Update handler.py
Browse filesReverting back to non-streaming vLLM implementation
- handler.py +271 -198
handler.py
CHANGED
@@ -1,21 +1,16 @@
|
|
1 |
-
import asyncio
|
2 |
-
import torch
|
3 |
import os
|
4 |
-
import
|
5 |
-
import queue
|
6 |
import numpy as np
|
7 |
-
import
|
|
|
|
|
8 |
import base64
|
9 |
import io
|
10 |
import wave
|
11 |
-
import librosa
|
12 |
-
import soundfile as sf
|
13 |
-
import random
|
14 |
|
15 |
-
from
|
16 |
-
from vllm.sampling_params import RequestOutputKind
|
17 |
-
from transformers import AutoTokenizer, pipeline
|
18 |
from snac import SNAC
|
|
|
19 |
|
20 |
class EndpointHandler:
|
21 |
def __init__(self, path=""):
|
@@ -31,15 +26,9 @@ class EndpointHandler:
|
|
31 |
self.END_OF_AI = 128262
|
32 |
self.AUDIO_TOKENS_START = 128266
|
33 |
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
max_model_len = 4096,
|
38 |
-
gpu_memory_utilization = 0.5,
|
39 |
-
)
|
40 |
-
self.engine = AsyncLLMEngine.from_engine_args(self.engine_args)
|
41 |
-
|
42 |
-
self.tokenizer = AutoTokenizer.from_pretrained("okezieowen/hypaai_orpheus")
|
43 |
|
44 |
# Move to devices
|
45 |
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
@@ -52,201 +41,180 @@ class EndpointHandler:
|
|
52 |
except Exception as e:
|
53 |
raise RuntimeError(f"Failed to load SNAC model: {e}")
|
54 |
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
torch.tensor([[self.START_OF_HUMAN]], dtype=torch.int64),
|
60 |
torch.tensor([[self.START_OF_TEXT]], dtype=torch.int64),
|
61 |
-
|
62 |
torch.tensor([[self.END_OF_TEXT]], dtype=torch.int64),
|
63 |
-
torch.tensor([[self.END_OF_HUMAN]], dtype=torch.int64)
|
|
|
|
|
|
|
|
|
64 |
torch.tensor([[self.START_OF_AI]], dtype=torch.int64),
|
65 |
torch.tensor([[self.START_OF_SPEECH]], dtype=torch.int64),
|
66 |
-
],
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
|
71 |
-
def
|
72 |
"""
|
73 |
-
|
|
|
|
|
|
|
|
|
74 |
"""
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
|
82 |
-
|
|
|
|
|
83 |
|
84 |
-
|
85 |
-
|
86 |
-
code_list = code_list[:n_codes]
|
87 |
|
88 |
-
|
89 |
-
|
90 |
-
layer_1.append(code_list[idx + 0])
|
91 |
-
layer_2.append(code_list[idx + 1])
|
92 |
-
layer_3.append(code_list[idx + 2])
|
93 |
-
layer_3.append(code_list[idx + 3])
|
94 |
-
layer_2.append(code_list[idx + 4])
|
95 |
-
layer_3.append(code_list[idx + 5])
|
96 |
-
layer_3.append(code_list[idx + 6])
|
97 |
|
98 |
-
|
99 |
-
torch.tensor(layer_1).unsqueeze(0).to(self.device),
|
100 |
-
torch.tensor(layer_2).unsqueeze(0).to(self.device),
|
101 |
-
torch.tensor(layer_3).unsqueeze(0).to(self.device),
|
102 |
-
]
|
103 |
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
|
108 |
-
# Decode audio
|
109 |
with torch.inference_mode():
|
110 |
-
|
111 |
-
return audio_hat
|
112 |
-
|
113 |
-
def _turn_token_into_id(self, token_string, index):
|
114 |
-
# Strip whitespace
|
115 |
-
token_string = token_string.strip()
|
116 |
-
|
117 |
-
# Find the last token in the string
|
118 |
-
last_token_start = token_string.rfind("<custom_token_")
|
119 |
-
|
120 |
-
if last_token_start == -1:
|
121 |
-
print("No token found in the string")
|
122 |
-
return None
|
123 |
|
124 |
-
|
125 |
-
|
126 |
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
