Spaces:
Paused
Paused
Update app.py
Browse files
app.py
CHANGED
@@ -6,46 +6,47 @@ from transformers import AutoTokenizer, AutoModelForCausalLM, LogitsProcessor
|
|
6 |
from transformers.generation.utils import Cache
|
7 |
from snac import SNAC
|
8 |
|
9 |
-
#
|
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 |
-
torch.backends.cuda.enable_flash_sdp(False)
|
16 |
-
|
17 |
-
#
|
18 |
-
REPO
|
19 |
-
CHUNK_TOKENS
|
20 |
-
START_TOKEN
|
21 |
-
NEW_BLOCK
|
22 |
-
EOS_TOKEN
|
23 |
-
AUDIO_BASE
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
|
|
28 |
def __init__(self, audio_ids: torch.Tensor, min_blocks:int=1):
|
29 |
super().__init__()
|
30 |
-
self.audio_ids
|
31 |
-
self.ctrl_ids
|
32 |
-
self.
|
33 |
-
self.
|
34 |
-
def __call__(self,
|
35 |
allow = torch.cat([self.audio_ids, self.ctrl_ids])
|
36 |
-
if self.blocks >= self.
|
37 |
allow = torch.cat([allow,
|
38 |
torch.tensor([EOS_TOKEN], device=scores.device)])
|
39 |
mask = torch.full_like(scores, float("-inf"))
|
40 |
mask[:, allow] = 0
|
41 |
return scores + mask
|
42 |
|
43 |
-
#
|
44 |
app = FastAPI()
|
45 |
|
46 |
@app.get("/")
|
47 |
async def root():
|
48 |
-
return {"msg": "Orpheus‑TTS
|
49 |
|
50 |
@app.on_event("startup")
|
51 |
async def load():
|
@@ -57,13 +58,14 @@ async def load():
|
|
57 |
REPO,
|
58 |
low_cpu_mem_usage=True,
|
59 |
device_map={"":0} if device=="cuda" else None,
|
60 |
-
torch_dtype=torch.bfloat16 if device=="cuda" else None
|
|
|
61 |
model.config.pad_token_id = model.config.eos_token_id
|
62 |
model.config.use_cache = True
|
63 |
-
masker =
|
64 |
print("✅ Modelle geladen")
|
65 |
|
66 |
-
#
|
67 |
def build_inputs(text:str, voice:str):
|
68 |
prompt = f"{voice}: {text}"
|
69 |
ids = tok(prompt, return_tensors="pt").input_ids.to(device)
|
@@ -72,66 +74,61 @@ def build_inputs(text:str, voice:str):
|
|
72 |
torch.tensor([[128009,128260]], device=device)],1)
|
73 |
return ids, torch.ones_like(ids)
|
74 |
|
75 |
-
def decode_block(
|
76 |
l1,l2,l3=[],[],[]
|
77 |
-
l1.append(
|
78 |
-
l2.append(
|
79 |
-
l3
|
80 |
-
l2.append(
|
81 |
-
l3
|
82 |
codes=[torch.tensor(x,device=device).unsqueeze(0) for x in (l1,l2,l3)]
|
83 |
audio=snac.decode(codes).squeeze().cpu().numpy()
|
84 |
return (audio*32767).astype("int16").tobytes()
|
85 |
|
86 |
-
#
|
87 |
@app.websocket("/ws/tts")
|
88 |
-
async def tts(ws:WebSocket):
|
89 |
await ws.accept()
|
90 |
try:
|
91 |
-
req
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
ids, attn = build_inputs(text, voice)
|
96 |
-
total_len = ids.shape[1] # Länge des Prompts
|
97 |
-
past = None
|
98 |
-
last_tok = None
|
99 |
-
buf = []
|
100 |
|
101 |
while True:
|
|
|
102 |
out = model.generate(
|
103 |
-
input_ids
|
104 |
-
attention_mask
|
105 |
-
past_key_values
|
106 |
-
max_new_tokens
|
107 |
-
logits_processor=
|
108 |
do_sample=True, temperature=0.7, top_p=0.95,
|
109 |
use_cache=True, return_dict_in_generate=True,
|
110 |
return_legacy_cache=True)
|
111 |
-
|
112 |
pkv = out.past_key_values
|
|
|
113 |
