eustlb’s gists

eustlb / benchmark_whisper_gen.py

Last active October 1, 2024 16:45

	import os
	import time
	import pickle
	from tqdm import tqdm

	import torch
	from transformers import WhisperForConditionalGeneration


	def benchmark_gen(

eustlb / benchmark_parler_static_cache_compile.py

Last active August 7, 2024 11:34

Benchmark parler-tts static cache and compilation.

	from dataclasses import dataclass, asdict, field
	import torch
	import soundfile as sf
	from parler_tts import ParlerTTSForConditionalGeneration
	from transformers import AutoTokenizer
	import torch._dynamo.config
	import torch._inductor.config
	import json
	import time
	import os

eustlb / test_parler.py

Last active August 2, 2024 12:18

Test generation of a parler tts branch for different combinaisons of parameters.

	from dataclasses import dataclass, asdict
	import torch
	import soundfile as sf
	from parler_tts import ParlerTTSForConditionalGeneration
	from transformers import AutoTokenizer
	import torch._dynamo.config
	import torch._inductor.config
	import numpy as np
	import time
	import os

eustlb / reproduce_bug_generation.py

Last active July 23, 2024 16:56

Reproduce generation error on dev branch

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

	torch._logging.set_logs(graph_breaks=True, recompiles=True)
	torch.manual_seed(0)

	CUDA_DEVICE = 0
	torch_device = f"cuda:{CUDA_DEVICE}"

eustlb / reproduce_bug_compile.py

Created July 23, 2024 16:30

Reproduce a bug happening with torch 2.3.1 and compile.

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

	torch.manual_seed(0)

	CUDA_DEVICE = 0
	torch_device = f"cuda:{CUDA_DEVICE}"

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")

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 / 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 / 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 / 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