Spaces:
Paused
Paused
Update app.py
Browse files
app.py
CHANGED
@@ -1,46 +1,47 @@
|
|
1 |
# app.py βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
2 |
-
import os, json,
|
3 |
from fastapi import FastAPI, WebSocket, WebSocketDisconnect
|
4 |
from huggingface_hub import login
|
5 |
-
from transformers import
|
6 |
from transformers.generation.utils import Cache
|
7 |
from snac import SNAC
|
8 |
|
9 |
-
# ββ 0. HFβ
|
10 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
11 |
if HF_TOKEN:
|
12 |
login(HF_TOKEN)
|
13 |
|
14 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
15 |
|
16 |
-
#
|
17 |
torch.backends.cuda.enable_flash_sdp(False)
|
18 |
|
19 |
# ββ 1. Konstanten ββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
20 |
REPO = "SebastianBodza/Kartoffel_Orpheus-3B_german_natural-v0.1"
|
21 |
-
CHUNK_TOKENS = 50
|
22 |
START_TOKEN = 128259
|
23 |
NEW_BLOCK_TOKEN = 128257
|
24 |
EOS_TOKEN = 128258
|
25 |
-
AUDIO_BASE = 128266
|
26 |
VALID_AUDIO_IDS = torch.arange(AUDIO_BASE, AUDIO_BASE + 4096)
|
27 |
|
28 |
# ββ 2. Dynamischer LogitβMasker ββββββββββββββββββββββββββββββββββββββ
|
29 |
class DynamicAudioMask(LogitsProcessor):
|
30 |
-
"""
|
31 |
-
|
32 |
-
"""
|
33 |
def __init__(self, audio_ids: torch.Tensor, min_audio_blocks: int = 1):
|
34 |
super().__init__()
|
35 |
-
self.audio_ids
|
36 |
-
self.ctrl_ids
|
37 |
-
self.min_blocks
|
38 |
-
self.blocks_done
|
39 |
|
40 |
def __call__(self, input_ids, scores):
|
41 |
allowed = torch.cat([self.audio_ids, self.ctrl_ids])
|
42 |
-
if self.blocks_done >= self.min_blocks:
|
43 |
-
allowed = torch.cat([allowed,
|
|
|
|
|
44 |
mask = torch.full_like(scores, float("-inf"))
|
45 |
mask[:, allowed] = 0
|
46 |
return scores + mask
|
@@ -50,7 +51,7 @@ app = FastAPI()
|
|
50 |
|
51 |
@app.get("/")
|
52 |
async def ping():
|
53 |
-
return {"msg": "OrpheusβTTS
|
54 |
|
55 |
@app.on_event("startup")
|
56 |
async def load_models():
|
@@ -79,30 +80,23 @@ def build_inputs(text: str, voice: str):
|
|
79 |
ids = torch.cat(
|
80 |
[ torch.tensor([[START_TOKEN]], device=device),
|
81 |
ids,
|
82 |
-
torch.tensor([[128009, 128260]], device=device) ],
|
83 |
-
|
84 |
-
)
|
85 |
-
attn = torch.ones_like(ids)
|
86 |
-
return ids, attn
|
87 |
|
88 |
def decode_block(block7: list[int]) -> bytes:
|
89 |
l1, l2, l3 = [], [], []
|
90 |
b = block7
|
91 |
l1.append(b[0])
|
92 |
l2.append(b[1] - 4096)
|
93 |
-
l3.extend([b[2] - 8192,
|
94 |
l2.append(b[4] - 16384)
|
95 |
l3.extend([b[5] - 20480, b[6] - 24576])
|
96 |
|
97 |
-
codes = [
|
98 |
-
torch.tensor(l1, device=device).unsqueeze(0),
|
99 |
-
torch.tensor(l2, device=device).unsqueeze(0),
|
100 |
-
torch.tensor(l3, device=device).unsqueeze(0),
|
101 |
-
]
|
102 |
audio = snac.decode(codes).squeeze().cpu().numpy()
|
103 |
return (audio * 32767).astype("int16").tobytes()
|
104 |
|
105 |
-
# ββ 5. WebSocketβ
|
106 |
@app.websocket("/ws/tts")
|
107 |
async def tts(ws: WebSocket):
