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

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=${SAMPLES} \
@@ -130,4 +136,4 @@ print_elapsed_time --no-separators "${POST_GEN_TS}" "${POST_AUGMENT_TS}" "Augmen
print_elapsed_time --no-separators "${POST_AUGMENT_TS}" "${END_TS}" "${TRAINING_STEPS} training steps"
python -c $'msg="="*54 ; print(f"{msg:>80s}")'
print_elapsed_time --no-separators "${START_TS}" "${END_TS}" "Total"
python -c $'print(f"{\'=\' * 80}")'
python -c $'print(f"{\'=\' * 80}")'