nemo_curator.models.asr.indic_canary

Adapter for an Indic Canary model exported as a TensorRT-LLM engine.

Module Contents

Classes

Name	Description
`IndicCanaryTRTLLMASR`	Run static-batch Indic Canary inference from a prebuilt engine directory.

Data

_DEFAULT_MAX_DURATION_SEC

_DEFAULT_MIN_DURATION_SEC

_MIN_DURATION_SAMPLES

_REQUIRED_ENGINE_FILES

_TARGET_SAMPLE_RATE

API

class nemo_curator.models.asr.indic_canary.IndicCanaryTRTLLMASR(
    engine_dir: str,
    num_beams: int = 4,
    max_new_tokens: int = 374,
    pnc: bool = False,
    max_duration_sec: float = _DEFAULT_MAX_DURATION_SEC,
    min_duration_sec: float = _DEFAULT_MIN_DURATION_SEC,
    kv_cache_free_gpu_memory_fraction: float = 0.2,
    cross_kv_cache_fraction: float = 0.2
)

Run static-batch Indic Canary inference from a prebuilt engine directory.

cross_kv_cache_fraction

= float(cross_kv_cache_fraction)

kv_cache_free_gpu_memory_fraction

= float(kv_cache_free_gpu_memory_fraction)

max_new_tokens

= int(max_new_tokens)

max_samples

= int(self.max_duration_sec * _TARGET_SAMPLE_RATE)

min_duration_sec

= min(min_duration_sec, self.max_duration_sec)

min_samples

= int(self.min_duration_sec * _TARGET_SAMPLE_RATE)

num_beams

= int(num_beams)

pnc

= bool(pnc)

nemo_curator.models.asr.indic_canary.IndicCanaryTRTLLMASR._normalize_language(
    language: str
) -> str | None

nemo_curator.models.asr.indic_canary.IndicCanaryTRTLLMASR._prompt_config(
    language: str
) -> dict[str, object]

nemo_curator.models.asr.indic_canary.IndicCanaryTRTLLMASR._transcribe_prepared(
    padded: list[typing.Any],
    durations: list[int],
    prompts: list[dict[str, object]]
) -> list[str]

Run engine-sized sub-batches without exposing a second batch control.

nemo_curator.models.asr.indic_canary.IndicCanaryTRTLLMASR.download_weights_on_node() -> None

Validate local engine artifacts without allocating GPU state.

nemo_curator.models.asr.indic_canary.IndicCanaryTRTLLMASR.load_model(
    num_gpus: int
) -> None

Load the TensorRT-LLM runtime on its one required GPU.

nemo_curator.models.asr.indic_canary.IndicCanaryTRTLLMASR.transcribe_batch(
    items: list[dict[str, typing.Any]]
) -> list[nemo_curator.models.asr.base.ASRResult]

Transcribe supported rows and preserve their original positions.

nemo_curator.models.asr.indic_canary.IndicCanaryTRTLLMASR.unload_model() -> None

Release the engine and its CUDA allocations.

nemo_curator.models.asr.indic_canary._DEFAULT_MAX_DURATION_SEC = 40.0

nemo_curator.models.asr.indic_canary._DEFAULT_MIN_DURATION_SEC = 0.5

nemo_curator.models.asr.indic_canary._MIN_DURATION_SAMPLES = 400

nemo_curator.models.asr.indic_canary._REQUIRED_ENGINE_FILES = ('encoder/encoder.plan', 'encoder/config.json', 'decoder/config.json', 'decoder/...

nemo_curator.models.asr.indic_canary._TARGET_SAMPLE_RATE = 16000