if isinstance(pkv, Cache): pkv = pkv.to_legacy()
|
114 |
past = pkv
|
|
|
115 |
|
116 |
-
seq
|
117 |
-
new
|
118 |
-
|
119 |
-
print("new tokens:", new[:32])
|
120 |
-
|
121 |
-
if not new: # nichts generiert
|
122 |
-
raise StopIteration
|
123 |
|
|
|
124 |
for t in new:
|
125 |
last_tok = t
|
126 |
if t == EOS_TOKEN: raise StopIteration
|
127 |
-
if t == NEW_BLOCK:
|
128 |
-
|
129 |
-
buf
|
130 |
-
if len(buf)==7:
|
131 |
await ws.send_bytes(decode_block(buf))
|
132 |
buf.clear()
|
133 |
masker.blocks += 1
|
134 |
-
|
|
|
135 |
|
136 |
except (StopIteration, WebSocketDisconnect):
|
137 |
pass
|
@@ -144,7 +141,7 @@ async def tts(ws:WebSocket):
|
|
144 |
try: await ws.close()
|
145 |
except RuntimeError: pass
|
146 |
|
147 |
-
#
|
148 |
if __name__ == "__main__":
|
149 |
import uvicorn
|
150 |
uvicorn.run("app:app", host="0.0.0.0", port=7860)
|
|
|
6 |
from transformers.generation.utils import Cache
|
7 |
from snac import SNAC
|
8 |
|
9 |
+
# 0 · Auth & Device ---------------------------------------------------
|
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 |
+
torch.backends.cuda.enable_flash_sdp(False) # SDP‑Assert fix
|
16 |
+
|
17 |
+
# 1 · Konstanten ------------------------------------------------------
|
18 |
+
REPO = "SebastianBodza/Kartoffel_Orpheus-3B_german_natural-v0.1"
|
19 |
+
CHUNK_TOKENS = 50
|
20 |
+
START_TOKEN = 128259
|
21 |
+
NEW_BLOCK = 128257
|
22 |
+
EOS_TOKEN = 128258
|
23 |
+
AUDIO_BASE = 128266
|
24 |
+
AUDIO_SPAN = 4096 * 7 # 28 672 Codes
|
25 |
+
VALID_AUDIO = torch.arange(AUDIO_BASE, AUDIO_BASE + AUDIO_SPAN)
|
26 |
+
|
27 |
+
# 2 · Logit‑Masker ----------------------------------------------------
|
28 |
+
class DynamicMask(LogitsProcessor):
|
29 |
def __init__(self, audio_ids: torch.Tensor, min_blocks:int=1):
|
30 |
super().__init__()
|
31 |
+
self.audio_ids = audio_ids
|
32 |
+
self.ctrl_ids = torch.tensor([NEW_BLOCK], device=audio_ids.device)
|
33 |
+
self.blocks = 0
|
34 |
+
self.min_blk = min_blocks
|
35 |
+
def __call__(self, inp_ids, scores):
|
36 |
allow = torch.cat([self.audio_ids, self.ctrl_ids])
|
37 |
+
if self.blocks >= self.min_blk:
|
38 |
allow = torch.cat([allow,
|
39 |
torch.tensor([EOS_TOKEN], device=scores.device)])
|
40 |
mask = torch.full_like(scores, float("-inf"))
|
41 |
mask[:, allow] = 0
|
42 |
return scores + mask
|
43 |
|
44 |
+
# 3 · FastAPI‑App -----------------------------------------------------
|
45 |
app = FastAPI()
|
46 |
|
47 |
@app.get("/")
|
48 |
async def root():
|
49 |
+
return {"msg": "Orpheus‑TTS online"}
|
50 |
|
51 |
@app.on_event("startup")
|
52 |
async def load():
|
|
|
58 |
REPO,
|
59 |
low_cpu_mem_usage=True,
|
60 |
device_map={"":0} if device=="cuda" else None,
|
61 |
+
torch_dtype=torch.bfloat16 if device=="cuda" else None,
|
62 |
+
)
|
63 |
model.config.pad_token_id = model.config.eos_token_id
|
64 |
model.config.use_cache = True
|
65 |
+
masker = DynamicMask(VALID_AUDIO.to(device))
|
66 |
print("✅ Modelle geladen")
|
67 |
|
68 |
+
# 4 · Hilfsfunktionen -------------------------------------------------
|
