chore: import upstream snapshot with attribution

scripts/snapdragon/qdc/tests/conftest.py

@@ -0,0 +1,20 @@
+"""Shared pytest fixtures for QDC on-device test runners."""
+
+import os
+
+import pytest
+from appium import webdriver
+
+from utils import options, write_qdc_log
+
+
+@pytest.fixture(scope="session", autouse=True)
+def driver():
+    return webdriver.Remote(command_executor="http://127.0.0.1:4723/wd/hub", options=options)
+
+
+def pytest_sessionfinish(session, exitstatus):
+    xml_path = getattr(session.config.option, "xmlpath", None) or "results.xml"
+    if os.path.exists(xml_path):
+        with open(xml_path) as f:
+            write_qdc_log("results.xml", f.read())

scripts/snapdragon/qdc/tests/linux/run_linux.sh

@@ -0,0 +1,232 @@
+#!/bin/bash
+# llama.cpp Hexagon test entry script for QDC Linux IoT (BASH framework).
+#
+# Placeholders substituted by run_qdc_jobs.py (--platform linux) before upload:
+#   {MODEL_URL}   direct URL to a .gguf model file
+#   {TEST_MODE}   bench | backend-ops | all
+#
+# QDC extracts the artifact zip to /data/local/tmp/TestContent/ and invokes
+# this script via: /bin/bash /data/local/tmp/TestContent/run_linux.sh
+# Any files written under /data/local/tmp/QDC_logs/ are auto-uploaded.
+
+set +e
+umask 022
+
+LOG_DIR=/data/local/tmp/QDC_logs
+BUNDLE_DIR=/data/local/tmp/TestContent/llama_cpp_bundle
+MODEL_DIR=/data/local/tmp/gguf
+MODEL_PATH="$MODEL_DIR/model.gguf"
+RESULTS_XML="$LOG_DIR/results.xml"
+
+mkdir -p "$LOG_DIR" "$MODEL_DIR"
+# Redirect all parent-shell output to script.log so QDC auto-uploads it;
+# per-case runs still capture their own stdout/stderr into dedicated logs.
+exec > "$LOG_DIR/script.log" 2>&1
+
+echo "=== env ==="
+date -u
+uname -a
+pwd
+
+mount -o rw,remount / 2>/dev/null || true
+
+cd "$BUNDLE_DIR" || { echo "FATAL: bundle missing at $BUNDLE_DIR"; exit 1; }
+chmod +x bin/* 2>/dev/null
+export LD_LIBRARY_PATH="$BUNDLE_DIR/lib:$LD_LIBRARY_PATH"
+export ADSP_LIBRARY_PATH="$BUNDLE_DIR/lib"
+export GGML_HEXAGON_EXPERIMENTAL=1
+
+echo "=== download model ==="
+MODEL_URL="{MODEL_URL}"
+if [ -z "$MODEL_URL" ]; then
+  echo "No model URL provided, skipping download"
+elif [ ! -f "$MODEL_PATH" ]; then
+  curl -L -fS --retry 3 --retry-delay 5 -o "$MODEL_PATH" "$MODEL_URL"
+  curl_rc=$?
+  if [ $curl_rc -ne 0 ]; then
+    echo "FATAL: model download failed (rc=$curl_rc)"
+    exit 1
+  fi
+  ls -la "$MODEL_PATH"
+fi
+
+# ---------------------------------------------------------------------------
+# JUnit XML helpers
+# ---------------------------------------------------------------------------
+
+xml_open() {
+  printf '%s\n' \
+    '<?xml version="1.0" encoding="utf-8"?>' \
+    "<testsuites>" \
+    "<testsuite name=\"llama_cpp_linux\">" \
+    > "$RESULTS_XML"
+}
+
+xml_close() {
+  printf '%s\n' '</testsuite>' '</testsuites>' >> "$RESULTS_XML"
+}
+
+xml_case_pass() {
+  local classname=$1 name=$2
+  printf '<testcase classname="%s" name="%s"/>\n' "$classname" "$name" >> "$RESULTS_XML"
+}
+
+xml_case_fail() {
+  local classname=$1 name=$2 rc=$3 logfile=$4
+  {
+    printf '<testcase classname="%s" name="%s">\n' "$classname" "$name"
+    printf '<failure message="exit %s"><![CDATA[\n' "$rc"
+    tail -c 4096 "$logfile" 2>/dev/null | sed 's/]]>/]] >/g'
+    printf '\n]]></failure>\n</testcase>\n'
+  } >> "$RESULTS_XML"
+}
+
+# Map backend name -> "NDEV --device" pair. "none" means no offload (CPU).
+backend_env() {
+  case "$1" in
+    cpu) echo "0 none" ;;
+    gpu) echo "0 GPUOpenCL" ;;
+    npu) echo "1 HTP0" ;;
+  esac
+}
+
+backend_log_name() {
+  case "$1" in
+    cpu) echo "cpu" ;;
+    gpu) echo "gpu" ;;
+    npu) echo "htp" ;;
+  esac
+}
+
+
+backend_device_name() {
+  case "$1" in
+    cpu) echo "none" ;;
+    gpu) echo "GPUOpenCL" ;;
+    npu) echo "HTP0" ;;
+  esac
+}
+
+# Append a diagnostic block when a per-case `timeout N` fires (rc=124). The
+# naked log file at that point usually just ends mid-OpenCL-init with no
