Update app.py
Browse files
app.py
CHANGED
@@ -70,7 +70,7 @@ from models.soundstream_hubert_new import SoundStream
|
|
70 |
from vocoder import build_codec_model, process_audio
|
71 |
from post_process_audio import replace_low_freq_with_energy_matched
|
72 |
|
73 |
-
device = "cuda"
|
74 |
|
75 |
model = AutoModelForCausalLM.from_pretrained(
|
76 |
"m-a-p/YuE-s1-7B-anneal-en-cot",
|
@@ -90,23 +90,18 @@ mmtokenizer = _MMSentencePieceTokenizer("./mm_tokenizer_v0.2_hf/tokenizer.model"
|
|
90 |
|
91 |
codectool = CodecManipulator("xcodec", 0, 1)
|
92 |
model_config = OmegaConf.load(basic_model_config)
|
93 |
-
# Load codec model
|
94 |
codec_model = eval(model_config.generator.name)(**model_config.generator.config).to(device)
|
95 |
parameter_dict = torch.load(resume_path, map_location='cpu')
|
96 |
codec_model.load_state_dict(parameter_dict['codec_model'])
|
97 |
-
codec_model
|
98 |
codec_model.eval()
|
99 |
|
100 |
-
# Preload and compile vocoders
|
101 |
vocal_decoder, inst_decoder = build_codec_model(config_path, vocal_decoder_path, inst_decoder_path)
|
102 |
vocal_decoder.to(device)
|
103 |
inst_decoder.to(device)
|
104 |
-
vocal_decoder = torch.compile(vocal_decoder)
|
105 |
-
inst_decoder = torch.compile(inst_decoder)
|
106 |
vocal_decoder.eval()
|
107 |
inst_decoder.eval()
|
108 |
|
109 |
-
cuda_idx = 0
|
110 |
|
111 |
def generate_music(
|
112 |
max_new_tokens=5,
|
@@ -117,14 +112,13 @@ def generate_music(
|
|
117 |
audio_prompt_path="",
|
118 |
prompt_start_time=0.0,
|
119 |
prompt_end_time=30.0,
|
|
|
120 |
rescale=False,
|
121 |
):
|
122 |
if use_audio_prompt and not audio_prompt_path:
|
123 |
raise FileNotFoundError("Please offer audio prompt filepath using '--audio_prompt_path', when you enable 'use_audio_prompt'!")
|
124 |
-
|
125 |
-
# ------------------------------------------
|
126 |
max_new_tokens = max_new_tokens * 100
|
127 |
-
stage1_output_data = {}
|
128 |
|
129 |
with tempfile.TemporaryDirectory() as output_dir:
|
130 |
stage1_output_dir = os.path.join(output_dir, f"stage1")
|
@@ -179,17 +173,7 @@ def generate_music(
|
|
179 |
# Format text prompt
|
180 |
run_n_segments = min(run_n_segments + 1, len(lyrics))
|
181 |
|
182 |
-
|
183 |
-
'top_p': 0.93,
|
184 |
-
'temperature': 1.0,
|
185 |
-
'repetition_penalty': 1.2,
|
186 |
-
'top_k': 50, # Faster than top_p alone
|
187 |
-
'num_beams': 1, # Disable beam search
|
188 |
-
'max_new_tokens': max_new_tokens,
|
189 |
-
'min_new_tokens': 100,
|
190 |
-
'do_sample': True,
|
191 |
-
'use_cache': True,
|
192 |
-
}
|
193 |
|
194 |
for i, p in enumerate(tqdm(prompt_texts[:run_n_segments])):
|
195 |
section_text = p.replace('[start_of_segment]', '').replace('[end_of_segment]', '')
|
@@ -226,7 +210,7 @@ def generate_music(
|
|
226 |
print(
|
227 |
f'Section {i}: output length {input_ids.shape[-1]} exceeding context length {max_context}, now using the last {max_context} tokens.')
