Merge pull request #1355 from 1carlito/batch_wrap

Batch wrap
This commit is contained in:
carlito authored and GitHub committed 2026-02-17 13:31:15 +00:00
commit 422c44f952
4 files changed
+268 -18

No files matched your search

+1
View File
@@ -60,6 +60,7 @@ def cli():
parser.add_argument("--initial_prompt", type=str, default=None, help="optional text to provide as a prompt for the first window.")
parser.add_argument("--hotwords", type=str, default=None, help="hotwords/hint phrases to the model (e.g. \"WhisperX, PyAnnote, GPU\"); improves recognition of rare/technical terms")
parser.add_argument("--condition_on_previous_text", type=str2bool, default=False, help="if True, provide the previous output of the model as a prompt for the next window; disabling may make the text inconsistent across windows, but the model becomes less prone to getting stuck in a failure loop")
parser.add_argument("--batch_context", action="store_true", help="use previous batch's transcription as context for the next batch (slower but more coherent across batches)")
parser.add_argument("--fp16", type=str2bool, default=True, help="whether to perform inference in fp16; True by default")
parser.add_argument("--temperature_increment_on_fallback", type=optional_float, default=0.2, help="temperature to increase when falling back when the decoding fails to meet either of the thresholds below")
+121 -18
View File
@@ -18,7 +18,6 @@ from whisperx.log_utils import get_logger
logger = get_logger(__name__)
def find_numeral_symbol_tokens(tokenizer):
numeral_symbol_tokens = []
for i in range(tokenizer.eot):
@@ -40,28 +39,48 @@ class WhisperModel(faster_whisper.WhisperModel):
tokenizer: Tokenizer,
options: TranscriptionOptions,
encoder_output=None,
use_batch_context: bool = False,
previous_batch_context_tokens: List[List[int]] = None,
):
batch_size = features.shape[0]
all_tokens = []
prompt_reset_since = 0
if previous_batch_context_tokens is None:
previous_batch_context_tokens = [[] for _ in range(batch_size)]
initial_prompt_tokens = []
if options.initial_prompt is not None:
initial_prompt = " " + options.initial_prompt.strip()
initial_prompt_tokens = tokenizer.encode(initial_prompt)
all_tokens.extend(initial_prompt_tokens)
previous_tokens = all_tokens[prompt_reset_since:]
prompt = self.get_prompt(
tokenizer,
previous_tokens,
without_timestamps=options.without_timestamps,
prefix=options.prefix,
hotwords=options.hotwords
)
batch_tokens = []
for i in range(batch_size):
all_tokens = list(initial_prompt_tokens)
if use_batch_context:
if i < len(previous_batch_context_tokens):
ctx = previous_batch_context_tokens[i]
if ctx:
# 223 is max prompt tokens
available = 223 - len(all_tokens)
if available > 0:
all_tokens.extend(ctx[-available:])
batch_tokens.append(all_tokens)
max_batch_tokens = max([len(t) for t in batch_tokens] + [0])
prompts = [
self.get_prompt(
tokenizer,
[tokenizer.eot] * (max_batch_tokens - len(t)) + t,
without_timestamps=options.without_timestamps,
prefix=options.prefix,
hotwords=options.hotwords
) for t in batch_tokens
]
encoder_output = self.encode(features)
result = self.model.generate(
encoder_output,
[prompt] * batch_size,
prompts,
beam_size=options.beam_size,
patience=options.patience,
length_penalty=options.length_penalty,
@@ -82,9 +101,9 @@ class WhisperModel(faster_whisper.WhisperModel):
return tokenizer.tokenizer.decode_batch(res)
text = decode_batch(tokens_batch)
return text
def encode(self, features: np.ndarray) -> ctranslate2.StorageView:
