Spaces:
Running on Zero
Running on Zero
Commit ·
2aceb54
1
Parent(s): a4a16d2
Buffer early WebRTC ICE candidates
Browse files- streaming_app.py +57 -1
streaming_app.py
CHANGED
|
@@ -1,5 +1,6 @@
|
|
| 1 |
from __future__ import annotations
|
| 2 |
|
|
|
|
| 3 |
import copy
|
| 4 |
import json
|
| 5 |
import multiprocessing
|
|
@@ -8,6 +9,7 @@ import tarfile
|
|
| 8 |
import threading
|
| 9 |
import time
|
| 10 |
import uuid
|
|
|
|
| 11 |
from pathlib import Path
|
| 12 |
|
| 13 |
import gradio as gr
|
|
@@ -16,6 +18,7 @@ import sentencepiece
|
|
| 16 |
import spaces
|
| 17 |
import torch
|
| 18 |
from fastrtc import StreamHandler, WebRTC, get_cloudflare_turn_credentials_async
|
|
|
|
| 19 |
from huggingface_hub import hf_hub_download
|
| 20 |
|
| 21 |
|
|
@@ -224,6 +227,50 @@ async def get_turn_configuration():
|
|
| 224 |
return configuration
|
| 225 |
|
| 226 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 227 |
def new_bridge(oauth_token: gr.OAuthToken | None):
|
| 228 |
remember_turn_token(oauth_token)
|
| 229 |
input_queue = _manager.Queue(INPUT_QUEUE_FRAMES)
|
|
@@ -284,6 +331,8 @@ class PersonaPlexQueueHandler(StreamHandler):
|
|
| 284 |
input_sample_rate=OUTPUT_SAMPLE_RATE,
|
| 285 |
fps=WEBRTC_FRAME_RATE,
|
| 286 |
)
|
|
|
|
|
|
|
| 287 |
|
| 288 |
def _bridge(self):
|
| 289 |
# FastRTC prepends the WebRTC component value to additional inputs.
|
|
@@ -301,6 +350,9 @@ class PersonaPlexQueueHandler(StreamHandler):
|
|
| 301 |
return
|
| 302 |
sample_rate, array = frame
|
| 303 |
audio = np.asarray(array, dtype=np.int16).reshape(-1)
|
|
|
|
|
|
|
|
|
|
| 304 |
if bounded_put(input_queue, (int(sample_rate), audio)):
|
| 305 |
counters["input_drops"] = int(counters.get("input_drops", 0)) + 1
|
| 306 |
|
|
@@ -315,6 +367,9 @@ class PersonaPlexQueueHandler(StreamHandler):
|
|
| 315 |
audio = output_queue.get_nowait()
|
| 316 |
except queue.Empty:
|
| 317 |
return None
|
|
|
|
|
|
|
|
|
|
| 318 |
return OUTPUT_SAMPLE_RATE, np.asarray(audio, dtype=np.int16).reshape(1, -1)
|
| 319 |
|
| 320 |
def copy(self):
|
|
@@ -407,6 +462,7 @@ def run_streaming_session(
|
|
| 407 |
drain_proxy_queue(input_queue)
|
| 408 |
ready_seconds = round(time.perf_counter() - session_started, 3)
|
| 409 |
ready_event.set()
|
|
|
|
| 410 |
live_started = time.perf_counter()
|
| 411 |
status = (
|
| 412 |
f"**Engine ready in {ready_seconds:.1f}s.** Start the WebRTC microphone and speak naturally. "
|
|
@@ -561,7 +617,7 @@ with gr.Blocks(title="RoleForge Streaming Spike", theme=gr.themes.Soft(), css=CS
|
|
| 561 |
with gr.Column(scale=2):
|
| 562 |
gr.Markdown("### 🎙️ Live Stage")
|
| 563 |
status = gr.Markdown("Sign in, choose the cue, then start the live engine.")
|
| 564 |
-
webrtc =
|
| 565 |
label="2. Connect microphone after Engine ready",
|
| 566 |
modality="audio",
|
| 567 |
mode="send-receive",
|
|
|
|
| 1 |
from __future__ import annotations
|
| 2 |
|
| 3 |
+
import asyncio
|
| 4 |
import copy
|
| 5 |
import json
|
| 6 |
import multiprocessing
|
|
|
|
| 9 |
import threading
|
| 10 |
import time
|
| 11 |
import uuid
|
| 12 |
+
from collections import defaultdict
|
| 13 |
from pathlib import Path
|
| 14 |
|
| 15 |
import gradio as gr
|
|
|
|
| 18 |
import spaces
|
| 19 |
import torch
|
| 20 |
from fastrtc import StreamHandler, WebRTC, get_cloudflare_turn_credentials_async
|
| 21 |
+
from gradio.components.base import server
|
| 22 |
from huggingface_hub import hf_hub_download
|
| 23 |
|
| 24 |
|
|
|
|
| 227 |
return configuration
|
| 228 |
|
| 229 |
|
| 230 |
+
class ReliableWebRTC(WebRTC):
|
| 231 |
+
"""Serialize FastRTC signaling and replay ICE candidates that beat the SDP offer."""