69 |
def build_inputs(text:str, voice:str):
|
70 |
prompt = f"{voice}: {text}"
|
71 |
ids = tok(prompt, return_tensors="pt").input_ids.to(device)
|
|
|
74 |
torch.tensor([[128009,128260]], device=device)],1)
|
75 |
return ids, torch.ones_like(ids)
|
76 |
|
77 |
+
def decode_block(b):
|
78 |
l1,l2,l3=[],[],[]
|
79 |
+
l1.append(b[0])
|
80 |
+
l2.append(b[1]-4096)
|
81 |
+
l3 += [b[2]-8192, b[3]-12288]
|
82 |
+
l2.append(b[4]-16384)
|
83 |
+
l3 += [b[5]-20480, b[6]-24576]
|
84 |
codes=[torch.tensor(x,device=device).unsqueeze(0) for x in (l1,l2,l3)]
|
85 |
audio=snac.decode(codes).squeeze().cpu().numpy()
|
86 |
return (audio*32767).astype("int16").tobytes()
|
87 |
|
88 |
+
# 5 · WebSocket‑Endpoint ---------------------------------------------
|
89 |
@app.websocket("/ws/tts")
|
90 |
+
async def tts(ws: WebSocket):
|
91 |
await ws.accept()
|
92 |
try:
|
93 |
+
req = json.loads(await ws.receive_text())
|
94 |
+
ids, attn = build_inputs(req.get("text",""), req.get("voice","Jakob"))
|
95 |
+
past, last_tok, buf = None, None, []
|
96 |
+
prompt_len = ids.shape[1]
|
|
|
|
|
|
|
|
|
|
|
97 |
|
98 |
while True:
|
99 |
+
print(f"DEBUG: Before generate - past is None: {past is None}") # Added logging
|
100 |
out = model.generate(
|
101 |
+
input_ids = ids if past is None else torch.tensor([[last_tok]], device=device),
|
102 |
+
attention_mask = attn if past is None else None,
|
103 |
+
past_key_values= past,
|
104 |
+
max_new_tokens = CHUNK_TOKENS,
|
105 |
+
logits_processor=[masker],
|
106 |
do_sample=True, temperature=0.7, top_p=0.95,
|
107 |
use_cache=True, return_dict_in_generate=True,
|
108 |
return_legacy_cache=True)
|
109 |
+
print(f"DEBUG: After generate - type of out.past_key_values: {type(out.past_key_values)}") # Added logging
|
110 |
pkv = out.past_key_values
|
111 |
+
print(f"DEBUG: After getting pkv - type of pkv: {type(pkv)}") # Added logging
|
112 |
if isinstance(pkv, Cache): pkv = pkv.to_legacy()
|
113 |
past = pkv
|
114 |
+
print(f"DEBUG: After cache handling - past is None: {past is None}") # Added logging
|
115 |
|
116 |
+
seq = out.sequences[0].tolist()
|
117 |
+
new = seq[prompt_len:]; prompt_len = len(seq)
|
118 |
+
print("new tokens:", new[:25])
|
|
|
|
|
|
|
|
|
119 |
|
120 |
+
if not new: raise StopIteration
|
121 |
for t in new:
|
122 |
last_tok = t
|
123 |
if t == EOS_TOKEN: raise StopIteration
|
124 |
+
if t == NEW_BLOCK: buf.clear(); continue
|
125 |
+
buf.append(t - AUDIO_BASE)
|
126 |
+
if len(buf) == 7:
|
|
|
127 |
await ws.send_bytes(decode_block(buf))
|
128 |
buf.clear()
|
129 |
masker.blocks += 1
|
130 |
+
|
131 |
+
ids, attn = None, None # ab jetzt 1‑Token‑Step
|
132 |
|
133 |
except (StopIteration, WebSocketDisconnect):
|
134 |
pass
|
|
|
141 |
try: await ws.close()
|
142 |
except RuntimeError: pass
|
143 |
|
144 |
+
# 6 · Local run -------------------------------------------------------
|
145 |
if __name__ == "__main__":
|
146 |
import uvicorn
|
147 |
uvicorn.run("app:app", host="0.0.0.0", port=7860)
|