caption-flow 0.1.0

Self-contained distributed community captioning system

captionflow

$ pip install caption-flow
$ caption-flow orchestrator|worker|monitor

scalable, fault-tolerant vllm-powered image captioning. this "first round" focuses on a blazing fast websocket orchestrator plus lightweight gpu workers that batch requests through vllm.

performance: consumer 4090s often outpace h100s on smaller models (3b-7b) due to higher clock speeds and lower overhead. we've seen 150+ images/sec on a single 4090 with qwen2.5-vl-3b.

• orchestrator: hands out work in chunked shards, collects captions, checkpoints progress, and keeps simple stats. handles 10k+ chunks/sec on commodity hardware.
• workers (vllm): connect to the orchestrator, stream in image samples, batch them, and generate 1..n captions per image using prompts supplied by the orchestrator.
• dataworkers (coming soon): separate non-gpu clients that fetch/preprocess images and feed them to the orchestrator, freeing gpu workers to focus purely on inference.
• config-driven: all components read yaml config; flags can override.
• tui monitor (optional): a monitor client is wired into the cli; ship a monitor module to enable it.

no conda. just venv + pip.

---

install

python -m venv .venv
source .venv/bin/activate  # windows: .venv\Scripts\activate
pip install --upgrade pip
pip install -e .  # installs the `caption-flow` command
# or: pip install -e git+ssh://[e-mail]/bghira/captionflow

quickstart (single box)

1. copy + edit the sample configs

cp examples/orchestrator.yaml config/orchestrator.yaml
cp examples/worker.yaml config/worker.yaml
cp examples/monitor.yaml config/monitor.yaml

set a unique shared token in both config/orchestrator.yaml and config/worker.yaml (see auth.worker_tokens in the orchestrator config and worker.token in the worker config). if you use private hugging face datasets/models, export HUGGINGFACE_HUB_TOKEN or use hf auth login (old style: huggingface-cli login) before starting workers.

2. start the orchestrator

caption-flow orchestrator

3. start one or more vllm workers

# gpu 0 on the same host
caption-flow worker --gpu-id 0

# your second gpu
caption-flow worker --gpu-id 1

4. (optional) start the monitor to check on status

caption-flow monitor

5. (optional) scan/fix chunks on disk if you had crashes or want to ensure you're actually receiving all captions correctly

caption-flow scan_chunks --data-dir ./caption_data --checkpoint-dir ./checkpoints --fix

---

how it's wired

orchestrator

• websocket server (default 0.0.0.0:8765) with four client roles: workers, dataworkers, monitors, and admin.
• blazing fast: handles 10,000+ chunks/sec, 100k+ concurrent connections. the bottleneck is always gpu inference, never the orchestrator.
• dataset control: the orchestrator centrally defines the dataset (huggingface or local) and version/name. it chunk-slices shards and assigns work.
• vllm config broadcast: model, tp size, dtype, max seq len, memory targets, batching, sampling params, and inference prompts are all pushed to workers; workers can apply many changes without a model reload.
• storage + checkpoints: captions buffer to disk with periodic checkpoints. chunk state is tracked so restarts don't double-work.
• auth: token lists for worker, dataworker, monitor, and admin roles.

start flags you'll likely use:

--config path                # yaml config for the orchestrator
--port int, --host str       # bind controls
--data-dir path              # overrides storage.data_dir
--cert path, --key path      # enable tls (or use --no-ssl for ws:// in dev)
--vllm                       # use the vllm-style orchestrator (webdataset/hf)

vllm worker

• one process per gpu. select the device with --gpu-id (or worker.gpu_id in yaml).
• gets its marching orders from the orchestrator: dataset info, model, prompts, batch size, and sampling.
• resilient: detects disconnects, abandons the current chunk cleanly, clears queues, reconnects, and resumes.
• batched generate(): images are resized down for consistent batching; each image can get multiple captions (one per prompt).
• optimized for consumer gpus: 4090s often beat h100s on 3b-7b models. higher boost clocks + lower kernel overhead = faster tokens/sec.