# When the model is running on multiple GPUs, the encoder output should be moved
# to the CPU since we don't know which GPU will handle the next job.
@@ -115,6 +134,7 @@ class FasterWhisperPipeline(Pipeline):
framework="pt",
language: Optional[str] = None,
suppress_numerals: bool = False,
use_batch_context: bool = False,
**kwargs,
):
self.model = model
@@ -122,6 +142,7 @@ class FasterWhisperPipeline(Pipeline):
self.options = options
self.preset_language = language
self.suppress_numerals = suppress_numerals
self.use_batch_context = use_batch_context
self._batch_size = kwargs.pop("batch_size", None)
self._num_workers = 1
self._preprocess_params, self._forward_params, self._postprocess_params = self._sanitize_parameters(**kwargs)
@@ -142,6 +163,8 @@ class FasterWhisperPipeline(Pipeline):
super(Pipeline, self).__init__()
self.vad_model = vad
self._vad_params = vad_params
self.previous_batch_context_tokens = []
def _sanitize_parameters(self, **kwargs):
preprocess_kwargs = {}
@@ -160,7 +183,35 @@ class FasterWhisperPipeline(Pipeline):
return {'inputs': features}
def _forward(self, model_inputs):
outputs = self.model.generate_segment_batched(model_inputs['inputs'], self.tokenizer, self.options)
current_batch_size = model_inputs['inputs'].shape[0]
# Ideally, batch[i] corresponds to stream[i].
# This holds if batch_size == number of streams.
valid_contexts = self.previous_batch_context_tokens[:current_batch_size]
outputs = self.model.generate_segment_batched(
model_inputs['inputs'],
self.tokenizer,
self.options,
use_batch_context=self.use_batch_context,
previous_batch_context_tokens=valid_contexts,
)
if self.use_batch_context:
initial_prompt_length = 0
if self.options.initial_prompt is not None:
initial_prompt = " " + self.options.initial_prompt.strip()
initial_prompt_length = len(self.tokenizer.encode(initial_prompt))
# Use 220 instead of 224 to be safe
max_context_window = max(0, 220 - initial_prompt_length)
for i, text in enumerate(outputs):
if i < len(self.previous_batch_context_tokens):
# Filter out special tokens (timestamps, SOT, EOT, etc.)
# We only want the text content for context.
tokens = [t for t in self.tokenizer.encode(text) if t < self.tokenizer.eot]
self.previous_batch_context_tokens[i].extend(tokens)
self.previous_batch_context_tokens[i] = self.previous_batch_context_tokens[i][-max_context_window:]
return {'text': outputs}
def postprocess(self, model_outputs):
@@ -201,6 +252,14 @@ class FasterWhisperPipeline(Pipeline):
) -> TranscriptionResult:
if isinstance(audio, str):
audio = load_audio(audio)
batch_size = batch_size or self._batch_size
# Initialize context for each stream.
# We have 'batch_size' concurrent streams.
if batch_size is None or batch_size < 1:
batch_size = 1
self.previous_batch_context_tokens = [[] for _ in range(batch_size)]
def data(audio, segments):
for seg in segments:
@@ -252,10 +311,33 @@ class FasterWhisperPipeline(Pipeline):
new_suppressed_tokens = numeral_symbol_tokens + self.options.suppress_tokens
new_suppressed_tokens = list(set(new_suppressed_tokens))
self.options = replace(self.options, suppress_tokens=new_suppressed_tokens)
segments: List[SingleSegment] = []
batch_size = batch_size or self._batch_size
total_segments = len(vad_segments)
if batch_size > 1 and self.use_batch_context:
num_streams = batch_size
# Distribute segments into streams
# Manual split
k, m = divmod(len(vad_segments), num_streams)
# lengths of each part: first m parts have k+1, rest have k
stream_segments = []
start_idx = 0
for i in range(num_streams):
part_len = k + 1 if i < m else k
stream_segments.append(vad_segments[start_idx : start_idx + part_len])
