chore: import upstream snapshot with attribution

scripts/snapdragon/adb/llama-cli.farf

@@ -0,0 +1 @@
+0xffff

scripts/snapdragon/adb/run-bench.sh

@@ -0,0 +1,49 @@
+#!/bin/sh
+#
+
+# Basedir on device
+basedir=/data/local/tmp/llama.cpp
+
+branch=.
+[ "$B" != "" ] && branch=$B
+
+adbserial=
+[ "$S" != "" ] && adbserial="-s $S"
+
+adbhost=
+[ "$H" != "" ] && adbhost="-H $H"
+
+model="Llama-3.2-3B-Instruct-Q4_0.gguf"
+[ "$M" != "" ] && model="$M"
+
+device="HTP0"
+[ "$D" != "" ] && device="$D"
+
+verbose=
+[ "$V" != "" ] && verbose="GGML_HEXAGON_VERBOSE=$V" cli_opts="$cli_opts -v"
+
+profile=
+[ "$PROF" != "" ] && profile="GGML_HEXAGON_PROFILE=$PROF" cli_opts="$cli_opts -v"
+
+opmask=
+[ "$OPSTAGE" != "" ] && opmask="GGML_HEXAGON_OPSTAGE=$OPSTAGE"
+
+nhvx=
+[ "$NHVX" != "" ] && nhvx="GGML_HEXAGON_NHVX=$NHVX"
+
+ndev=
+[ "$NDEV" != "" ] && ndev="GGML_HEXAGON_NDEV=$NDEV"
+
+hb=
+[ "$HB" != "" ] && hb="GGML_HEXAGON_HOSTBUF=$HB"
+
+set -x
+
+adb $adbserial $adbhost shell " \
+  cd $basedir;         \
+  LD_LIBRARY_PATH=$basedir/$branch/lib   \
+  ADSP_LIBRARY_PATH=$basedir/$branch/lib \
+    $ndev $nhvx $opmask $verbose $profile $hb ./$branch/bin/llama-bench --device $device --mmap 0 -m $basedir/../gguf/$model \
+        --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 \
+        --ubatch-size 1024 -fa 1 -ngl 99 $cli_opts $@   \
+"

scripts/snapdragon/adb/run-cli.sh

@@ -0,0 +1,78 @@
+#!/bin/sh
+#
+
+# Basedir on device
+basedir=/data/local/tmp/llama.cpp
+
+cli_opts=
+
+branch=.
+[ "$B" != "" ] && branch=$B
+
+adbserial=
+[ "$S" != "" ] && adbserial="-s $S"
+
+adbhost=
+[ "$H" != "" ] && adbhost="-H $H"
+
+model="Llama-3.2-3B-Instruct-Q4_0.gguf"
+[ "$M" != "" ] && model="$M"
+
+device="HTP0"
+[ "$D" != "" ] && device="$D"
+
+verbose=
+[ "$V" != "" ] && verbose="GGML_HEXAGON_VERBOSE=$V" cli_opts="$cli_opts -v"
+
+sched=
+[ "$SCHED" != "" ] && sched="GGML_SCHED_DEBUG=2" cli_opts="$cli_opts -v"
+
+profile=
+[ "$PROF" != "" ] && profile="GGML_HEXAGON_PROFILE=$PROF" cli_opts="$cli_opts -v"
+
+opmask=
+[ "$OPSTAGE" != "" ] && opmask="GGML_HEXAGON_OPSTAGE=$OPSTAGE"
+
+nhvx=
+[ "$NHVX" != "" ] && nhvx="GGML_HEXAGON_NHVX=$NHVX"
+
+hmx=
+[ "$HMX" != "" ] && hmx="GGML_HEXAGON_USE_HMX=$HMX"
+
+ndev=
+[ "$NDEV" != "" ] && ndev="GGML_HEXAGON_NDEV=$NDEV"
+
+hb=
+[ "$HB" != "" ] && hb="GGML_HEXAGON_HOSTBUF=$HB"
+
+opbatch=
+[ "$OB" != "" ] && opbatch="GGML_HEXAGON_OPBATCH=$OB"
+
+opqueue=
+[ "$OQ" != "" ] && opqueue="GGML_HEXAGON_OPQUEUE=$OQ"
+
+opflt=
+[ "$OF" != "" ] && opflt="GGML_HEXAGON_OPFILTER=$OF"
+
+vmem=
+[ "$VM" != "" ] && opflt="GGML_HEXAGON_VMEM=$VM"
+
+mbuf=
+[ "$MB" != "" ] && opflt="GGML_HEXAGON_MBUF=$MB"
+vmem=
+[ "$VM" != "" ] && vmem="GGML_HEXAGON_VMEM=$VM"
+
+mbuf=
+[ "$MB" != "" ] && mbuf="GGML_HEXAGON_MBUF=$MB"
+set -x
+
+adb $adbserial $adbhost shell " \
+  cd $basedir; ulimit -c unlimited;        \
+    LD_LIBRARY_PATH=$basedir/$branch/lib   \
+    ADSP_LIBRARY_PATH=$basedir/$branch/lib \
+    $verbose $sched $opmask $profile $nhvx $hmx $ndev $hb $opbatch $opqueue $opflt $vmem $mbuf \
+      ./$branch/bin/llama-cli --no-mmap -m $basedir/../gguf/$model \
+         --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1           \
+         --ctx-size 8192 --ubatch-size 1024 -fa on                 \
+         -ngl 99 --device $device $cli_opts $@                     \
+"

scripts/snapdragon/adb/run-completion.sh

@@ -0,0 +1,86 @@
+#!/bin/sh
+#
+
+# Basedir on device
+basedir=/data/local/tmp/llama.cpp
+
+cli_opts=
+
+branch=.
+[ "$B" != "" ] && branch=$B
+
+adbserial=
+[ "$S" != "" ] && adbserial="-s $S"
+
+adbhost=
+[ "$H" != "" ] && adbhost="-H $H"
+
+model="Llama-3.2-3B-Instruct-Q4_0.gguf"
+[ "$M" != "" ] && model="$M"
+
+device="HTP0"
+[ "$D" != "" ] && device="$D"
+
+verbose=
+[ "$V" != "" ] && verbose="GGML_HEXAGON_VERBOSE=$V" cli_opts="$cli_opts -v"
+
+sched=
+[ "$SCHED" != "" ] && sched="GGML_SCHED_DEBUG=2" cli_opts="$cli_opts -v"
+
+profile=
+[ "$PROF" != "" ] && profile="GGML_HEXAGON_PROFILE=$PROF" cli_opts="$cli_opts -v"
+
+opmask=
+[ "$OPSTAGE" != "" ] && opmask="GGML_HEXAGON_OPSTAGE=$OPSTAGE"
+
+nhvx=
+[ "$NHVX" != "" ] && nhvx="GGML_HEXAGON_NHVX=$NHVX"
+
+hmx=
+[ "$HMX" != "" ] && hmx="GGML_HEXAGON_USE_HMX=$HMX"
+
+ndev=
+[ "$NDEV" != "" ] && ndev="GGML_HEXAGON_NDEV=$NDEV"
+
+hb=
+[ "$HB" != "" ] && hb="GGML_HEXAGON_HOSTBUF=$HB"
+
+opbatch=
+[ "$OB" != "" ] && opbatch="GGML_HEXAGON_OPBATCH=$OB"
+
+opqueue=
+[ "$OQ" != "" ] && opqueue="GGML_HEXAGON_OPQUEUE=$OQ"
+
+oppoll=
+[ "$OP" != "" ] && oppoll="GGML_HEXAGON_OPPOLL=$OP"
+
+opflt=
+[ "$OF" != "" ] && opflt="GGML_HEXAGON_OPFILTER=$OF"
+
+opfuse=
+[ "$OC" != "" ] && opfuse="GGML_HEXAGON_OPFUSION=$OC"
+
+vmem=
+[ "$VM" != "" ] && vmem="GGML_HEXAGON_VMEM=$VM"
+
+mbuf=
+[ "$MB" != "" ] && mbuf="GGML_HEXAGON_MBUF=$MB"
+
+mmsel=
+[ "$MM" != "" ] && mmsel="GGML_HEXAGON_MM_SELECT=$MM"
+
+fasel=
+[ "$FA" != "" ] && fasel="GGML_HEXAGON_FA_SELECT=$FA"
+
+set -x
+
+adb $adbserial $adbhost shell " \
+  cd $basedir; ulimit -c unlimited;        \
+    LD_LIBRARY_PATH=$basedir/$branch/lib   \
+    ADSP_LIBRARY_PATH=$basedir/$branch/lib \
+    $verbose $sched $opmask $profile $nhvx $hmx $ndev $hb $opbatch $opqueue $oppoll $opflt $opfuse $vmem $mbuf $mmsel $fasel \
+      ./$branch/bin/llama-completion --no-mmap -m $basedir/../gguf/$model \
+         --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1                  \
+         --ctx-size 8192 --ubatch-size 1024 -fa on                        \
+         -ngl 99 --device $device $cli_opts $@                            \
+"

scripts/snapdragon/adb/run-mtmd.sh

@@ -0,0 +1,71 @@
+#!/bin/sh
+#
+
+# Basedir on device
+basedir=/data/local/tmp/llama.cpp
+
+cli_opts=
+
+branch=.
+[ "$B" != "" ] && branch=$B
+
+adbserial=
+[ "$S" != "" ] && adbserial="-s $S"
+
+adbhost=
+[ "$H" != "" ] && adbhost="-H $H"
+
+model="gemma-3-4b-it-Q4_0.gguf"
+[ "$M" != "" ] && model="$M"
+
+mmproj="mmproj-F16.gguf"
+[ "$MMPROJ" != "" ] && mmproj="$MMPROJ"
+
+image=
+[ "$IMG" != "" ] && image="$IMG"
+
+device="HTP0"
+[ "$D" != "" ] && device="$D"
+
+verbose=
+[ "$V" != "" ] && verbose="GGML_HEXAGON_VERBOSE=$V"
+
+experimental="GGML_HEXAGON_EXPERIMENTAL=1"
+[ "$E" != "" ] && experimental="GGML_HEXAGON_EXPERIMENTAL=$E"
+
+sched=
+[ "$SCHED" != "" ] && sched="GGML_SCHED_DEBUG=2" cli_opts="$cli_opts -v"
+
+profile=
+[ "$PROF" != "" ] && profile="GGML_HEXAGON_PROFILE=$PROF"
+
+opmask=
+[ "$OPSTAGE" != "" ] && opmask="GGML_HEXAGON_OPSTAGE=$OPSTAGE"
+
+nhvx=
+[ "$NHVX" != "" ] && nhvx="GGML_HEXAGON_NHVX=$NHVX"
+
+hmx=
+[ "$HMX" != "" ] && hmx="GGML_HEXAGON_USE_HMX=$HMX"
+
+ndev=
+[ "$NDEV" != "" ] && ndev="GGML_HEXAGON_NDEV=$NDEV"
+
+# MTMD backend device for vision model (defaults to CPU if not set)
+mtmd_backend=
+[ "$MTMD_DEVICE" != "" ] && mtmd_backend="MTMD_BACKEND_DEVICE=$MTMD_DEVICE"
+
+set -x
+
+adb $adbserial $adbhost shell " \
+  cd $basedir; ulimit -c unlimited;        \
+    LD_LIBRARY_PATH=$basedir/$branch/lib   \
+    ADSP_LIBRARY_PATH=$basedir/$branch/lib \
+    $verbose $experimental $sched $opmask $profile $hmx $nhvx $ndev $mtmd_backend \
+      ./$branch/bin/llama-mtmd-cli --no-mmap -m $basedir/../gguf/$model      \
+         --mmproj $basedir/../gguf/$mmproj                                   \
+         --image $basedir/../gguf/$image                                     \
+         --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1                     \
+         --ctx-size 8192 --ubatch-size 1024 -fa on                           \
+         -ngl 99 --device $device -v $cli_opts $@                            \
+"