else:
|
135 |
-
return None
|
136 |
-
|
137 |
-
async def _generate_token(self, prompt_string, sampling_params, request_id):
|
138 |
-
async for ro in self.engine.generate(prompt=prompt_string, sampling_params=sampling_params, request_id=request_id):
|
139 |
-
token = ro.outputs[0].text
|
140 |
-
yield token
|
141 |
-
|
142 |
-
async def _generate_token_buffer(self, token_gen, audio_frame_width, audio_frame_overlap):
|
143 |
-
last_emit = 0
|
144 |
-
buffer = []
|
145 |
-
count = 0
|
146 |
-
hop_length = (audio_frame_width - audio_frame_overlap) * 7
|
147 |
-
token_frame_width = audio_frame_width * 7
|
148 |
-
|
149 |
-
async for token in token_gen:
|
150 |
-
token_id = self._turn_token_into_id(token, count)
|
151 |
-
if token_id is None:
|
152 |
-
continue
|
153 |
-
|
154 |
-
# Accept only token IDs in [0, 4095]
|
155 |
-
if 0 <= token_id < 4096:
|
156 |
-
buffer.append(token_id)
|
157 |
-
count += 1
|
158 |
-
else:
|
159 |
-
continue
|
160 |
-
|
161 |
-
while count - last_emit >= hop_length and count >= token_frame_width:
|
162 |
-
buffer_to_process = buffer[-token_frame_width:]
|
163 |
-
yield buffer_to_process
|
164 |
-
last_emit += hop_length
|
165 |
-
|
166 |
-
# After the vLLM engine finishes, yield any remaining tokens.
|
167 |
-
if count > last_emit:
|
168 |
-
# Pad the final buffer to be a multiple of 7 before yielding.
|
169 |
-
remaining_len = len(buffer) % token_frame_width
|
170 |
-
if remaining_len != 0:
|
171 |
-
padding_needed = token_frame_width - remaining_len
|
172 |
-
buffer.extend([0] * padding_needed)
|
173 |
-
|
174 |
-
# Process and yield the final, potentially incomplete but padded buffer.
|
175 |
-
buffer_to_process = buffer[-token_frame_width:]
|
176 |
-
yield buffer_to_process
|
177 |
-
|
178 |
-
async def _decode_tokens(self, token_buffer_generator):
|
179 |
-
async for audio_token_buffer in token_buffer_generator:
|
180 |
-
audio_samples = self._convert_codes_to_audio_array(audio_token_buffer)
|
181 |
-
yield audio_samples
|
182 |
-
|
183 |
-
async def _convert_audio_tensor_to_audio_numpy(self, audio_tensor_generator):
|
184 |
-
async for audio_tensor in audio_tensor_generator:
|
185 |
-
audio_numpy = audio_tensor.detach().squeeze().cpu().numpy()
|
186 |
-
|
187 |
-
# # Convert float32 array to int16 for WAV format
|
188 |
-
# audio_int16 = (audio_numpy * 32767).astype(np.int16)
|
189 |
-
|
190 |
-
# # Write to WAV in memory (float32 or int16 depending on your preference)
|
191 |
-
# buffer = io.BytesIO()
|
192 |
-
# sf.write(buffer, audio_numpy, samplerate=24000, format='WAV', subtype='PCM_16') # or PCM_32
|
193 |
-
# buffer.seek(0)
|
194 |
-
|
195 |
-
# # Encode WAV bytes as base64
|
196 |
-
# audio_b64 = base64.b64encode(buffer.read()).decode('utf-8')
|
197 |
-
|
198 |
-
yield audio_numpy
|
199 |
-
|
200 |
-
async def _generate_speech(self, prompt_string, sampling_params, request_id, audio_frame_width, audio_frame_overlap):
|
201 |
-
|
202 |
-
# Step 1: Generate tokens from prompt
|
203 |
-
token_gen = self._generate_token(
|
204 |
-
prompt_string=prompt_string,
|
205 |
-
sampling_params=sampling_params,
|
206 |
-
request_id=request_id,
|
207 |
-
)
|
208 |
-
|
209 |
-
# Step 2 : Buffer tokens
|
210 |
