File size: 53,166 Bytes
d1afbc8 2bd198e d1afbc8 87147f5 d1afbc8 1a7ea3c af5a7b2 1a7ea3c 2bd198e 6896250 d54b5ce 2bd198e 9a1b4dc 1a7ea3c d1afbc8 b1bc032 d1afbc8 7c14c0e d1afbc8 87147f5 b1bc032 87147f5 d1afbc8 af5a7b2 d1afbc8 af5a7b2 8cedcd0 af5a7b2 8cedcd0 af5a7b2 8cedcd0 d1afbc8 1a7ea3c e7bec6e 1a7ea3c f04acc3 1a7ea3c 2bd198e 6896250 2bd198e d54b5ce 2bd198e d54b5ce 2bd198e 6896250 2bd198e 6896250 2bd198e 6896250 2bd198e 6896250 2bd198e 6896250 2bd198e d54b5ce 2bd198e 6896250 2bd198e 6896250 2bd198e 6896250 d54b5ce 2bd198e d54b5ce 2bd198e d54b5ce 6896250 d54b5ce 6896250 d54b5ce 6896250 d54b5ce 2bd198e 6896250 2bd198e d54b5ce 2bd198e d54b5ce 2bd198e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 |
from magenta_rt import system, audio as au
import numpy as np
from fastapi import FastAPI, UploadFile, File, Form, Body, HTTPException, Response, Request, WebSocket, WebSocketDisconnect
import tempfile, io, base64, math, threading
from fastapi.middleware.cors import CORSMiddleware
from contextlib import contextmanager
import soundfile as sf
from math import gcd
from scipy.signal import resample_poly
from utils import (
match_loudness_to_reference, stitch_generated, hard_trim_seconds,
apply_micro_fades, make_bar_aligned_context, take_bar_aligned_tail,
resample_and_snap, wav_bytes_base64
)
from jam_worker import JamWorker, JamParams, JamChunk
import uuid, threading
import os
import logging
import gradio as gr
from typing import Optional
import json, asyncio, base64
import time
from starlette.websockets import WebSocketState
try:
from uvicorn.protocols.utils import ClientDisconnected # uvicorn >= 0.20
except Exception:
class ClientDisconnected(Exception): # fallback
pass
async def send_json_safe(ws: WebSocket, obj) -> bool:
"""Try to send. Returns False if the socket is (or becomes) closed."""
if ws.client_state == WebSocketState.DISCONNECTED or ws.application_state == WebSocketState.DISCONNECTED:
return False
try:
await ws.send_text(json.dumps(obj))
return True
except (WebSocketDisconnect, ClientDisconnected, RuntimeError):
return False
except Exception:
return False
# --- Patch T5X mesh helpers for GPUs on JAX >= 0.7 (coords present, no core_on_chip) ---
def _patch_t5x_for_gpu_coords():
try:
import jax
from t5x import partitioning as _t5x_part
old_bounds = getattr(_t5x_part, "bounds_from_last_device", None)
old_getcoords = getattr(_t5x_part, "get_coords", None)
def _bounds_from_last_device_gpu_safe(last_device):
# TPU: coords + core_on_chip
core = getattr(last_device, "core_on_chip", None)
coords = getattr(last_device, "coords", None)
if coords is not None and core is not None:
x, y, z = coords
return x + 1, y + 1, z + 1, core + 1
# Non-TPU (or GPU lacking core_on_chip): hosts x local_devices
return jax.host_count(), jax.local_device_count()
def _get_coords_gpu_safe(device):
core = getattr(device, "core_on_chip", None)
coords = getattr(device, "coords", None)
if coords is not None and core is not None:
return (*coords, core)
# Fallback that works on CPU/GPU
return (device.process_index, device.id % jax.local_device_count())
_t5x_part.bounds_from_last_device = _bounds_from_last_device_gpu_safe
_t5x_part.get_coords = _get_coords_gpu_safe
import logging; logging.info("Patched t5x.partitioning for GPU coords without core_on_chip.")
except Exception as e:
import logging; logging.exception("t5x GPU-coords patch failed: %s", e)
# Call the patch immediately at import time (before MagentaRT init)
_patch_t5x_for_gpu_coords()
def create_documentation_interface():
"""Create a Gradio interface for documentation and transparency"""
with gr.Blocks(title="MagentaRT Research API", theme=gr.themes.Soft()) as interface:
gr.Markdown("""
# 🎵 MagentaRT Live Music Generation Research API
**Research-only implementation for iOS app development**
This API uses Google's [MagentaRT](https://github.com/magenta/magenta-realtime) to generate
continuous music based on input audio loops for experimental iOS app development.