|
| 232 |
+
|
| 233 |
+
def __init__(self, *args, **kwargs):
|
| 234 |
+
super().__init__(*args, **kwargs)
|
| 235 |
+
self._signaling_lock = None
|
| 236 |
+
self._pending_ice = defaultdict(list)
|
| 237 |
+
|
| 238 |
+
@server
|
| 239 |
+
async def offer(self, body):
|
| 240 |
+
if self._signaling_lock is None:
|
| 241 |
+
self._signaling_lock = asyncio.Lock()
|
| 242 |
+
|
| 243 |
+
async with self._signaling_lock:
|
| 244 |
+
webrtc_id = body.get("webrtc_id")
|
| 245 |
+
is_candidate = body.get("type") == "ice-candidate" and "candidate" in body
|
| 246 |
+
|
| 247 |
+
if is_candidate and webrtc_id not in self.pcs:
|
| 248 |
+
pending = self._pending_ice[webrtc_id]
|
| 249 |
+
if len(pending) < 32:
|
| 250 |
+
pending.append(body)
|
| 251 |
+
print(f"Buffered early ICE candidate for pending connection: {webrtc_id}")
|
| 252 |
+
return {"status": "success"}
|
| 253 |
+
|
| 254 |
+
response = await self.handle_offer(
|
| 255 |
+
body,
|
| 256 |
+
self.set_additional_outputs(webrtc_id),
|
| 257 |
+
)
|
| 258 |
+
|
| 259 |
+
if not is_candidate and webrtc_id in self.pcs:
|
| 260 |
+
pending = self._pending_ice.pop(webrtc_id, [])
|
| 261 |
+
for candidate in pending:
|
| 262 |
+
await self.handle_offer(
|
| 263 |
+
candidate,
|
| 264 |
+
self.set_additional_outputs(webrtc_id),
|
| 265 |
+
)
|
| 266 |
+
print(
|
| 267 |
+
f"Registered WebRTC offer and replayed {len(pending)} early ICE candidates: "
|
| 268 |
+
f"{webrtc_id}"
|
| 269 |
+
)
|
| 270 |
+
|
| 271 |
+
return response
|
| 272 |
+
|
| 273 |
+
|
| 274 |
def new_bridge(oauth_token: gr.OAuthToken | None):
|
| 275 |
remember_turn_token(oauth_token)
|
| 276 |
input_queue = _manager.Queue(INPUT_QUEUE_FRAMES)
|
|
|
|
| 331 |
input_sample_rate=OUTPUT_SAMPLE_RATE,
|
| 332 |
fps=WEBRTC_FRAME_RATE,
|
| 333 |
)
|
| 334 |
+
self._logged_input = False
|
| 335 |
+
self._logged_output = False
|
| 336 |
|
| 337 |
def _bridge(self):
|
| 338 |
# FastRTC prepends the WebRTC component value to additional inputs.
|
|
|
|
| 350 |
return
|
| 351 |
sample_rate, array = frame
|
| 352 |
audio = np.asarray(array, dtype=np.int16).reshape(-1)
|
| 353 |
+
if not self._logged_input:
|
| 354 |
+
print("WebRTC audio bridge received its first input frame.")
|
| 355 |
+
self._logged_input = True
|
| 356 |
if bounded_put(input_queue, (int(sample_rate), audio)):
|
| 357 |
counters["input_drops"] = int(counters.get("input_drops", 0)) + 1
|
| 358 |
|
|
|
|
| 367 |
audio = output_queue.get_nowait()
|
| 368 |
except queue.Empty:
|
| 369 |
return None
|
| 370 |
+
if not self._logged_output:
|
| 371 |
+
print("WebRTC audio bridge emitted its first output frame.")
|
| 372 |
+
self._logged_output = True
|
| 373 |
return OUTPUT_SAMPLE_RATE, np.asarray(audio, dtype=np.int16).reshape(1, -1)
|
| 374 |
|
| 375 |
def copy(self):
|
|
|
|
| 462 |
drain_proxy_queue(input_queue)
|
| 463 |
ready_seconds = round(time.perf_counter() - session_started, 3)
|
| 464 |
ready_event.set()
|
| 465 |
+
print("PersonaPlex GPU session is ready and waiting for WebRTC audio.")
|
| 466 |
live_started = time.perf_counter()
|
| 467 |
status = (
|
| 468 |
f"**Engine ready in {ready_seconds:.1f}s.** Start the WebRTC microphone and speak naturally. "
|
|
|
|
| 617 |
with gr.Column(scale=2):
|
| 618 |
gr.Markdown("### 🎙️ Live Stage")
|
| 619 |
status = gr.Markdown("Sign in, choose the cue, then start the live engine.")
|
| 620 |
+
webrtc = ReliableWebRTC(
|
| 621 |
label="2. Connect microphone after Engine ready",
|
| 622 |
modality="audio",
|
| 623 |
mode="send-receive",
|