+# stderr, which is hard to read in CI summaries.
+note_timeout_if_triggered() {
+  local rc=$1 budget=$2 log=$3
+  [ "$rc" -eq 124 ] || return 0
+  {
+    printf '\n'
+    printf '=== TIMEOUT after %ss ===\n' "$budget"
+    printf 'uptime: '; uptime 2>/dev/null
+    printf 'free -m:\n'; free -m 2>/dev/null
+    printf 'loadavg: '; cat /proc/loadavg 2>/dev/null
+  } >> "$log"
+}
+
+completion_extra_args() {
+  case "$1" in
+    cpu) echo "--device none --ctx-size 128 -no-cnv -n 32 --seed 42 --batch-size 128" ;;
+    gpu) echo "--device GPUOpenCL --ctx-size 128 -no-cnv -n 32 --seed 42 --ubatch-size 512" ;;
+    npu) echo "--device HTP0 --ctx-size 128 -no-cnv -n 32 --seed 42 --ubatch-size 1024" ;;
+  esac
+}
+
+run_completion_case() {
+  local name=$1
+  local parts=($(backend_env "$name"))
+  local ndev=${parts[0]} device=${parts[1]}
+  local device_log_name=$(backend_device_name "$name")
+  local log="$LOG_DIR/llama_completion_${device_log_name}.log"
+  local prompt="$LOG_DIR/bench_prompt.txt"
+  echo 'What is the capital of France?' > "$prompt"
+  local extra
+  extra=$(completion_extra_args "$name")
+  echo "=== [completion:$name] llama-completion --device $device (NDEV=$ndev) ==="
+  timeout 600 env GGML_HEXAGON_NDEV=$ndev ./bin/llama-completion \
+      -m "$MODEL_PATH" \
+      -f "$prompt" \
+      $extra \
+      > "$log" 2>&1 < /dev/null
+  local rc=$?
+  note_timeout_if_triggered "$rc" 600 "$log"
+  if [ $rc -eq 0 ]; then
+    xml_case_pass "tests.test_bench_posix" "test_llama_completion[$name]"
+  else
+    xml_case_fail "tests.test_bench_posix" "test_llama_completion[$name]" "$rc" "$log"
+  fi
+}
+
+run_bench_case() {
+  local name=$1
+  local parts=($(backend_env "$name"))
+  local ndev=${parts[0]} device=${parts[1]}
+  local log_suffix=$(backend_log_name "$name")
+  local log="$LOG_DIR/llama_bench_${log_suffix}.log"
+  echo "=== [bench:$name] llama-bench --device $device (NDEV=$ndev) ==="
+  timeout 600 env GGML_HEXAGON_NDEV=$ndev ./bin/llama-bench \
+      -m "$MODEL_PATH" \
+      --device "$device" \
+      -ngl 99 \
+      --batch-size 128 \
+      -t 4 \
+      -p 128 \
+      -n 32 \
+      > "$log" 2>&1
+  local rc=$?
+  note_timeout_if_triggered "$rc" 600 "$log"
+  if [ $rc -eq 0 ]; then
+    xml_case_pass "tests.test_bench_posix" "test_llama_bench[$name]"
+  else
+    xml_case_fail "tests.test_bench_posix" "test_llama_bench[$name]" "$rc" "$log"
+  fi
+}
+
+run_backend_ops_case() {
+  local dtype=$1
+  local log="$LOG_DIR/backend_ops_${dtype}.log"
+  local pattern
+  case "$dtype" in
+    q4_0)
+      # Matches Android: exclude a known-broken shape on NPU.
+      pattern='^(?=.*type_a=q4_0)(?!.*type_b=f32,m=576,n=512,k=576).*$'
+      ;;
+    *)
+      pattern="type_a=${dtype}"
+      ;;
+  esac
+  echo "=== [backend-ops:$dtype] test-backend-ops -b HTP0 -o MUL_MAT ==="
+  timeout 600 env GGML_HEXAGON_NDEV=1 GGML_HEXAGON_HOSTBUF=0 ./bin/test-backend-ops \
+      -b HTP0 -o MUL_MAT -p "$pattern" \
+      > "$log" 2>&1
+  local rc=$?
+  note_timeout_if_triggered "$rc" 600 "$log"
+  if [ $rc -eq 0 ]; then
+    xml_case_pass "tests.test_backend_ops_posix" "test_backend_ops_htp0[$dtype]"
+  else
+    xml_case_fail "tests.test_backend_ops_posix" "test_backend_ops_htp0[$dtype]" "$rc" "$log"
+  fi
+}
+
+xml_open
+
+case "{TEST_MODE}" in
+  bench)
+    for b in cpu gpu npu; do run_completion_case "$b"; done
+    for b in cpu gpu npu; do run_bench_case "$b"; done
+    ;;
+  backend-ops)
+    for d in mxfp4 fp16 q4_0; do run_backend_ops_case "$d"; done
+    ;;
+  all)
+    for b in cpu gpu npu; do run_completion_case "$b"; done
+    for b in cpu gpu npu; do run_bench_case "$b"; done
+    for d in mxfp4 fp16 q4_0; do run_backend_ops_case "$d"; done
+    ;;
+  *)
+    echo "FATAL: unsupported TEST_MODE={TEST_MODE}"
+    ;;
+esac
+
+xml_close
+echo "=== done ==="
+# Host parses results.xml to decide pass/fail.
+exit 0