-
token_buffer_gen = self._generate_token_buffer(
|
211 |
-
token_gen,
|
212 |
-
audio_frame_width,
|
213 |
-
audio_frame_overlap,
|
214 |
-
)
|
215 |
|
216 |
-
|
217 |
-
audio_tensor_gen = self._decode_tokens(token_buffer_gen)
|
218 |
|
219 |
-
|
220 |
-
audio_numpy_gen = self._convert_audio_tensor_to_audio_numpy(audio_tensor_gen)
|
221 |
|
222 |
-
#
|
223 |
-
async for audio_numpy in audio_numpy_gen:
|
224 |
-
yield audio_numpy
|
225 |
|
226 |
-
|
227 |
|
228 |
-
|
229 |
-
|
230 |
-
|
231 |
parameters = data.get("parameters", {})
|
|
|
232 |
|
233 |
temperature = float(parameters.get("temperature", 0.6))
|
234 |
top_p = float(parameters.get("top_p", 0.95))
|
235 |
max_new_tokens = int(parameters.get("max_new_tokens", 1200))
|
236 |
repetition_penalty = float(parameters.get("repetition_penalty", 1.1))
|
237 |
-
audio_frame_width = int(parameters.get("audio_frame_width", 10))
|
238 |
-
audio_frame_overlap = int(parameters.get("audio_frame_overlap", 5))
|
239 |
|
240 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
241 |
|
242 |
return {
|
243 |
-
"
|
244 |
"temperature": temperature,
|
245 |
"top_p": top_p,
|
246 |
"max_new_tokens": max_new_tokens,
|
247 |
"repetition_penalty": repetition_penalty,
|
248 |
-
"audio_frame_width": audio_frame_width,
|
249 |
-
"audio_frame_overlap": audio_frame_overlap,
|
250 |
}
|
251 |
|
252 |
def inference(self, inputs):
|
@@ -254,34 +222,139 @@ class EndpointHandler:
|
|
254 |
Run model inference on the preprocessed inputs
|
255 |
"""
|
256 |
# Extract parameters
|
257 |
-
|
258 |
|
259 |
-
|
260 |
temperature = inputs["temperature"],
|
261 |
top_p = inputs["top_p"],
|
262 |
-
max_tokens = inputs["max_new_tokens"],
|
263 |
repetition_penalty = inputs["repetition_penalty"],
|
264 |
stop_token_ids = [self.END_OF_SPEECH],
|
265 |
)
|
|
|
|
|
266 |
|
267 |
-
|
268 |
-
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
|
275 |
|
276 |
-
# Main entry point for the handler
|
277 |
-
async def __call__(self, data):
|
278 |
try:
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
283 |
|
284 |
# Catch that error, baby
|
285 |
except Exception as e:
|
286 |
traceback.print_exc()
|
287 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import os
|
2 |
+
import torch
|
|
|
3 |
import numpy as np
|
4 |
+
import librosa
|
5 |
+
import soundfile as sf
|
6 |
+
import traceback
|
7 |
import base64
|
8 |
import io
|
9 |
import wave
|
|
|
|
|
|
|
10 |
|
11 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
|
12 |
from snac import SNAC
|
13 |
+
from vllm import LLM, SamplingParams
|
14 |
|
15 |
class EndpointHandler:
|
16 |
def __init__(self, path=""):
|
|
|
26 |
self.END_OF_AI = 128262
|
27 |
self.AUDIO_TOKENS_START = 128266
|
28 |
|
29 |
+
# Load the models and tokenizer
|
30 |
+
self.model = LLM(path, max_model_len = 4096, gpu_memory_utilization = 0.3)
|
31 |
+
self.tokenizer = AutoTokenizer.from_pretrained(path)
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
|
33 |
# Move to devices
|
34 |
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
|
|
41 |
except Exception as e:
|
42 |
raise RuntimeError(f"Failed to load SNAC model: {e}")
|
43 |
|
44 |
+
# Set up functions to format and encode text/audio
|
45 |
+