|
108 |
await ws.accept()
|
@@ -111,9 +105,9 @@ async def tts(ws: WebSocket):
|
|
111 |
text = req.get("text", "")
|
112 |
voice = req.get("voice", "Jakob")
|
113 |
|
114 |
-
ids, attn = build_inputs(text, voice)
|
115 |
past = None
|
116 |
-
last_tok = None
|
117 |
buf = []
|
118 |
|
119 |
while True:
|
@@ -126,24 +120,22 @@ async def tts(ws: WebSocket):
|
|
126 |
do_sample=True, temperature=0.7, top_p=0.95,
|
127 |
return_dict_in_generate=True,
|
128 |
use_cache=True,
|
129 |
-
return_legacy_cache=True
|
130 |
)
|
131 |
|
132 |
-
# ----- Cache & neue Token --------------------------------------
|
133 |
pkv = out.past_key_values
|
134 |
-
if isinstance(pkv, Cache):
|
135 |
pkv = pkv.to_legacy()
|
136 |
past = pkv
|
137 |
|
138 |
-
|
139 |
-
print("new tokens:",
|
140 |
|
141 |
-
if not
|
142 |
raise StopIteration
|
143 |
|
144 |
-
|
145 |
-
|
146 |
-
last_tok = t # speichern fΓΌr nΓ€chste Runde
|
147 |
|
148 |
if t == EOS_TOKEN:
|
149 |
raise StopIteration
|
@@ -156,9 +148,9 @@ async def tts(ws: WebSocket):
|
|
156 |
if len(buf) == 7:
|
157 |
await ws.send_bytes(decode_block(buf))
|
158 |
buf.clear()
|
159 |
-
masker.blocks_done += 1
|
160 |
|
161 |
-
# ab
|
162 |
ids, attn = None, None
|
163 |
|
164 |
except (StopIteration, WebSocketDisconnect):
|
@@ -174,7 +166,7 @@ async def tts(ws: WebSocket):
|
|
174 |
except RuntimeError:
|
175 |
pass
|
176 |
|
177 |
-
# ββ 6. Lokaler Start
|
178 |
if __name__ == "__main__":
|
179 |
import uvicorn
|
180 |
uvicorn.run("app:app", host="0.0.0.0", port=7860)
|
|
|
1 |
# app.py βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
2 |
+
import os, json, torch, asyncio
|
3 |
from fastapi import FastAPI, WebSocket, WebSocketDisconnect
|
4 |
from huggingface_hub import login
|
5 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, LogitsProcessor
|
6 |
from transformers.generation.utils import Cache
|
7 |
from snac import SNAC
|
8 |
|
9 |
+
# ββ 0. HFβAuth & Device ββββββββββββββββββββββββββββββββββββββββββββββ
|
10 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
11 |
if HF_TOKEN:
|
12 |
login(HF_TOKEN)
|
13 |
|
14 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
15 |
|
16 |
+
# FlashβSDPβBug (PyTorch 2.2) deaktivieren
|
17 |
torch.backends.cuda.enable_flash_sdp(False)
|
18 |
|
19 |
# ββ 1. Konstanten ββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
20 |
REPO = "SebastianBodza/Kartoffel_Orpheus-3B_german_natural-v0.1"
|
21 |
+
CHUNK_TOKENS = 50 # βMiniβGenerateββLΓ€nge
|
22 |
START_TOKEN = 128259
|
23 |
NEW_BLOCK_TOKEN = 128257
|
24 |
EOS_TOKEN = 128258
|
25 |
+
AUDIO_BASE = 128266
|
26 |
VALID_AUDIO_IDS = torch.arange(AUDIO_BASE, AUDIO_BASE + 4096)
|
27 |
|
28 |
# ββ 2. Dynamischer LogitβMasker ββββββββββββββββββββββββββββββββββββββ
|
29 |
class DynamicAudioMask(LogitsProcessor):
|
30 |
+
"""LΓ€sst zu Beginn nur Audioβ und NEW_BLOCKβTokens zu;
|
31 |
+
EOS erst, wenn min_audio_blocks fertig sind."""
