# ------------------------------------------------------------------------- # Copyright (c) Microsoft Corporation. All rights reserved. # Licensed under the MIT License. See License.txt in the project root for # license information. # -------------------------------------------------------------------------- import logging import os from pathlib import Path import numpy as np import torch from convert_generation import add_cache_indirection_to_mha, add_output_qk_to_mha, fix_past_sequence_length from optimizer import optimize_model from transformers import WhisperConfig, WhisperForConditionalGeneration, WhisperProcessor from whisper_decoder import WhisperDecoder from whisper_encoder import WhisperEncoder from whisper_encoder_decoder_init import WhisperEncoderDecoderInit from whisper_jump_times import WhisperJumpTimes from onnxruntime import InferenceSession logger = logging.getLogger(__name__) PRETRAINED_WHISPER_MODELS = [ "whisper-tiny", "whisper-tiny.en", "whisper-base", "whisper-base.en", "whisper-small", "whisper-small.en", "whisper-medium", "whisper-medium.en", "whisper-large", "whisper-large-v2", "whisper-large-v3", "whisper-large-v3-turbo", ] class WhisperHelper: @staticmethod def get_onnx_path( output_dir: str, model_name_or_path: str, suffix: str = "", new_folder: bool = False, ) -> str: """Build onnx path Args: output_dir (str): output directory model_name_or_path (str): pretrained model name, or path to the model checkpoint suffix (str, optional): suffix like "_encoder" or "_decoder_fp16" will be appended to file name. Defaults to None. new_folder (bool, optional): create a new directory for the model. Defaults to False. Returns: str: path of onnx model """ model_name = model_name_or_path if os.path.isdir(model_name_or_path): model_name = Path(model_name_or_path).parts[-1] else: model_name = model_name.split("/")[-1] model_name += suffix directory = os.path.join(output_dir, model_name) if new_folder else output_dir return os.path.join(directory, model_name + ".onnx") @staticmethod def load_model( model_name_or_path: str, model_impl: str, cache_dir: str, device: torch.device, dtype: torch.dtype, merge_encoder_and_decoder_init: bool = True, no_beam_search_op: bool = False, output_qk: bool = False, ) -> dict[str, torch.nn.Module]: """Load model given a pretrained name or path, then build models for ONNX conversion. Args: model_name_or_path (str): pretrained model name or path model_impl (str): library to load model from cache_dir (str): cache directory device (torch.device): device to run the model dtype (torch.dtype): dtype to run the model merge_encoder_and_decoder_init (bool, optional): Whether merge encoder and decoder initialization into one ONNX model. Defaults to True. no_beam_search_op (bool, optional): Whether to use beam search op or not. Defaults to False. output_qk (bool, optional): Whether to output QKs to calculate batched jump times for word-level timestamps. Defaults to False. Returns: Dict[str, torch.nn.Module]: mapping from name to modules for ONNX conversion. """ # Load PyTorch model if model_impl == "hf": # Load from Hugging Face model = WhisperForConditionalGeneration.from_pretrained( model_name_or_path, cache_dir=cache_dir, attn_implementation="eager" ) else: # Load from OpenAI import whisper if not os.path.exists(model_name_or_path): name_or_path = model_name_or_path.split("/")[-1][8:] else: name_or_path = model_name_or_path model = whisper.load_model(name_or_path, device, download_root=cache_dir, in_memory=True) # Set PyTorch model properties model.eval().to(device=device) if model_impl == "hf": model.to(dtype=dtype) config = WhisperConfig.from_pretrained(model_name_or_path, cache_dir=cache_dir) # Load each component of PyTorch model decoder = WhisperDecoder(config, model, model_impl, no_beam_search_op).eval() components = {"decoder": decoder} if merge_encoder_and_decoder_init: encoder_decoder_init = WhisperEncoderDecoderInit(config, model, model_impl, no_beam_search_op).eval() components.update({"encoder": encoder_decoder_init}) else: encoder = WhisperEncoder(config, model, model_impl).eval() components.update({"encoder": encoder, "decoder_init": decoder}) if