def encode_text(self, text):
|
46 |
+
return self.tokenizer.encode(text, return_tensors="pt", add_special_tokens=False)
|
47 |
+
|
48 |
+
def encode_audio(self, base64_audio_str):
|
49 |
+
audio_bytes = base64.b64decode(base64_audio_str)
|
50 |
+
audio_buffer = io.BytesIO(audio_bytes)
|
51 |
+
waveform, sr = sf.read(audio_buffer, dtype='float32')
|
52 |
+
|
53 |
+
if waveform.ndim > 1:
|
54 |
+
waveform = np.mean(waveform, axis=1)
|
55 |
+
if sr != 24000:
|
56 |
+
waveform = librosa.resample(waveform, orig_sr=sr, target_sr=24000)
|
57 |
+
return self.tokenize_audio(waveform)
|
58 |
+
|
59 |
+
def format_text_block(self, text_ids):
|
60 |
+
return [
|
61 |
torch.tensor([[self.START_OF_HUMAN]], dtype=torch.int64),
|
62 |
torch.tensor([[self.START_OF_TEXT]], dtype=torch.int64),
|
63 |
+
text_ids,
|
64 |
torch.tensor([[self.END_OF_TEXT]], dtype=torch.int64),
|
65 |
+
torch.tensor([[self.END_OF_HUMAN]], dtype=torch.int64)
|
66 |
+
]
|
67 |
+
|
68 |
+
def format_audio_block(self, audio_codes):
|
69 |
+
return [
|
70 |
torch.tensor([[self.START_OF_AI]], dtype=torch.int64),
|
71 |
torch.tensor([[self.START_OF_SPEECH]], dtype=torch.int64),
|
72 |
+
torch.tensor([audio_codes], dtype=torch.int64),
|
73 |
+
torch.tensor([[self.END_OF_SPEECH]], dtype=torch.int64),
|
74 |
+
torch.tensor([[self.END_OF_AI]], dtype=torch.int64)
|
75 |
+
]
|
76 |
|
77 |
+
def enroll_user(self, enrollment_pairs):
|
78 |
"""
|
79 |
+
Parameters:
|
80 |
+
- enrollment_pairs: List of tuples (text, audio_data), where audio_data is
|
81 |
+
base64-encoded audio data
|
82 |
+
Returns:
|
83 |
+
- cloning_features (str): serialized enrollment data
|
84 |
"""
|
85 |
+
enrollment_data = []
|
86 |
+
|
87 |
+
for text, base64_audio in enrollment_pairs:
|
88 |
+
text_ids = self.encode_text(text).cpu()
|
89 |
+
audio_codes = self.encode_audio(base64_audio)
|
90 |
+
enrollment_data.append({
|
91 |
+
"text_ids": text_ids,
|
92 |
+
"audio_codes": audio_codes
|
93 |
+
})
|
94 |
+
|
95 |
+
# Serialize enrollment data
|
96 |
+
buffer = io.BytesIO()
|
97 |
+
torch.save(enrollment_data, buffer)
|
98 |
+
buffer.seek(0)
|
99 |
+
|
100 |
+
# Encode as base64 string and assign to attribute
|
101 |
+
cloning_features = base64.b64encode(buffer.read()).decode('utf-8')
|
102 |
+
return cloning_features
|
103 |
+
|
104 |
+
def prepare_audio_tokens_for_decoder(self, audio_codes_list):
|
105 |
+
"""
|
106 |
+
Given a list containing sequences of generated audio codes, do the following:
|
107 |
+
1. Trim length to a multiple of 7 (SNAC decoder requires 7 tokens per audio frame)
|
108 |
+
2. Adjust token values to SNAC decoder's expected range
|
109 |
+
"""
|
110 |
+
modified_audio_codes_list = []
|
111 |
+
for audio_codes in audio_codes_list:
|
112 |
|
113 |
+
# Trim length to a multiple of 7
|
114 |
+
length = (audio_codes.size(0) // 7) * 7
|
115 |
+
trimmed = audio_codes[:length]
|
116 |
|
117 |
+
# Adjust token values to SNAC decoder's expected range
|
118 |
+
audio_codes = trimmed - self.AUDIO_TOKENS_START
|
|
|
119 |
|
120 |
+
# Add modified audio codes to list
|
121 |
+
modified_audio_codes_list.append(audio_codes)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
122 |
|
123 |
+
return modified_audio_codes_list
