Spaces:
Paused
Paused
Update app.py
Browse files
app.py
CHANGED
|
@@ -107,37 +107,44 @@ def decode_block(block7: list[int]) -> bytes:
|
|
| 107 |
async def tts(ws: WebSocket):
|
| 108 |
await ws.accept()
|
| 109 |
try:
|
| 110 |
-
req
|
| 111 |
text = req.get("text", "")
|
| 112 |
voice = req.get("voice", "Jakob")
|
| 113 |
|
| 114 |
-
ids, attn = build_inputs(text, voice)
|
| 115 |
past = None
|
|
|
|
| 116 |
buf = []
|
| 117 |
|
| 118 |
while True:
|
| 119 |
out = model.generate(
|
| 120 |
-
input_ids = ids if past is None else
|
| 121 |
attention_mask = attn if past is None else None,
|
| 122 |
past_key_values = past,
|
| 123 |
max_new_tokens = CHUNK_TOKENS,
|
| 124 |
-
logits_processor= [masker],
|
| 125 |
do_sample=True, temperature=0.7, top_p=0.95,
|
| 126 |
return_dict_in_generate=True,
|
| 127 |
use_cache=True,
|
|
|
|
| 128 |
)
|
| 129 |
|
| 130 |
-
# Cache & neue
|
| 131 |
pkv = out.past_key_values
|
| 132 |
-
if isinstance(pkv, Cache):
|
| 133 |
pkv = pkv.to_legacy()
|
| 134 |
past = pkv
|
| 135 |
|
| 136 |
new = out.sequences[0, -out.num_generated_tokens :].tolist()
|
| 137 |
-
print("new tokens:", new[:20])
|
| 138 |
|
| 139 |
-
#
|
|
|
|
|
|
|
|
|
|
| 140 |
for t in new:
|
|
|
|
|
|
|
| 141 |
if t == EOS_TOKEN:
|
| 142 |
raise StopIteration
|
| 143 |
|
|
@@ -149,9 +156,9 @@ async def tts(ws: WebSocket):
|
|
| 149 |
if len(buf) == 7:
|
| 150 |
await ws.send_bytes(decode_block(buf))
|
| 151 |
buf.clear()
|
| 152 |
-
masker.blocks_done += 1
|
| 153 |
|
| 154 |
-
#
|
| 155 |
ids, attn = None, None
|
| 156 |
|
| 157 |
except (StopIteration, WebSocketDisconnect):
|
|
@@ -165,7 +172,7 @@ async def tts(ws: WebSocket):
|
|
| 165 |
try:
|
| 166 |
await ws.close()
|
| 167 |
except RuntimeError:
|
| 168 |
-
pass
|
| 169 |
|
| 170 |
# ── 6. Lokaler Start (uvicorn) ───────────────────────────────────────
|
| 171 |
if __name__ == "__main__":
|
|
|
|
| 107 |
async def tts(ws: WebSocket):
|
| 108 |
await ws.accept()
|
| 109 |
try:
|
| 110 |
+
req = json.loads(await ws.receive_text())
|
| 111 |
text = req.get("text", "")
|
| 112 |
voice = req.get("voice", "Jakob")
|
| 113 |
|
| 114 |
+
ids, attn = build_inputs(text, voice) # vollständiger Prompt
|
| 115 |
past = None
|
| 116 |
+
last_tok = None # <- NEU
|
| 117 |
buf = []
|
| 118 |
|
| 119 |
while True:
|
| 120 |
out = model.generate(
|
| 121 |
+
input_ids = ids if past is None else torch.tensor([[last_tok]], device=device),
|
| 122 |
attention_mask = attn if past is None else None,
|
| 123 |
past_key_values = past,
|
| 124 |
max_new_tokens = CHUNK_TOKENS,
|
| 125 |
+
logits_processor= [masker],
|
| 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, # <- Warnung unterdrücken
|
| 130 |
)
|
| 131 |
|
| 132 |
+
# ----- Cache & neue Token --------------------------------------
|
| 133 |
pkv = out.past_key_values
|
| 134 |
+
if isinstance(pkv, Cache): # HF >= 4.47
|
| 135 |
pkv = pkv.to_legacy()
|
| 136 |
past = pkv
|
| 137 |
|
| 138 |
new = out.sequences[0, -out.num_generated_tokens :].tolist()
|
| 139 |
+
print("new tokens:", new[:20]) # Debug‑Print
|
| 140 |
|
| 141 |
+
if not new: # Safety – nichts erzeugt
|
| 142 |
+
raise StopIteration
|
| 143 |
+
|
| 144 |
+
# ----- Token‑Handling ------------------------------------------
|
| 145 |
for t in new:
|
| 146 |
+
last_tok = t # speichern für nächste Runde
|
| 147 |
+
|
| 148 |
if t == EOS_TOKEN:
|
| 149 |
raise StopIteration
|
| 150 |
|
|
|
|
| 156 |
if len(buf) == 7:
|
| 157 |
await ws.send_bytes(decode_block(buf))
|
| 158 |
buf.clear()
|
| 159 |
+
masker.blocks_done += 1 # nach 1. Block darf EOS
|
| 160 |
|
| 161 |
+
# ab nächster Runde nur 1 Token + Cache
|
| 162 |
ids, attn = None, None
|
| 163 |
|
| 164 |
except (StopIteration, WebSocketDisconnect):
|
|
|
|
| 172 |
try:
|
| 173 |
await ws.close()
|
| 174 |
except RuntimeError:
|
| 175 |
+
pass
|
| 176 |
|
| 177 |
# ── 6. Lokaler Start (uvicorn) ───────────────────────────────────────
|
| 178 |
if __name__ == "__main__":
|