scripts/snapdragon/adb/run-tool.sh

@@ -0,0 +1,72 @@
+#!/bin/sh
+#
+
+# Basedir on device
+basedir=/data/local/tmp/llama.cpp
+
+cli_opts=
+
+branch=.
+[ "$B" != "" ] && branch=$B
+
+adbserial=
+[ "$S" != "" ] && adbserial="-s $S"
+
+adbhost=
+[ "$H" != "" ] && adbhost="-H $H"
+
+device="HTP0"
+[ "$D" != "" ] && device="$D"
+
+verbose=
+[ "$V" != "" ] && verbose="GGML_HEXAGON_VERBOSE=$V"
+
+sched=
+[ "$SCHED" != "" ] && sched="GGML_SCHED_DEBUG=2" cli_opts="$cli_opts -v"
+
+profile=
+[ "$PROF" != "" ] && profile="GGML_HEXAGON_PROFILE=$PROF"
+
+opmask=
+[ "$OPSTAGE" != "" ] && opmask="GGML_HEXAGON_OPSTAGE=$OPSTAGE"
+
+nhvx=
+[ "$NHVX" != "" ] && nhvx="GGML_HEXAGON_NHVX=$NHVX"
+
+hmx=
+[ "$HMX" != "" ] && hmx="GGML_HEXAGON_USE_HMX=$HMX"
+
+ndev=
+[ "$NDEV" != "" ] && ndev="GGML_HEXAGON_NDEV=$NDEV"
+
+hb=
+[ "$HB" != "" ] && hb="GGML_HEXAGON_HOSTBUF=$HB"
+
+opbatch=
+[ "$OB" != "" ] && opbatch="GGML_HEXAGON_OPBATCH=$OB"
+
+opqueue=
+[ "$OQ" != "" ] && opqueue="GGML_HEXAGON_OPQUEUE=$OQ"
+
+oppoll=
+[ "$OP" != "" ] && oppoll="GGML_HEXAGON_OPPOLL=$OP"
+
+opfuse=
+[ "$OC" != "" ] && opfuse="GGML_HEXAGON_OPFUSION=$OC"
+
+mmsel=
+[ "$MM" != "" ] && mmsel="GGML_HEXAGON_MM_SELECT=$MM"
+
+fasel=
+[ "$FA" != "" ] && fasel="GGML_HEXAGON_FA_SELECT=$FA"
+
+set -x
+
+tool=$1; shift
+
+adb $adbserial $adbhost shell " \
+  cd $basedir; ulimit -c unlimited;        \
+    LD_LIBRARY_PATH=$basedir/$branch/lib   \
+    ADSP_LIBRARY_PATH=$basedir/$branch/lib \
+    $verbose $sched $opmask $profile $nhvx $hmx $ndev $hb $opbatch $opqueue $oppoll $opfuse $mmsel $fasel ./$branch/bin/$tool $@ \
+"