|
|
|
|
|
|
|
|
|
124 |
|
125 |
+
# Convert audio sample to codes and reconstruct
|
126 |
+
def tokenize_audio(self, waveform):
|
127 |
+
waveform = torch.from_numpy(waveform).unsqueeze(0).unsqueeze(0).to(self.device)
|
128 |
|
|
|
129 |
with torch.inference_mode():
|
130 |
+
codes = self.snac_model.encode(waveform)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
131 |
|
132 |
+
all_codes = []
|
133 |
+
for i in range(codes[0].shape[1]):
|
134 |
|
135 |
+
all_codes.append(codes[0][0][(1 * i) + 0].item() + self.AUDIO_TOKENS_START + (0 * 4096))
|
136 |
+
all_codes.append(codes[1][0][(2 * i) + 0].item() + self.AUDIO_TOKENS_START + (1 * 4096))
|
137 |
+
all_codes.append(codes[2][0][(4 * i) + 0].item() + self.AUDIO_TOKENS_START + (2 * 4096))
|
138 |
+
all_codes.append(codes[2][0][(4 * i) + 1].item() + self.AUDIO_TOKENS_START + (3 * 4096))
|
139 |
+
all_codes.append(codes[1][0][(2 * i) + 1].item() + self.AUDIO_TOKENS_START + (4 * 4096))
|
140 |
+
all_codes.append(codes[2][0][(4 * i) + 2].item() + self.AUDIO_TOKENS_START + (5 * 4096))
|
141 |
+
all_codes.append(codes[2][0][(4 * i) + 3].item() + self.AUDIO_TOKENS_START + (6 * 4096))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
142 |
|
143 |
+
return all_codes
|
|
|
144 |
|
145 |
+
def preprocess(self, data):
|
|
|
146 |
|
147 |
+
# Preprocess input data before inference
|
|
|
|
|
148 |
|
149 |
+
self.voice_cloning = data.get("clone", False)
|
150 |
|
151 |
+
# Extract parameters from request
|
152 |
+
target_text = data["inputs"]
|
|
|
153 |
parameters = data.get("parameters", {})
|
154 |
+
cloning_features = data.get("cloning_features", None)
|
155 |
|
156 |
temperature = float(parameters.get("temperature", 0.6))
|
157 |
top_p = float(parameters.get("top_p", 0.95))
|
158 |
max_new_tokens = int(parameters.get("max_new_tokens", 1200))
|
159 |
repetition_penalty = float(parameters.get("repetition_penalty", 1.1))
|
|
|
|
|
160 |
|
161 |
+
if self.voice_cloning:
|
162 |
+
"""Handle voice cloning using cloning features"""
|
163 |
+
|
164 |
+
if not cloning_features:
|
165 |
+
raise ValueError("No cloning features were provided")
|
166 |
+
else:
|
167 |
+
# Decode back into tensors
|
168 |
+
enrollment_data = torch.load(io.BytesIO(base64.b64decode(cloning_features)))
|
169 |
+
|
170 |
+
# Process pre-tokenized enrollment_data
|
171 |
+
input_sequence = []
|
172 |
+
for item in enrollment_data:
|
173 |
+
text_ids = item["text_ids"]
|
174 |
+
audio_codes = item["audio_codes"]
|
175 |
+
input_sequence.extend(self.format_text_block(text_ids))
|
176 |
+
input_sequence.extend(self.format_audio_block(audio_codes))
|
177 |
+
|
178 |
+
# Append target text whose audio we want
|
179 |
+
target_text_ids = self.encode_text(target_text)
|
180 |
+
input_sequence.extend(self.format_text_block(target_text_ids))
|
181 |
+
|
182 |
+
# Start of target audio - audio codes to be completed by model
|
183 |
+
input_sequence.extend([
|
184 |
+
torch.tensor([[self.START_OF_AI]], dtype=torch.int64),
|
185 |
+
torch.tensor([[self.START_OF_SPEECH]], dtype=torch.int64)
|
186 |
+
])
|
187 |
+
|
188 |
+
# Final input tensor
|
189 |
+
input_ids = torch.cat(input_sequence, dim=1)
|
190 |
+
|
191 |
+