output_qk: batched_jump_times = WhisperJumpTimes(config, device, cache_dir).eval() components.update({"jump_times": batched_jump_times}) return components @staticmethod def export_onnx( model: WhisperEncoder | WhisperEncoderDecoderInit | WhisperDecoder, onnx_model_path: str, provider: str, verbose: bool, use_external_data_format: bool, use_fp16_inputs: bool, use_int32_inputs: bool, use_encoder_hidden_states: bool, use_kv_cache_inputs: bool, ): """Export model component to ONNX Args: model (class): PyTorch class to export onnx_model_path (str): path to save ONNX model provider (str): provider to use for verifying parity on ONNX model verbose (bool): print verbose information. use_external_data_format (bool): use external data format or not. use_fp16_inputs (bool): use float16 inputs for the audio_features, encoder_hidden_states, logits, and KV caches. use_int32_inputs (bool): use int32 inputs for the decoder_input_ids. use_encoder_hidden_states (bool): use encoder_hidden_states as model input for decoder-init/decoder-without-past models. use_kv_cache_inputs (bool): use KV caches as model inputs for decoder-with-past models. """ if isinstance(model, WhisperEncoder): model.export_onnx( onnx_model_path, provider, verbose, use_external_data_format, use_fp16_inputs, ) elif isinstance(model, WhisperEncoderDecoderInit): model.export_onnx( onnx_model_path, provider, verbose, use_external_data_format, use_fp16_inputs, use_int32_inputs, ) elif isinstance(model, WhisperDecoder): model.export_onnx( onnx_model_path, provider, verbose, use_external_data_format, use_fp16_inputs, use_int32_inputs, use_encoder_hidden_states, use_kv_cache_inputs, ) elif isinstance(model, WhisperJumpTimes): model.export_onnx( onnx_model_path, provider, verbose, use_external_data_format, use_fp16_inputs, use_int32_inputs, ) else: raise ValueError(f"Unknown instance for model detected: {type(model)}") @staticmethod def optimize_onnx( onnx_model_path: str, optimized_model_path: str, is_float16: bool, num_attention_heads: int, hidden_size: int, num_layers: int, use_external_data_format: bool = False, use_gpu: bool = False, provider: str = "cpu", is_decoder: bool = False, no_beam_search_op: bool = False, output_qk: bool = False, ): """Optimize ONNX model with an option to convert it to use mixed precision.""" from fusion_options import FusionOptions optimization_options = FusionOptions("bart") optimization_options.use_multi_head_attention = True optimization_options.disable_multi_head_attention_bias = provider == "rocm" m = optimize_model( onnx_model_path, model_type="bart", num_heads=num_attention_heads, hidden_size=hidden_size, opt_level=0, optimization_options=optimization_options, use_gpu=use_gpu, only_onnxruntime=False, ) # Add `past_sequence_length`, `cache_indirection`, and `output_qk` to `MultiHeadAttention` ops if is_decoder and no_beam_search_op: if provider == "cuda": # FP32 CPU can be supported here once the DMMHA CPU kernel bugs are fixed # FP16 CUDA, FP32 CUDA, and FP32 CPU use the `DecoderMaskedMultiHeadAttention` kernel # via `MultiHeadAttention`, which requires the `past_sequence_length` and # `cache_indirection` inputs m, past_seq_len_name = fix_past_sequence_length(m) m = add_cache_indirection_to_mha(m, past_seq_len_name) if output_qk: m = add_output_qk_to_mha(m, skip_node_idxs=list(range(0, 2 * num_layers, 2))) m.save_model_to_file(optimized_model_path, use_external_data_format, all_tensors_to_one_file=True) @staticmethod def pt_transcription_for_verify_onnx( processor: WhisperProcessor, pt_model: torch.nn.Module, device: torch.device, batch_size: int = 1, prompt_mode: bool = False, ): # Try to import `datasets` pip package try: from datasets import load_dataset except Exception as e: logger.error(f"An error occurred while importing `datasets`: {e}", exc_info=True) # noqa: G201 install_cmd = "pip install datasets" logger.warning(f"Could not import `datasets`. Attempting to install `datasets` via `{install_cmd}`.") os.system(install_cmd) from datasets import load_dataset ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation") input_features_ = [] if batch_size == 1: input_features = processor([ds[0]["audio"]["array"]], return_tensors="pt").input_features else: input_features_ = [ processor([ds[3]["audio"]["array"]], return_tensors="pt").input_features, processor([ds[3]["audio"]["array"]], return_tensors="pt").input_features, ] assert len(input_features_) == batch_size input_features = torch.cat((input_features_[0], input_features_[1])) max_length, min_length, num_beams, num_return_sequences = 30, 0, 1, 1 length_penalty, repetition_penalty = 1.0, 1.0 inputs = { "input_features": input_features.to(device), "max_length": max_length, "min_length": min_length, "num_beams": num_beams, "num_return_sequences": num_return_sequences, "length_penalty": length_penalty, "repetition_penalty": repetition_penalty, "early_stopping": True, "use_cache": True, } if prompt_mode: prompts = ["John has doubts", "Maria has grave doubts"] prompt_ids = [processor.get_prompt_ids(p) for p in prompts] pt_transcription = [] pt_outputs = [] # The looping for model.generate is necessary here due to the limitation as per # https://huggingface.co/docs/transformers/model_doc/whisper#transformers.WhisperForConditionalGeneration.generate.prompt_ids # prompt_ids input requires a tensor of rank 1 for i in range(batch_size): inputs["prompt_ids"] = torch.from_numpy(prompt_ids[i]).to(device=device) inputs["input_features"] = input_features_[i].to(device) pt_output = pt_model.generate(**inputs).detach().cpu().numpy() pt_outputs.append(pt_output) pt_transcription.append(processor.batch_decode(pt_output, skip_special_tokens=True)[0]) inputs["input_features"] = input_features del inputs["prompt_ids"] else: prompt_ids = [] pt_outputs = pt_model.generate(**inputs).detach().cpu().numpy() pt_transcription = [processor.batch_decode(pt_outputs, skip_special_tokens=True)[0]] pt_outputs = list(pt_outputs) del inputs["early_stopping"] del inputs["use_cache"] return inputs, pt_transcription, pt_outputs, prompt_ids @staticmethod def select_transcription_options( batch_size: int, prompt_mode: bool, ): if batch_size > 1 and prompt_mode: expected_transcription_no_comma_prompt1 = " John has doubts whether Sir Frederick Layton's work is really Greek after all and can discover in it but little of Rocky I" expected_transcription_misspelled_prompt1 = " John has doubts whether Sir Frederick Latins work is really Greek after all and can discover in it but little of Rocky I" expected_transcription_no_comma_prompt2 = " Maria has grave doubts whether Sir Frederick Layton's work is really Greek after all and can discover in it but little of Rocky" expected_transcription_misspelled_prompt2 = " Maria has grave doubts whether Sir Frederick Latins work is really Greek after all and can discover in it but little of Rocky I" expected_transcription_options = { expected_transcription_no_comma_prompt1, expected_transcription_no_comma_prompt2, expected_transcription_misspelled_prompt1, expected_transcription_misspelled_prompt2, } else: expected_transcription_no_comma = ( " Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel." ) expected_transcription_with_comma = ( " Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel." ) expected_transcription_with_quote_and_comma = ( ' "Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.' ) expected_transcription_options = { expected_transcription_no_comma, expected_transcription_with_comma, expected_transcription_with_quote_and_comma, } return expected_transcription_options @staticmethod def get_outputs( pt_outputs: np.ndarray, ort_outputs: np.ndarray, i: int, ): """Get PyTorch and ONNX Runtime output token ids at index i""" pt_output, ort_output = pt_outputs[i], ort_outputs[i] pt_shape, ort_shape = pt_output.shape, ort_output.shape # Hugging Face impl. + Beam Search op: PyTorch = (26,) and ORT = (30,) # OpenAI impl. + Beam Search op: PyTorch = (1, 30) and ORT = (30,) if pt_shape != ort_shape: if len(pt_shape) > 1: pt_output = pt_output[0] pt_shape = pt_output.shape if len(ort_shape) > 1: ort_output = ort_output[0] ort_shape = ort_output.shape if pt_shape[0] != ort_shape[0]: min_len = min(pt_shape[0], ort_shape[0]) pt_output = pt_output[:min_len] ort_output = ort_output[:min_len] assert pt_output.shape == ort_output.shape return