""")
with gr.Tabs():
with gr.Tab("📖 About This Research"):
gr.Markdown("""
## What This API Does
We're exploring AI-assisted loop-based music creation for mobile apps. Websockets are notoriously annoying in ios-swift apps, so I tried to come up with an http version tailored to the loop based nature of an existing swift app. This API provides:
### 🎹 Single Generation (`/generate`)
- Upload audio loop + BPM + style parameters
- Returns 4-8 bars of AI-generated continuation
- **Performance**: 4 bars in ~9s, 8 bars in ~16s (L40S GPU)
### 🔄 Continuous Jamming (`/jam/*`)
- `/jam/start` - Begin continuous generation session
- `/jam/next` - Get next bar-aligned chunk
- `/jam/stop` - End session
- **Performance**: Real-time 8-bar chunks after warmup
## Technical Specs
- **Model**: MagentaRT (800M parameter transformer)
- **Quality**: 48kHz stereo output
- **Context**: 10-second audio analysis window
- **Styles**: Text descriptions (e.g., "acid house, techno")
## Research Goals
- Seamless AI music generation for loop-based composition
- Real-time parameter adjustment during generation
- Mobile-optimized music creation workflows
""")
with gr.Tab("🔧 API Documentation"):
gr.Markdown("""
## Single Generation Example
```bash
curl -X POST "/generate" \\
-F "loop_audio=@drum_loop.wav" \\
-F "bpm=120" \\
-F "bars=8" \\
-F "styles=acid house,techno" \\
-F "guidance_weight=5.0" \\
-F "temperature=1.1"
```
## Continuous Jamming Example
```bash
# 1. Start session
SESSION=$(curl -X POST "/jam/start" \\
-F "loop_audio=@loop.wav" \\
-F "bpm=120" \\
-F "bars_per_chunk=8" | jq -r .session_id)
# 2. Get chunks in real-time
curl "/jam/next?session_id=$SESSION"
# 3. Stop when done
curl -X POST "/jam/stop" \\
-H "Content-Type: application/json" \\
-d "{\\"session_id\\": \\"$SESSION\\"}"
```
## Key Parameters
- **bpm**: 60-200 (beats per minute)
- **bars**: 1-16 (bars to generate)
- **styles**: Text descriptions, comma-separated
- **guidance_weight**: 0.1-10.0 (style adherence)
- **temperature**: 0.1-2.0 (randomness)
- **intro_bars_to_drop**: Skip N bars from start
## Response Format
```json
{
"audio_base64": "...",
"metadata": {
"bpm": 120,
"bars": 8,
"sample_rate": 48000,
"loop_duration_seconds": 16.0
}
}
```
""")
with gr.Tab("📱 iOS App Integration"):
gr.Markdown("""
## How Our iOS App Uses This API
### User Flow
1. **Record/Import**: User provides drum or instrument loop
2. **Parameter Setup**: Set BPM, style, generation settings
3. **Continuous Generation**: App calls `/jam/start`
4. **Real-time Playback**: App fetches chunks via `/jam/next`
5. **Seamless Mixing**: Generated audio mixed into live stream
### Technical Implementation
- **Audio Format**: 48kHz WAV for consistency
- **Chunk Size**: 8 bars (~16 seconds at 120 BPM)
- **Buffer Management**: 3-5 chunks ahead for smooth playback
- **Style Updates**: Real-time parameter adjustment via `/jam/update`
### Networking Considerations
- **Latency**: ~2-3 seconds per chunk after warmup
- **Bandwidth**: ~500KB per 8-bar chunk (compressed)
- **Reliability**: Automatic retry with exponential backoff
- **Caching**: Local buffer for offline resilience
""")
with gr.Tab("⚖️ Licensing & Legal"):
gr.Markdown("""
## MagentaRT Licensing
This project uses Google's MagentaRT model under:
- **Source Code**: Apache License 2.0
- **Model Weights**: Creative Commons Attribution 4.0 International
- **Usage Terms**: [See MagentaRT repository](https://github.com/magenta/magenta-realtime)
### Key Requirements
- ✅ **Attribution**: Credit MagentaRT in derivative works
- ✅ **Responsible Use**: Don't infringe copyrights
- ✅ **No Warranties**: Use at your own risk
- ✅ **Patent License**: Explicit patent grants included
## Our Implementation
- **Purpose**: Research and development only
- **Non-Commercial**: Experimental iOS app development
- **Open Source**: Will release implementation under Apache 2.0
- **Attribution**: Proper credit to Google Research team
### Required Attribution
```
Generated using MagentaRT
Copyright 2024 Google LLC
Licensed under Apache 2.0 and CC-BY 4.0
Implementation for research purposes
```
""")
with gr.Tab("📊 Performance & Limits"):
gr.Markdown("""
## Current Performance (L40S 48GB)
### ⚡ Single Generation
- **4 bars @ 100 BPM**: ~9 seconds
- **8 bars @ 100 BPM**: ~16 seconds
- **Memory usage**: ~40GB VRAM during generation
### 🔄 Continuous Jamming
- **Warmup**: ~10-15 seconds first chunk
- **8-bar chunks @ 120 BPM**: Real-time delivery
- **Buffer ahead**: 3-5 chunks for smooth playback
## Known Limitations
### 🎵 Model Limitations (MagentaRT)
- **Context**: 10-second maximum memory
- **Training**: Primarily Western instrumental music
- **Vocals**: Non-lexical only, no lyric conditioning
- **Structure**: No long-form song arrangement
- **Inside Swift**: After a few turns of continuous chunks, the swift app works best if you restart the jam from the combined audio again. In this way you might end up with a real jam.
### 🖥️ Infrastructure Limitations
- **Concurrency**: Single user jam sessions only
- **GPU Memory**: 40GB+ VRAM required for stable operation
- **Latency**: 2+ second minimum for style changes
- **Uptime**: Research setup, no SLA guarantees
## Resource Requirements
- **Minimum**: 24GB VRAM (basic operation, won't operate realtime enough for new chunks coming in)
- **Recommended**: 48GB VRAM (stable performance)
- **CPU**: 8+ cores
- **System RAM**: 32GB+
- **Storage**: 50GB+ for model weights
""")
gr.Markdown("""
---
**🔬 Research Project** | **📱 iOS Development** | **🎵 Powered by MagentaRT**
This API is part of ongoing research into AI-assisted music creation for mobile devices.
For technical details, see the API documentation tabs above.
