run-cli.sh 1.4 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859
  1. #!/bin/sh
  2. #
  3. # Basedir on device
  4. basedir=/data/local/tmp/llama.cpp
  5. cli_opts=
  6. branch=.
  7. [ "$B" != "" ] && branch=$B
  8. adbserial=
  9. [ "$S" != "" ] && adbserial="-s $S"
  10. adbhost=
  11. [ "$H" != "" ] && adbhost="-H $H"
  12. model="Llama-3.2-3B-Instruct-Q4_0.gguf"
  13. [ "$M" != "" ] && model="$M"
  14. device="HTP0"
  15. [ "$D" != "" ] && device="$D"
  16. experimental=
  17. [ "$E" != "" ] && experimental="GGML_HEXAGON_EXPERIMENTAL=$E"
  18. verbose=
  19. [ "$V" != "" ] && verbose="GGML_HEXAGON_VERBOSE=$V" cli_opts="$cli_opts -v"
  20. sched=
  21. [ "$SCHED" != "" ] && sched="GGML_SCHED_DEBUG=2" cli_opts="$cli_opts -v"
  22. profile=
  23. [ "$PROF" != "" ] && profile="GGML_HEXAGON_PROFILE=$PROF GGML_HEXAGON_OPSYNC=1" cli_opts="$cli_opts -v"
  24. opmask=
  25. [ "$OPMASK" != "" ] && opmask="GGML_HEXAGON_OPMASK=$OPMASK"
  26. nhvx=
  27. [ "$NHVX" != "" ] && nhvx="GGML_HEXAGON_NHVX=$NHVX"
  28. ndev=
  29. [ "$NDEV" != "" ] && ndev="GGML_HEXAGON_NDEV=$NDEV"
  30. hb=
  31. [ "$HB" != "" ] && hb="GGML_HEXAGON_HOSTBUF=$HB"
  32. set -x
  33. adb $adbserial $adbhost shell " \
  34. cd $basedir; ulimit -c unlimited; \
  35. LD_LIBRARY_PATH=$basedir/$branch/lib \
  36. ADSP_LIBRARY_PATH=$basedir/$branch/lib \
  37. $verbose $experimental $sched $opmask $profile $nhvx $ndev $hb \
  38. ./$branch/bin/llama-cli --no-mmap -m $basedir/../gguf/$model \
  39. --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 \
  40. --ctx-size 8192 --batch-size 128 -fa on \
  41. -ngl 99 --device $device $cli_opts $@ \
  42. "