scripts/snapdragon/ggml-hexagon-profile.py

@@ -0,0 +1,475 @@
+#!/usr/bin/env python3
+
+import sys
+import os
+import re
+import argparse
+import statistics
+import logging
+from typing import Any, Dict, List, Optional
+
+from collections import defaultdict
+
+# Mapping of cli-friendly names to (internal_data_key, Display Header, numeric_sort_key)
+COL_MAP = {
+    "tot-usec":   ("tot_usec",   "Tot usec",   "_sort_tot_usec"),
+    "op":         ("op",         "Op",         "op"),
+    "dims":       ("dims",       "Dims",       "dims"),
+    "dtypes":     ("dtypes",     "DTypes",     "dtypes"),
+    "count":      ("count",      "Count",      "_sort_count"),
+    "max-usec":   ("max_usec",   "Max usec",   "_sort_max_usec"),
+    "avg-usec":   ("avg_usec",   "Avg usec",   "_sort_avg_usec"),
+    "max-cycles": ("max_cycles", "Max Cycles", "_sort_max_cycles"),
+    "avg-cycles": ("avg_cycles", "Avg Cycles", "_sort_avg_cycles"),
+    "max-pmu":    ("max_pmu",    "Max PMU",    "_sort_max_pmu"),
+    "avg-pmu":    ("avg_pmu",    "Avg PMU",    "_sort_avg_pmu"),
+}
+
+op_pattern = re.compile(
+    r"profile-op\s+(?P<op_name>[A-Z_0-9+]+):\s+.*?\s+:\s+(?P<dims>[\d:x\s\->!]+)\s+:\s+(?P<types>[a-z\d_\s\->x]+)\s+:\s+.*?\s+:\s+(?:op-)?usec\s+(?P<usec>\d+)\s+(?:op-)?cycles\s+(?P<cycles>\d+)(?:\s+start\s+(?P<start>\d+))?(?:\s+mhz\s+(?P<mhz>[\d.]+))?(?:\s+pmu\s+\[(?P<pmu>[\d,\s]+)\])?(?:\s+evt\s+\[(?P<evt>[\d,\s]+)\])?"
+)
+
+trace_pattern = re.compile(
+    r"trace-op\s+(?P<op_name>[A-Z_0-9+]+):\s+thread\s+(?P<thread>\d+)\s+event\s+(?P<event>[A-Z_0-9\-]+)\s+info\s+(?P<info>\d+)\s+(?P<state>start|stop)\s+(?P<cycles>\d+)"
+)
+
+logger = logging.getLogger("ggml-hexagon-profile")
+
+
+def normalize_event_name(evt_type):
+    if evt_type == "HVX_COMP":
+        return "V-COMP"
+    if evt_type == "HMX_COMP":
+        return "M-COMP"
+
+    # Strip HVX_ or HMX_ prefixes
+    name = evt_type
+    if name.startswith("HVX_") or name.startswith("HMX_"):
+        name = name[4:]
+    return name.replace("_", "-")
+
+
+class CycleUnwrapper:
+    def __init__(self):
+        self.last_raw = None
+        self.high_part = 0
+
+    def unwrap(self, raw):
+        if self.last_raw is None:
+            self.last_raw = raw
+            return raw
+        diff = raw - self.last_raw
+        if diff < -0x80000000:
+            self.high_part += 0x100000000
+        elif diff > 0x80000000:
+            self.high_part -= 0x100000000
+        self.last_raw = raw
+        return raw + self.high_part
+
+
+def parse_log(file_path, pmu_index=None):
+    try:
+        if file_path != "-":
+            f = open(file_path, 'r', encoding='utf-8', errors='ignore')
+        else:
+            f = os.fdopen(0, 'r', encoding='utf-8', errors='ignore')
+    except FileNotFoundError:
+        logger.error(f"file '{file_path}' not found.")
+        sys.exit(1)
+
+    all_ops: List[Dict[str, Any]] = []
+    current_op: Optional[Dict[str, Any]] = None
+
+    timestamp_pattern = re.compile(r"^(?P<min>\d+)\.(?P<sec>\d+)\.(?P<ms>\d+)\.(?P<us>\d+)\s+[A-Z]\s+")
+    unwrapper = CycleUnwrapper()
+
+    for line in f:
+        ts_match = timestamp_pattern.match(line)
+        abs_usec = 0
+        if ts_match:
+            abs_usec = (
+                (int(ts_match.group('min')) * 60 + int(ts_match.group('sec'))) * 1000000
+                + int(ts_match.group('ms')) * 1000
+                + int(ts_match.group('us'))
+            )
+
+        if "|" in line and "profile-op" in line:
+            parts = [p.strip() for p in line.split("|")]
+            prefix = parts[0]
+            prefix_match = re.search(r"profile-op\s+(?P<op_name>[A-Z_0-9+]+)", prefix)
+            if not prefix_match:
+                continue
+
+            if len(parts) == 7:
+                dims, types, timings = parts[2], parts[3], parts[6]
+            elif len(parts) == 6:
+                dims, types, timings = parts[2], parts[3], parts[5]
+            else:
+                continue
+
+            timing_match = re.search(
+                r"(?:op-)?usec\s+(?P<usec>\d+)\s+(?:op-)?cycles\s+(?P<cycles>\d+)(?:\s+start\s+(?P<start>\d+))?(?:\s+mhz\s+(?P<mhz>[\d.]+))?(?:\s+pmu\s+\[(?P<pmu>[\d,\s]+)\])?(?:\s+evt\s+\[(?P<evt>[\d,\s]+)\])?",
+                timings
+            )
+            if not timing_match:
+                continue
+
+            op_match = timing_match
+            op_name = prefix_match.group("op_name")
+        else:
+            op_match = op_pattern.search(line)
+            if op_match:
+                op_name = op_match.group('op_name')
+                dims = op_match.group('dims').strip()
+                types = op_match.group('types').strip()
+            else:
+                op_match = None
+
+        if op_match:
+            pmu_raw = op_match.group('pmu') if 'pmu' in op_match.groupdict() else None
+            pmu_val = None
+            if pmu_raw and pmu_index is not None:
+                try:
+                    pmu_list = [int(x.strip()) for x in pmu_raw.split(',')]
+                    if len(pmu_list) > pmu_index:
+                        pmu_val = pmu_list[pmu_index]
+                except (ValueError, IndexError):
+                    pmu_val = None
+
+            evt_raw = op_match.group('evt') if 'evt' in op_match.groupdict() else None
+            evt_val = None
+            if evt_raw:
+                try:
+                    evt_val = [int(x.strip()) for x in evt_raw.split(',')]
+                except ValueError:
+                    evt_val = None
+
+            cycles_start_raw = op_match.group('start')
+            unwrapped_cycles_start = None
+            if cycles_start_raw:
+                unwrapped_cycles_start = unwrapper.unwrap(int(cycles_start_raw))
+
+            idx = line.find("profile-op ")
+            op_text = line[idx + 11:].strip() if idx != -1 else line.strip()
+
+            current_op = {
+                'name':         op_name,
+                'dims':         dims,
+                'types':        types,
+                'op_text':      op_text,
+                'usec':         int(op_match.group('usec')),
+                'cycles':       int(op_match.group('cycles')),
+                'cycles_start': int(cycles_start_raw) if cycles_start_raw else None,
+                'unwrapped_cycles_start': unwrapped_cycles_start,
+                'pmu_val':      pmu_val,
+                'evt_val':      evt_val,
+                'abs_usec':     abs_usec,
+                'trace_events': []
+            }
+            all_ops.append(current_op)
+            continue
+
+        trace_match = trace_pattern.search(line)
+        if trace_match and current_op:
+            if trace_match.group('op_name') == current_op['name']:
+                raw_cyc = int(trace_match.group('cycles'))
+                current_op['trace_events'].append({
+                    'thread': int(trace_match.group('thread')),
+                    'event':  trace_match.group('event'),
+                    'info':   int(trace_match.group('info')),
+                    'cycles': raw_cyc,
+                    'unwrapped_cycles': unwrapper.unwrap(raw_cyc),
+                    'state':  trace_match.group('state')
+                })
+
+    f.close()
+    return all_ops
+
+
+def print_ascii_timeline(op_name, dims, types, usec, cycles, events, evt_val=None):
+    evt_str = ""
+    if evt_val:
+        evt_str = " - evt [" + ",".join(str(x) for x in evt_val) + "]"
+    logger.info("=" * 100)
+    logger.info(f"{op_name} ({dims} : {types}) - {usec} usec {cycles} cycles{evt_str}")
+    logger.info("=" * 100)
+
+    events = sorted(events, key=lambda e: e['cycles'])
+    if not events:
+        logger.info("  No trace events recorded.")
+        return
+
+    min_cycles = events[0]['cycles']
+
+    logger.info("Cycles      %-30s" % "EventDetails" + " ".join(f"T{i:<2}" for i in range(10)) + " HMX")
+    logger.info("-" * 100)
+
+    thread_stacks = [[] for _ in range(11)]
+
+    for e in events:
+        t = e['thread']
+        if t < 0 or t > 10:
+            continue
+
+        if e['cycles'] >= min_cycles:
+            rel_cycles = e['cycles'] - min_cycles
+        else:
+            rel_cycles = (e['cycles'] + 0x100000000) - min_cycles
+
+        state = e['state']
+        evt_type = e['event']
+
+        # Determine char representing the event
+        norm_evt = normalize_event_name(evt_type)
+        char = '?'
+        if norm_evt == 'V-COMP':
+            char = 'V'
+        elif norm_evt == 'M-COMP':
+            char = 'H'
+        elif norm_evt == 'A-QUANT':
+            char = 'Q'
+        elif norm_evt == 'A-PREP':
+            char = 'A'
+        elif norm_evt == 'Q-PREP':
+            char = 'q'
+        elif norm_evt == 'K-PREP':
+            char = 'k'
+        elif norm_evt == 'V-PREP':
+            char = 'v'
+        elif norm_evt == 'W-DEQUANT':
+            char = 'D'
+        elif norm_evt == 'O-PROC':
+            char = 'O'
+        elif norm_evt == 'W-PREP':
+            char = 'P'
+        elif norm_evt == 'DMA':
+            char = 'M'
+
+        if state == 'start':
+            thread_stacks[t].append(char)
+        elif state == 'stop':
+            if thread_stacks[t]:
+                if thread_stacks[t][-1] == char:
+                    thread_stacks[t].pop()
+                elif char in thread_stacks[t]:
+                    thread_stacks[t].remove(char)
+                else:
+                    thread_stacks[t].pop()
+
+        cols = []
+        for i in range(11):
+            if thread_stacks[i]:
+                cols.append(f"[{thread_stacks[i][-1]}]")
+            else:
+                cols.append(" | ")
+
+        evt_desc = f"T{t}: {evt_type} {state} ({e['info']})"
+        logger.info(f"{rel_cycles:10d}  %-30s" % evt_desc + " ".join(cols[:10]) + "  " + cols[10])
+    logger.info("-" * 100)
+
+
+def print_ascii_summary(op_name, dims, types, usec, cycles, events, evt_val=None):
+    evt_str = ""
+    if evt_val:
+        evt_str = " - evt [" + ",".join(str(x) for x in evt_val) + "]"
+    logger.info("=" * 100)
+    logger.info(f"{op_name} ({dims} : {types}) - {usec} usec {cycles} cycles{evt_str}")
+    logger.info("=" * 100)
+
+    events = sorted(events, key=lambda e: e['cycles'])
+    if not events:
+        logger.info("  No trace events recorded.")
+        return
+
+    active_starts = {}
+    thread_totals = defaultdict(lambda: defaultdict(int))
+
+    for e in events:
+        t = e['thread']
+        evt = e['event']
+        info = e['info']
+        cyc = e['cycles']
+        state = e['state']
+
+        key = (t, evt, info)
+        if state == 'start':
+            active_starts[key] = cyc
+        elif state == 'stop':
+            if key in active_starts:
+                start_cyc = active_starts[key]
+                del active_starts[key]
+
+                if cyc >= start_cyc:
+                    dur = cyc - start_cyc
+                else:
+                    dur = (cyc + 0x100000000) - start_cyc
+
+                norm_evt = normalize_event_name(evt)
+                thread_totals[t][norm_evt] += dur
+
+    for t in sorted(thread_totals.keys()):
+        thread_name = f"Thread {t} (HVX)" if t != 10 else "Thread 10 (HMX)"
+        sorted_evts = sorted(thread_totals[t].items(), key=lambda item: item[0])
+
+        evt_strs = []
+        for evt, dur in sorted_evts:
+            pct = (dur / cycles * 100) if cycles > 0 else 0
+            evt_strs.append(f"{evt} {dur} ({pct:.1f}%)")
+
+        logger.info(f"  {thread_name:<16}: " + " | ".join(evt_strs))
+
+
+def generate_report(ops, top_n, width_overrides, sort_col, pmu_name=None):
+    if not ops:
+        logger.info("No valid records found.")
+        return
+
+    grouped = defaultdict(list)
+    for op in ops:
+        key = (op['name'], op['dims'], op['types'])
+        grouped[key].append(op)
+
+    group_stats = []
+    for (name, dims, types), group_ops in grouped.items():
+        usecs = [o['usec'] for o in group_ops]
+        cycles = [o['cycles'] for o in group_ops]
+        pmu_vals = [o['pmu_val'] for o in group_ops if o['pmu_val'] is not None]
+
+        avg_usec_val = statistics.mean(usecs)
+        count_val = len(group_ops)
+        tot_usec_val = avg_usec_val * count_val
+
+        group_stats.append({
+            'op':               name,
+            'dims':             dims,
+            'dtypes':           types,
+            'count':            str(count_val),
+            'max_usec':         str(max(usecs)),
+            'avg_usec':         f"{avg_usec_val:.2f}",
+            'tot_usec':         f"{tot_usec_val:.2f}",
+            'max_cycles':       str(max(cycles)),
+            'avg_cycles':       f"{statistics.mean(cycles):.2f}",
+            'max_pmu':          str(max(pmu_vals)) if pmu_vals else "0",
+            'avg_pmu':          f"{statistics.mean(pmu_vals):.2f}" if pmu_vals else "0.00",
+            # Numeric values for accurate sorting
+            '_sort_count':      count_val,
+            '_sort_max_usec':   max(usecs),
+            '_sort_avg_usec':   avg_usec_val,
+            '_sort_tot_usec':   tot_usec_val,
+            '_sort_max_cycles': max(cycles),
+            '_sort_avg_cycles': statistics.mean(cycles),
+            '_sort_max_pmu':    max(pmu_vals) if pmu_vals else 0,
+            '_sort_avg_pmu':    statistics.mean(pmu_vals) if pmu_vals else 0
+        })
+
+    # Sorting logic
+    actual_sort_key = COL_MAP[sort_col][2]
+    is_numeric    = actual_sort_key.startswith("_") or actual_sort_key == "count"
+    sorted_groups = sorted(group_stats, key=lambda x: x[actual_sort_key], reverse=is_numeric)[:top_n]
+
+    # Define initial column order
+    active_cols = ["op", "dims", "dtypes"]
+    if pmu_name:
+        active_cols += ["max-pmu", "avg-pmu"]
+    active_cols += ["tot-usec", "avg-usec", "avg-cycles", "max-usec", "max-cycles", "count"]
+
+    final_headers, final_keys, final_widths = [], [], []
+
+    for col_name in active_cols:
+        data_key, header_text, _ = COL_MAP[col_name]
+        if "pmu" in col_name and pmu_name:
+            header_text = header_text.replace("PMU", pmu_name)
+
+        natural_width = max([len(str(row[data_key])) for row in sorted_groups] + [len(header_text)])
+        target_width  = width_overrides.get(col_name, natural_width)
+
+        if target_width == 0:
+            continue
+
+        final_headers.append(header_text)
+        final_keys.append(data_key)
+        final_widths.append(target_width)
+
+    # Print Report
+    logger.info(f"\n# Profile Report (Top {top_n} Ops sorted by {sort_col})\n")
+    header_line = "| " + " | ".join(f"{h:<{final_widths[i]}}" for i, h in enumerate(final_headers)) + " |"
+    sep_line    = "| " + " | ".join("-" * final_widths[i] for i in range(len(final_headers))) + " |"
+    logger.info(header_line)
+    logger.info(sep_line)
+
+    for group in sorted_groups:
+        row_vals = []
+        for i, key in enumerate(final_keys):
+            val = str(group[key])
+            if len(val) > final_widths[i]:
+                val = val[:final_widths[i] - 3] + "..."
+            row_vals.append(f"{val:<{final_widths[i]}}")
+        logger.info("| " + " | ".join(row_vals) + " |")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Post-process Op profile info.")
+    parser.add_argument("logfile")
+    parser.add_argument("-n", "--top", type=int, default=100)
+    parser.add_argument("--sort", type=str, default="tot-usec", choices=list(COL_MAP.keys()))
+    parser.add_argument("--pmu-index", type=int)
+    parser.add_argument("--pmu-name", type=str)
+    parser.add_argument("--width", action='append', default=['dims:40'], help="Override column width, e.g. --width dims:50")
+    parser.add_argument("--timeline", type=str, nargs='?', const='summary', choices=["summary", "diagram"],
+                        help="Output ASCII art event summary or timing diagram (default: summary)")
+    parser.add_argument("--filter", type=str, help="Regex filter matching against the original profile-op line")
+
+    group = parser.add_mutually_exclusive_group()
+    group.add_argument("--head", type=int, help="Limit to first N ops")
+    group.add_argument("--tail", type=int, help="Limit to last N ops")
+
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.INFO, format='%(message)s')
+
+    if "pmu" in args.sort and args.pmu_index is None:
+        logger.error(f"Cannot sort by '{args.sort}' without --pmu-index.")
+        sys.exit(1)
+
+    overrides = {}
+    if args.width:
+        for w in args.width:
+            try:
+                name, val = w.split(':')
+                overrides[name.lower()] = int(val)
+            except ValueError:
+                logger.warning(f"Invalid width format '{w}'")
+
+    final_pmu_name = (args.pmu_name or f"#{args.pmu_index}") if args.pmu_index is not None else None
+    ops = parse_log(args.logfile, pmu_index=args.pmu_index)
+
+    if args.filter:
+        try:
+            filter_re = re.compile(args.filter)
+        except re.error as e:
+            logger.error(f"Invalid regex filter: {e}")
+            sys.exit(1)
+        ops = [op for op in ops if filter_re.search(op['op_text'])]
+
+    if args.head is not None:
+        ops = ops[:args.head]
+    elif args.tail is not None:
+        ops = ops[-args.tail:]
+
+    if args.timeline:
+        logger.info(f"\n# ASCII Timing {args.timeline.capitalize()}\n")
+        printed_cnt = 0
+        for op in ops:
+            if args.timeline == "summary":
+                print_ascii_summary(op['name'], op['dims'], op['types'], op['usec'], op['cycles'], op['trace_events'], op.get('evt_val'))
+            elif args.timeline == "diagram":
+                print_ascii_timeline(op['name'], op['dims'], op['types'], op['usec'], op['cycles'], op['trace_events'], op.get('evt_val'))
+            printed_cnt += 1
+            if printed_cnt >= args.top:
+                break
+    else:
+        generate_report(ops, args.top, overrides, args.sort, pmu_name=final_pmu_name)
+
+
+if __name__ == "__main__":
+    main()