# Heuristic to determine max_new_tokens based on empirical relationship
|
192 |
+
# between the length of the prompt ids and the length of the generated ids
|
193 |
+
prompt_ids = self.encode_text(target_text)
|
194 |
+
max_new_tokens = int(prompt_ids.size()[1] * 20 + 200)
|
195 |
+
|
196 |
+
input_ids = input_ids.to(self.device)
|
197 |
+
|
198 |
+
else:
|
199 |
+
# Handle standard text-to-speech
|
200 |
+
|
201 |
+
# Extract parameters from request
|
202 |
+
voice = parameters.get("voice", "Eniola")
|
203 |
+
prompt = f"{voice}: {target_text}"
|
204 |
+
input_ids = self.tokenizer(prompt, return_tensors="pt").input_ids
|
205 |
+
|
206 |
+
# Add special tokens
|
207 |
+
input_ids = torch.cat(self.format_text_block(input_ids), dim=1)
|
208 |
+
|
209 |
+
# No need for padding as we're processing a single sequence
|
210 |
+
input_ids = input_ids.to(self.device)
|
211 |
|
212 |
return {
|
213 |
+
"input_ids": input_ids,
|
214 |
"temperature": temperature,
|
215 |
"top_p": top_p,
|
216 |
"max_new_tokens": max_new_tokens,
|
217 |
"repetition_penalty": repetition_penalty,
|
|
|
|
|
218 |
}
|
219 |
|
220 |
def inference(self, inputs):
|
|
|
222 |
Run model inference on the preprocessed inputs
|
223 |
"""
|
224 |
# Extract parameters
|
225 |
+
input_ids = inputs["input_ids"]
|
226 |
|
227 |
+
sampling_params = SamplingParams(
|
228 |
temperature = inputs["temperature"],
|
229 |
top_p = inputs["top_p"],
|
230 |
+
max_tokens = inputs["max_new_tokens"],
|
231 |
repetition_penalty = inputs["repetition_penalty"],
|
232 |
stop_token_ids = [self.END_OF_SPEECH],
|
233 |
)
|
234 |
+
|
235 |
+
prompt_string = self.tokenizer.decode(input_ids[0])
|
236 |
|
237 |
+
# Forward pass through the model
|
238 |
+
generated_ids = self.model.generate(prompt_string, sampling_params)
|
239 |
+
|
240 |
+
return torch.tensor(generated_ids[0].outputs[0].token_ids).unsqueeze(0)
|
241 |
+
|
242 |
+
def __call__(self, data):
|
243 |
+
|
244 |
+
# Main entry point for the handler
|
245 |
|
|
|
|
|
246 |
try:
|
247 |
+
enroll_user = data.get("enroll_user", False)
|
248 |
+
|
249 |
+
if enroll_user:
|
250 |
+
# We extract cloning features for enrollment
|
251 |
+
enrollment_pairs = data.get("enrollments", [])
|
252 |
+
cloning_features = self.enroll_user(enrollment_pairs)
|
253 |
+
return {"cloning_features": cloning_features}
|
254 |
+
else:
|
255 |
+
# We want to generate speech using preset cloning features
|
256 |
+
preprocessed_inputs = self.preprocess(data)
|
257 |
+
model_outputs = self.inference(preprocessed_inputs)
|
258 |
+
response = self.postprocess(model_outputs)
|
259 |
+
return response
|
260 |
|
261 |
# Catch that error, baby
|
262 |
except Exception as e:
|
263 |
traceback.print_exc()
|
264 |
+
return {"error": str(e)}
|
265 |
+
|
266 |
+
# Postprocess generated ids
|
267 |
+
def convert_codes_to_waveform(self, code_list):
|
268 |
+
"""
|
269 |
+
Reorganize tokens for SNAC decoding
|
270 |
+
"""
|
271 |
+
layer_1 = [] # Coarsest layer
|
272 |
+
layer_2 = [] # Intermediate layer
|
273 |
+
layer_3 = [] # Finest layer
|
274 |
+
|
275 |
+
num_groups = len(code_list) // 7
|
276 |
+
for i in range(num_groups):
|
277 |
+
idx = 7 * i
|
278 |
+
layer_1.append(code_list[7 * i + 0] - (0 * 4096))
|
279 |
+