""")
return interface
jam_registry: dict[str, JamWorker] = {}
jam_lock = threading.Lock()
@contextmanager
def mrt_overrides(mrt, **kwargs):
"""Temporarily set attributes on MRT if they exist; restore after."""
old = {}
try:
for k, v in kwargs.items():
if hasattr(mrt, k):
old[k] = getattr(mrt, k)
setattr(mrt, k, v)
yield
finally:
for k, v in old.items():
setattr(mrt, k, v)
# loudness utils
try:
import pyloudnorm as pyln
_HAS_LOUDNORM = True
except Exception:
_HAS_LOUDNORM = False
# ----------------------------
# Main generation (single combined style vector)
# ----------------------------
def generate_loop_continuation_with_mrt(
mrt,
input_wav_path: str,
bpm: float,
extra_styles=None,
style_weights=None,
bars: int = 8,
beats_per_bar: int = 4,
loop_weight: float = 1.0,
loudness_mode: str = "auto",
loudness_headroom_db: float = 1.0,
intro_bars_to_drop: int = 0, # <— NEW
):
# Load & prep (unchanged)
loop = au.Waveform.from_file(input_wav_path).resample(mrt.sample_rate).as_stereo()
# Use tail for context (your recent change)
codec_fps = float(mrt.codec.frame_rate)
ctx_seconds = float(mrt.config.context_length_frames) / codec_fps
loop_for_context = take_bar_aligned_tail(loop, bpm, beats_per_bar, ctx_seconds)
tokens_full = mrt.codec.encode(loop_for_context).astype(np.int32)
tokens = tokens_full[:, :mrt.config.decoder_codec_rvq_depth]
# Bar-aligned token window (unchanged)
context_tokens = make_bar_aligned_context(
tokens, bpm=bpm, fps=float(mrt.codec.frame_rate),
ctx_frames=mrt.config.context_length_frames, beats_per_bar=beats_per_bar
)
state = mrt.init_state()
state.context_tokens = context_tokens
# STYLE embed (optional: switch to loop_for_context if you want stronger “recent” bias)
loop_embed = mrt.embed_style(loop_for_context)
embeds, weights = [loop_embed], [float(loop_weight)]
if extra_styles:
for i, s in enumerate(extra_styles):
if s.strip():
embeds.append(mrt.embed_style(s.strip()))
w = style_weights[i] if (style_weights and i < len(style_weights)) else 1.0
weights.append(float(w))
wsum = float(sum(weights)) or 1.0
weights = [w / wsum for w in weights]
combined_style = np.sum([w * e for w, e in zip(weights, embeds)], axis=0).astype(loop_embed.dtype)
# --- Length math ---
seconds_per_bar = beats_per_bar * (60.0 / bpm)
total_secs = bars * seconds_per_bar
drop_bars = max(0, int(intro_bars_to_drop))
drop_secs = min(drop_bars, bars) * seconds_per_bar # clamp to <= bars
gen_total_secs = total_secs + drop_secs # generate extra
# Chunk scheduling to cover gen_total_secs
chunk_secs = mrt.config.chunk_length_frames * mrt.config.frame_length_samples / mrt.sample_rate # ~2.0
steps = int(math.ceil(gen_total_secs / chunk_secs)) + 1 # pad then trim
# Generate
chunks = []
for _ in range(steps):
wav, state = mrt.generate_chunk(state=state, style=combined_style)
chunks.append(wav)
# Stitch continuous audio
stitched = stitch_generated(chunks, mrt.sample_rate, mrt.config.crossfade_length).as_stereo()
# Trim to generated length (bars + dropped bars)
stitched = hard_trim_seconds(stitched, gen_total_secs)
# 👉 Drop the intro bars
if drop_secs > 0:
n_drop = int(round(drop_secs * stitched.sample_rate))
stitched = au.Waveform(stitched.samples[n_drop:], stitched.sample_rate)
# Final exact-length trim to requested bars
out = hard_trim_seconds(stitched, total_secs)
# Final polish AFTER drop
out = out.peak_normalize(0.95)
apply_micro_fades(out, 5)
# Loudness match to input (after drop) so bar 1 sits right
out, loud_stats = match_loudness_to_reference(
ref=loop, target=out,
method=loudness_mode, headroom_db=loudness_headroom_db
)
return out, loud_stats
# ----------------------------
# FastAPI app with lazy, thread-safe model init
# ----------------------------
app = FastAPI()
app.add_middleware(
CORSMiddleware,
allow_origins=["*"], # or lock to your domain(s)
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
_MRT = None
_MRT_LOCK = threading.Lock()
def get_mrt():
global _MRT
if _MRT is None:
with _MRT_LOCK:
if _MRT is None:
_MRT = system.MagentaRT(tag="large", guidance_weight=5.0, device="gpu", lazy=False)
return _MRT
_WARMED = False
_WARMUP_LOCK = threading.Lock()
def _mrt_warmup():
"""
Build a minimal, bar-aligned silent context and run one 2s generate_chunk
to trigger XLA JIT & autotune so first real request is fast.
