Fix installer and ROCm environment setup

.gitattributes

@@ -3,4 +3,5 @@
 *.sh text eol=lf
 *.md text eol=lf
 *.png filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
 Ling-3.0-Flash-CIRU-int4-Strix-native.tar.gz filter=lfs diff=lfs merge=lfs -text

Ling-3.0-Flash-CIRU-int4-Strix-native.tar.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ccb017c7c323085dd424903f3de3b3bbdca9ad092ec20bab5fe2fb3444c16e30
-size 265541
+oid sha256:5fbc813c26bb470d7a38c6fbfa08ca099cd4455b1af3f2db32117133a3fb5515
+size 268772

README.md

@@ -30,7 +30,7 @@ The model weights were **not** changed, requantized, merged, fine-tuned, or rena
 
 The runtime is based on vLLM commit `d35eb6c44071ea806018841c490f0d2f3219c485` plus a 15-commit CIRU branch ending at `388d82de2abe418e73380ba250270d3abc62ac12`. Its net source delta is six files, 365 insertions, and 26 deletions.
 
-- Fixes a Wave32 LDS overflow in the ROCm top-k merge launch.
+- Includes [`liminfei-amd`'s Wave32 LDS fix](https://github.com/vllm-project/vllm/pull/46012) for the ROCm top-k merge launch, preserving the AMD author's commit credit.
 - Adds an opt-in safe PyTorch attention-state merge for a `gfx1151` Triton/HSA fault, including both supported MLA LSE layouts.
 - Routes causal uniform multi-token MLA verification blocks through Triton decode with correct per-token sequence lengths.
 - Adds narrowly guarded W4A16 MoE decode/verification fast paths for Ling's exact small-token geometry.
@@ -99,7 +99,8 @@ The native row predates the release's final SiLU kernel commit, so this table is
 
 - `runtime/Ling-3.0-Flash-CIRU-int4-Strix-native-vllm.bundle`: exact Git objects for the CIRU branch, based on the pinned upstream commit.
 - `runtime/Ling-3.0-Flash-CIRU-int4-Strix-native-vllm.patch`: reviewable net source diff.
-- `scripts/download-model.sh`: downloads the unchanged official checkpoint into a directory named `Ling-3.0-Flash-CIRU-int4-Strix-native`.
+- `install.sh`: guided human installer plus a flag-driven interface for agents and automation.
+- `scripts/download-model.sh`: downloads the unchanged official checkpoint to a parent directory or exact `--model-path`.
 - `scripts/install-host-deps.sh`: distro-aware Linux/WSL host dependency installer.
 - `scripts/build-vllm-gfx1151.sh`: reproducible build entry point for the pinned AMD wheel stack.
 - `scripts/run-256k.sh`: native 256K, native-MTP K1 server profile with five active sequence slots.
@@ -112,16 +113,32 @@ The named release archive is `Ling-3.0-Flash-CIRU-int4-Strix-native.tar.gz`. It
 
 The exact validated environment is Python 3.12, Torch `2.13.0+rocm7.15.0a20260728`, HIP `7.15.0`, Triton `3.8.0+git4cff872c.rocm7.15.0a20260728`, FlashAttention `2.8.3`, and the CIRU vLLM source at `388d82de2`.
 
+For a guided install, run this from the repository root:
+
 ```bash
-bash scripts/install-host-deps.sh
-bash scripts/build-vllm-gfx1151.sh "$HOME/ciru-ling-runtime"
-bash scripts/download-model.sh "$HOME/models"
+bash install.sh --interactive
+```
 
-VENV="$HOME/ciru-ling-runtime/.venv" \
-MODEL_PATH="$HOME/models/Ling-3.0-Flash-CIRU-int4-Strix-native" \
+The runtime path defaults to the current directory. The installer asks for the
+exact model directory, makes the 77 GB checkpoint download opt-in, validates
+the completed runtime, and writes a sourceable `ling3-runtime.env` containing
+`VENV` and `MODEL_PATH`. It prints the launch command when finished.
+
+Agents and automation can avoid prompts:
+
+```bash
+bash install.sh \
+  --install-root "$PWD" \
+  --model-path "/path/to/Ling-checkpoint"
+
+source "$PWD/ling3-runtime.env"
 bash scripts/run-256k.sh
 ```
 
