPY ../scripts /scripts
 RUN chmod +x /scripts/*.sh && \
     /scripts/build_llama_and_whisper.sh "asahi" "$LLAMA_CPP_SHA" \
-      "$WHISPER_CPP_SHA" "/usr" "-DGGML_VULKAN=1"
+      "$WHISPER_CPP_SHA"
 

container-images/cuda/Containerfile

@@ -1,15 +1,14 @@
 # Base image with CUDA for compilation
 FROM docker.io/nvidia/cuda:12.6.2-devel-ubi9 AS builder
 
-ARG LLAMA_CPP_SHA=af148c9386da825a60c7038549c121c35ca56b50
+ARG LLAMA_CPP_SHA=0827b2c1da299805288abbd556d869318f2b121e
 # renovate: datasource=git-refs depName=ggerganov/whisper.cpp packageName=https://github.com/ggerganov/whisper.cpp gitRef=master versioning=loose type=digest
-ARG WHISPER_CPP_SHA=6266a9f9e56a5b925e9892acf650f3eb1245814d
+ARG WHISPER_CPP_SHA=3de9deead5759eb038966990e3cb5d83984ae467
 
 COPY ../scripts /scripts
 RUN chmod +x /scripts/*.sh && \
     /scripts/build_llama_and_whisper.sh "cuda" "$LLAMA_CPP_SHA" \
-      "$WHISPER_CPP_SHA" "/tmp/install" \
-      "-DGGML_CUDA=ON" "-DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined"
+      "$WHISPER_CPP_SHA"
 
 # Final runtime image
 FROM docker.io/nvidia/cuda:12.6.2-runtime-ubi9

container-images/ramalama/Containerfile

@@ -1,13 +1,13 @@
 FROM registry.access.redhat.com/ubi9/ubi:9.4-1214.1729773476
 
-ARG LLAMA_CPP_SHA=af148c9386da825a60c7038549c121c35ca56b50
+ARG LLAMA_CPP_SHA=0827b2c1da299805288abbd556d869318f2b121e
 # renovate: datasource=git-refs depName=ggerganov/whisper.cpp packageName=https://github.com/ggerganov/whisper.cpp gitRef=master versioning=loose type=digest
-ARG WHISPER_CPP_SHA=6266a9f9e56a5b925e9892acf650f3eb1245814d
+ARG WHISPER_CPP_SHA=3de9deead5759eb038966990e3cb5d83984ae467
 
 COPY ../scripts /scripts
 RUN chmod +x /scripts/*.sh && \
     /scripts/build_llama_and_whisper.sh "ramalama" "$LLAMA_CPP_SHA" \
-      "$WHISPER_CPP_SHA" "/usr" "-DGGML_KOMPUTE=1"
+      "$WHISPER_CPP_SHA"
 
 ENV WHISPER_CPP_SHA=${WHISPER_CPP_SHA}
 ENV LLAMA_CPP_SHA=${LLAMA_CPP_SHA}

container-images/rocm/Containerfile

@@ -8,5 +8,5 @@ COPY rocm/rocm.repo /etc/yum.repos.d/
 COPY scripts /scripts
 RUN chmod +x /scripts/*.sh && \
     /scripts/build_llama_and_whisper.sh "rocm" "$LLAMA_CPP_SHA" \
-      "$WHISPER_CPP_SHA" "/usr" "-DGGML_HIPBLAS=1"
+      "$WHISPER_CPP_SHA"
 

container-images/scripts/build_llama_and_whisper.sh

@@ -9,8 +9,7 @@ dnf_install() {
   if [ "$containerfile" = "ramalama" ]; then
     local url="https://dl.fedoraproject.org/pub/epel/epel-release-latest-9.noarch.rpm"
     dnf install -y "$url"
-    crb enable
-    dnf install -y epel-release
+    crb enable # this is in epel-release, can only install epel-release via url
     dnf --enablerepo=ubi-9-appstream-rpms install -y "${rpm_list[@]}"
     local uname_m
     uname_m="$(uname -m)"
@@ -30,54 +29,97 @@ dnf_install() {
     dnf install -y rocm-dev hipblas-devel rocblas-devel
   elif [ "$containerfile" = "cuda" ]; then
     dnf install -y "${rpm_list[@]}" gcc-toolset-12
-    source /opt/rh/gcc-toolset-12/enable
+    # shellcheck disable=SC1091
+    . /opt/rh/gcc-toolset-12/enable
   fi
 