|
228 |
input_ids = input_ids[:, -(max_context):]
|
229 |
-
with torch.inference_mode(), torch.autocast(device_type=
|
230 |
output_seq = model.generate(
|
231 |
input_ids=input_ids,
|
232 |
max_new_tokens=max_new_tokens,
|
@@ -390,7 +374,8 @@ def generate_music(
|
|
390 |
def infer(genre_txt_content, lyrics_txt_content, num_segments=2, max_new_tokens=10):
|
391 |
# Execute the command
|
392 |
try:
|
393 |
-
audio_data = generate_music(genre_txt=genre_txt_content, lyrics_txt=lyrics_txt_content, run_n_segments=num_segments,
|
|
|
394 |
return audio_data
|
395 |
except Exception as e:
|
396 |
gr.Warning("An Error Occured: " + str(e))
|
|
|
70 |
from vocoder import build_codec_model, process_audio
|
71 |
from post_process_audio import replace_low_freq_with_energy_matched
|
72 |
|
73 |
+
device = "cuda:0"
|
74 |
|
75 |
model = AutoModelForCausalLM.from_pretrained(
|
76 |
"m-a-p/YuE-s1-7B-anneal-en-cot",
|
|
|
90 |
|
91 |
codectool = CodecManipulator("xcodec", 0, 1)
|
92 |
model_config = OmegaConf.load(basic_model_config)
|
|
|
93 |
codec_model = eval(model_config.generator.name)(**model_config.generator.config).to(device)
|
94 |
parameter_dict = torch.load(resume_path, map_location='cpu')
|
95 |
codec_model.load_state_dict(parameter_dict['codec_model'])
|
96 |
+
codec_model.to(device)
|
97 |
codec_model.eval()
|
98 |
|
|
|
99 |
vocal_decoder, inst_decoder = build_codec_model(config_path, vocal_decoder_path, inst_decoder_path)
|
100 |
vocal_decoder.to(device)
|
101 |
inst_decoder.to(device)
|
|
|
|
|
102 |
vocal_decoder.eval()
|
103 |
inst_decoder.eval()
|
104 |
|
|
|
105 |
|
106 |
def generate_music(
|
107 |
max_new_tokens=5,
|
|
|
112 |
audio_prompt_path="",
|
113 |
prompt_start_time=0.0,
|
114 |
prompt_end_time=30.0,
|
115 |
+
cuda_idx=0,
|
116 |
rescale=False,
|
117 |
):
|
118 |
if use_audio_prompt and not audio_prompt_path:
|
119 |
raise FileNotFoundError("Please offer audio prompt filepath using '--audio_prompt_path', when you enable 'use_audio_prompt'!")
|
120 |
+
cuda_idx = cuda_idx
|
|
|
121 |
max_new_tokens = max_new_tokens * 100
|
|
|
122 |
|
123 |
with tempfile.TemporaryDirectory() as output_dir:
|
124 |
stage1_output_dir = os.path.join(output_dir, f"stage1")
|
|
|
173 |
# Format text prompt
|
174 |
run_n_segments = min(run_n_segments + 1, len(lyrics))
|
175 |
|
176 |
+
print(list(enumerate(tqdm(prompt_texts[:run_n_segments]))))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
177 |
|
178 |
for i, p in enumerate(tqdm(prompt_texts[:run_n_segments])):
|
179 |
section_text = p.replace('[start_of_segment]', '').replace('[end_of_segment]', '')
|
|
|
210 |
print(
|
211 |
f'Section {i}: output length {input_ids.shape[-1]} exceeding context length {max_context}, now using the last {max_context} tokens.')
|
212 |
input_ids = input_ids[:, -(max_context):]
|
213 |
+
with torch.inference_mode(), torch.autocast(device_type='cuda', dtype=torch.float16):
|
214 |
output_seq = model.generate(
|
215 |
input_ids=input_ids,
|
216 |
max_new_tokens=max_new_tokens,
|
|
|
374 |
def infer(genre_txt_content, lyrics_txt_content, num_segments=2, max_new_tokens=10):
|
375 |
# Execute the command
|
376 |
try:
|
377 |
+
audio_data = generate_music(genre_txt=genre_txt_content, lyrics_txt=lyrics_txt_content, run_n_segments=num_segments,
|
378 |
+
cuda_idx=0, max_new_tokens=max_new_tokens)
|
379 |
return audio_data
|
380 |
except Exception as e:
|
381 |
gr.Warning("An Error Occured: " + str(e))
|