+Add `--install-host-deps` when the supported distro packages are not already
+installed, or `--download-model` when the exact model path does not yet contain
+the checkpoint. Use `bash install.sh --help` for the complete agent interface.
+
 The API advertises model ID `Ling-3.0-Flash-CIRU-int4-Strix-native` on `http://127.0.0.1:18081/v1`.
 
 The default launcher provides the checkpoint's native 256K context (`262144` tokens), five active sequence slots, and `gpu_memory_utilization=0.72`. Requests beyond five wait in vLLM's scheduler.
@@ -190,6 +207,7 @@ Other W4A16 models will run only if already supported by this vLLM base; they sh
 
 - Model and official INT4 weights: [InclusionAI/Ling-3.0-flash-int4](https://huggingface.co/inclusionAI/Ling-3.0-flash-int4), MIT license.
 - Inference engine: [vLLM](https://github.com/vllm-project/vllm), Apache-2.0.
+- Wave32 top-k merge fix: [`liminfei-amd`](https://github.com/liminfei-amd), [vLLM PR #46012](https://github.com/vllm-project/vllm/pull/46012). The release history retains the original AMD authorship.
 - ROCm, Torch, and Triton packages: AMD and their respective upstream projects.
 - Strix Halo kernel, TTM/GTT, and WSL/ROCDXG guidance: [AMD ROCm documentation](https://rocm.docs.amd.com/en/docs-7.2.0/how-to/system-optimization/strixhalo.html) and [ROCDXG](https://github.com/ROCm/librocdxg).
 - Runtime integration, profiling, kernels, and validation: CIRU.