scripts/snapdragon/qdc/tests/run_backend_ops_posix.py

@@ -0,0 +1,51 @@
+"""
+On-device test-backend-ops runner for llama.cpp (HTP0 backend).
+
+On Android: executed by QDC's Appium test framework on the QDC runner.
+The runner has ADB access to the allocated device.
+On Linux: runs test-backend-ops directly via run_linux.sh (BASH framework).
+"""
+
+import os
+import sys
+
+import pytest
+
+from utils import (
+    BIN_PATH,
+    push_bundle_if_needed,
+    run_script,
+    write_qdc_log,
+)
+
+
+@pytest.fixture(scope="session", autouse=True)
+def install(driver):
+    push_bundle_if_needed(f"{BIN_PATH}/test-backend-ops")
+
+
+@pytest.mark.parametrize("type_a", ["mxfp4", "fp16", "q4_0"])
+def test_backend_ops_htp0(type_a):
+    if type_a == "q4_0":
+        pattern = r'^(?=.*type_a=q4_0)(?!.*type_b=f32,m=576,n=512,k=576).*$'
+    else:
+        pattern = f"type_a={type_a}"
+
+    quoted_pattern = f'"{pattern}"' if type_a == "q4_0" else pattern
+    result = run_script(
+        "run-tool.sh",
+        extra_env={"HB": "0"},
+        extra_args=["test-backend-ops", "-b", "HTP0", "-o", "MUL_MAT", "-p", quoted_pattern],
+    )
+    write_qdc_log(f"backend_ops_{type_a}.log", result.stdout or "")
+    assert result.returncode == 0, (
+        f"test-backend-ops type_a={type_a} failed (exit {result.returncode})"
+    )
+
+
+if __name__ == "__main__":
+    ret = pytest.main(["-s", "--junitxml=results.xml", os.path.realpath(__file__)])
+    if os.path.exists("results.xml"):
+        with open("results.xml") as f:
+            write_qdc_log("results.xml", f.read())
+    sys.exit(ret)