scripts/snapdragon/ggml-hexagon-trace.py

@@ -0,0 +1,499 @@
+#!/usr/bin/env python3
+
+import sys
+import os
+import re
+import argparse
+import statistics
+import logging
+from typing import Any, Dict, List, Optional
+from collections import defaultdict
+
+logger = logging.getLogger("ggml-hexagon-trace")
+
+op_pattern = re.compile(
+    r"profile-op\s+(?P<op_name>[A-Z_0-9+]+):\s+.*?\s+:\s+(?P<dims>[\d:x\s\->!]+)\s+:\s+(?P<types>[a-z\d_\s\->x]+)\s+:\s+(?P<strides>[\d:x\s\->!]+?)\s+:\s+(?:(?P<params>.*?)\s+:\s+)?(?:op-)?usec\s+(?P<usec>\d+)\s+(?:op-)?cycles\s+(?P<cycles>\d+)(?:\s+start\s+(?P<start>\d+))?(?:\s+mhz\s+(?P<mhz>[\d.]+))?(?:\s+pmu\s+\[(?P<pmu>[\d,\s]+)\])?(?:\s+evt\s+\[(?P<evt>[\d,\s]+)\])?"
+)
+
+trace_pattern = re.compile(
+    r"trace-op\s+(?P<op_name>[A-Z_0-9+]+):\s+thread\s+(?P<thread>\d+)\s+event\s+(?P<event>[A-Z_0-9\-]+)\s+info\s+(?P<info>\d+)\s+(?P<state>start|stop)\s+(?P<cycles>\d+)"
+)
+
+
+def normalize_event_name(evt_type):
+    if evt_type == "HVX_COMP":
+        return "V-COMP"
+    if evt_type == "HMX_COMP":
+        return "M-COMP"
+    name = evt_type
+    if name.startswith("HVX_") or name.startswith("HMX_"):
+        name = name[4:]
+    return name.replace("_", "-")
+
+
+class CycleUnwrapper:
+    def __init__(self):
+        self.last_raw = None
+        self.high_part = 0
+
+    def unwrap(self, raw):
+        if self.last_raw is None:
+            self.last_raw = raw
+            return raw
+        diff = raw - self.last_raw
+        if diff < -0x80000000:
+            self.high_part += 0x100000000
+        elif diff > 0x80000000:
+            self.high_part -= 0x100000000
+        self.last_raw = raw
+        return raw + self.high_part
+
+
+def parse_log(file_path):
+    try:
+        if file_path != "-":
+            f = open(file_path, 'r', encoding='utf-8', errors='ignore')
+        else:
+            f = os.fdopen(0, 'r', encoding='utf-8', errors='ignore')
+    except FileNotFoundError:
+        logger.error(f"file '{file_path}' not found.")
+        sys.exit(1)
+
+    all_ops: List[Dict[str, Any]] = []
+    current_op: Optional[Dict[str, Any]] = None
+    unwrapper = CycleUnwrapper()
+    line_idx = 0
+
+    for line in f:
+        line_idx += 1
+        if "|" in line and "profile-op" in line:
+            parts = [p.strip() for p in line.split("|")]
+            prefix = parts[0]
+            prefix_match = re.search(r"profile-op\s+(?P<op_name>[A-Z_0-9+]+)", prefix)
+            if not prefix_match:
+                continue
+
+            if len(parts) == 7:
+                dims, types, strides, params, timings = parts[2], parts[3], parts[4], parts[5], parts[6]
+            elif len(parts) == 6:
+                dims, types, strides, params, timings = parts[2], parts[3], parts[4], "", parts[5]
+            else:
+                continue
+
+            timing_match = re.search(
+                r"(?:op-)?usec\s+(?P<usec>\d+)\s+(?:op-)?cycles\s+(?P<cycles>\d+)(?:\s+start\s+(?P<start>\d+))?(?:\s+mhz\s+(?P<mhz>[\d.]+))?(?:\s+pmu\s+\[(?P<pmu>[\d,\s]+)\])?(?:\s+evt\s+\[(?P<evt>[\d,\s]+)\])?",
+                timings
+            )
+            if not timing_match:
+                continue
+
+            op_match = timing_match
+            op_name = prefix_match.group("op_name")
+        else:
+            op_match = op_pattern.search(line)
+            if op_match:
+                op_name = op_match.group('op_name')
+                dims = op_match.group('dims').strip() if op_match.group('dims') else ''
+                types = op_match.group('types').strip() if op_match.group('types') else ''
+                strides = op_match.group('strides').strip() if op_match.group('strides') else ''
+                params = op_match.group('params').strip() if ('params' in op_match.groupdict() and op_match.group('params')) else ''
+            else:
+                op_match = None
+
+        if op_match:
+            cycles_start_raw = op_match.group('start')
+            unwrapped_cycles_start = None
+            if cycles_start_raw:
+                unwrapped_cycles_start = unwrapper.unwrap(int(cycles_start_raw))
+
+            idx = line.find("profile-op ")
+            op_text = line[idx + 11:].strip() if idx != -1 else line.strip()
+
+            current_op = {
+                'name':         op_name,
+                'dims':         dims,
+                'types':        types,
+                'strides':      strides,
+                'params':       params,
+                'op_text':      op_text,
+                'usec':         int(op_match.group('usec')),
+                'cycles':       int(op_match.group('cycles')),
+                'cycles_start': int(cycles_start_raw) if cycles_start_raw else None,
+                'unwrapped_cycles_start': unwrapped_cycles_start,
+                'trace_events': [],
+                'line_num':     line_idx
+            }
+            all_ops.append(current_op)
+            continue
+
+        trace_match = trace_pattern.search(line)
+        if trace_match and current_op:
+            if trace_match.group('op_name') == current_op['name']:
+                raw_cyc = int(trace_match.group('cycles'))
+                current_op['trace_events'].append({
+                    'thread': int(trace_match.group('thread')),
+                    'event':  trace_match.group('event'),
+                    'info':   int(trace_match.group('info')),
+                    'cycles': raw_cyc,
+                    'unwrapped_cycles': unwrapper.unwrap(raw_cyc),
+                    'state':  trace_match.group('state')
+                })
+
+    f.close()
+    return all_ops
+
+# --- Simple protobuf encoder ---
+
+
+def write_varint(val):
+    if val < 0:
+        val = (1 << 64) + val
+    res = bytearray()
+    while True:
+        towrite = val & 0x7f
+        val >>= 7
+        if val > 0:
+            res.append(towrite | 0x80)
+        else:
+            res.append(towrite)
+            break
+    return bytes(res)
+
+
+def pb_field(num, wire, data):
+    return write_varint((num << 3) | wire) + data
+
+
+def pb_varint(num, val):
+    return pb_field(num, 0, write_varint(val))
+
+
+def pb_length_delimited(num, data):
+    return pb_field(num, 2, write_varint(len(data)) + data)
+
+
+def pb_string(num, text):
+    return pb_length_delimited(num, text.encode('utf-8'))
+
+
+# Message Encoders
+def make_process_descriptor(pid, name):
+    return pb_varint(1, pid) + pb_string(6, name)
+
+
+def make_thread_descriptor(pid, tid, name, sort_index=None):
+    payload = pb_varint(1, pid) + pb_varint(2, tid) + pb_string(5, name)
+    if sort_index is not None:
+        payload += pb_varint(3, sort_index)
+    return payload
+
+
+def make_track_descriptor(uuid, name=None, parent_uuid=None, thread=None, process=None, sibling_merge_behavior=None, child_ordering=None, sibling_order_rank=None):
+    payload = pb_varint(1, uuid)
+    if name is not None:
+        payload += pb_string(2, name)
+    if parent_uuid is not None:
+        payload += pb_varint(5, parent_uuid)
+    if process is not None:
+        payload += pb_length_delimited(3, process)
+    if thread is not None:
+        payload += pb_length_delimited(4, thread)
+    if sibling_merge_behavior is not None:
+        payload += pb_varint(15, sibling_merge_behavior)
+    if child_ordering is not None:
+        payload += pb_varint(11, child_ordering)
+    if sibling_order_rank is not None:
+        payload += pb_varint(12, sibling_order_rank)
+    return payload
+
+
+def make_debug_annotation(name, string_val=None, int_val=None):
+    payload = pb_string(10, name)
+    if string_val is not None:
+        payload += pb_string(6, string_val)
+    elif int_val is not None:
+        payload += pb_varint(4, int_val)
+    return payload
+
+
+def make_track_event(event_type, track_uuid, name=None, category=None, debug_annotations=None):
+    payload = pb_varint(9, event_type)
+    payload += pb_varint(11, track_uuid)
+    if name is not None:
+        payload += pb_string(23, name)
+    if category is not None:
+        payload += pb_string(22, category)
+    if debug_annotations is not None:
+        for da in debug_annotations:
+            payload += pb_length_delimited(4, da)
+    return payload
+
+
+def make_trace_packet(timestamp, track_event=None, track_descriptor=None, seq_id=1):
+    payload = pb_varint(8, timestamp)
+    payload += pb_varint(10, seq_id)
+    if track_event is not None:
+        payload += pb_length_delimited(11, track_event)
+    if track_descriptor is not None:
+        payload += pb_length_delimited(60, track_descriptor)
+    return payload
+
+
+def write_trace_packet_to_file(f, packet_bytes):
+    # Write as field 1 of top-level Trace message
+    f.write(pb_length_delimited(1, packet_bytes))
+
+# --- End Protobuf Encoder ---
+
+
+def generate_perfetto_trace(filtered_ops, output_path):
+    if not filtered_ops:
+        logger.warning("No operators found after filtering.")
+        return
+
+    # Compute average frequency
+    frequencies = []
+    for op in filtered_ops:
+        if op['usec'] > 0 and op['cycles'] > 0:
+            frequencies.append(op['cycles'] / op['usec'])
+    avg_freq_mhz = statistics.mean(frequencies) if frequencies else 1000.0
+    if avg_freq_mhz <= 0:
+        avg_freq_mhz = 1000.0
+
+    # Assign start and end cycles to each operator
+    for op in filtered_ops:
+        op['start_cycles'] = op['unwrapped_cycles_start']
+        op['end_cycles'] = op['start_cycles'] + op['cycles']
+
+    global_min_cyc = min(op['start_cycles'] for op in filtered_ops if op['start_cycles'] is not None)
+
+    # Process events
+    completed_events = []
+    for op in filtered_ops:
+        events = op['trace_events']
+        if not events:
+            continue
+        events = sorted(events, key=lambda e: e['unwrapped_cycles'])
+
+        active_starts = {}
+        for e in events:
+            t = e['thread']
+            evt = e['event']
+            info = e['info']
+            state = e['state']
+            cyc = e['unwrapped_cycles']
+
+            key = (t, evt, info)
+            if state == 'start':
+                active_starts[key] = cyc
+            elif state == 'stop':
+                if key in active_starts:
+                    start_cyc = active_starts[key]
+                    del active_starts[key]
+                    completed_events.append({
+                        'thread': t,
+                        'event': evt,
+                        'info': info,
+                        'start_cyc': start_cyc,
+                        'end_cyc': cyc,
+                        'op_name': op['name']
+                    })
+
+    completed_events.sort(key=lambda e: e['start_cyc'])
+
+    # Convert event times to microseconds and apply clamp rounded to 1ns resolution (3 decimals)
+    for e in completed_events:
+        start_us = (e['start_cyc'] - global_min_cyc) / avg_freq_mhz
+        dur_us = (e['end_cyc'] - e['start_cyc']) / avg_freq_mhz
+        e['ts_ns'] = int(round(start_us * 1000))
+        e['dur_ns'] = int(round(max(dur_us, 0.1) * 1000))
+
+    # Allocate slots (sub-tracks) to prevent overlaps on same virtual track
+    active_slots = defaultdict(list)
+    for e in completed_events:
+        t = e['thread']
+        evt = e['event']
+        ts = e['ts_ns']
+        dur = e['dur_ns']
+
+        norm_evt = normalize_event_name(evt)
+        if norm_evt == "DMA":
+            track_key = (t, "DMA")
+        elif t == 10:
+            track_key = (t, "HMX")
+        else:
+            track_key = (t, "HVX")
+
+        slots = active_slots[track_key]
+        allocated_slot = -1
+        for idx, slot_end_ns in enumerate(slots):
+            if ts >= slot_end_ns:
+                slots[idx] = ts + dur
+                allocated_slot = idx
+                break
+        if allocated_slot == -1:
+            slots.append(ts + dur)
+            allocated_slot = len(slots) - 1
+        e['slot'] = allocated_slot
+
+    # Generate Track IDs and track definitions
+    used_tracks = {}
+    for e in completed_events:
+        t = e['thread']
+        evt = e['event']
+        slot = e['slot']
+
+        norm_evt = normalize_event_name(evt)
+        if norm_evt == "DMA":
+            track_evt = "DMA"
+            evt_id = 1
+        elif t == 10:
+            track_evt = "HMX"
+            evt_id = 3
+        else:
+            track_evt = "HVX"
+            evt_id = 2
+
+        t_sort = 1 if t == 10 else t + 2
+        # Unique UUID for each sub-track
+        if t == 10:
+            uuid = 20  # HMX thread track UUID
+        else:
+            uuid = int(t_sort * 1000000 + evt_id * 1000 + slot)
+        e['uuid'] = uuid
+        used_tracks[uuid] = (t, track_evt, slot)
+
+    with open(output_path, "wb") as f:
+        # Define Process with EXPLICIT child sorting
+        proc_desc = make_process_descriptor(1, "HTP NPU")
+        proc_packet = make_trace_packet(0, track_descriptor=make_track_descriptor(1, process=proc_desc, child_ordering=3))
+        write_trace_packet_to_file(f, proc_packet)
+
+        # Define Operators Track (UUID = 2) as a thread track at rank 1, tid 8
+        op_thread_desc = make_thread_descriptor(1, 8, "Ops", sort_index=1)
+        op_packet = make_trace_packet(0, track_descriptor=make_track_descriptor(2, parent_uuid=1, thread=op_thread_desc))
+        write_trace_packet_to_file(f, op_packet)
+
+        # Define HMX Thread Track (UUID = 20) at rank 2, tid 9
+        hmx_thread_desc = make_thread_descriptor(1, 9, "HMX", sort_index=2)
+        hmx_packet = make_trace_packet(0, track_descriptor=make_track_descriptor(20, parent_uuid=1, thread=hmx_thread_desc))
+        write_trace_packet_to_file(f, hmx_packet)
+
+        # Define Thread Tracks (T0, T1, ..., T9)
+        unique_threads = sorted(list(set(t for (t, _, _) in used_tracks.values() if t != 10)))
+        for t in unique_threads:
+            thread_uuid = 10 + t
+            thread_name = f"T{t}"
+            # Sort order starts from index 3 (T0 -> 3, T1 -> 4, etc.)
+            sort_index = 3 + t
+            tid = 10 + t
+            thread_desc = make_thread_descriptor(1, tid, thread_name, sort_index=sort_index)
+            thread_packet = make_trace_packet(0, track_descriptor=make_track_descriptor(
+                thread_uuid,
+                parent_uuid=1,
+                thread=thread_desc,
+                sibling_order_rank=sort_index,
+                child_ordering=3  # Explicit child sorting for sub-tracks
+            ))
+            write_trace_packet_to_file(f, thread_packet)
+
+        # Define Track descriptors for sub-tracks parented to thread tracks
+        for uuid in sorted(used_tracks.keys()):
+            if uuid == 20:
+                continue
+            t, evt, slot = used_tracks[uuid]
+            name = f"T{t} {evt}"
+            rank = 0 if evt == "HVX" else 1
+            parent_thread_uuid = 10 + t
+            # Sibling merge behavior: 1 (SIBLING_MERGE_BEHAVIOR_BY_TRACK_NAME)
+            track_desc = make_track_descriptor(
+                uuid=uuid,
+                name=name,
+                parent_uuid=parent_thread_uuid,
+                sibling_merge_behavior=1,
+                sibling_order_rank=rank
+            )
+            track_packet = make_trace_packet(0, track_descriptor=track_desc)
+            write_trace_packet_to_file(f, track_packet)
+
+        # Emit Operators
+        last_op_end_ns = 0
+        for op in filtered_ops:
+            op_start_ns = int(round(((op['start_cycles'] - global_min_cyc) / avg_freq_mhz) * 1000))
+            op_dur_ns = int(round((op['cycles'] / avg_freq_mhz) * 1000))
+            if op_start_ns < last_op_end_ns:
+                op_start_ns = last_op_end_ns
+            clamped_dur = max(op_dur_ns, 100) # Clamp to 100ns (0.1us)
+
+            # Debug annotations for Ops
+            debug_annots = []
+            if 'line_num' in op:
+                debug_annots.append(make_debug_annotation("line", int_val=op['line_num']))
+            if 'strides' in op and op['strides']:
+                debug_annots.append(make_debug_annotation("strides", string_val=op['strides']))
+            if 'params' in op and op['params'] and op['params'] != '----':
+                debug_annots.append(make_debug_annotation("params", string_val=op['params']))
+
+            # Slice Begin
+            evt_begin = make_track_event(1, 2, name=f"{op['name']} ({op['dims']})", category="operator", debug_annotations=debug_annots)
+            packet_begin = make_trace_packet(op_start_ns, track_event=evt_begin)
+            write_trace_packet_to_file(f, packet_begin)
+
+            # Slice End
+            evt_end = make_track_event(2, 2)
+            packet_end = make_trace_packet(op_start_ns + clamped_dur, track_event=evt_end)
+            write_trace_packet_to_file(f, packet_end)
+
+            last_op_end_ns = op_start_ns + clamped_dur
+
+        # Emit Thread Trace Events
+        for e in completed_events:
+            norm_name = normalize_event_name(e['event'])
+            name = f"DMA {e['info']}" if norm_name == "DMA" else norm_name
+
+            # Slice Begin
+            evt_begin = make_track_event(1, e['uuid'], name=name, category="trace")
+            packet_begin = make_trace_packet(e['ts_ns'], track_event=evt_begin)
+            write_trace_packet_to_file(f, packet_begin)
+
+            # Slice End
+            evt_end = make_track_event(2, e['uuid'])
+            packet_end = make_trace_packet(e['ts_ns'] + e['dur_ns'], track_event=evt_end)
+            write_trace_packet_to_file(f, packet_end)
+
+    logger.info(f"Successfully generated Perfetto trace at {output_path}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Convert Hexagon Op profile logs to native Perfetto Protobuf traces.")
+    parser.add_argument("logfile", help="Path to hex-log profile file")
+    parser.add_argument("-o", "--output", default="optrace.perfetto-trace", help="Output trace file path (default: optrace.perfetto-trace)")
+    parser.add_argument("--filter", type=str, help="Regex filter matching against the original profile-op line")
+
+    group = parser.add_mutually_exclusive_group()
+    group.add_argument("--head", type=int, help="Limit to first N ops")
+    group.add_argument("--tail", type=int, help="Limit to last N ops")
+
+    args = parser.parse_args()
+    logging.basicConfig(level=logging.INFO, format='%(message)s')
+
+    ops = parse_log(args.logfile)
+
+    if args.filter:
+        try:
+            filter_re = re.compile(args.filter)
+        except re.error as e:
+            logger.error(f"Invalid regex filter: {e}")
+            sys.exit(1)
+        ops = [op for op in ops if filter_re.search(op['op_text'])]
+
+    if args.head is not None:
+        ops = ops[:args.head]
+    elif args.tail is not None:
+        ops = ops[-args.tail:]
+
+    generate_perfetto_trace(ops, args.output)
+
+
+if __name__ == "__main__":
+    main()