SHA256SUMS

@@ -1,21 +1,22 @@
-eb208bf38f31d7b068af8e8351378615017b911eb7d0302264d5398cde5d403e  .gitattributes
+7b77dda00cc6bd6d0ce15f7c055f9e997f56e1cf1709c61cca5a57283ce3f008  .gitattributes
 411ac21b4849cb3bf65f84a55263550dcd658e0623c2b117233ba1fced1c17ff  assets/ling30int4.png
 65cf5b48f9df164c40fbac98e21c6f9c5ced98660d653630cb58f5ab4b28919f  CHANGELOG.md
 16213dfa6abe90226ab92f9407f3bb92f9ed729e30bb8c37ffd92c7dbc999c73  config/experimental-yarn-1m.env
 c3cdeda6c1bdd47f2b947eac7c4c7182e27249220ddd6eb372c5d12c2c6c405f  config/native-256k.env
 0bb41680a6f0c8366e8f883ca7d76ce1a2ccd3e42541f6d0616df33fe867f55a  docs/BENCHMARKS.md
-165705668ab267829936d77d62585bd4bb3c00b4e1fdfcf5bcd43145720338e1  docs/UBUNTU_BUILD.md
+482e0b0b86622890f4e5db69636ea6427d4d061c6939fc2456a5f7581bda3bd9  docs/UBUNTU_BUILD.md
 23336099c40774a77a555165da7aee562d5e603017201668178f35f6a8e5cd0d  docs/VALIDATED_ENVIRONMENT.md
 3176eba46cccbd60d33589230190e530eaace8c96466c9621772a1f2c50c07e4  docs/WINDOWS_WSL2.md
-852b2422c20355b533884dae00f330aea8e836efecf0b070d59b3e848540c169  MANIFEST.json
-a676f974838f3f2886cfd3109a1b991a1f4e984ea068b08107b8f206852c0294  README.md
+b85d98b8c4f66ff3e69696a47470d11d010b36358d1b6e184ec3c8edb7ab81fb  install.sh
+852b2422c20355b533884dae00f330aea8e836efecf0b070d59b3e848540c169  MANIFEST.json
+a4288054963d206dc4978658e12b4dc9e7a77d0f5e4717dfd4cd486ba4c7b993  README.md
 15e48a8df3b65a73c81cf73c820ba12b4ab9474697d6b3f876c26c0b0cd3d455  runtime/COMMITS.txt
 6b03af3a1685a1969ff96870ea776537f5f67fd3674fc353880bd088bcef9a9c  runtime/Ling-3.0-Flash-CIRU-int4-Strix-native-vllm.bundle
 8f307e9f2f16207856ca0b064a765a73bbebc029ad0b0fed61e5975a17e07ab6  runtime/Ling-3.0-Flash-CIRU-int4-Strix-native-vllm.patch
 f4caeda80ed751a40fa976aa7853357b88075e6eb6d2d49e68b0bc8599754c34  runtime/SHA256SUMS.runtime
-25ab169703ed6c39260f9336f3d484c9a11b1f9b31c7f9072866b158365de983  scripts/build-vllm-gfx1151.sh
-6a8530b9b38c52d6fe8efab5dd3a6bbf02a3f24a0b9a800b86a6dde3a3103071  scripts/download-model.sh
-d6aa5efee74c850f2ab65efeb4dfb66142c440d66276983282de6aec6a855270  scripts/install-host-deps.sh
+2c066e583d3b24d4c0caa7d04263228dff80e464c61a557736e9c5c5f79b9869  scripts/build-vllm-gfx1151.sh
+ef474884f7d52d1643e0209da4fb9407f42ba4f7d1e19b91cf7d1112849f74cc  scripts/download-model.sh
+8142ef953e2c9618df885756eb897303e6443a4d139a09fe9a882c80d054f466  scripts/install-host-deps.sh
 27d3cfba9f4178f83cbd2d7e44db098ab69e515f76e5c98c40ac219b7b0597ca  scripts/run-1m-yarn-experimental.sh
 296ee94a925db5b97e21afb6ffd4ae3c0cc1c2be67af81e21350f43efaa59ead  scripts/run-256k.sh
-aa5f1137c666abe07c9fc37a64547dc535d461958db58f7da80beea4133367f9  scripts/run-common.sh
+718bd922b3bb0a07fb9e2bb0c5ca9a67b799582ac6f614c17af291793e2a69e0  scripts/run-common.sh

docs/UBUNTU_BUILD.md

@@ -39,6 +39,23 @@ ls -l /dev/kfd /dev/dri/renderD*
 
 If `/dev/kfd` is absent, stop here. The Python wheels cannot replace a missing kernel driver.
 
+## Guided installer
+
+After completing the host-driver and unified-memory setup, the recommended
+human interface is:
+
+```bash
+bash install.sh --interactive
+```
+
+The runtime install directory defaults to the current directory. The installer
+asks for the exact checkpoint directory, makes host-package installation and
+the approximately 77 GB model download explicit choices, builds the runtime,
+validates `vllm`, and writes `ling3-runtime.env` with the launch environment.
+
+For non-interactive agents, see `bash install.sh --help`. The numbered steps
+below remain available for manual diagnosis and custom automation.
+
 ## 2. Configure unified-memory access
 
 Strix Halo maps system memory dynamically through GTT/TTM. AMD recommends keeping the fixed BIOS GPU reservation small and increasing the shared TTM limit instead. On CIRU's 128 GB host, the validated mapping limit was 124 GiB. That limit is not a permanent reservation, but the model can still leave very little free RAM during long-context requests.
@@ -87,8 +104,14 @@ A successful build ends with:
 
 ```text
 CIRU vLLM gfx1151 build: PASS
+Runtime virtual environment ready: /home/USER/ciru-ling-runtime/.venv
 ```
 