"""
global _WARMED
with _WARMUP_LOCK:
if _WARMED:
return
try:
mrt = get_mrt()
# --- derive timing from model config ---
codec_fps = float(mrt.codec.frame_rate)
ctx_seconds = float(mrt.config.context_length_frames) / codec_fps
sr = int(mrt.sample_rate)
# We'll align to 120 BPM, 4/4, and generate one ~2s chunk
bpm = 120.0
beats_per_bar = 4
# --- build a silent, stereo context of ctx_seconds ---
import numpy as np, soundfile as sf
samples = int(max(1, round(ctx_seconds * sr)))
silent = np.zeros((samples, 2), dtype=np.float32)
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
sf.write(tmp.name, silent, sr, subtype="PCM_16")
tmp_path = tmp.name
try:
# Load as Waveform and take a tail of exactly ctx_seconds
loop = au.Waveform.from_file(tmp_path).resample(sr).as_stereo()
seconds_per_bar = beats_per_bar * (60.0 / bpm)
ctx_tail = take_bar_aligned_tail(loop, bpm, beats_per_bar, ctx_seconds)
# Tokens for context window
tokens_full = mrt.codec.encode(ctx_tail).astype(np.int32)
tokens = tokens_full[:, :mrt.config.decoder_codec_rvq_depth]
context_tokens = make_bar_aligned_context(
tokens,
bpm=bpm,
fps=float(mrt.codec.frame_rate),
ctx_frames=mrt.config.context_length_frames,
beats_per_bar=beats_per_bar,
)
# Init state and a basic style vector (text token is fine)
state = mrt.init_state()
state.context_tokens = context_tokens
style_vec = mrt.embed_style("warmup")
# --- one throwaway chunk (~2s) ---
_wav, _state = mrt.generate_chunk(state=state, style=style_vec)
logging.info("MagentaRT warmup complete.")
finally:
try:
os.unlink(tmp_path)
except Exception:
pass
_WARMED = True
except Exception as e:
# Never crash on warmup errors; log and continue serving
logging.exception("MagentaRT warmup failed (continuing without warmup): %s", e)
# Kick it off in the background on server start
@app.on_event("startup")
def _kickoff_warmup():
if os.getenv("MRT_WARMUP", "1") != "0":
threading.Thread(target=_mrt_warmup, name="mrt-warmup", daemon=True).start()
@app.post("/generate")
def generate(
loop_audio: UploadFile = File(...),
bpm: float = Form(...),
bars: int = Form(8),
beats_per_bar: int = Form(4),
styles: str = Form("acid house"),
style_weights: str = Form(""),
loop_weight: float = Form(1.0),
loudness_mode: str = Form("auto"),
loudness_headroom_db: float = Form(1.0),
guidance_weight: float = Form(5.0),
temperature: float = Form(1.1),
topk: int = Form(40),
target_sample_rate: int | None = Form(None),
intro_bars_to_drop: int = Form(0), # <— NEW
):
# Read file
data = loop_audio.file.read()
if not data:
return {"error": "Empty file"}
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
tmp.write(data)
tmp_path = tmp.name
# Parse styles + weights
extra_styles = [s for s in (styles.split(",") if styles else []) if s.strip()]
weights = [float(x) for x in style_weights.split(",")] if style_weights else None
mrt = get_mrt() # warm once, in this worker thread
# Temporarily override MRT inference knobs for this request
with mrt_overrides(mrt,
guidance_weight=guidance_weight,
temperature=temperature,
topk=topk):
wav, loud_stats = generate_loop_continuation_with_mrt(
mrt,
input_wav_path=tmp_path,
bpm=bpm,
extra_styles=extra_styles,
style_weights=weights,
bars=bars,
beats_per_bar=beats_per_bar,
loop_weight=loop_weight,
loudness_mode=loudness_mode,
loudness_headroom_db=loudness_headroom_db,
intro_bars_to_drop=intro_bars_to_drop, # <— pass through
)
# 1) Figure out the desired SR
inp_info = sf.info(tmp_path)
input_sr = int(inp_info.samplerate)
target_sr = int(target_sample_rate or input_sr)
# 2) Convert to target SR + snap to exact bars
cur_sr = int(mrt.sample_rate)
x = wav.samples if wav.samples.ndim == 2 else wav.samples[:, None]
seconds_per_bar = (60.0 / float(bpm)) * int(beats_per_bar)
expected_secs = float(bars) * seconds_per_bar
x = resample_and_snap(x, cur_sr=cur_sr, target_sr=target_sr, seconds=expected_secs)
# 3) Encode WAV once (no extra write)
audio_b64, total_samples, channels = wav_bytes_base64(x, target_sr)
loop_duration_seconds = total_samples / float(target_sr)
# 4) Metadata
metadata = {
"bpm": int(round(bpm)),
"bars": int(bars),
"beats_per_bar": int(beats_per_bar),
"styles": extra_styles,
"style_weights": weights,
"loop_weight": loop_weight,
"loudness": loud_stats,
"sample_rate": int(target_sr),
"channels": int(channels),
"crossfade_seconds": mrt.config.crossfade_length,
"total_samples": int(total_samples),
"seconds_per_bar": seconds_per_bar,
"loop_duration_seconds": loop_duration_seconds,
"guidance_weight": guidance_weight,
"temperature": temperature,
"topk": topk,
}
return {"audio_base64": audio_b64, "metadata": metadata}
# ----------------------------
# the 'keep jamming' button
# ----------------------------
@app.post("/jam/start")
def jam_start(
loop_audio: UploadFile = File(...),
bpm: float = Form(...),
bars_per_chunk: int = Form(4),
beats_per_bar: int = Form(4),
styles: str = Form(""),
style_weights: str = Form(""),
loop_weight: float = Form(1.0),
loudness_mode: str = Form("auto"),
loudness_headroom_db: float = Form(1.0),
guidance_weight: float = Form(1.1),
temperature: float = Form(1.1),
topk: int = Form(40),
target_sample_rate: int | None = Form(None),
):
# enforce single active jam per GPU
with jam_lock:
for sid, w in list(jam_registry.items()):
if w.is_alive():
raise HTTPException(status_code=429, detail="A jam is already running. Try again later.")