scripts/snapdragon/qdc/requirements.txt

@@ -0,0 +1,22 @@
+Appium-Python-Client==5.2.4
+attrs==25.4.0
+certifi==2025.10.5
+exceptiongroup==1.3.0
+h11==0.16.0
+idna==3.11
+iniconfig==2.1.0
+outcome==1.3.0.post0
+packaging==25.0
+pluggy==1.6.0
+PySocks==1.7.1
+pytest==8.4.2
+selenium==4.36.0
+sniffio==1.3.1
+sortedcontainers==2.4.0
+tomli==2.3.0
+trio==0.31.0
+trio-websocket==0.12.2
+typing_extensions==4.15.0
+urllib3==2.5.0
+websocket-client==1.9.0
+wsproto==1.2.0

scripts/snapdragon/qdc/run_qdc_jobs.py

@@ -0,0 +1,684 @@
+"""Run llama.cpp Hexagon tests in a single QDC job.
+
+Bundles test scripts into one artifact and submits a single QDC job:
+
+  1. run_bench_tests_posix.py — llama-cli and llama-bench on CPU / GPU / NPU
+                                (from scripts/snapdragon/qdc/)
+
+Results are written to $GITHUB_STEP_SUMMARY when set (GitHub Actions).
+
+Prerequisites:
+  pip install /path/to/qualcomm_device_cloud_sdk*.whl
+
+Platform is inferred from --device:
+  android  Appium + pytest (Android phones: SM8750 / SM8650 / SM8850)
+  linux    BASH (Linux IoT: QCS9075M)
+
+Required environment variables:
+  QDC_API_KEY   API key from QDC UI -> Users -> Settings -> API Keys
+
+Usage:
+  python run_qdc_jobs.py \\
+      --pkg-dir    pkg-snapdragon/llama.cpp \\
+      --model-url  https://.../Llama-3.2-1B-Instruct-Q4_0.gguf \\
+      --device     SM8750
+"""
+
+from __future__ import annotations
+
+import argparse
+import enum
+import logging
+import os
+import re
+import shutil
+import sys
+import tempfile
+import time
+import urllib.request
+import xml.etree.ElementTree as ET
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Callable
+
+from qualcomm_device_cloud_sdk.api import qdc_api
+from qualcomm_device_cloud_sdk.logging import configure_logging
+from qualcomm_device_cloud_sdk.models import (
+    ArtifactType,
+    JobMode,
+    JobState,
+    JobSubmissionParameter,
+    JobType,
+    TestFramework,
+)
+
+# configure_logging only sets up the SDK logger; basicConfig is needed for
+# our own log.info to reach stdout.
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s %(name)s %(levelname)s - %(message)s",
+    handlers=[logging.StreamHandler()],
+)
+configure_logging(level=logging.INFO, handlers=[logging.StreamHandler()])
+# Silence per-poll GET/status spam from the SDK and its HTTP client.
+logging.getLogger("qualcomm_device_cloud").setLevel(logging.WARNING)
+logging.getLogger("httpx").setLevel(logging.WARNING)
+logging.getLogger("httpcore").setLevel(logging.WARNING)
+log = logging.getLogger(__name__)
+
+POLL_INTERVAL        = 30
+JOB_TIMEOUT          = 3600
+LOG_UPLOAD_TIMEOUT   = 600
+CAPACITY_TIMEOUT     = 1800
+CAPACITY_POLL        = 60
+MAX_CONCURRENT_JOBS  = 5
+DEFAULT_RETRIES      = 0
+RETRY_DELAY          = 300
+TERMINAL_STATES     = {JobState.COMPLETED, JobState.CANCELED}
+NON_TERMINAL_STATES = {JobState.DISPATCHED, JobState.RUNNING, JobState.SETUP, JobState.SUBMITTED}
+
+
+class DeviceUnavailableError(Exception):
+    """Raised when the QDC device resource is not available (retryable)."""
+
+
+_SCRIPTS_DIR = Path(__file__).parent
+_TESTS_DIR = _SCRIPTS_DIR / "tests"
+
+# --- Shared test assets -------------------------------------------------------
+_UTILS = _TESTS_DIR / "utils.py"
+_CONFTEST = _TESTS_DIR / "conftest.py"
+_PYTEST_LINE_RE = re.compile(
+    r"(?:[\w/]+\.py::)?(?:\w+::)?([\w\[\].-]+)\s+(PASSED|FAILED|ERROR|SKIPPED)"
+)
+_EXCLUDED_LOGS = {
+    "qdc_android_whole_host-000.log",
+    "qdc_kernel_host-000.log",
+    "qdc_LE_whole_host-000.log",
+    "qdc_LE_kernel_host-000.log",
+    "script.log",
+}
+_NON_TERMINAL_STATE_VALUES = {s.value for s in NON_TERMINAL_STATES}
+
+# --- Android (Appium + pytest) assets ----------------------------------------
+_RUN_BENCH = _TESTS_DIR / "run_bench_tests_posix.py"
+_RUN_BACKEND_OPS = _TESTS_DIR / "run_backend_ops_posix.py"
+_REQUIREMENTS = _SCRIPTS_DIR / "requirements.txt"
+_UPSTREAM_ADB_SCRIPTS = (
+    "https://raw.githubusercontent.com/ggml-org/llama.cpp/master/scripts/snapdragon/adb"
+)
+_ADB_SCRIPT_NAMES = [
+    "run-bench.sh",
+    "run-cli.sh",
+    "run-completion.sh",
+    "run-tool.sh",
+]
+
+# --- Linux (BASH) assets ------------------------------------------------------
+_RUN_LINUX_TEMPLATE = _TESTS_DIR / "linux" / "run_linux.sh"
+_LINUX_ENTRY_SCRIPT = "/bin/bash /data/local/tmp/TestContent/run_linux.sh"
+
+# =============================================================================
+# Artifact builders (per platform)
+# =============================================================================
+
+
+@dataclass
+class JobResult:
+    passed: bool
+    tests: dict[str, bool] = field(default_factory=dict)
+    raw_logs: dict[str, str] = field(default_factory=dict)
+    failure_details: dict[str, str] = field(default_factory=dict)
+
+
+def _write_lf(path: Path, content: str) -> None:
+    """Write text with LF line endings (required by /bin/bash on Linux)."""
+    with open(path, "w", encoding="utf-8", newline="\n") as f:
+        f.write(content)
+
+
+def _build_android_artifact(
+    pkg_dir: Path,
+    stage_dir: Path,
+    test_mode: str,
+    model_url: str | None,
+) -> Path:
+    """Android zip (Appium/pytest). Extracted by QDC under /qdc/appium/.
+
+    Zip structure:
+      llama_cpp_bundle/            installed package (adb pushed to /data/local/tmp/)
+      run-{bench,cli,completion,tool}.sh  upstream adb wrappers (patched)
+      tests/
+        utils.py                   shared adb helpers
+        conftest.py                Appium pytest fixtures
+        test_bench_posix.py        bench + cli tests (for --test bench or all)
+        test_backend_ops_posix.py  test-backend-ops on HTP0
+      requirements.txt
+      pytest.ini                   addopts = --junitxml=results.xml
+    """
+    bundle_dir = stage_dir / "llama_cpp_bundle"
+    shutil.copytree(pkg_dir, bundle_dir)
+
+    # Download upstream adb scripts so they land at /qdc/appium/ on the QDC
+    # runner. They wrap `adb shell` internally. Patch in `chmod +x bin/* lib/*`
+    # right after `cd $basedir` so device binaries are executable.
+    for name in _ADB_SCRIPT_NAMES:
+        url = f"{_UPSTREAM_ADB_SCRIPTS}/{name}"
+        dest = stage_dir / name
+        log.info("Downloading %s", url)
+        urllib.request.urlretrieve(url, str(dest))
+        content = dest.read_text()
+        content = content.replace(
+            "cd $basedir;",
+            "cd $basedir; chmod +x bin/* lib/* 2>/dev/null;",
+        )
+        dest.write_text(content)
+        dest.chmod(0o755)
+
+    tests_dir = stage_dir / "tests"
+    tests_dir.mkdir()
+
+    shutil.copy(_UTILS, tests_dir / "utils.py")
+    shutil.copy(_CONFTEST, tests_dir / "conftest.py")
+
+    if test_mode in ("bench", "all"):
+        assert model_url is not None
+        (tests_dir / "test_bench_posix.py").write_text(
+            _RUN_BENCH.read_text().replace("<<MODEL_URL>>", model_url)
+        )
+    if test_mode in ("backend-ops", "all"):
+        shutil.copy(_RUN_BACKEND_OPS, tests_dir / "test_backend_ops_posix.py")
+
+    shutil.copy(_REQUIREMENTS, stage_dir / "requirements.txt")
+    (stage_dir / "pytest.ini").write_text(
+        "[pytest]\naddopts = --junitxml=results.xml\n"
+    )
+
+    zip_base = str(stage_dir / "artifact")
+    shutil.make_archive(zip_base, "zip", stage_dir)
+    return Path(f"{zip_base}.zip")
+
+
+def _build_linux_artifact(
+    pkg_dir: Path,
+    stage_dir: Path,
+    test_mode: str,
+    model_url: str | None,
+) -> Path:
+    """Linux IoT zip (BASH framework). Extracted by QDC to /data/local/tmp/TestContent/.
+
+    Zip structure:
+      run_linux.sh               entry script (placeholder-substituted, LF line endings)
+      llama_cpp_bundle/          installed package
+    """
+    bundle_dir = stage_dir / "llama_cpp_bundle"
+    shutil.copytree(pkg_dir, bundle_dir)
+
+    template = _RUN_LINUX_TEMPLATE.read_text(encoding="utf-8")
+    rendered = template.replace("{MODEL_URL}", model_url or "").replace(
+        "{TEST_MODE}", test_mode
+    )
+    script_path = stage_dir / "run_linux.sh"
+    _write_lf(script_path, rendered)
+    script_path.chmod(0o755)
+
+    zip_base = str(stage_dir / "artifact")
+    shutil.make_archive(zip_base, "zip", stage_dir)
+    return Path(f"{zip_base}.zip")
+
+
+# =============================================================================
+# Platform enum + strategy table
+# =============================================================================
+
+
+class Platform(enum.Enum):
+    ANDROID = "android"
+    LINUX = "linux"
+
+
+@dataclass(frozen=True)
+class PlatformSpec:
+    test_framework: TestFramework
+    entry_script: str | None
+    build_artifact: Callable[[Path, Path, str, str | None], Path]
+    job_name_fmt: str
+
+
+PLATFORM_SPECS: dict[Platform, PlatformSpec] = {
+    Platform.ANDROID: PlatformSpec(
+        test_framework=TestFramework.APPIUM,
+        entry_script=None,
+        build_artifact=_build_android_artifact,
+        job_name_fmt="{base}",
+    ),
+    Platform.LINUX: PlatformSpec(
+        test_framework=TestFramework.BASH,
+        entry_script=_LINUX_ENTRY_SCRIPT,
+        build_artifact=_build_linux_artifact,
+        job_name_fmt="{base} (Linux)",
+    ),
+}
+
+DEVICE_PLATFORM: dict[str, Platform] = {
+    "SM8750": Platform.ANDROID,
+    "SM8650": Platform.ANDROID,
+    "SM8850": Platform.ANDROID,
+    "QCS9075M": Platform.LINUX,
+}
+
+
+# =============================================================================
+# Shared QDC job plumbing
+# =============================================================================
+
+
+def wait_for_job(client, job_id: str, timeout: int) -> str:
+    elapsed = 0
+    last_state = None
+    consecutive_errors = 0
+    max_consecutive_errors = 5
+    while elapsed < timeout:
+        try:
+            raw = qdc_api.get_job_status(client, job_id)
+            consecutive_errors = 0
+        except Exception as e:
+            consecutive_errors += 1
+            log.warning(
+                "Transient error polling job %s (%d/%d): %s",
+                job_id,
+                consecutive_errors,
+                max_consecutive_errors,
+                e,
+            )
+            if consecutive_errors >= max_consecutive_errors:
+                raise
+            time.sleep(POLL_INTERVAL)
+            elapsed += POLL_INTERVAL
+            continue
+        try:
+            status = JobState(raw)
+        except ValueError:
+            status = raw
+        if status in TERMINAL_STATES:
+            return raw.lower()
+        if raw != last_state:
+            log.info("Job %s: %s", job_id, raw)
+            last_state = raw
+        time.sleep(POLL_INTERVAL)
+        elapsed += POLL_INTERVAL
+    # Abort to free the QDC concurrency slot instead of leaking it.
+    try:
+        qdc_api.abort_job(client, job_id)
+        log.warning("Aborted job %s after timeout to free concurrency slot", job_id)
+    except Exception as e:
+        log.warning("Failed to abort job %s: %s", job_id, e)
+    raise TimeoutError(f"Job {job_id} did not finish within {timeout}s")
+
+
+def wait_for_log_upload(client, job_id: str) -> None:
+    elapsed = 0
+    while elapsed <= LOG_UPLOAD_TIMEOUT:
+        try:
+            status = (qdc_api.get_job_log_upload_status(client, job_id) or "").lower()
+        except Exception as e:
+            log.warning("get_job_log_upload_status failed: %s — will retry", e)
+            status = ""
+        if status in {"completed", "failed"}:
+            return
+        log.info("Waiting for log upload (status=%s) ...", status)
+        time.sleep(POLL_INTERVAL)
+        elapsed += POLL_INTERVAL
+    log.warning("Timed out waiting for log upload after %ds", LOG_UPLOAD_TIMEOUT)
+
+
+def wait_for_capacity(client, max_jobs: int = MAX_CONCURRENT_JOBS) -> None:
+    """Block until the user's active (non-terminal) QDC job count is below max_jobs."""
+    elapsed = 0
+    while elapsed < CAPACITY_TIMEOUT:
+        jobs_page = qdc_api.get_jobs_list(client, page_number=0, page_size=50)
+        if jobs_page is None:
+            log.warning(
+                "Could not retrieve job list; proceeding without capacity check"
+            )
+            return
+        items = getattr(jobs_page, "data", []) or []
+        active = sum(
+            1 for j in items if getattr(j, "state", None) in _NON_TERMINAL_STATE_VALUES
+        )
+        if active < max_jobs:
+            log.info("Active QDC jobs: %d / %d — proceeding", active, max_jobs)
+            return
+        log.info(
