taco/microWakeWord-Trainer-Nvidia-Docker
1
0
Fork 0
mirror of https://github.com/TaterTotterson/microWakeWord-Trainer-Nvidia-Docker.git synced 2026-06-12 20:10:19 -06:00
Code Issues Packages Projects Releases Wiki Activity

fixes

Browse Source
This commit is contained in:
MasterPhooey
2026-01-21 06:14:32 -06:00
parent 423bbd15f5
commit 98a087c87d
2 changed files with 32 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
cli/setup_python_venv
View File

@@ -178,14 +178,20 @@ echo " ===== Installing keras ====="
# keras 3.13 has "issues" so we need to back down to 3.12.
pip_install "keras==3.12.0"
# -----------------------------------------------------------------------------
# Optional CUDA data dir (GPU-only)
# Some stacks expect a CUDA "nvvm/libdevice" tree. We create one in /data/cuda
# and link Triton's libdevice if it exists. This is safe and does NOT enable
# any extra XLA flags by itself.
# -----------------------------------------------------------------------------
if ${GPU} ; then
CUDA_DATA_DIR="${DATA_DIR}/cuda"
LIBDEVICE_DIR="${CUDA_DATA_DIR}/nvvm/libdevice"
mkdir -p "${LIBDEVICE_DIR}"
TRITON_LIBDEVICE="$(
python - <<'PY'
import glob
import sys
paths = glob.glob("**/site-packages/triton/backends/nvidia/lib/libdevice.10.bc", recursive=True)
print(paths[0] if paths else "", end="")
PY
@@ -195,7 +201,8 @@ if [ -n "${TRITON_LIBDEVICE}" ] ; then
ln -sf "${TRITON_LIBDEVICE}" "${LIBDEVICE_DIR}/libdevice.10.bc"
echo " Linked Triton libdevice.10.bc to ${LIBDEVICE_DIR}"
else
echo " ⚠️ Triton libdevice.10.bc not found; XLA may require --xla_gpu_cuda_data_dir"
echo " ℹ️ Triton libdevice.10.bc not found (ok)"
fi
fi
"${PROGDIR}/test_python" --data-dir="${DATA_DIR}"

14
train_wake_word
View File

@@ -67,17 +67,23 @@ echo "===== Running '${WAKE_WORD}(${WAKE_WORD_TITLE})' generation, augmentation
echo
START_TS=$EPOCHSECONDS
# -----------------------------------------------------------------------------
# TensorFlow / XLA environment (known-good, portable)
# -----------------------------------------------------------------------------
export TF_CPP_MIN_LOG_LEVEL=9
export TF_FORCE_GPU_ALLOW_GROWTH=true
export TF_GPU_ALLOCATOR=cuda_malloc_async
DEFAULT_XLA_FLAGS="--tf_xla_auto_jit=0 --xla_gpu_unsafe_fallback_to_driver_on_ptxas_not_found --xla_gpu_cuda_data_dir=${DATA_DIR}/cuda"
DEFAULT_XLA_RUNTIME_FLAGS="--xla_gpu_unsafe_fallback_to_driver_on_ptxas_not_found --xla_gpu_cuda_data_dir=${DATA_DIR}/cuda"
export TF_XLA_FLAGS="${TF_XLA_FLAGS:+${TF_XLA_FLAGS} }${DEFAULT_XLA_FLAGS}"
export XLA_FLAGS="${XLA_FLAGS:+${XLA_FLAGS} }${DEFAULT_XLA_RUNTIME_FLAGS}"
# Hard-set TF_XLA_FLAGS to ONLY what we know this build supports.
# Do NOT append user environment flags (can cause hard failures).
export TF_XLA_FLAGS="--tf_xla_auto_jit=0"
unset XLA_FLAGS
export NVIDIA_TF32_OVERRIDE=1
export TF_CUDNN_WORKSPACE_LIMIT_IN_MB=512
export GLOG_minloglevel=2
export GRPC_VERBOSITY=ERROR
# -----------------------------------------------------------------------------
"${CLIDIR}/wake_word_sample_generator" \
--samples=${SAMPLES} \