   # For Vulkan image, we don't need to install anything extra but rebuild with
   # -DGGML_VULKAN
 }
 
 cmake_steps() {
-  local flag="$1"
-  cmake -B build "${common_flags[@]}" "$flag"
+  local cmake_flags=("${!1}")
+  cmake -B build "${cmake_flags[@]}"
   cmake --build build --config Release -j"$(nproc)"
   cmake --install build
 }
 
-main() {
-  set -e
+set_install_prefix() {
+  if [ "$containerfile" = "cuda" ]; then
+    install_prefix="/tmp/install"
+  else
+    install_prefix="/usr"
+  fi
+}
 
+configure_common_flags() {
   local containerfile="$1"
-  local llama_cpp_sha="$2"
-  local whisper_cpp_sha="$3"
-  local install_prefix="$4"
-  local build_flag_1="$5"
-  local build_flag_2="$6"
-  local common_flags=("-DGGML_CCACHE=0" \
-                      "-DCMAKE_INSTALL_PREFIX=$install_prefix" "$build_flag_1")
-  if [ -n "$build_flag_2" ]; then
-    common_flags+=("$build_flag_2")
-  fi
+  local -n common_flags_ref=$2
 
-  dnf_install
+  common_flags_ref=("-DGGML_NATIVE=OFF")
+  case "$containerfile" in
+    rocm)
+      common_flags_ref+=("-DGGML_HIPBLAS=1")
+      ;;
+    cuda)
+      common_flags_ref+=("-DGGML_CUDA=ON" "-DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined")
+      ;;
+    vulkan | asahi)
+      common_flags_ref+=("-DGGML_VULKAN=1")
+      ;;
+  esac
+}
 
-  git clone https://github.com/ggerganov/llama.cpp
-  cd llama.cpp
-  git reset --hard "$llama_cpp_sha"
-  cmake_steps
-  cd ..
+clone_and_build_whisper_cpp() {
+  local whisper_cpp_sha="$1"
+  local install_prefix="$2"
+  local whisper_flags=("${!3}")
+  whisper_flags+=("-DBUILD_SHARED_LIBS=NO")
 
   git clone https://github.com/ggerganov/whisper.cpp
   cd whisper.cpp
+  git submodule update --init --recursive
   git reset --hard "$whisper_cpp_sha"
-  cmake_steps "-DBUILD_SHARED_LIBS=NO"
+  cmake_steps whisper_flags
+  mkdir -p "$install_prefix/bin"
   mv build/bin/main "$install_prefix/bin/whisper-main"
   mv build/bin/server "$install_prefix/bin/whisper-server"
   cd ..
+}
+
+clone_and_build_llama_cpp() {
+  local llama_cpp_sha="$1"
+  local common_flags=("${!2}")
 
-  CMAKE_ARGS="${common_flags[*]}" FORCE_CMAKE=1 \
-    pip install --prefix="$install_prefix" 'llama-cpp-python[server]'
+  git clone https://github.com/ggerganov/llama.cpp
+  cd llama.cpp
+  git submodule update --init --recursive
+  git reset --hard "$llama_cpp_sha"
+  cmake_steps common_flags
+  cd ..
+}
+
+main() {
+  set -ex
+
+  local containerfile="$1"
+  local llama_cpp_sha="$2"
+  local whisper_cpp_sha="$3"
+  local install_prefix
+  set_install_prefix
+  local common_flags
+  configure_common_flags "$containerfile" common_flags
+
+  common_flags+=("-DGGML_CCACHE=0" "-DCMAKE_INSTALL_PREFIX=$install_prefix")
+  dnf_install
+  clone_and_build_whisper_cpp "$whisper_cpp_sha" "$install_prefix" common_flags[@]
+  case "$containerfile" in
+    ramalama)
+      common_flags+=("-DGGML_KOMPUTE=ON" "-DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON")
+      ;;
+  esac
 
+  clone_and_build_llama_cpp "$llama_cpp_sha" common_flags[@]
   dnf clean all
   rm -rf /var/cache/*dnf* /opt/rocm-*/lib/llvm \
     /opt/rocm-*/lib/rocblas/library/*gfx9* llama.cpp whisper.cpp

container-images/vulkan/Containerfile

@@ -3,5 +3,5 @@ FROM quay.io/ramalama/ramalama:latest
 COPY ../scripts /scripts
 RUN chmod +x /scripts/*.sh && \
     /scripts/build_llama_and_whisper.sh "vulkan" "$LLAMA_CPP_SHA" \
-      "$WHISPER_CPP_SHA" "/usr" "-DGGML_VULKAN=1"
+      "$WHISPER_CPP_SHA"
 

ramalama/cli.py

@@ -215,6 +215,7 @@ def configure_arguments(parser):
     )
     parser.add_argument("-v", "--version", dest="version", action="store_true", help="show RamaLama version")
 
