# Default deploy knobs for Qwen/Qwen3-ForcedAligner-0.6B, the TTS word-timestamp
# aligner. These configure the pooling stage that --forced-aligner appends to
# the pipeline tail (see utils.forced_aligner.inject_forced_aligner_stage); they are
# NOT a standalone engine. Pass --forced-aligner-config to override this file;
# --forced-aligner sets just the model path.
#
# The stage runs the model with runner="pooling" + pooling_task="token_classify".
# Prompt tokens, chat template and word segmentation are Qwen-specific and live
# in qwen3_force_align_processor (not configurable here).

forced_aligner:
  model: Qwen/Qwen3-ForcedAligner-0.6B
  runner: pooling
  architecture: Qwen3ASRForcedAlignerForTokenClassification
  pooling_task: token_classify

  gpu_memory_utilization: 0.30

  dtype: bfloat16
  max_model_len: 4096
  trust_remote_code: true