+The command argument is the install root. Use `.venv` beneath that exact same
+path as `VENV` when launching. If the build exits before the two lines above,
+the environment is partial even if ROCm tools such as `hipcc` already exist;
+rerun the corrected build command against the same install root.
+
 ## 5. Download the unchanged checkpoint and serve
 
 ```bash

install.sh

@@ -0,0 +1,236 @@
+#!/usr/bin/env bash
+set -Eeuo pipefail
+
+script_dir="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
+release_name=Ling-3.0-Flash-CIRU-int4-Strix-native
+export PATH="$HOME/.local/bin:$PATH"
+
+usage() {
+  cat <<EOF
+Install the CIRU Ling 3.0 Flash gfx1151 runtime.
+
+Human-guided install:
+  bash install.sh --interactive
+
+Non-interactive install for agents and automation:
+  bash install.sh --install-root PATH --model-path PATH [OPTIONS]
+
+Options:
+  --interactive          Ask for paths and optional actions.
+  --install-root PATH    Runtime root; defaults to the current directory.
+                         The vLLM environment is created at PATH/.venv.
+  --model-path PATH      Exact checkpoint directory; defaults to
+                         ./Ling-3.0-Flash-CIRU-int4-Strix-native.
+  --install-host-deps    Run scripts/install-host-deps.sh (may use sudo).
+  --download-model       Download the pinned official checkpoint to MODEL_PATH.
+                         The checkpoint is approximately 77 GB.
+  --skip-build           Validate/configure an already-built runtime.
+  --dry-run              Print the commands and resolved paths without running them.
+  --print-env            Print launch exports for the resolved paths and exit.
+  -h, --help             Show this help.
+
+The installer writes PATH/ling3-runtime.env after validation. Agents can use:
+  source PATH/ling3-runtime.env
+  bash scripts/run-256k.sh
+
+Environment variables can always be supplied directly:
+  VENV=PATH/.venv MODEL_PATH=MODEL_PATH bash scripts/run-256k.sh
+EOF
+}
+
+die() {
+  printf 'install.sh: %s\n' "$*" >&2
+  exit 2
+}
+
+absolute_path() {
+  local path="$1"
+  case "$path" in
+    '~') path="$HOME" ;;
+    '~/'*) path="$HOME/${path#\~/}" ;;
+  esac
+  if command -v realpath >/dev/null 2>&1; then
+    realpath -m -- "$path"
+  elif [[ "$path" = /* ]]; then
+    printf '%s\n' "$path"
+  else
+    printf '%s/%s\n' "$PWD" "${path#./}"
+  fi
+}
+
+prompt_value() {
+  local variable_name="$1"
+  local label="$2"
+  local default_value="$3"
+  local reply
+  read -r -p "$label [$default_value]: " reply
+  printf -v "$variable_name" '%s' "${reply:-$default_value}"
+}
+
+ask_yes_no() {
+  local label="$1"
+  local default_answer="$2"
+  local suffix reply
+  if [ "$default_answer" = yes ]; then
+    suffix='[Y/n]'
+  else
+    suffix='[y/N]'
+  fi
+  while true; do
+    read -r -p "$label $suffix: " reply
+    reply="${reply:-$default_answer}"
+    case "${reply,,}" in
+      y|yes) return 0 ;;
+      n|no) return 1 ;;
+      *) printf 'Please answer yes or no.\n' >&2 ;;
+    esac
+  done
+}
+
+interactive=0
+install_root=
+model_path=
+install_host_deps=0
+download_model=0
+build_runtime=1