scripts/snapdragon/qdc/tests/run_bench_tests_posix.py

@@ -0,0 +1,95 @@
+"""
+On-device bench and completion test runner for llama.cpp (CPU, GPU, NPU backends).
+
+On Android: calls upstream run-*.sh scripts from llama.cpp/scripts/snapdragon/adb/
+on the QDC runner host (scripts wrap commands in ``adb shell`` internally).
+
+On Linux: runs llama-bench directly via run_linux.sh (BASH framework).
+
+Placeholders replaced at artifact creation time by run_qdc_jobs.py:
+  <<MODEL_URL>>  Direct URL to the GGUF model file (downloaded on-device)
+"""
+
+import os
+import subprocess
+import sys
+
+import pytest
+
+from utils import (
+    BIN_PATH,
+    MODEL_DEVICE_PATH,
+    MODEL_NAME,
+    PROMPT_DIR,
+    push_bundle_if_needed,
+    run_adb_command,
+    run_script,
+    write_qdc_log,
+)
+
+MODEL_URL = "<<MODEL_URL>>"
+
+
+@pytest.fixture(scope="session", autouse=True)
+def install(driver):
+    push_bundle_if_needed(f"{BIN_PATH}/llama-cli")
+    run_adb_command(f"mkdir -p /data/local/tmp/gguf {PROMPT_DIR}")
+    run_adb_command(f"echo 'What is the capital of France?' > {PROMPT_DIR}/bench_prompt.txt")
+    check = subprocess.run(
+        ["adb", "shell", f"ls {MODEL_DEVICE_PATH}"],
+        text=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT,
+    )
+    if check.returncode != 0:
+        run_adb_command(f'curl -L -J --output {MODEL_DEVICE_PATH} "{MODEL_URL}"')
+
+
+@pytest.mark.parametrize(
+    "device",
+    [
+        pytest.param("none", id="cpu"),
+        pytest.param("GPUOpenCL", id="gpu"),
+        pytest.param("HTP0", id="npu"),
+    ],
+)
+def test_llama_completion(device):
+    result = run_script(
+        "run-completion.sh",
+        extra_env={"D": device, "M": MODEL_NAME},
+        extra_args=["--batch-size", "128", "-n", "128", "--seed", "42",
+                    "-f", f"{PROMPT_DIR}/bench_prompt.txt"],
+    )
+    write_qdc_log(f"llama_completion_{device}.log", result.stdout or "")
+    assert result.returncode == 0, (
+        f"llama-completion {device} failed (exit {result.returncode})"
+    )
+
+
+_DEVICE_LOG_NAME = {"none": "cpu", "GPUOpenCL": "gpu", "HTP0": "htp"}
+
+
+@pytest.mark.parametrize(
+    "device",
+    [
+        pytest.param("none", id="cpu"),
+        pytest.param("GPUOpenCL", id="gpu"),
+        pytest.param("HTP0", id="npu"),
+    ],
+)
+def test_llama_bench(device):
+    result = run_script(
+        "run-bench.sh",
+        extra_env={"D": device, "M": MODEL_NAME},
+        extra_args=["--batch-size", "128", "-p", "128", "-n", "32"],
+    )
+    write_qdc_log(f"llama_bench_{_DEVICE_LOG_NAME[device]}.log", result.stdout or "")
+    assert result.returncode == 0, (
+        f"llama-bench {device} failed (exit {result.returncode})"
+    )
+
+
+if __name__ == "__main__":
+    ret = pytest.main(["-s", "--junitxml=results.xml", os.path.realpath(__file__)])
+    if os.path.exists("results.xml"):
+        with open("results.xml") as f:
+            write_qdc_log("results.xml", f.read())
+    sys.exit(ret)