start flags you'll likely use:

--config path                 # yaml for the worker
--server url                  # ws(s)://host:port
--token str                   # must match an allowed worker token on the orchestrator
--batch-size int              # override vllm batch size
--vllm                        # use the vllm worker implementation
--gpu-id int                  # which gpu to use
--precision str, --model str  # optional overrides for dtype/model
--no-verify-ssl               # accept self-signed certs in dev

dataworker (coming soon)

• cpu-only image fetching: separate clients that handle dataset i/o, image loading, and preprocessing
• frees gpu workers: gpu workers receive pre-loaded images, spending 100% of time on inference
• scales horizontally: spin up dozens of dataworkers on cpu nodes to saturate gpu throughput
• smart prefetching: predictive loading keeps gpu workers fed with zero wait time

(optional) monitor

• a cli entry exists for a tui monitor; wire in a monitor module to enable it. config lives in monitor.yaml or inside orchestrator.yaml under monitor:.

---

configuration

config discovery order

for any component, the cli looks for config in this order (first match wins):

1. --config /path/to/file.yaml
2. ./<component>.yaml (current directory)
3. ./config/<component>.yaml (config subdirectory)
4. ~/.caption-flow/<component>.yaml
5. $XDG_CONFIG_HOME/caption-flow/<component>.yaml
6. /etc/caption-flow/<component>.yaml
7. any $XDG_CONFIG_DIRS entries under caption-flow/
8. ./examples/<component>.yaml (fallback)

orchestrator.yaml (highlights)

orchestrator:
  host: 0.0.0.0
  port: 8765
  # ssl:
  #   cert: /path/fullchain.pem
  #   key:  /path/privkey.pem

  dataset:
    type: huggingface   # or "local"
    path: <hf-dataset-or-local-path>
    name: <logical-name>
    version: "1.0"

  vllm:
    model: qwen/qwen2.5-vl-3b-instruct
    tensor_parallel_size: 1
    max_model_len: 16384
    dtype: float16
    gpu_memory_utilization: 0.92
    enforce_eager: true
    disable_mm_preprocessor_cache: true
    limit_mm_per_prompt: { image: 1 }

    batch_size: 8

    sampling:
      temperature: 0.7
      top_p: 0.95
      max_tokens: 256
      repetition_penalty: 1.05
      skip_special_tokens: true
      stop: ["<|end|>", "<|endoftext|>", "<|im_end|>"]

    inference_prompts:
      - "describe this image in detail"
      - "provide a comprehensive description of the visual content"
      - "what are the key elements in this image?"

  storage:
    data_dir: ./caption_data
    checkpoint_dir: ./checkpoints
    caption_buffer_size: 100
    checkpoint_interval: 1000

  # chunking/queueing
  chunk_size: 1000
  chunks_per_request: 2
  chunk_buffer_multiplier: 3
  min_chunk_buffer: 10

  auth:
    worker_tokens:
      - { token: "example-worker-token", name: "example worker" }
    dataworker_tokens:
      - { token: "dataworker-token", name: "data feeder 1" }
    monitor_tokens:
      - { token: "letmein", name: "default monitor" }
    admin_tokens:
      - { token: "admin-secret-2024", name: "admin" }

worker.yaml (highlights)

worker:
  server: ws://localhost:8765   # use wss:// in prod
  token: example-worker-token
  name: local-gpu
  gpu_id: 0
  vllm: true

  # local queues
  readahead_size: 256
  inference_queue_size: 128

monitor.yaml (optional)

monitor:
  server: ws://localhost:8765
  token: letmein
  refresh_rate: 1.0
  show_contributors: true
  show_quality_metrics: true
  max_activity_items: 20
  show_chunk_progress: true
  show_worker_queues: true
  show_throughput_graph: true

---

performance notes

consumer gpus shine on smaller models where CPU bottlenecks arise:

• 4090 @ 3b model: 8-15 images/sec
• 4090 @ 7b model: 8-12 images/sec
• h100 @ 3b model: 2-10 images/sec (lower CPU clocks)
• h100 @ 70b model: 2-10 images/sec (where the H100 belongs)

orchestrator throughput:

• 10,000+ chunks/sec on a typical Ryzen / Intel virtual machine
• 10,000+ concurrent websocket connections
• sub-millisecond chunk assignment latency
• bottleneck is always gpu inference, never the orchestrator

scaling tips:

• use smaller models (3b-7b) for first-pass captioning
• consumer gpus (4090/4080) offer best perf/$ on these models
• add dataworkers to prefetch and saturate gpu throughput
• run multiple workers per node (one per gpu)
• for B200, RTX 6000 Pro, and other fast GPUs, using two worker processes per GPU (two tokens required) can provide added GPU utilisation

---

tls / certificates

use the built-in helpers during development:

# self-signed certs for quick local testing
caption-flow generate_cert --self-signed --domain localhost --output-dir ./certs

# inspect any certificate file
caption-flow inspect_cert ./certs/fullchain.pem

then point the orchestrator at the resulting cert/key (or run --no-ssl for dev-only ws://).

---

tips & notes

• multi-gpu: start one worker process per gpu (set --gpu-id or worker.gpu_id).
• throughput: tune vllm.batch_size in the orchestrator config (or override with --batch-size at worker start). higher isn't always better; watch vram.
• prompts: add more strings under vllm.inference_prompts to get multiple captions per image; the worker returns only non-empty generations.
• private hf: if your dataset/model needs auth, export HUGGINGFACE_HUB_TOKEN before caption-flow worker ....
• self-signed ssl: pass --no-verify-ssl to workers/monitors in dev.
• recovery: if you hard-crash mid-run and want to verify your database, caption-flow scan_chunks --fix can help but is basically never needed.

---

architecture

                                   ┌──────────────┐
                                   │              │
┌─────────────┐     websocket      │              │      ┌──────────────┐
│ gpu worker  │◄───────────────────┤              ├─────►│arrow/parquet │
└─────────────┘                    │              │      │   storage    │
                                   │ orchestrator │      └──────────────┘
┌─────────────┐                    │              │
│ gpu worker  │◄───────────────────┤   10k+       │      ┌──────────────┐
└─────────────┘                    │ chunks/sec   ├─────►│ checkpoints  │
                                   │              │      └──────────────┘
┌─────────────┐                    │              │
│ dataworker  │◄───────────────────┤              │
└─────────────┘                    │              │
                                   │              │
┌─────────────┐                    │              │
│   monitor   │◄───────────────────┤              │
└─────────────┘                    └──────────────┘

storage schema

captions.parquet

• job_id: unique job identifier
• dataset: dataset name
• shard: shard identifier
• item_key: item within shard
• caption: generated caption text
• contributor_id: worker who generated it
• timestamp: generation time
• quality_score: optional quality metric

jobs.parquet

• job_id: unique identifier
• dataset: dataset name
• shard: shard identifier
• status: pending/processing/completed/failed
• assigned_to: worker id
• timestamp: status change time

contributors.parquet

• contributor_id: unique identifier
• name: display name
• total_captions: lifetime count
• trust_level: quality tier (0-5)

development

# install with dev dependencies
pip install -e ".[dev]"

# run tests
pytest

# format code
black src/
ruff --fix src/

# type checking
mypy src/

community contribution

to contribute compute:

1. install caption-flow: pip install caption-flow
2. get a worker token from the project maintainer
3. run: caption-flow worker --server wss://project.domain.com:8765 --token YOUR_TOKEN

your contributions will be tracked and attributed in the final dataset!

roadmap

in no particular order:

• dataworker implementation for cpu-based data feeding
• video captioning
• hot config reload via the admin websocket path
• richer monitor tui with real-time graphs
• web interface
• automatic huggingface hub dataset continuous exports
• sequence-parallel inference for large vision models
• discord interface
• more in-depth integration for non-wds datasets
• support chaining of workflows, for 2nd/3rd pass after use of initial tag model etc
• distributed orchestrator clustering for planet-scale captioning

prs welcome. keep it simple and fast.

license

AGPLv3