+dry_run=0
+print_env=0
+
+if [ "$#" -eq 0 ]; then
+  usage
+  exit 0
+fi
+
+while [ "$#" -gt 0 ]; do
+  case "$1" in
+    --interactive)
+      interactive=1
+      shift
+      ;;
+    --install-root)
+      [ "$#" -ge 2 ] || die '--install-root requires a path'
+      install_root="$2"
+      shift 2
+      ;;
+    --model-path)
+      [ "$#" -ge 2 ] || die '--model-path requires a path'
+      model_path="$2"
+      shift 2
+      ;;
+    --install-host-deps)
+      install_host_deps=1
+      shift
+      ;;
+    --download-model)
+      download_model=1
+      shift
+      ;;
+    --skip-build)
+      build_runtime=0
+      shift
+      ;;
+    --dry-run)
+      dry_run=1
+      shift
+      ;;
+    --print-env)
+      print_env=1
+      shift
+      ;;
+    -h|--help)
+      usage
+      exit 0
+      ;;
+    *)
+      die "unknown option: $1"
+      ;;
+  esac
+done
+
+if [ "$interactive" = 1 ]; then
+  [ -t 0 ] || die '--interactive requires a terminal; use explicit flags for automation'
+  prompt_value install_root 'Runtime install directory' "${install_root:-$PWD}"
+  prompt_value model_path 'Ling checkpoint directory' \
+    "${model_path:-$PWD/$release_name}"
+  if ask_yes_no 'Run the distro host-dependency installer (may use sudo)?' no; then
+    install_host_deps=1
+  fi
+  if [ ! -f "$model_path/config.json" ]; then
+    if ask_yes_no 'Checkpoint not found there. Download the pinned model (~77 GB)?' no; then
+      download_model=1
+    fi
+  fi
+fi
+
+install_root="$(absolute_path "${install_root:-$PWD}")"
+model_path="$(absolute_path "${model_path:-$PWD/$release_name}")"
+venv="$install_root/.venv"
+env_file="$install_root/ling3-runtime.env"
+
+if [ "$print_env" = 1 ]; then
+  printf 'export VENV=%q\n' "$venv"
+  printf 'export MODEL_PATH=%q\n' "$model_path"
+  exit 0
+fi
+
+printf 'CIRU Ling runtime installation\n'
+printf '  package: %s\n  install root: %s\n  venv: %s\n  model: %s\n' \
+  "$script_dir" "$install_root" "$venv" "$model_path"
+printf '  host dependencies: %s\n  download model: %s\n  build runtime: %s\n' \
+  "$install_host_deps" "$download_model" "$build_runtime"
+
+if [ "$interactive" = 1 ] && ! ask_yes_no 'Continue with these settings?' yes; then
+  printf 'Installation cancelled.\n'
+  exit 0
+fi
+
+if [ "$dry_run" = 1 ]; then
+  if [ "$install_host_deps" = 1 ]; then
+    printf 'DRY RUN: bash %q\n' "$script_dir/scripts/install-host-deps.sh"
+  fi
+  if [ "$download_model" = 1 ]; then
+    printf 'DRY RUN: bash %q --model-path %q\n' \
+      "$script_dir/scripts/download-model.sh" "$model_path"
+  fi
+  if [ "$build_runtime" = 1 ]; then
+    printf 'DRY RUN: bash %q %q\n' \
+      "$script_dir/scripts/build-vllm-gfx1151.sh" "$install_root"
+  fi
+  printf 'DRY RUN: write %q\n' "$env_file"
+  exit 0
+fi
+
+if [ "$download_model" != 1 ] && [ ! -f "$model_path/config.json" ]; then
+  die "model config not found at $model_path/config.json; choose an existing checkpoint or pass --download-model"
+fi
+
+if [ "$install_host_deps" = 1 ]; then
+  bash "$script_dir/scripts/install-host-deps.sh"
+fi
+
+if [ "$download_model" = 1 ]; then
