qlabeler/.env.example at main · qforge-dev/qlabeler · GitHub

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
# Hugging Face token with access to gated models.
# facebook/sam-audio-large requires approval on the HF account behind this token.
HF_TOKEN=hf_your_token_here

# Repo and runtime paths. Defaults work for RunPod /workspace machines.
# ROOT_DIR defaults to the repo root inferred from scripts/setup_model_apis.sh.
# WORKSPACE_DIR defaults to /workspace when it exists, otherwise the repo root.
# ROOT_DIR=/workspace/qlabeler
# WORKSPACE_DIR=/workspace
# VENV_DIR=/workspace/venvs
# OUTPUT_DIR=/workspace/outputs
# LOG_DIR=/workspace/logs/qlabeler

# API ports.
PIPELINE_PORT=8000
AFNEXT_PORT=8001
SAM_AUDIO_PORT=8002

# Pipeline queue/dashboard.
# The RunPod setup uses real model endpoints. scripts/run_pipeline_dev.sh overrides this to mock.
PIPELINE_BACKEND=real
PIPELINE_DB_PATH=/workspace/pipeline.sqlite3
PIPELINE_WORKER_ENABLED=1
PIPELINE_CHUNK_SECONDS=30
PIPELINE_OVERLAP_SECONDS=5
PIPELINE_SOUND_GATE_MIN_DBFS=-50
PIPELINE_SOUND_GATE_MIN_PEAK_DBFS=-55
PIPELINE_SOUND_GATE_WINDOW_MS=100
PIPELINE_SOUND_GATE_MIN_ACTIVE_MS=250
PIPELINE_SOUND_GATE_MIN_ACTIVE_RATIO=0.01
AFNEXT_ENDPOINT=http://127.0.0.1:8001/v1/audio-flamingo/ask
AFNEXT_BATCH_ENDPOINT=http://127.0.0.1:8001/v1/audio-flamingo/ask_batch
SAM_AUDIO_ENDPOINT=http://127.0.0.1:8002/v1/sam-audio/separate
SAM_AUDIO_BATCH_ENDPOINT=http://127.0.0.1:8002/v1/sam-audio/separate_batch
# How many tasks each queue worker claims per round. One worker thread runs per
# queue (sound_gate, audio_flamingo, sam_audio), so the models work in parallel.
# SAM batches are sent to the batch endpoint in one inference call.
PIPELINE_SAM_BATCH_SIZE=4
# Batched Audio Flamingo generation is opt-in (1 = single requests). Validate
# output quality with batches before raising this.
PIPELINE_AFNEXT_BATCH_SIZE=1

# Artifact storage.
# Local dev/tests use local. In production, set PIPELINE_STORAGE_BACKEND=s3
# and configure the bucket. AWS credentials are read from the standard
# environment/instance role chain used by boto3.
PIPELINE_STORAGE_BACKEND=local
# S3_BUCKET=your-artifact-bucket
# S3_PREFIX=qlabeler
# AWS_REGION=us-east-1
# S3_ENDPOINT_URL=
# S3_PUBLIC_BASE_URL=
# S3_PRESIGN_SECONDS=0

# Model ids.
AFNEXT_MODEL_ID=nvidia/audio-flamingo-next-think-hf
SAM_AUDIO_MODEL_ID=facebook/sam-audio-large

# PyTorch CUDA wheel index. cu124 works on the tested A100 RunPod image.
PYTORCH_INDEX_URL=https://download.pytorch.org/whl/cu124
REQUIRE_CUDA=1

# Bootstrap behavior. Set LOAD_MODELS=0 to start APIs without preloading weights.
LOAD_MODELS=1

# Audio Flamingo generation defaults.
AFNEXT_LOAD_ON_STARTUP=0
AFNEXT_MAX_NEW_TOKENS=1024
AFNEXT_REPETITION_PENALTY=1.2
# Server-side cap on items processed per generate() call on /ask_batch.
AFNEXT_MAX_BATCH=4

# SAM-Audio defaults. Keep clips short unless the service code is changed.
SAM_AUDIO_LOAD_ON_STARTUP=0
SAM_AUDIO_MAX_AUDIO_SECONDS=35
SAM_AUDIO_PREDICT_SPANS=0
SAM_AUDIO_RERANKING_CANDIDATES=1
# Server-side cap on items processed per separate() call on /separate_batch.
# Lower this if peak_cuda_allocated_gb gets close to the GPU limit.
SAM_AUDIO_MAX_BATCH=4