|
|
|
32 |
def __init__(self, audio_ids: torch.Tensor, min_audio_blocks: int = 1):
|
33 |
super().__init__()
|
34 |
+
self.audio_ids = audio_ids
|
35 |
+
self.ctrl_ids = torch.tensor([NEW_BLOCK_TOKEN], device=audio_ids.device)
|
36 |
+
self.min_blocks = min_audio_blocks
|
37 |
+
self.blocks_done = 0
|
38 |
|
39 |
def __call__(self, input_ids, scores):
|
40 |
allowed = torch.cat([self.audio_ids, self.ctrl_ids])
|
41 |
+
if self.blocks_done >= self.min_blocks:
|
42 |
+
allowed = torch.cat([allowed,
|
43 |
+
torch.tensor([EOS_TOKEN],
|
44 |
+
device=scores.device)])
|
45 |
mask = torch.full_like(scores, float("-inf"))
|
46 |
mask[:, allowed] = 0
|
47 |
return scores + mask
|
|
|
51 |
|
52 |
@app.get("/")
|
53 |
async def ping():
|
54 |
+
return {"msg": "OrpheusβTTS OK"}
|
55 |
|
56 |
@app.on_event("startup")
|
57 |
async def load_models():
|
|
|
80 |
ids = torch.cat(
|
81 |
[ torch.tensor([[START_TOKEN]], device=device),
|
82 |
ids,
|
83 |
+
torch.tensor([[128009, 128260]], device=device) ], dim=1)
|
84 |
+
return ids, torch.ones_like(ids)
|
|
|
|
|
|
|
85 |
|
86 |
def decode_block(block7: list[int]) -> bytes:
|
87 |
l1, l2, l3 = [], [], []
|
88 |
b = block7
|
89 |
l1.append(b[0])
|
90 |
l2.append(b[1] - 4096)
|
91 |
+
l3.extend([b[2] - 8192, b[3] - 12288])
|
92 |
l2.append(b[4] - 16384)
|
93 |
l3.extend([b[5] - 20480, b[6] - 24576])
|
94 |
|
95 |
+
codes = [torch.tensor(x, device=device).unsqueeze(0) for x in (l1,l2,l3)]
|
|
|
|
|
|
|
|
|
96 |
audio = snac.decode(codes).squeeze().cpu().numpy()
|
97 |
return (audio * 32767).astype("int16").tobytes()
|
98 |
|
99 |
+
# ββ 5. WebSocketβEndpoint ββββββββββββββββββββββββββββββββββββββββββββ
|
100 |
@app.websocket("/ws/tts")
|
101 |
async def tts(ws: WebSocket):
|
102 |
await ws.accept()
|
|
|
105 |
text = req.get("text", "")
|
106 |
voice = req.get("voice", "Jakob")
|
107 |
|
108 |
+
ids, attn = build_inputs(text, voice)
|
109 |
past = None
|
110 |
+
last_tok = None
|
111 |
buf = []
|
112 |
|
113 |
while True:
|
|
|
120 |
do_sample=True, temperature=0.7, top_p=0.95,
|
121 |
return_dict_in_generate=True,
|
122 |
use_cache=True,
|
123 |
+
return_legacy_cache=True # verhindert CacheβWarnung
|
124 |
)
|
125 |
|
|
|
126 |
pkv = out.past_key_values
|
127 |
+
if isinstance(pkv, Cache): # HFΒ β₯Β 4.47
|
128 |
pkv = pkv.to_legacy()
|
129 |
past = pkv
|
130 |
|
131 |
+
new_toks = out.sequences[0, -out.num_generated_tokens:].tolist()
|
132 |
+
print("new tokens:", new_toks[:32]) # DebugβAusgabe
|
133 |
|
134 |
+
if not new_toks:
|
135 |
raise StopIteration
|
136 |
|
137 |
+
for t in new_toks:
|
138 |
+
last_tok = t
|
|
|
139 |
|
140 |
if t == EOS_TOKEN:
|
141 |
raise StopIteration
|
|
|
148 |
if len(buf) == 7:
|
149 |
await ws.send_bytes(decode_block(buf))
|
150 |
buf.clear()
|
151 |
+
masker.blocks_done += 1
|
152 |
|
153 |
+
# ab jetzt nur noch 1Β Token + Cache
|
154 |
ids, attn = None, None
|
155 |
|
156 |
except (StopIteration, WebSocketDisconnect):
|
|
|
166 |
except RuntimeError:
|
167 |
pass
|
168 |
|
169 |
+
# ββ 6. Lokaler TestβStart βββββββββββββββββββββββββββββββββββββββββββ
|
170 |
if __name__ == "__main__":
|
171 |
import uvicorn
|
172 |
uvicorn.run("app:app", host="0.0.0.0", port=7860)
|