layer_2.append(code_list[7 * i + 1] - (1 * 4096))
|
280 |
+
layer_3.append(code_list[7 * i + 2] - (2 * 4096))
|
281 |
+
layer_3.append(code_list[7 * i + 3] - (3 * 4096))
|
282 |
+
layer_2.append(code_list[7 * i + 4] - (4 * 4096))
|
283 |
+
layer_3.append(code_list[7 * i + 5] - (5 * 4096))
|
284 |
+
layer_3.append(code_list[7 * i + 6] - (6 * 4096))
|
285 |
+
|
286 |
+
codes = [
|
287 |
+
torch.tensor(layer_1).unsqueeze(0).to(self.device),
|
288 |
+
torch.tensor(layer_2).unsqueeze(0).to(self.device),
|
289 |
+
torch.tensor(layer_3).unsqueeze(0).to(self.device)
|
290 |
+
]
|
291 |
+
|
292 |
+
# Decode audio
|
293 |
+
audio_hat = self.snac_model.decode(codes)
|
294 |
+
return audio_hat
|
295 |
+
|
296 |
+
def postprocess(self, generated_ids):
|
297 |
+
|
298 |
+
if self.voice_cloning:
|
299 |
+
"""
|
300 |
+
For cloning applications, use this postprocess function to get generated audio samples
|
301 |
+
"""
|
302 |
+
# Modify audio codes to be digestible byb SNAC decoder
|
303 |
+
code_lists = self.prepare_audio_tokens_for_decoder(generated_ids)
|
304 |
+
|
305 |
+
# Generate audio from codes
|
306 |
+
temp = self.convert_codes_to_waveform(code_lists[0])
|
307 |
+
audio_sample = temp.detach().squeeze().to("cpu").numpy()
|
308 |
+
|
309 |
+
else:
|
310 |
+
"""
|
311 |
+
Process generated tokens into audio
|
312 |
+
"""
|
313 |
+
# Find Start of Audio token
|
314 |
+
token_indices = (generated_ids == self.START_OF_SPEECH).nonzero(as_tuple=True)
|
315 |
+
|
316 |
+
if len(token_indices[1]) > 0:
|
317 |
+
last_occurrence_idx = token_indices[1][-1].item()
|
318 |
+
cropped_tensor = generated_ids[:, last_occurrence_idx+1:]
|
319 |
+
else:
|
320 |
+
cropped_tensor = generated_ids
|
321 |
+
|
322 |
+
# Remove End of Audio tokens
|
323 |
+
processed_rows = []
|
324 |
+
for row in cropped_tensor:
|
325 |
+
masked_row = row[row != self.END_OF_SPEECH]
|
326 |
+
processed_rows.append(masked_row)
|
327 |
+
|
328 |
+
code_lists = self.prepare_audio_tokens_for_decoder(processed_rows)
|
329 |
+
|
330 |
+
# Generate audio from codes
|
331 |
+
audio_samples = []
|
332 |
+
for code_list in code_lists:
|
333 |
+
if len(code_list) > 0:
|
334 |
+
audio = self.convert_codes_to_waveform(code_list)
|
335 |
+
audio_samples.append(audio)
|
336 |
+
else:
|
337 |
+
raise ValueError("Empty code list, no audio to generate")
|
338 |
+
|
339 |
+
if not audio_samples:
|
340 |
+
return {"error": "No audio samples generated"}
|
341 |
+
|
342 |
+
# Return first (and only) audio sample
|
343 |
+
audio_sample = audio_samples[0].detach().squeeze().cpu().numpy()
|
344 |
+
|
345 |
+
# Convert float32 array to int16 for WAV format
|
346 |
+
audio_int16 = (audio_sample * 32767).astype(np.int16)
|
347 |
+
|
348 |
+
# Write to WAV in memory (float32 or int16 depending on your preference)
|
349 |
+
buffer = io.BytesIO()
|
350 |
+
sf.write(buffer, audio_sample, samplerate=24000, format='WAV', subtype='PCM_16') # or PCM_32
|
351 |
+
buffer.seek(0)
|
352 |
+
|
353 |
+
# Encode WAV bytes as base64
|
354 |
+
audio_b64 = base64.b64encode(buffer.read()).decode('utf-8')
|
355 |
+
|
356 |
+
return {
|
357 |
+
"audio_sample": audio_sample,
|
358 |
+
"audio_b64": audio_b64,
|
359 |
+
"sample_rate": 24000,
|
360 |
+
}
|