# read input + prep context/style (reuse your existing code)
data = loop_audio.file.read()
if not data: raise HTTPException(status_code=400, detail="Empty file")
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
tmp.write(data); tmp_path = tmp.name
mrt = get_mrt()
loop = au.Waveform.from_file(tmp_path).resample(mrt.sample_rate).as_stereo()
# build tail context + style vec (tail-biased)
codec_fps = float(mrt.codec.frame_rate)
ctx_seconds = float(mrt.config.context_length_frames) / codec_fps
loop_tail = take_bar_aligned_tail(loop, bpm, beats_per_bar, ctx_seconds)
# style vec = normalized mix of loop_tail + extra styles
embeds, weights = [mrt.embed_style(loop_tail)], [float(loop_weight)]
extra = [s for s in (styles.split(",") if styles else []) if s.strip()]
sw = [float(x) for x in style_weights.split(",")] if style_weights else []
for i, s in enumerate(extra):
embeds.append(mrt.embed_style(s.strip()))
weights.append(sw[i] if i < len(sw) else 1.0)
wsum = sum(weights) or 1.0
weights = [w / wsum for w in weights]
style_vec = np.sum([w * e for w, e in zip(weights, embeds)], axis=0).astype(embeds[0].dtype)
# target SR (default input SR)
inp_info = sf.info(tmp_path)
input_sr = int(inp_info.samplerate)
target_sr = int(target_sample_rate or input_sr)
params = JamParams(
bpm=bpm,
beats_per_bar=beats_per_bar,
bars_per_chunk=bars_per_chunk,
target_sr=target_sr,
loudness_mode=loudness_mode,
headroom_db=loudness_headroom_db,
style_vec=style_vec,
ref_loop=loop_tail, # For loudness matching
combined_loop=loop, # NEW: Full loop for context setup
guidance_weight=guidance_weight,
temperature=temperature,
topk=topk
)
worker = JamWorker(mrt, params)
sid = str(uuid.uuid4())
with jam_lock:
jam_registry[sid] = worker
worker.start()
return {"session_id": sid}
@app.get("/jam/next")
def jam_next(session_id: str):
"""
Get the next sequential chunk in the jam session.
This ensures chunks are delivered in order without gaps.
"""
with jam_lock:
worker = jam_registry.get(session_id)
if worker is None or not worker.is_alive():
raise HTTPException(status_code=404, detail="Session not found")
# Get the next sequential chunk (this blocks until ready)
chunk = worker.get_next_chunk()
if chunk is None:
raise HTTPException(status_code=408, detail="Chunk not ready within timeout")
return {
"chunk": {
"index": chunk.index,
"audio_base64": chunk.audio_base64,
"metadata": chunk.metadata
}
}
@app.post("/jam/consume")
def jam_consume(session_id: str = Form(...), chunk_index: int = Form(...)):
"""
Mark a chunk as consumed by the frontend.
This helps the worker manage its buffer and generation flow.
"""
with jam_lock:
worker = jam_registry.get(session_id)
if worker is None or not worker.is_alive():
raise HTTPException(status_code=404, detail="Session not found")
worker.mark_chunk_consumed(chunk_index)
return {"consumed": chunk_index}
@app.post("/jam/stop")
def jam_stop(session_id: str = Body(..., embed=True)):
with jam_lock:
worker = jam_registry.get(session_id)
if worker is None:
raise HTTPException(status_code=404, detail="Session not found")
worker.stop()
worker.join(timeout=5.0)
if worker.is_alive():
# It’s daemon=True, so it won’t block process exit, but report it
print(f"⚠️ JamWorker {session_id} did not stop within timeout")
with jam_lock:
jam_registry.pop(session_id, None)
return {"stopped": True}
@app.post("/jam/update") # consolidated
def jam_update(
session_id: str = Form(...),
# knobs (all optional)
guidance_weight: Optional[float] = Form(None),
temperature: Optional[float] = Form(None),
topk: Optional[int] = Form(None),
# styles (all optional)
styles: str = Form(""),
style_weights: str = Form(""),
loop_weight: Optional[float] = Form(None), # None means "don’t change"
use_current_mix_as_style: bool = Form(False),
):
with jam_lock:
worker = jam_registry.get(session_id)
if worker is None or not worker.is_alive():
raise HTTPException(status_code=404, detail="Session not found")
# --- 1) Apply knob updates (atomic under lock)
if any(v is not None for v in (guidance_weight, temperature, topk)):
worker.update_knobs(
guidance_weight=guidance_weight,
temperature=temperature,
topk=topk
)
# --- 2) Apply style updates only if requested
wants_style_update = use_current_mix_as_style or (styles.strip() != "")
if wants_style_update:
embeds, weights = [], []
# optional: include current mix as a style component
if use_current_mix_as_style and worker.params.combined_loop is not None:
lw = 1.0 if loop_weight is None else float(loop_weight)
embeds.append(worker.mrt.embed_style(worker.params.combined_loop))
weights.append(lw)
# extra text styles
extra = [s for s in (styles.split(",") if styles else []) if s.strip()]
sw = [float(x) for x in style_weights.split(",")] if style_weights else []
for i, s in enumerate(extra):
embeds.append(worker.mrt.embed_style(s.strip()))
weights.append(sw[i] if i < len(sw) else 1.0)
if embeds: # only swap if we actually built something
wsum = sum(weights) or 1.0
weights = [w / wsum for w in weights]
style_vec = np.sum([w * e for w, e in zip(weights, embeds)], axis=0).astype(np.float32)
# install atomically
with worker._lock:
worker.params.style_vec = style_vec
return {"ok": True}
@app.post("/jam/reseed")
def jam_reseed(session_id: str = Form(...), loop_audio: UploadFile = File(None)):
with jam_lock:
worker = jam_registry.get(session_id)
if worker is None or not worker.is_alive():
raise HTTPException(status_code=404, detail="Session not found")
# Option 1: use uploaded new “combined” bounce from the app
if loop_audio is not None:
data = loop_audio.file.read()
if not data:
raise HTTPException(status_code=400, detail="Empty file")
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
tmp.write(data); path = tmp.name
wav = au.Waveform.from_file(path).resample(worker.mrt.sample_rate).as_stereo()
else:
# Option 2: reseed from what we’ve been streaming (the model side)
# (Usually better to reseed from the Swift-side “combined” mix you trust.)
s = getattr(worker, "_stream", None)
if s is None or s.shape[0] == 0:
raise HTTPException(status_code=400, detail="No internal stream to reseed from")
wav = au.Waveform(s.astype(np.float32, copy=False), int(worker.mrt.sample_rate)).as_stereo()
worker.reseed_from_waveform(wav)
return {"ok": True}
@app.post("/jam/reseed_splice")
def jam_reseed_splice(
session_id: str = Form(...),
anchor_bars: float = Form(2.0), # how much of the original to re-inject
combined_audio: UploadFile = File(None), # preferred: Swift supplies the current combined mix
):
worker = jam_registry.get(session_id)
if worker is None or not worker.is_alive():
raise HTTPException(status_code=404, detail="Session not found")
# Build a waveform to reseed from
wav = None
if combined_audio is not None:
data = combined_audio.file.read()
if not data:
raise HTTPException(status_code=400, detail="Empty combined_audio")
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
tmp.write(data)
path = tmp.name
wav = au.Waveform.from_file(path).resample(worker.mrt.sample_rate).as_stereo()
else:
# Fallback: reseed from the model’s internal stream (less ideal than the Swift-side bounce)
s = getattr(worker, "_stream", None)
if s is None or s.shape[0] == 0:
raise HTTPException(status_code=400, detail="No audio available to reseed from")
wav = au.Waveform(s.astype(np.float32, copy=False), int(worker.mrt.sample_rate)).as_stereo()
# Perform the splice reseed
worker.reseed_splice(wav, anchor_bars=float(anchor_bars))
return {"ok": True, "anchor_bars": float(anchor_bars)}
@app.get("/jam/status")
def jam_status(session_id: str):
with jam_lock:
worker = jam_registry.get(session_id)
if worker is None:
raise HTTPException(status_code=404, detail="Session not found")
running = worker.is_alive()
# Snapshot safely
with worker._lock:
last_generated = int(worker.idx)
last_delivered = int(worker._last_delivered_index)
queued = len(worker.outbox)
buffer_ahead = last_generated - last_delivered
p = worker.params
spb = p.beats_per_bar * (60.0 / p.bpm)
chunk_secs = p.bars_per_chunk * spb
return {
"running": running,
"last_generated_index": last_generated, # Last chunk that finished generating
"last_delivered_index": last_delivered, # Last chunk sent to frontend
"buffer_ahead": buffer_ahead, # How many chunks ahead we are
"queued_chunks": queued, # Total chunks in outbox
"bpm": p.bpm,
"beats_per_bar": p.beats_per_bar,
"bars_per_chunk": p.bars_per_chunk,
"seconds_per_bar": spb,
"chunk_duration_seconds": chunk_secs,
"target_sample_rate": p.target_sr,
"last_chunk_started_at": worker.last_chunk_started_at,
"last_chunk_completed_at": worker.last_chunk_completed_at,
}
@app.get("/health")
def health():
return {"ok": True}
@app.middleware("http")
async def log_requests(request: Request, call_next):
rid = request.headers.get("X-Request-ID", "-")
print(f"📥 {request.method} {request.url.path}?{request.url.query} [rid={rid}]")
try:
response = await call_next(request)
except Exception as e:
print(f"💥 exception for {request.url.path} [rid={rid}]: {e}")
raise
print(f"📤 {response.status_code} {request.url.path} [rid={rid}]")
return response
@app.get("/ping")
def ping():
return {"ok": True}
@app.get("/", response_class=Response)
def read_root():
"""Root endpoint that explains what this API does"""
html_content = """
<!DOCTYPE html>
<html>
<head><title>MagentaRT Research API</title></head>
<body style="font-family: Arial; max-width: 800px; margin: 50px auto; padding: 20px;">
<h1>🎵 MagentaRT Research API</h1>
<p><strong>Purpose:</strong> AI music generation for iOS app research using Google's MagentaRT</p>
<h2>Available Endpoints:</h2>
<ul>
<li><code>POST /generate</code> - Generate 4-8 bars of music</li>
<li><code>POST /jam/start</code> - Start continuous jamming</li>
<li><code>GET /jam/next</code> - Get next chunk</li>
<li><code>POST /jam/consume</code> - confirm a chunk as consumed</li>
<li><code>POST /jam/stop</code> - End session</li>
<li><code>GET /docs</code> - API documentation</li>
</ul>
<p><strong>Research Only:</strong> Experimental implementation for iOS app development.</p>
<p><strong>Licensing:</strong> Uses MagentaRT (Apache 2.0 + CC-BY 4.0). Users responsible for outputs.</p>
<p>Visit <a href="/docs">/docs</a> for detailed API documentation.</p>
</body>
</html>
"""
return Response(content=html_content, media_type="text/html")
# ----------------------------
# websockets route
# ----------------------------
def _combine_styles(mrt, styles_str: str = "", weights_str: str = ""):
extra = [s.strip() for s in (styles_str or "").split(",") if s.strip()]
if not extra:
return mrt.embed_style("warmup")
sw = [float(x) for x in (weights_str or "").split(",") if x.strip()]
embeds, weights = [], []
for i, s in enumerate(extra):
embeds.append(mrt.embed_style(s))
weights.append(sw[i] if i < len(sw) else 1.0)
wsum = sum(weights) or 1.0
weights = [w/wsum for w in weights]