pt_output, ort_output @staticmethod def verify_onnx( model_name_or_path: str, cache_dir: str, ort_session: InferenceSession, device: torch.device, batch_size: int = 1, prompt_mode: bool = False, ): """Compare the result from PyTorch and ONNX Runtime to verify the ONNX model is good.""" pt_model = WhisperForConditionalGeneration.from_pretrained( model_name_or_path, cache_dir=cache_dir, attn_implementation="eager" ).to(device) processor = WhisperProcessor.from_pretrained(model_name_or_path, cache_dir=cache_dir) config = WhisperConfig.from_pretrained(model_name_or_path, cache_dir=cache_dir) inputs, pt_transcription, pt_outputs, decoder_prompt_ids = WhisperHelper.pt_transcription_for_verify_onnx( processor, pt_model, device, batch_size=batch_size, prompt_mode=prompt_mode, ) start_id = [config.decoder_start_token_id] # ex: [50258] prompt_ids = processor.get_decoder_prompt_ids(language="english", task="transcribe") prompt_ids = [token[1] for token in prompt_ids] # ex: [50259, 50358, 50363] forced_decoder_ids = start_id + prompt_ids # ex: [50258, 50259, 50358, 50363] ort_names = [entry.name for entry in ort_session.get_inputs()] ort_dtypes = [entry.type for entry in ort_session.get_inputs()] ort_to_np = { "tensor(float)": np.float32, "tensor(float16)": np.float16, "tensor(int64)": np.int64, "tensor(int32)": np.int32, "tensor(int8)": np.int8, "tensor(uint8)": np.uint8, } use_extra_decoding_ids = "extra_decoding_ids" in ort_names for name, dtype in zip(ort_names, ort_dtypes, strict=False): if name == "input_features": inputs[name] = inputs[name].detach().cpu().numpy() elif name == "vocab_mask": inputs[name] = np.ones(config.vocab_size, dtype=ort_to_np[dtype]) elif name == "prefix_vocab_mask": inputs[name] = np.ones((batch_size, config.vocab_size), dtype=ort_to_np[dtype]) elif name == "decoder_input_ids": if not prompt_mode: raw_input_ids = [start_id] if use_extra_decoding_ids else [forced_decoder_ids] inputs[name] = np.array(raw_input_ids, dtype=ort_to_np[dtype]) else: # This logic handles the scenario for when prompts are not of the same size # For example if our prompt ids are [p1_id_1, p1_id_2] and [p2_id_1] # The final decoder_input_ids will look as such after padding # [prev_token, p1_id_1, p1_id_2, start_token, lang_token, transcribe_token] # [prev_token, p2_id_1, PAD_TOKEN, start_token, lang_token, transcribe_token] ort_prompts = [] for i in range(batch_size): ort_prompts.append(decoder_prompt_ids[i].tolist()) max_len = max(len(p) for p in ort_prompts) padded_prompts = [] for p in ort_prompts: padded_prompt = [*p, *([config.pad_token_id] * (max_len - len(p)))] padded_prompts.append(padded_prompt + forced_decoder_ids) inputs[name] = np.array(padded_prompts, dtype=ort_to_np[dtype]) elif name == "logits_processor": inputs[name] = np.array([1], dtype=ort_to_np[dtype]) elif name == "cross_qk_layer_head": inputs[name] = np.array([[0, 0]], dtype=ort_to_np[dtype]) elif name == "extra_decoding_ids": inputs[name] = np.repeat(np.array([prompt_ids], dtype=ort_to_np[dtype]), batch_size, 0) elif name == "temperature": inputs[name] = np.array([1.0], dtype=ort_to_np[dtype]) else: inputs[name] = np.array([inputs[name]], dtype=ort_to_np[dtype]) ort_outputs = ort_session.run(None, inputs)[0][:, 0, :] ort_transcription = processor.batch_decode(ort_outputs, skip_special_tokens=True) expected_transcription_options = WhisperHelper.select_transcription_options(batch_size, prompt_mode) parity = 1 for i in range(batch_size): pt_output, ort_output = WhisperHelper.get_outputs(pt_outputs, ort_outputs, i) # Check if token ids match parity *= np.allclose(pt_output, ort_output) # Check if transcribed outputs match parity *= ( pt_transcription[i] in expected_transcription_options and ort_transcription[i] in expected_transcription_options ) max_diff = 0 if not parity: for i in range(batch_size): pt_output, ort_output = WhisperHelper.get_outputs(pt_outputs, ort_outputs, i) diff = pt_output - ort_output max_diff_i = max(diff.min(), diff.max(), key=abs) max_diff = max(max_diff, max_diff_i) if max_diff != 0: logger.warning(f"PyTorch outputs: {pt_transcription}") logger.warning(f"ONNX Runtime outputs: {ort_transcription}") return 0