start_idx += part_len
# Interleave
# We need to pick [s0[0], s1[0], s2[0]... s0[1], s1[1]...]
interleaved_segments = []
max_len = max(len(s) for s in stream_segments)
for i in range(max_len):
for stream in stream_segments:
if i < len(stream):
interleaved_segments.append(stream[i])
vad_segments = interleaved_segments
for idx, out in enumerate(self.__call__(data(audio, vad_segments), batch_size=batch_size, num_workers=num_workers)):
if print_progress:
base_progress = ((idx + 1) / total_segments) * 100
@@ -274,6 +356,25 @@ class FasterWhisperPipeline(Pipeline):
}
)
if self.use_batch_context and batch_size > 1:
last_stream_index = (total_segments - 1) % batch_size
final_context = self.previous_batch_context_tokens[last_stream_index]
# Prepare context for the wrap-around re-run
# ONLY Stream 0 (which processes the start of the file) should get the context (which comes from the end of the file).
# All other streams should have EMPTY context for this re-run to avoid self-referencing loops (feeding Segment N to Segment N).
new_rerun_context = [[] for _ in range(batch_size)]
new_rerun_context[0] = final_context
# Temporarily overwrite previous_batch_context_tokens for the re-run
self.previous_batch_context_tokens = new_rerun_context
first_batch_segments = vad_segments[:batch_size]
# Runs the model again just on 'first_batch_segments'
for i, out in enumerate(self.__call__(data(audio, first_batch_segments), batch_size=batch_size, num_workers=num_workers)):
text = out['text']
# L398: Overwrite the existing text with the new wrap-around text
segments[i]['text'] = text
# Sort segments by start time to restore original order
segments.sort(key=lambda x: x['start'])
# revert the tokenizer if multilingual inference is enabled
if self.preset_language is None:
self.tokenizer = None
@@ -289,8 +390,8 @@ class FasterWhisperPipeline(Pipeline):
logger.warning("Audio is shorter than 30s, language detection may be inaccurate")
model_n_mels = self.model.feat_kwargs.get("feature_size")
segment = log_mel_spectrogram(audio[: N_SAMPLES],
n_mels=model_n_mels if model_n_mels is not None else 80,
padding=0 if audio.shape[0] >= N_SAMPLES else N_SAMPLES - audio.shape[0])
n_mels=model_n_mels if model_n_mels is not None else 80,
padding=0 if audio.shape[0] >= N_SAMPLES else N_SAMPLES - audio.shape[0])
encoder_output = self.model.encode(segment)
results = self.model.model.detect_language(encoder_output)
language_token, language_probability = results[0][0]
@@ -315,6 +416,7 @@ def load_model(
local_files_only=False,
threads=4,
use_auth_token: Optional[Union[str, bool]] = None,
use_batch_context: bool = False,
) -> FasterWhisperPipeline:
"""Load a Whisper model for inference.
Args:
@@ -421,4 +523,5 @@ def load_model(
language=language,
suppress_numerals=suppress_numerals,
vad_params=default_vad_options,
use_batch_context=use_batch_context,
)
+144
View File
@@ -0,0 +1,144 @@
import argparse
import os
import time
import torch
import torchaudio
import jiwer
import whisperx
import numpy as np
from typing import Tuple
def load_tedlium(root: str, download: bool = False, subset: str = "test"):
print(f"Loading TEDLIUM dataset ({subset}) from {root}...")
try:
dataset = torchaudio.datasets.TEDLIUM(
root=root,
release="release3",
subset=subset,
download=download
)
return dataset
except Exception as e:
print(f"Error loading dataset: {e}")
return None
def normalize_text(text: str) -> str:
"""
Simple normalization: lower case, remove punctuation.
"""
import string
text = text.lower()
text = text.translate(str.maketrans('', '', string.punctuation))
return " ".join(text.split())
def benchmark(dataset, model_size="large-v2", device="cuda", compute_type="float16", batch_size=4, limit=None):
print(f"Loading WhisperX model: {model_size} on {device} ({compute_type})...")
try:
model = whisperx.load_model(model_size, device, compute_type=compute_type)
except Exception as e:
print(f"Failed to load model: {e}")
return
print("Model loaded.")
total_wer = 0
total_cer = 0
total_latency = 0
total_audio_duration = 0
count = 0
print(f"\nBenchmarking on {limit if limit else len(dataset)} samples...")