+            "Active QDC jobs: %d / %d — waiting %ds ...",
+            active,
+            max_jobs,
+            CAPACITY_POLL,
+        )
+        time.sleep(CAPACITY_POLL)
+        elapsed += CAPACITY_POLL
+    raise TimeoutError(
+        f"Capacity wait timed out after {CAPACITY_TIMEOUT}s"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Log parsing helpers
+# ---------------------------------------------------------------------------
+
+
+def _parse_junit_xml(content: str) -> tuple[dict[str, bool], dict[str, str]]:
+    try:
+        root = ET.fromstring(content)
+    except ET.ParseError:
+        return {}, {}
+    results: dict[str, bool] = {}
+    failures: dict[str, str] = {}
+    for tc in root.iter("testcase"):
+        name = tc.get("name", "")
+        if classname := tc.get("classname", ""):
+            name = f"{classname}.{name}"
+        failure_el = tc.find("failure")
+        if failure_el is None:
+            failure_el = tc.find("error")
+        results[name] = failure_el is None
+        if failure_el is not None:
+            parts = [failure_el.get("message", ""), failure_el.text or ""]
+            failures[name] = "\n".join(p for p in parts if p).strip()
+    return results, failures
+
+
+def _parse_pytest_output(content: str) -> dict[str, bool]:
+    results: dict[str, bool] = {}
+    for m in _PYTEST_LINE_RE.finditer(content):
+        results[m.group(1)] = m.group(2) == "PASSED"
+    return results
+
+
+def fetch_logs_and_parse_tests(
+    client, job_id: str, max_retries: int = 5, retry_delay: int = 30
+) -> tuple[dict[str, bool], dict[str, str], dict[str, str]]:
+    """Returns (test_results, raw_logs, failure_details)."""
+    log_files = None
+    for attempt in range(1, max_retries + 1):
+        try:
+            log_files = qdc_api.get_job_log_files(client, job_id)
+            break
+        except Exception as e:
+            if attempt < max_retries:
+                log.warning(
+                    "get_job_log_files failed (attempt %d/%d): %s — retrying in %ds",
+                    attempt, max_retries, e, retry_delay,
+                )
+                time.sleep(retry_delay)
+            else:
+                log.error(
+                    "get_job_log_files failed after %d attempts: %s", max_retries, e
+                )
+                return {}, {}, {}
+    if not log_files:
+        log.warning("No log files returned for job %s", job_id)
+        return {}, {}, {}
+
+    test_results: dict[str, bool] = {}
+    pytest_fallback: dict[str, bool] = {}
+    raw_logs: dict[str, str] = {}
+    failure_details: dict[str, str] = {}
+
+    with tempfile.TemporaryDirectory() as tmpdir:
+        for lf in log_files:
+            zip_path = os.path.join(tmpdir, "log.zip")
+            log.info("Downloading log file: %s", lf.filename)
+            qdc_api.download_job_log_files(client, lf.filename, zip_path)
+            try:
+                shutil.unpack_archive(zip_path, tmpdir, "zip")
+            except Exception as e:
+                log.warning("Could not unpack %s as zip: %s", lf.filename, e)
+
+        for root_dir, _, files in os.walk(tmpdir):
+            for fname in sorted(files):
+                fpath = os.path.join(root_dir, fname)
+                content = Path(fpath).read_text(errors="replace")
+                if fname.endswith(".xml"):
+                    results, failures = _parse_junit_xml(content)
+                    test_results.update(results)
+                    failure_details.update(failures)
+                elif fname.endswith(".log"):
+                    if fname in _EXCLUDED_LOGS:
+                        continue
+                    log.info("--- %s ---\n%s", fname, content)
+                    raw_logs[fname] = content
+                    pytest_fallback.update(_parse_pytest_output(content))
+
+    return (
+        (test_results if test_results else pytest_fallback),
+        raw_logs,
+        failure_details,
+    )
+
+
+def write_summary(result: JobResult, title: str = "QDC Test Results") -> None:
+    summary_path = os.environ.get("GITHUB_STEP_SUMMARY")
+    if not summary_path:
+        return
+
+    icon = "✅" if result.passed else "❌"
+
+    lines = [
+        f"## {title}\n",
+        f"Overall: {icon} {'PASSED' if result.passed else 'FAILED'}\n",
+    ]
+    reportable = {n: ok for n, ok in result.tests.items() if "test_install" not in n}
+    if reportable:
+        lines += ["| Test | Result |", "| ---- | ------ |"]
+        for name, ok in reportable.items():
+            lines.append(f"| `{name}` | {'✅' if ok else '❌'} |")
+        passed_n = sum(1 for v in reportable.values() if v)
+        failed_n = sum(1 for v in reportable.values() if not v)
+        lines += ["", f"**{passed_n} passed, {failed_n} failed**"]
+    else:
+        lines.append("_No per-test data available._")
+
+    failed_names = [n for n, ok in reportable.items() if not ok]
+    if failed_names:
+        lines += ["", "### Failures"]
+        for name in failed_names:
+            detail = result.failure_details.get(name)
+            if detail:
+                lines += [
+                    f"<details><summary><code>{name}</code></summary>",
+                    "",
+                    "```",
+                    detail,
+                    "```",
+                    "",
+                    "</details>",
+                ]
+
+    if result.raw_logs:
+        lines += ["", "### Raw Logs"]
+        for fname, content in sorted(result.raw_logs.items()):
+            lines += [
+                f"<details><summary>{fname}</summary>",
+                "",
+                "```",
+                content.rstrip(),
+                "```",
+                "",
+                "</details>",
+            ]
+
+    with open(summary_path, "a") as f:
+        f.write("\n".join(lines) + "\n")
+
+
+# =============================================================================
+# CLI + main
+# =============================================================================
+
+def parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(
+        description=__doc__,
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    p.add_argument("--pkg-dir", required=True, type=Path,
+                   help="Installed llama.cpp package directory (contains bin/ and lib/)")
+    p.add_argument("--model-url",
+                   help="Direct URL to the GGUF model file (required for --test bench)")
+    p.add_argument("--device", required=True,
+                   help="QDC chipset name, e.g. SM8750")
+    p.add_argument("--test", choices=["bench", "backend-ops", "all"], default="bench",
+                   help="Test suite to run (default: bench)")
+    p.add_argument("--job-timeout", type=int, default=JOB_TIMEOUT, metavar="SECONDS",
+                   help=f"Max seconds to wait for job completion (default: {JOB_TIMEOUT})")
+    p.add_argument("--retries", type=int, default=DEFAULT_RETRIES, metavar="N",
+                   help="Number of retries when device is unavailable (default: 0)")
+    p.add_argument("--retry-delay", type=int, default=RETRY_DELAY, metavar="SECONDS",
+                   help=f"Seconds to wait between retries (default: {RETRY_DELAY})")
+    args = p.parse_args()
+    if args.test in ("bench", "all") and not args.model_url:
+        p.error("--model-url is required when --test bench or --test all")
+    return args
+
+
+def _submit_and_run_job(client, args, spec, target_id, artifact_id) -> JobResult:
+    """Submit a QDC job and wait for results.
+
+    Raises DeviceUnavailableError for transient device/resource issues that
+    are worth retrying. Returns JobResult for definitive outcomes (pass or
+    test failure).
+    """
+    try:
+        wait_for_capacity(client)
+    except TimeoutError:
+        raise DeviceUnavailableError("Capacity wait timed out — device busy")
+
+    job_name = spec.job_name_fmt.format(base="llama.cpp Hexagon tests")
+
+    job_id = qdc_api.submit_job(
+        public_api_client=client,
+        target_id=target_id,
+        job_name=job_name,
+        external_job_id=None,
+        job_type=JobType.AUTOMATED,
+        job_mode=JobMode.APPLICATION,
+        timeout=max(1, args.job_timeout // 60),
+        test_framework=spec.test_framework,
+        entry_script=spec.entry_script,
+        job_artifacts=[artifact_id],
+        monkey_events=None,
+        monkey_session_timeout=None,
+        job_parameters=[JobSubmissionParameter.WIFIENABLED],
+    )
+    if job_id is None:
+        raise DeviceUnavailableError("Job submission failed — device may be unavailable")
+    log.info("Job submitted: %s  (device=%s)", job_id, args.device)
+
+    try:
+        job_status = wait_for_job(client, job_id, timeout=args.job_timeout)
+    except TimeoutError as e:
+        raise DeviceUnavailableError(str(e))
+    log.info("Job %s finished: %s", job_id, job_status)
+
+    wait_for_log_upload(client, job_id)
+    tests, raw_logs, failure_details = fetch_logs_and_parse_tests(client, job_id)
+
+    job_ok = job_status == JobState.COMPLETED.value.lower()
+
+    if not job_ok and not tests:
+        raise DeviceUnavailableError(
+            f"Job did not complete (status={job_status}) and produced no test results"
+        )
+
+    passed = job_ok and all(tests.values()) if tests else job_ok
+    if spec.test_framework == TestFramework.BASH and not tests:
+        log.error("No test results recovered (state=%s). Script likely never ran.", job_status)
+        passed = False
+    if not passed:
+        log.error("Job did not complete successfully or tests failed (status=%s)", job_status)
+
+    return JobResult(passed=passed, tests=tests, raw_logs=raw_logs, failure_details=failure_details)
+
+
+def main() -> int:
+    args = parse_args()
+
+    platform = DEVICE_PLATFORM.get(args.device)
+    if platform is None:
+        log.error(
+            "Unknown device %r. Known: %s",
+            args.device, ", ".join(sorted(DEVICE_PLATFORM.keys())),
+        )
+        return 1
+    spec = PLATFORM_SPECS[platform]
+
+    api_key = os.environ.get("QDC_API_KEY")
+    if not api_key:
+        log.error("QDC_API_KEY environment variable must be set")
+        return 1
+    if not args.pkg_dir.is_dir():
+        log.error("--pkg-dir %s does not exist", args.pkg_dir)
+        return 1
+
+    client = qdc_api.get_public_api_client_using_api_key(
+        api_key_header=api_key,
+        app_name_header="llama-cpp-ci",
+        on_behalf_of_header="llama-cpp-ci",
+        client_type_header="Python",
+    )
+
+    target_id = qdc_api.get_target_id(client, args.device)
+    if target_id is None:
+        log.error("Could not find QDC target for device %r", args.device)
+        return 1
+
+    with tempfile.TemporaryDirectory() as tmpdir:
+        log.info("Building %s artifact (test=%s) ...", platform.value, args.test)
+        zip_path = spec.build_artifact(
+            args.pkg_dir, Path(tmpdir), args.test, args.model_url
+        )
+        log.info("Uploading artifact (%d MB) ...", zip_path.stat().st_size // 1_000_000)
+        artifact_id = qdc_api.upload_file(client, str(zip_path), ArtifactType.TESTSCRIPT)
+
+    if artifact_id is None:
+        log.error("Artifact upload failed")
+        return 1
+
+    max_attempts = 1 + args.retries
+    for attempt in range(1, max_attempts + 1):
+        try:
+            result = _submit_and_run_job(client, args, spec, target_id, artifact_id)
+            break
+        except DeviceUnavailableError as e:
+            if attempt < max_attempts:
+                log.warning(
+                    "Attempt %d/%d failed (device unavailable): %s — retrying in %ds",
+                    attempt, max_attempts, e, args.retry_delay,
+                )
+                time.sleep(args.retry_delay)
+            else:
+                log.error(
+                    "Attempt %d/%d failed (device unavailable): %s — no retries left",
+                    attempt, max_attempts, e,
+                )
+                write_summary(
+                    JobResult(passed=False, tests={}),
+                    title=f"QDC Device Unavailable ({args.device})",
+                )
+                return 1
+    else:
+        return 1
+
+    if args.test == "backend-ops":
+        title = f"Backend Ops — HTP0 ({args.device})"
+    elif args.test == "all":
+        title = f"QDC Tests ({args.device})"
+    else:
+        title = f"QDC Test Results ({args.device})"
+    write_summary(result, title=title)
+
+    return 0 if result.passed else 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())