import numpy as np
return np.sum([w*e for w, e in zip(weights, embeds)], axis=0).astype(np.float32)
@app.websocket("/ws/jam")
async def ws_jam(websocket: WebSocket):
await websocket.accept()
sid = None
worker = None
binary_audio = False
mode = "rt" # or "bar"
# NEW: capture ws in closure
async def send_json(obj):
return await send_json_safe(websocket, obj)
try:
while True:
raw = await websocket.receive_text()
msg = json.loads(raw)
mtype = msg.get("type")
# --- START ---
if mtype == "start":
binary_audio = bool(msg.get("binary_audio", False))
mode = msg.get("mode", "bar")
params = msg.get("params", {}) or {}
sid = msg.get("session_id")
# attach or create
if sid:
with jam_lock:
worker = jam_registry.get(sid)
if worker is None or not worker.is_alive():
await send_json({"type":"error","error":"Session not found"})
continue
else:
# optionally accept base64 loop and start a new worker (bar-mode)
if mode == "bar":
loop_b64 = msg.get("loop_audio_b64")
if not loop_b64:
await send_json({"type":"error","error":"loop_audio_b64 required for mode=bar when no session_id"})
continue
loop_bytes = base64.b64decode(loop_b64)
# mimic /jam/start
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
tmp.write(loop_bytes); tmp_path = tmp.name
# build JamParams similar to /jam/start
mrt = get_mrt()
model_sr = int(mrt.sample_rate) # typically 48000
# Defaults for WS: raw loudness @ model SR, unless overridden by client:
target_sr = int(params.get("target_sr", model_sr))
loudness_mode = params.get("loudness_mode", "none")
headroom_db = float(params.get("headroom_db", 1.0))
loop = au.Waveform.from_file(tmp_path).resample(mrt.sample_rate).as_stereo()
codec_fps = float(mrt.codec.frame_rate)
ctx_seconds = float(mrt.config.context_length_frames) / codec_fps
bpm = float(params.get("bpm", 120.0))
bpb = int(params.get("beats_per_bar", 4))
loop_tail = take_bar_aligned_tail(loop, bpm, bpb, ctx_seconds)
# style vector (loop + extra styles)
embeds, weights = [mrt.embed_style(loop_tail)], [float(params.get("loop_weight", 1.0))]
extra = [s for s in (params.get("styles","").split(",")) if s.strip()]
sw = [float(x) for x in params.get("style_weights","").split(",") if x.strip()]
for i, s in enumerate(extra):
embeds.append(mrt.embed_style(s.strip()))
weights.append(sw[i] if i < len(sw) else 1.0)
wsum = sum(weights) or 1.0
weights = [w/wsum for w in weights]
style_vec = np.sum([w*e for w, e in zip(weights, embeds)], axis=0).astype(np.float32)
# target SR fallback: input SR
inp_info = sf.info(tmp_path)
target_sr = int(params.get("target_sr", int(inp_info.samplerate)))
# Build JamParams for WS bar-mode
jp = JamParams(
bpm=bpm, beats_per_bar=bpb, bars_per_chunk=int(params.get("bars_per_chunk", 8)),
target_sr=target_sr,
loudness_mode=loudness_mode, headroom_db=headroom_db,
style_vec=style_vec,
ref_loop=None if loudness_mode == "none" else loop_tail, # disable match by default
combined_loop=loop,
guidance_weight=float(params.get("guidance_weight", 1.1)),
temperature=float(params.get("temperature", 1.1)),
topk=int(params.get("topk", 40)),
)
worker = JamWorker(get_mrt(), jp)
sid = str(uuid.uuid4())
with jam_lock:
# single active jam per GPU, mirroring /jam/start
for _sid, w in list(jam_registry.items()):
if w.is_alive():
await send_json({"type":"error","error":"A jam is already running"})
worker = None; sid = None
break
if worker is not None:
jam_registry[sid] = worker
worker.start()
else:
# mode == "rt" (Colab-style, no loop context)
# seed a fresh state with a silent context like warmup
mrt = get_mrt()
state = mrt.init_state()
codec_fps = float(mrt.codec.frame_rate)
ctx_seconds = float(mrt.config.context_length_frames) / codec_fps
sr = int(mrt.sample_rate)
samples = int(max(1, round(ctx_seconds * sr)))
silent = au.Waveform(np.zeros((samples,2), np.float32), sr)
tokens = mrt.codec.encode(silent).astype(np.int32)[:, :mrt.config.decoder_codec_rvq_depth]
state.context_tokens = tokens
websocket._mrt = mrt
websocket._state = state
websocket._style = _combine_styles(mrt,
params.get("styles","warmup"),
params.get("style_weights",""))
websocket._rt_running = True
websocket._rt_sr = sr
websocket._rt_topk = int(params.get("topk", 40))
websocket._rt_temp = float(params.get("temperature", 1.1))
websocket._rt_guid = float(params.get("guidance_weight", 1.1))
websocket._pace = params.get("pace", "asap") # "realtime" | "asap"
await send_json({"type":"started","mode":"rt"})
# kick off a background task to stream ~2s chunks
async def _rt_loop():
try:
mrt = websocket._mrt
chunk_secs = (mrt.config.chunk_length_frames * mrt.config.frame_length_samples) / float(mrt.sample_rate)
target_next = time.perf_counter()
while websocket._rt_running:
# read knobs (already set by update)
mrt.guidance_weight = websocket._rt_guid
mrt.temperature = websocket._rt_temp
mrt.topk = websocket._rt_topk
wav, new_state = mrt.generate_chunk(state=websocket._state, style=websocket._style)
websocket._state = new_state
x = wav.samples.astype(np.float32, copy=False)
buf = io.BytesIO()