+  bash "$script_dir/scripts/download-model.sh" --model-path "$model_path"
+fi
+
+if [ "$build_runtime" = 1 ]; then
+  bash "$script_dir/scripts/build-vllm-gfx1151.sh" "$install_root"
+fi
+
+[ -x "$venv/bin/python" ] || die "runtime Python is missing after installation: $venv/bin/python"
+[ -x "$venv/bin/vllm" ] || die "vLLM launcher is missing after installation: $venv/bin/vllm"
+[ -f "$model_path/config.json" ] || die "model config is missing after installation: $model_path/config.json"
+
+mkdir -p "$install_root"
+{
+  printf '# Generated by %s/install.sh\n' "$script_dir"
+  printf 'export VENV=%q\n' "$venv"
+  printf 'export MODEL_PATH=%q\n' "$model_path"
+} >"$env_file"
+
+printf '\nInstallation complete.\n'
+printf 'Environment file: %s\n\n' "$env_file"
+printf 'Launch the native 256K profile:\n'
+printf '  source %q\n  bash %q\n\n' "$env_file" "$script_dir/scripts/run-256k.sh"
+printf 'For agents or automation:\n'
+printf '  export VENV=%q\n  export MODEL_PATH=%q\n' "$venv" "$model_path"
+printf '  bash %q\n' "$script_dir/scripts/run-256k.sh"

scripts/build-vllm-gfx1151.sh

@@ -12,6 +12,10 @@ head=388d82de2abe418e73380ba250270d3abc62ac12
 bundle="$package_root/runtime/Ling-3.0-Flash-CIRU-int4-Strix-native-vllm.bundle"
 flash_source="$install_root/flash-attention"
 flash_commit=0e60e39473e8df549a20fb5353760f7a65b30e2d
+export PATH="$HOME/.local/bin:$PATH"
+
+printf 'CIRU build install root: %s\n' "$install_root"
+printf 'CIRU runtime virtual environment: %s\n' "$venv"
 
 command -v git >/dev/null
 command -v uv >/dev/null || {
@@ -56,7 +60,6 @@ UV_CACHE_DIR="${UV_CACHE_DIR:-$install_root/uv-cache}" uv pip install \
   'rocm[libraries,devel,device-gfx1151]==7.15.0a20260728'
 
 "$venv/bin/rocm-sdk" init
-"$venv/bin/rocm-sdk" test
 
 if [ ! -d "$source_dir/.git" ]; then
   git clone https://github.com/vllm-project/vllm.git "$source_dir"
@@ -111,6 +114,10 @@ export MAX_JOBS="${MAX_JOBS:-16}"
 export CMAKE_BUILD_PARALLEL_LEVEL="$MAX_JOBS"
 export CMAKE_ARGS="-DROCM_PATH=$sdk -DCMAKE_HIP_ARCHITECTURES=gfx1151 ${CMAKE_ARGS:-}"
 
+# Run the SDK checks against this virtual environment's development root, not
+# an inherited system ROCm installation.
+"$venv/bin/rocm-sdk" test
+
 UV_CACHE_DIR="${UV_CACHE_DIR:-$install_root/uv-cache}" uv pip install \
   --python "$venv/bin/python" \
   'cmake>=3.26.1,<4' ninja 'packaging>=24.2' \
@@ -124,6 +131,11 @@ UV_CACHE_DIR="${UV_CACHE_DIR:-$install_root/uv-cache}" uv pip install \
 cd "$source_dir"
 UV_CACHE_DIR="${UV_CACHE_DIR:-$install_root/uv-cache}" uv pip install \
   --python "$venv/bin/python" --no-build-isolation --no-deps -e .
+if [ ! -x "$venv/bin/vllm" ]; then
+  printf 'vLLM installation did not create the expected launcher: %s\n' \
+    "$venv/bin/vllm" >&2
+  exit 2
+fi
 
 if [ ! -d "$flash_source/.git" ]; then
   git clone https://github.com/Dao-AILab/flash-attention.git "$flash_source"
@@ -153,3 +165,4 @@ for name in (
     importlib.import_module(name)
 print("CIRU vLLM gfx1151 build: PASS")
 PY
+printf 'Runtime virtual environment ready: %s\n' "$venv"

scripts/download-model.sh

@@ -1,10 +1,40 @@
 #!/usr/bin/env bash
 set -Eeuo pipefail
 
