eustlb’s gists

eustlb / benchmark_moonshine_tiny_artif.py

Last active January 21, 2025 16:13

Benchmark moonshine/ whisper for varying audio length and artificial number of generate tokens.

	import torch
	from transformers import MoonshineForConditionalGeneration, AutoProcessor
	from tqdm import tqdm
	import json

	device = "cuda:0" if torch.cuda.is_available() else "cpu"
	torch_dtype = torch.float32
	attn_implementation = "sdpa"

	model = MoonshineForConditionalGeneration.from_pretrained("UsefulSensors/moonshine-tiny", attn_implementation=attn_implementation).to(device, torch_dtype)

eustlb / benchmark_moonshine_tiny.py

Last active January 21, 2025 13:40

Benchmark moonshine/ whisper for varying batch sizes (FLEURS test set)

	import torch
	import evaluate
	from transformers.models.whisper.english_normalizer import EnglishTextNormalizer
	from transformers import MoonshineForConditionalGeneration, AutoProcessor, WhisperProcessor
	from datasets import load_dataset, Audio
	from tqdm import tqdm
	import json

	wer_metric = evaluate.load("wer")
	device = "cuda:0" if torch.cuda.is_available() else "cpu"

eustlb / benchmark_openai_whisper_myfork.py

Last active November 22, 2024 10:38

	OPENAI_SRC_PATH = "/admin/home/eustache_lebihan/dev/benchmark-whisper/whisper-myfork"

	import sys
	sys.path.insert(0, OPENAI_SRC_PATH)


	import wandb
	from tqdm import tqdm
	import evaluate
	import os

eustlb / benchmark_openai_whisper.py

Last active November 21, 2024 22:40

Benchmark WER and RTFx for openai whisper.

	OPENAI_SRC_PATH = "/admin/home/eustache_lebihan/dev/benchmark-whisper/whisper"

	import sys
	sys.path.insert(0, OPENAI_SRC_PATH)

	import wandb
	from tqdm import tqdm
	import evaluate
	import os
	import torch

eustlb / benchmark_transformers_whisper.py

Last active November 22, 2024 10:36

Benchmark WER and RTFx for transformers whisper.

	TRANSFORMERS_SRC_PATH = "/admin/home/eustache_lebihan/dev/benchmark-whisper/transformers-fix/src"

	import sys
	sys.path.insert(0, TRANSFORMERS_SRC_PATH)

	import wandb
	from tqdm import tqdm
	import evaluate
	import os
	import torch

eustlb / benchmark_parler_streaming.py

Last active September 19, 2024 09:45

Benchmark ParlerTTS + streaming time to first audio.

	import os
	import torch
	import time
	from parler_tts import ParlerTTSForConditionalGeneration, ParlerTTSStreamer
	from transformers import AutoTokenizer
	from threading import Thread


	# caching allows ~50% compilation time reduction
	# see https://docs.google.com/document/d/1y5CRfMLdwEoF1nTk9q8qEu1mgMUuUtvhklPKJ2emLU8/edit#heading=h.o2asbxsrp1ma

eustlb / test_compile_parler.py

Created September 5, 2024 16:38

Test compile on ParlerTTS + streaming

	import os
	import torch
	import time
	from parler_tts import ParlerTTSForConditionalGeneration, ParlerTTSStreamer
	from transformers import AutoTokenizer
	from threading import Thread

	# caching allows ~50% compilation time reduction
	# see https://docs.google.com/document/d/1y5CRfMLdwEoF1nTk9q8qEu1mgMUuUtvhklPKJ2emLU8/edit#heading=h.o2asbxsrp1ma
	CURRENT_DIR = os.path.dirname(os.path.realpath(__file__))

eustlb / reproduce_bug_description.py

Created September 5, 2024 14:00

Reproduces a bug when changing the description when using a compiled model

	import os
	import torch
	import soundfile as sf
	from parler_tts import ParlerTTSForConditionalGeneration
	from transformers import AutoTokenizer

	# caching allows ~50% compilation time reduction
	# see https://docs.google.com/document/d/1y5CRfMLdwEoF1nTk9q8qEu1mgMUuUtvhklPKJ2emLU8/edit#heading=h.o2asbxsrp1ma
	CURRENT_DIR = os.path.dirname(os.path.realpath(__file__))
	os.environ["TORCHINDUCTOR_CACHE_DIR"] = os.path.join(CURRENT_DIR, "tmp")

eustlb / reproducer_bug_jenny.py

Created July 29, 2024 09:45

Reproduce generation bug for a long prompt with the Jenny model (Parler-TTS).

	import torch
	import soundfile as sf
	from parler_tts import ParlerTTSForConditionalGeneration
	from transformers import AutoTokenizer

	model_name = "ylacombe/parler-tts-mini-jenny-30H"
	torch_device = "cuda:0"
	torch_dtype = torch.bfloat16
	attn_implementation = "eager"

eustlb / parler_compile_multiple.py

Last active September 5, 2024 16:08

	import os
	import torch
	import soundfile as sf
	from parler_tts import ParlerTTSForConditionalGeneration
	from transformers import AutoTokenizer

	# caching allows ~50% compilation time reduction
	# see https://docs.google.com/document/d/1y5CRfMLdwEoF1nTk9q8qEu1mgMUuUtvhklPKJ2emLU8/edit#heading=h.o2asbxsrp1ma
	CURRENT_DIR = os.path.dirname(os.path.realpath(__file__))
	os.environ["TORCHINDUCTOR_CACHE_DIR"] = os.path.join(CURRENT_DIR, "tmp")