sf.write(buf, x, mrt.sample_rate, subtype="FLOAT", format="WAV")
# send bytes / json best-effort
ok = True
if binary_audio:
try:
await websocket.send_bytes(buf.getvalue())
ok = await send_json({"type":"chunk_meta","metadata":{"sample_rate":mrt.sample_rate}})
except Exception:
ok = False
else:
b64 = base64.b64encode(buf.getvalue()).decode("utf-8")
ok = await send_json({"type":"chunk","audio_base64":b64,
"metadata":{"sample_rate":mrt.sample_rate}})
if not ok:
# client went away — exit cleanly
break
# pacing (use captured flag from start)
if getattr(websocket, "_pace", "asap") == "realtime":
t1 = time.perf_counter()
target_next += chunk_secs
sleep_s = max(0.0, target_next - t1 - 0.02)
if sleep_s > 0:
await asyncio.sleep(sleep_s)
except asyncio.CancelledError:
# normal on stop/close — just exit
pass
except Exception:
# don't try to send an error; socket may be closed
pass
websocket._rt_task = asyncio.create_task(_rt_loop())
continue # skip the “bar-mode started” message below
await send_json({"type":"started","session_id": sid, "mode": mode})
# if we’re in bar-mode, begin pushing chunks as they arrive
if mode == "bar" and worker is not None:
async def _pump():
while True:
if not worker.is_alive():
break
chunk = worker.get_next_chunk(timeout=60.0)
if chunk is None:
continue
if binary_audio:
await websocket.send_bytes(base64.b64decode(chunk.audio_base64))
await send_json({"type":"chunk_meta","index":chunk.index,"metadata":chunk.metadata})
else:
await send_json({"type":"chunk","index":chunk.index,
"audio_base64":chunk.audio_base64,"metadata":chunk.metadata})
asyncio.create_task(_pump())
# --- UPDATES (bar or rt) ---
elif mtype == "update":
if mode == "bar":
if not sid:
await send_json({"type":"error","error":"No session_id yet"}); return
# fan values straight into your existing HTTP handler:
res = jam_update(
session_id=sid,
guidance_weight=msg.get("guidance_weight"),
temperature=msg.get("temperature"),
topk=msg.get("topk"),
styles=msg.get("styles",""),
style_weights=msg.get("style_weights",""),
loop_weight=msg.get("loop_weight"),
use_current_mix_as_style=bool(msg.get("use_current_mix_as_style", False)),
)
await send_json({"type":"status", **res}) # {"ok": True}
else:
# rt-mode: there’s no JamWorker; update the local knobs/state
websocket._rt_temp = float(msg.get("temperature", websocket._rt_temp))
websocket._rt_topk = int(msg.get("topk", websocket._rt_topk))
websocket._rt_guid = float(msg.get("guidance_weight", websocket._rt_guid))
if ("styles" in msg) or ("style_weights" in msg):
websocket._style = _combine_styles(
websocket._mrt,
msg.get("styles", ""),
msg.get("style_weights", "")
)
await send_json({"type":"status","updated":"rt-knobs"})
elif mtype == "consume" and mode == "bar":
with jam_lock:
worker = jam_registry.get(msg.get("session_id"))
if worker is not None:
worker.mark_chunk_consumed(int(msg.get("chunk_index", -1)))
elif mtype == "reseed" and mode == "bar":
with jam_lock:
worker = jam_registry.get(msg.get("session_id"))
if worker is None or not worker.is_alive():
await send_json({"type":"error","error":"Session not found"}); continue
loop_b64 = msg.get("loop_audio_b64")
if not loop_b64:
await send_json({"type":"error","error":"loop_audio_b64 required"}); continue
loop_bytes = base64.b64decode(loop_b64)
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
tmp.write(loop_bytes); path = tmp.name
wav = au.Waveform.from_file(path).resample(worker.mrt.sample_rate).as_stereo()
worker.reseed_from_waveform(wav)
await send_json({"type":"status","reseeded":True})
elif mtype == "reseed_splice" and mode == "bar":
with jam_lock:
worker = jam_registry.get(msg.get("session_id"))
if worker is None or not worker.is_alive():
await send_json({"type":"error","error":"Session not found"}); continue
anchor = float(msg.get("anchor_bars", 2.0))
b64 = msg.get("combined_audio_b64")
if b64:
data = base64.b64decode(b64)
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
tmp.write(data); path = tmp.name
wav = au.Waveform.from_file(path).resample(worker.mrt.sample_rate).as_stereo()
worker.reseed_splice(wav, anchor_bars=anchor)
else:
# fallback: model-side stream splice
worker.reseed_splice(worker.params.combined_loop, anchor_bars=anchor)
await send_json({"type":"status","splice":anchor})
elif mtype == "stop":
if mode == "rt":
websocket._rt_running = False
task = getattr(websocket, "_rt_task", None)
if task is not None:
task.cancel()
try: await task
except asyncio.CancelledError: pass
await send_json({"type":"stopped"})
break # <- add this if you want to end the socket after stop
elif mtype == "ping":
await send_json({"type":"pong"})
else:
await send_json({"type":"error","error":f"Unknown type {mtype}"})
except WebSocketDisconnect:
# best-effort cleanup for bar-mode sessions started within this socket (optional)
pass
except Exception as e:
try:
await send_json({"type":"error","error":str(e)})
except Exception:
pass
finally:
try:
if websocket.client_state != WebSocketState.DISCONNECTED:
await websocket.close()
except Exception:
pass |