# Clear CUDA cache for accurate VRAM measurement
if torch.cuda.is_available():
torch.cuda.empty_cache()
torch.cuda.reset_peak_memory_stats()
initial_vram = torch.cuda.memory_allocated() / 1024**3
print(f"Initial VRAM usage: {initial_vram:.2f} GB")
for i, item in enumerate(dataset):
if limit and i >= limit:
break
waveform, sample_rate, transcript, talk_id, speaker_id, identifier = item
# WhisperX expects audio as a numpy array, float32, mono, 16kHz
# TEDLIUM is likely 16kHz, but let's verify/resample if needed
# waveform is (channels, time)
if sample_rate != 16000:
resampler = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=16000)
waveform = resampler(waveform)
audio_np = waveform.squeeze().numpy()
duration = len(audio_np) / 16000
total_audio_duration += duration
# Measure Latency
start_time = time.time()
result = model.transcribe(audio_np, batch_size=batch_size)
end_time = time.time()
latency = end_time - start_time
total_latency += latency
# Combine segments for full transcript
hypothesis = " ".join([seg['text'] for seg in result['segments']])
# Normalize
ref_norm = normalize_text(transcript)
hyp_norm = normalize_text(hypothesis)
if not ref_norm.strip():
# Skip empty references to avoid division by zero in WER
continue
# Measure WER/CER
wer = jiwer.wer(ref_norm, hyp_norm)
cer = jiwer.cer(ref_norm, hyp_norm)
total_wer += wer
total_cer += cer
count += 1
print(f"Sample {i}: WER={wer:.2f}, CER={cer:.2f}, Latency={latency:.2f}s, Dur={duration:.2f}s, RTF={latency/duration:.2f}")
if count == 0:
print("No samples processed.")
return
avg_wer = total_wer / count
avg_cer = total_cer / count
avg_rtf = total_latency / total_audio_duration
print("\n--- Benchmark Results ---")
print(f"Average WER: {avg_wer:.4f}")
print(f"Average CER: {avg_cer:.4f}")
print(f"Average RTF (Real Time Factor): {avg_rtf:.4f}")
print(f"Total Latency: {total_latency:.2f}s for {total_audio_duration:.2f}s audio")
if torch.cuda.is_available():
peak_vram = torch.cuda.max_memory_allocated() / 1024**3
print(f"Peak VRAM Usage: {peak_vram:.2f} GB")
else:
print("VRAM Usage: N/A (CPU only)")
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Benchmark WhisperX on TEDLIUM")
parser.add_argument("--root", type=str, default="./data", help="Root directory for dataset")
parser.add_argument("--download", action="store_true", help="Download dataset if not found")
parser.add_argument("--limit", type=int, default=None, help="Limit number of samples")
parser.add_argument("--model", type=str, default="large-v2", help="Whisper model size")
parser.add_argument("--device", type=str, default="cuda" if torch.cuda.is_available() else "cpu", help="Device")
parser.add_argument("--batch_size", type=int, default=4, help="Batch size")
args = parser.parse_args()
# Create data dir
os.makedirs(args.root, exist_ok=True)
ds = load_tedlium(args.root, download=args.download)
if ds:
benchmark(ds, model_size=args.model, device=args.device, batch_size=args.batch_size, limit=args.limit)
+2
View File
@@ -63,6 +63,7 @@ def transcribe_task(args: dict, parser: argparse.ArgumentParser):
diarize_model_name: str = args.pop("diarize_model")
print_progress: bool = args.pop("print_progress")
return_speaker_embeddings: bool = args.pop("speaker_embeddings")
batch_context: bool = args.pop("batch_context", False)
if return_speaker_embeddings and not diarize:
warnings.warn("--speaker_embeddings has no effect without --diarize")
@@ -142,6 +143,7 @@ def transcribe_task(args: dict, parser: argparse.ArgumentParser):
local_files_only=model_cache_only,
threads=faster_whisper_threads,
use_auth_token=hf_token,
use_batch_context=batch_context,
)
for audio_path in args.pop("audio"):