-destination_root="${1:-$PWD/models}"
 release_name=Ling-3.0-Flash-CIRU-int4-Strix-native
 revision=ca3ea63b0255d212c4fe6020db9e0a51ce136006
-destination="$destination_root/$release_name"
+destination_root="$PWD/models"
+destination=
+export PATH="$HOME/.local/bin:$PATH"
+
+usage() {
+  cat <<EOF
+Download the pinned official Ling checkpoint.
+
+Usage:
+  bash scripts/download-model.sh [DESTINATION_ROOT]
+  bash scripts/download-model.sh --model-path EXACT_DIRECTORY
+
+The legacy positional form downloads to DESTINATION_ROOT/$release_name.
+EOF
+}
+
+if [ "${1:-}" = --model-path ]; then
+  [ "$#" -eq 2 ] || {
+    echo '--model-path requires exactly one directory.' >&2
+    exit 2
+  }
+  destination="$2"
+elif [ "${1:-}" = -h ] || [ "${1:-}" = --help ]; then
+  usage
+  exit 0
+elif [ "$#" -le 1 ]; then
+  destination_root="${1:-$destination_root}"
+  destination="$destination_root/$release_name"
+else
+  usage >&2
+  exit 2
+fi
 
 if command -v hf >/dev/null 2>&1; then
   hf_cmd=(hf)
@@ -15,7 +45,7 @@ else
   exit 2
 fi
 
-mkdir -p "$destination_root"
+mkdir -p "$destination"
 "${hf_cmd[@]}" download inclusionAI/Ling-3.0-flash-int4 \
   --revision "$revision" \
   --local-dir "$destination"

scripts/install-host-deps.sh

@@ -44,4 +44,3 @@ command -v uv >/dev/null || {
 }
 
 printf 'Host dependencies ready on %s.\n' "${PRETTY_NAME:-$distro}"
-printf 'Next: bash scripts/build-vllm-gfx1151.sh "$HOME/ciru-ling-runtime"\n'

scripts/run-common.sh

@@ -8,8 +8,23 @@ host="${HOST:-127.0.0.1}"
 port="${PORT:-18081}"
 cache_root="${CACHE_ROOT:-$HOME/.cache/ciru-ling3-vllm}"
 
-test -x "$venv/bin/vllm"
-test -d "$model_path"
+if [ ! -x "$venv/bin/python" ]; then
+  printf 'Runtime virtual environment is missing Python: %s\n' \
+    "$venv/bin/python" >&2
+  printf 'Set VENV to the install root .venv created by scripts/build-vllm-gfx1151.sh.\n' >&2
+  exit 2
+fi
+if [ ! -x "$venv/bin/vllm" ]; then
+  printf 'Runtime virtual environment is missing the vLLM launcher: %s\n' \
+    "$venv/bin/vllm" >&2
+  printf 'Re-run scripts/build-vllm-gfx1151.sh for this VENV.\n' >&2
+  exit 2
+fi
+if [ ! -d "$model_path" ]; then
+  printf 'Model checkpoint directory does not exist: %s\n' "$model_path" >&2
+  printf 'Set MODEL_PATH to the directory containing the downloaded Ling checkpoint.\n' >&2
+  exit 2
+fi
 site="$($venv/bin/python - <<'PY'
 import site
 print(site.getsitepackages()[0])
@@ -17,7 +32,11 @@ PY
 )"
 sdk="$site/_rocm_sdk_devel"
 core="$site/_rocm_sdk_core"
-test -x "$sdk/bin/hipcc"
+if [ ! -x "$sdk/bin/hipcc" ]; then
+  printf 'Private ROCm SDK is missing hipcc: %s\n' "$sdk/bin/hipcc" >&2
+  printf 'Re-run scripts/build-vllm-gfx1151.sh to install and initialize the SDK.\n' >&2
+  exit 2
+fi
 mkdir -p "$cache_root/vllm" "$cache_root/triton" "$cache_root/huggingface-modules"
 
 export PATH="$sdk/bin:$sdk/llvm/bin:$venv/bin:$PATH"
@@ -74,4 +93,7 @@ if [ "${DRY_RUN:-0}" = 1 ]; then
   exit 0
 fi
 
+printf 'Launching %s\n' "$release_name"
+printf '  vLLM: %s\n  model: %s\n  listen: %s:%s\n' \
+  "$venv/bin/vllm" "$model_path" "$host" "$port"
 exec "$venv/bin/vllm" "${args[@]}"