+
 def configure_subcommands(parser):
     """Add subcommand parsers to the main argument parser."""
     subparsers = parser.add_subparsers(dest="subcommand")

ramalama/common.py

@@ -59,11 +59,6 @@ def exec_cmd(args, stderr=True, debug=False):
     if debug:
         perror("exec_cmd: ", *args)
 
-    if not stderr:
-        # Redirecting stderr to /dev/null
-        with open(os.devnull, "w") as devnull:
-            os.dup2(devnull.fileno(), sys.stderr.fileno())
-
     try:
         return os.execvp(args[0], args)
     except Exception:

ramalama/model.py

@@ -233,15 +233,15 @@ def run(self, args):
             if not args.container:
                 raise KeyError("--nocontainer and --name options conflict. --name requires a container.")
 
-        prompt = "You are a helpful assistant"
+        prompt = ""
         if args.ARGS:
             prompt = " ".join(args.ARGS)
 
         # Build a prompt with the stdin text that prepend the prompt passed as
         # an argument to ramalama cli
         if not sys.stdin.isatty():
-            input = sys.stdin.read()
-            prompt = input + "\n\n" + prompt
+            inp = sys.stdin.read()
+            prompt = inp + "\n\n" + prompt
 
         if args.dryrun:
             model_path = "/path/to/model"
@@ -254,36 +254,22 @@ def run(self, args):
         if not args.container:
             exec_model_path = model_path
 
-        exec_args = [
-            "llama-cli",
-            "-m",
-            exec_model_path,
-            "--in-prefix",
-            "",
-            "--in-suffix",
-            "",
-            "-c",
-            f"{args.context}",
-            "--temp",
-            f"{args.temp}",
-        ]
+        exec_args = ["llama-run", "-c", f"{args.context}", "--temp", f"{args.temp}"]
 
         if args.seed:
             exec_args += ["--seed", args.seed]
 
-        if not args.debug:
-            exec_args += ["--no-display-prompt"]
-        exec_args += [
-            "-p",
-            prompt,
-        ]
-
-        if not args.ARGS and sys.stdin.isatty():
-            exec_args.append("-cnv")
+        if args.debug:
+            exec_args += ["-v"]
 
         if args.gpu:
             exec_args.extend(self.gpu_args())
 
+        exec_args += [
+            exec_model_path,
+            prompt,
+        ]
+
         try:
             if self.exec_model_in_container(model_path, exec_args, args):
                 return
@@ -329,8 +315,7 @@ def serve(self, args):
             exec_args += ["--seed", args.seed]
 
         if args.runtime == "vllm":
-            if not (exec_model_path.endswith(".GGUF") or exec_model_path.endswith(".gguf")):
-                exec_model_path = os.path.dirname(exec_model_path)
+            exec_model_path = os.path.dirname(exec_model_path)
             exec_args = ["vllm", "serve", "--port", args.port, exec_model_path]
         else:
             if args.gpu:

scripts/replace-shas.sh

@@ -0,0 +1,15 @@
+#!/bin/bash
+
+set -euo pipefail
+
+find_files() {
+  grep -rl "$1_CPP_SHA=" container-images/
+}
+
+sed_files() {
+  xargs sed -i "s/ARG $1_CPP_SHA=.*/ARG $1_CPP_SHA=$2/g"
+}
+
+find_files "LLAMA" | sed_files "LLAMA" "$1"
+find_files "WHISPER" | sed_files "WHISPER" "$2"
+

test/system/030-run.bats

@@ -34,7 +34,7 @@ load helpers
 	is "$output" ".*${image} /bin/sh -c" "verify image name"
     else
 	run_ramalama --dryrun run -c 4096 ${model}
-	is "$output" 'llama-cli -m /path/to/model --in-prefix --in-suffix -c 4096 --temp 0.8 --no-display-prompt -p.*' "dryrun correct"
+	is "$output" 'llama-run -c 4096 --temp 0.8 /path/to/model.*' "dryrun correct"
 	is "$output" ".*-c 4096" "verify model name"
 
 	run_ramalama 1 run --ctx-size=4096 --name foobar tiny