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,
+        )

scripts/snapdragon/windows/run-bench.ps1

@@ -0,0 +1,48 @@
+
+#!/usr/bin/env pwsh
+
+# Basedir on device
+$basedir=".\pkg-snapdragon"
+
+$cli_opts=$args
+
+$model="Llama-3.2-3B-Instruct-Q4_0.gguf"
+if ($null -ne $env:M) {
+    $model=$env:M
+}
+
+$device="HTP0"
+if ($null -ne $env:D) {
+    $device=$env:D
+}
+
+if ($null -ne $env:V) {
+    $env:GGML_HEXAGON_VERBOSE=$env:V
+}
+
+if ($null -ne $env:PROF) {
+    $env:GGML_HEXAGON_PROFILE=$env:PROF
+}
+
+if ($null -ne $env:OPSTAGE) {
+    $env:GGML_HEXAGON_OPSTAGE=$env:OPSTAGE
+}
+
+if ($null -ne $env:NHVX) {
+    $env:GGML_HEXAGON_NHVX=$env:NHVX
+}
+
+if ($null -ne $env:NDEV) {
+    $env:GGML_HEXAGON_NDEV=$env:NDEV
+}
+
+if ($null -ne $env:HB) {
+    $env:GGML_HEXAGON_HOSTBUF=$env:HB
+}
+
+$env:ADSP_LIBRARY_PATH="$basedir\lib"
+
+& "$basedir\bin\llama-bench.exe" `
+    --mmap 0 -m $basedir\..\..\gguf\$model `
+    --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 `
+    --ubatch-size 1024 -ngl 99 --device $device $cli_opts

scripts/snapdragon/windows/run-cli.ps1

@@ -0,0 +1,53 @@
+
+#!/usr/bin/env pwsh
+
+# Basedir on device
+$basedir=".\pkg-snapdragon"
+
+$cli_opts=$args
+
+$model="Llama-3.2-3B-Instruct-Q4_0.gguf"
+if ($null -ne $env:M) {
+    $model=$env:M
+}
+
+$device="HTP0"
+if ($null -ne $env:D) {
+    $device=$env:D
+}
+
+if ($null -ne $env:V) {
+    $env:GGML_HEXAGON_VERBOSE=$env:V
+}
+
+if ($null -ne $env:SCHED) {
+    $env:GGML_SCHED_DEBUG=$env:SCHED; $cli_opts="$cli_opts -v"
+}
+
+if ($null -ne $env:PROF) {
+    $env:GGML_HEXAGON_PROFILE=$env:PROF
+}
+
+if ($null -ne $env:OPSTAGE) {
+    $env:GGML_HEXAGON_OPSTAGE=$env:OPSTAGE
+}
+
+if ($null -ne $env:NHVX) {
+    $env:GGML_HEXAGON_NHVX=$env:NHVX
+}
+
+if ($null -ne $env:NDEV) {
+    $env:GGML_HEXAGON_NDEV=$env:NDEV
+}
+
+if ($null -ne $env:HB) {
+    $env:GGML_HEXAGON_HOSTBUF=$env:HB
+}
+
+$env:ADSP_LIBRARY_PATH="$basedir\lib"
+
+& "$basedir\bin\llama-cli.exe" `
+    --no-mmap -m $basedir\..\..\gguf\$model `
+    --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 `
+    --ctx-size 8192 --ubatch-size 1024 -fa on `
+    -ngl 99 --device $device $cli_opts