scripts/snapdragon/qdc/tests/utils.py

@@ -0,0 +1,143 @@
+"""Shared helpers for QDC on-device test runners."""
+
+from __future__ import annotations
+
+import logging
+import os
+import subprocess
+import tempfile
+
+from appium.options.common import AppiumOptions
+
+log = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# On-device paths
+# ---------------------------------------------------------------------------
+
+BUNDLE_PATH = "/data/local/tmp/llama.cpp"
+BIN_PATH = f"{BUNDLE_PATH}/bin"
+LIB_PATH = f"{BUNDLE_PATH}/lib"
+QDC_LOGS_PATH = "/data/local/tmp/QDC_logs"
+SCRIPTS_DIR = "/qdc/appium"
+MODEL_NAME = "model.gguf"
+MODEL_DEVICE_PATH = "/data/local/tmp/gguf/model.gguf"
+PROMPT_DIR = "/data/local/tmp/scorecard_prompts"
+
+# ---------------------------------------------------------------------------
+# Appium session options
+# ---------------------------------------------------------------------------
+
+options = AppiumOptions()
+options.set_capability("automationName", "UiAutomator2")
+options.set_capability("platformName", "Android")
+options.set_capability("deviceName", os.getenv("ANDROID_DEVICE_VERSION"))
+
+# ---------------------------------------------------------------------------
+# Shell / process helpers
+# ---------------------------------------------------------------------------
+
+
+def write_qdc_log(filename: str, content: str) -> None:
+    """Write content as a log file for QDC log collection."""
+    subprocess.run(
+        ["adb", "shell", f"mkdir -p {QDC_LOGS_PATH}"],
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+    )
+    with tempfile.NamedTemporaryFile(mode="w", suffix=".log", delete=False) as f:
+        f.write(content)
+        tmp_path = f.name
+    try:
+        subprocess.run(
+            ["adb", "push", tmp_path, f"{QDC_LOGS_PATH}/{filename}"],
+            stdout=subprocess.PIPE,
+            stderr=subprocess.STDOUT,
+        )
+    finally:
+        os.unlink(tmp_path)
+
+
+def ensure_bundle(check_binary: str | None = None) -> None:
+    """Ensure the llama_cpp_bundle is available on the target device."""
+    push_bundle_if_needed(check_binary or f"{BIN_PATH}/llama-cli")
+
+
+# ---------------------------------------------------------------------------
+# Android / Linux host helpers
+# ---------------------------------------------------------------------------
+
+
+def run_adb_command(cmd: str, *, check: bool = True) -> subprocess.CompletedProcess:
+    """Run a command on-device via ``adb shell`` with exit-code sentinel."""
+    raw = subprocess.run(
+        ["adb", "shell", f"{cmd}; echo __RC__:$?"],
+        text=True,
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+    )
+    stdout = raw.stdout
+    returncode = raw.returncode
+    if stdout:
+        lines = stdout.rstrip("\n").split("\n")
+        if lines and lines[-1].startswith("__RC__:"):
+            try:
+                returncode = int(lines[-1][7:])
+                stdout = "\n".join(lines[:-1]) + "\n"
+            except ValueError:
+                pass
+    log.info(stdout)
+    result = subprocess.CompletedProcess(raw.args, returncode, stdout=stdout)
+    if check:
+        assert returncode == 0, f"Command failed (exit {returncode})"
+    return result
+
+
+def run_script(
+    script: str,
+    extra_env: dict[str, str] | None = None,
+    extra_args: list[str] | None = None,
+) -> subprocess.CompletedProcess:
+    """Run an upstream shell script from /qdc/appium/ on the QDC runner host."""
+    env = os.environ.copy()
+    env["GGML_HEXAGON_EXPERIMENTAL"] = "1"
+    if extra_env:
+        env.update(extra_env)
+    cmd = [f"{SCRIPTS_DIR}/{script}"] + (extra_args or [])
+    result = subprocess.run(
+        cmd, env=env,
+        text=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT,
+    )
+    log.info(result.stdout)
+    return result
+
+
+def adb_shell(cmd: str) -> None:
+    """Run a command via adb shell (fire-and-forget, no error check)."""
+    subprocess.run(
+        ["adb", "shell", "sh", "-c", cmd],
+        capture_output=True, encoding="utf-8", errors="replace", check=False,
+    )
+
+
+def push_bundle_if_needed(check_binary: str) -> None:
+    """Push llama_cpp_bundle to the device if check_binary is not already present."""
+    result = subprocess.run(
+        ["adb", "shell", f"ls {check_binary}"],
+        text=True,
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+    )
+    if result.returncode != 0:
+        subprocess.run(
+            ["adb", "push", "/qdc/appium/llama_cpp_bundle/", BUNDLE_PATH],
+            text=True,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.STDOUT,
+        )
+        subprocess.run(
+            ["adb", "shell", f"find {BUNDLE_PATH}/bin -type f -exec chmod 755 {{}} +"],
+            text=True,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.STDOUT,
+        )