scripts/snapdragon/windows/run-completion.ps1

@@ -0,0 +1,53 @@
+
+#!/usr/bin/env pwsh
+
+# Basedir on device
+$basedir=".\pkg-snapdragon"
+
+$cli_opts=$args
+
+$model="Llama-3.2-3B-Instruct-Q4_0.gguf"
+if ($null -ne $env:M) {
+    $model=$env:M
+}
+
+$device="HTP0"
+if ($null -ne $env:D) {
+    $device=$env:D
+}
+
+if ($null -ne $env:V) {
+    $env:GGML_HEXAGON_VERBOSE=$env:V
+}
+
+if ($null -ne $env:SCHED) {
+    $env:GGML_SCHED_DEBUG=$env:SCHED; $cli_opts="$cli_opts -v"
+}
+
+if ($null -ne $env:PROF) {
+    $env:GGML_HEXAGON_PROFILE=$env:PROF
+}
+
+if ($null -ne $env:OPSTAGE) {
+    $env:GGML_HEXAGON_OPSTAGE=$env:OPSTAGE
+}
+
+if ($null -ne $env:NHVX) {
+    $env:GGML_HEXAGON_NHVX=$env:NHVX
+}
+
+if ($null -ne $env:NDEV) {
+    $env:GGML_HEXAGON_NDEV=$env:NDEV
+}
+
+if ($null -ne $env:HB) {
+    $env:GGML_HEXAGON_HOSTBUF=$env:HB
+}
+
+$env:ADSP_LIBRARY_PATH="$basedir\lib"
+
+& "$basedir\bin\llama-completion.exe" `
+    --no-mmap -m $basedir\..\..\gguf\$model `
+    --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 `
+    --ctx-size 8192 --ubatch-size 1024 -fa on `
+    -ngl 99 -no-cnv --device $device $cli_opts

scripts/snapdragon/windows/run-mtmd.ps1

@@ -0,0 +1,68 @@
+#!/usr/bin/env pwsh
+
+# Basedir on device
+$basedir=".\pkg-snapdragon"
+
+$cli_opts=$args
+
+$model="gemma-3-4b-it-Q4_0.gguf"
+if ($null -ne $env:M) {
+    $model=$env:M
+}
+
+$mmproj="mmproj-F16.gguf"
+if ($null -ne $env:MMPROJ) {
+    $mmproj=$env:MMPROJ
+}
+
+$image=""
+if ($null -ne $env:IMG) {
+    $image=$env:IMG
+}
+
+$device="HTP0"
+if ($null -ne $env:D) {
+    $device=$env:D
+}
+
+if ($null -ne $env:V) {
+    $env:GGML_HEXAGON_VERBOSE=$env:V
+}
+
+if ($null -ne $env:SCHED) {
+    $env:GGML_SCHED_DEBUG=$env:SCHED; $cli_opts="$cli_opts -v"
+}
+
+if ($null -ne $env:PROF) {
+    $env:GGML_HEXAGON_PROFILE=$env:PROF
+}
+
+if ($null -ne $env:OPSTAGE) {
+    $env:GGML_HEXAGON_OPSTAGE=$env:OPSTAGE
+}
+
+if ($null -ne $env:NHVX) {
+    $env:GGML_HEXAGON_NHVX=$env:NHVX
+}
+
+if ($null -ne $env:NDEV) {
+    $env:GGML_HEXAGON_NDEV=$env:NDEV
+}
+
+if ($null -ne $env:HB) {
+    $env:GGML_HEXAGON_HOSTBUF=$env:HB
+}
+
+if ($null -ne $env:MTMD_DEVICE) {
+    $env:MTMD_BACKEND_DEVICE=$env:MTMD_DEVICE
+}
+
+$env:ADSP_LIBRARY_PATH="$basedir\lib"
+
+& "$basedir\bin\llama-mtmd-cli.exe" `
+    --no-mmap -m $basedir\..\..\gguf\$model `
+    --mmproj $basedir\..\..\gguf\$mmproj `
+    --image $basedir\..\..\gguf\$image `
+    --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 `
+    --ctx-size 8192 --ubatch-size 1024 -fa on `
+    -ngl 99 --device $device -v $cli_opts

scripts/snapdragon/windows/run-tool.ps1

@@ -0,0 +1,56 @@
+
+#!/usr/bin/env pwsh
+
+# Basedir on device
+$basedir=".\pkg-snapdragon"
+
+if ($args.Count -eq 0) {
+    Write-Host "No arguments provided.Expected the tool and argument to run."
+    exit -1
+}
+
+$tool=$args[0]
+$cli_opts=@()
+
+if ($args.Count -gt 1) {
+    $cli_opts=$args[1..($args.Count - 1)]
+    $remainingArgs = $args[1..($args.Count - 1)]
+}
+
+$device="HTP0"
+if ($null -ne $env:D) {
+    $device=$env:D
+}
+
+if ($null -ne $env:V) {
+    $env:GGML_HEXAGON_VERBOSE=$env:V
+}
+
+if ($null -ne $env:SCHED) {
+    $env:GGML_SCHED_DEBUG=$env:SCHED; $cli_opts="$cli_opts -v"
+}
+
+if ($null -ne $env:PROF) {
+    $env:GGML_HEXAGON_PROFILE=$env:PROF
+}
+
+if ($null -ne $env:OPSTAGE) {
+    $env:GGML_HEXAGON_OPSTAGE=$env:OPSTAGE
+}
+
+if ($null -ne $env:NHVX) {
+    $env:GGML_HEXAGON_NHVX=$env:NHVX
+}
+
+if ($null -ne $env:NDEV) {
+    $env:GGML_HEXAGON_NDEV=$env:NDEV
+}
+
+if ($null -ne $env:HB) {
+    $env:GGML_HEXAGON_HOSTBUF=$env:HB
+}
+
+$env:ADSP_LIBRARY_PATH="$basedir\lib"
+
+& "$basedir\bin\$tool" `
+    $cli_opts

scripts/snapdragon/windows/setup-build.ps1

@@ -0,0 +1,105 @@
+# Requires Run as Administrator is NOT strictly necessary for User-scope env vars,
+# but recommended for creating directories in C:\ root if permissions are restricted.
+
+$ErrorActionPreference = "Stop"
+
+# --- Configuration ---
+$BaseDir = "C:\Qualcomm"
+
+# SDK 1: Hexagon
+$HexagonUrl     = "https://github.com/snapdragon-toolchain/hexagon-sdk/releases/download/v6.6.0.0/hexagon-sdk-v6.6.0.0-arm64-wos.tar.xz"
+$HexagonParent  = Join-Path $BaseDir "Hexagon_SDK"
+$HexagonSdkVersion   = "6.6.0.0"
+$HexagonToolsVersion = "19.0.07"
+$HexagonSdkTarget    = Join-Path $HexagonParent $HexagonSdkVersion
+$HexagonToolsTarget  = Join-Path $HexagonSdkTarget "\tools\HEXAGON_Tools\$HexagonToolsVersion"
+
+# SDK 2: OpenCL
+$OpenCLUrl      = "https://github.com/snapdragon-toolchain/opencl-sdk/releases/download/v2.3.2/adreno-opencl-sdk-v2.3.2-arm64-wos.tar.xz"
+$OpenCLParent   = Join-Path $BaseDir "OpenCL_SDK"
+$OpenCLVersion  = "2.3.2"
+$OpenCLTarget   = Join-Path $OpenCLParent $OpenCLVersion
+
+# --- Helper Function ---
+function Install-QualcommSDK {
+    param (
+        [string]$Url,
+        [string]$ParentDir,
+        [string]$TargetDir,
+        [string]$Name
+    )
+
+    # 1. Create Parent Directory
+    if (-not (Test-Path -Path $ParentDir)) {
+        Write-Host "Creating directory: $ParentDir" -ForegroundColor Cyan
+        New-Item -Path $ParentDir -ItemType Directory -Force | Out-Null
+    }
+
+    # 2. Check for Specific Version Directory
+    if (Test-Path -Path $TargetDir) {
+        Write-Host "$Name ($TargetDir) already exists. Skipping download." -ForegroundColor Green
+    }
+    else {
+        Write-Host "$Name not found. preparing to download..." -ForegroundColor Yellow
+
+        # Create the target directory to extract into
+        New-Item -Path $TargetDir -ItemType Directory -Force | Out-Null
+
+        # Define temporary archive path
+        $TempFile = Join-Path $ParentDir "temp_sdk.tar.xz"
+
+        try {
+            # Download
+            Write-Host "Downloading from: $Url"
+            Invoke-WebRequest -Uri $Url -OutFile $TempFile
+
+            # Untar
+            # Note: We assume Windows includes tar.exe (Win 10 build 17063+)
+            Write-Host "Extracting archive to $TargetDir..."
+
+            # We use -C to extract contents INTO the target directory created above
+            tar -xJvf $TempFile -C $TargetDir\..
+
+            Write-Host "Extraction complete." -ForegroundColor Green
+        }
+        catch {
+            Write-Error "Failed to download or extract $Name. Error: $_"
+            # Cleanup target dir if failed so script tries again next time
+            Remove-Item -Path $TargetDir -Recurse -Force -ErrorAction SilentlyContinue
+        }
+        finally {
+            # Cleanup Archive
+            if (Test-Path $TempFile) { Remove-Item $TempFile -Force }
+        }
+    }
+}
+
+# --- Execution ---
+
+# 1. Ensure Base C:\Qualcomm exists
+if (-not (Test-Path $BaseDir)) {
+    New-Item -Path $BaseDir -ItemType Directory -Force | Out-Null
+}
+
+# 2. Run Install Logic
+Install-QualcommSDK -Url $HexagonUrl -ParentDir $HexagonParent -TargetDir $HexagonSdkTarget -Name "Hexagon SDK"
+Install-QualcommSDK -Url $OpenCLUrl -ParentDir $OpenCLParent -TargetDir $OpenCLTarget -Name "OpenCL SDK"
+
+# --- Environment Variables ---
+
+Write-Host "`nSetting Environment Variables..." -ForegroundColor Cyan
+
+# Set OPENCL_SDK_ROOT
+[System.Environment]::SetEnvironmentVariable('OPENCL_SDK_ROOT', $OpenCLTarget, [System.EnvironmentVariableTarget]::User)
+$env:OPENCL_SDK_ROOT = $OpenCLTarget # Set for current session as well
+Write-Host "OPENCL_SDK_ROOT set to:  $OpenCLTarget"
+
+# Set HEXAGON_SDK_ROOT
+[System.Environment]::SetEnvironmentVariable('HEXAGON_SDK_ROOT', $HexagonSdkTarget, [System.EnvironmentVariableTarget]::User)
+$env:HEXAGON_SDK_ROOT = $HexagonSdkTarget # Set for current session as well
+Write-Host "HEXAGON_SDK_ROOT set to: $HexagonSdkTarget"
+
+# Set HEXAGON_SDK_ROOT
+[System.Environment]::SetEnvironmentVariable('HEXAGON_TOOLS_ROOT', $HexagonToolsTarget, [System.EnvironmentVariableTarget]::User)
+$env:HEXAGON_TOOLS_ROOT = $HexagonToolsTarget # Set for current session as well
+Write-Host "HEXAGON_TOOLS_ROOT set to: $HexagonToolsTarget"