hao-ai-lab-FastVideo/v2/examples.py

"""Worked examples, runnable as a demo:

    python3 -m v2.examples

Each prints what it demonstrates. This doubles as living documentation of the API.
"""
from __future__ import annotations

from typing import Any

from v2.recipes import build_default_engine, build_omni_engine
from v2.recipes.wan21 import build_wan21_card
from v2.core.request import DiffusionParams, OutputSpec, Request, SamplingParams, TaskType, make_request


def _t2v(mid: str, prompt: str, seed: int, steps: int = 4, **kw: Any) -> Request:
    return make_request(TaskType.T2V, mid, prompt, diffusion=DiffusionParams(num_steps=steps, seed=seed), **kw)


def example_a_text_to_video(eng) -> None:
    print("\n(a) Text → video, one instance (Wan2.1-1.3B)")
    out = eng.run(_t2v("wan2.1-1.3b", "a cat surfing a wave", 7))
    print(f"    video {out.artifacts['video'].frames.shape}  "
          f"denoise_steps={out.metrics['denoise_steps']:.0f}")


def example_b_ltx2_two_stage(eng) -> None:
    print("\n(b) LTX-2 two-stage distilled (base 8-step → upsample → refine 3-step), shared transformer")
    out = eng.run(_t2v("ltx2-2stage-distilled", "a neon city at night", 2))
    print(f"    video {out.artifacts['video'].frames.shape}  "
          f"base={out.metrics['base_steps']:.0f} refine={out.metrics['refine_steps']:.0f}")


def example_c_causal_streaming(eng) -> None:
    print("\n(c) Causal streaming (Wan-causal): chunk rollout + slab-KV, streamable by chunk")
    out = eng.run(
        _t2v("wan-causal-sf-1.3b", "a drone flight over mountains", 3, outputs=OutputSpec(stream={"video": True})))
    print(f"    latents {out.artifacts['latents'].latent.shape}  chunks={out.metrics['chunks']:.0f}  "
          f"streamed_chunks={out.metrics.get('stream_chunks', 0)}")


def example_g_omni_mot() -> None:
    print("\n(g) Omni / MoT (§16): ONE resident instance runs AR + diffusion loops on shared weights")
    eng = build_omni_engine()
    o = eng.run(
        make_request(TaskType.T2V,
                     "cosmos3-vfm",
                     "a phoenix",
                     sampling=SamplingParams(max_tokens=6, seed=1),
                     diffusion=DiffusionParams(num_steps=4, seed=1)))
    print(f"    Cosmos3 (reason→joint denoise): text={o.artifacts['text'].text} "
          f"video={o.artifacts['video'].frames.shape}")
    o2 = eng.run(
        make_request(TaskType.T2I,
                     "bagel-mot",
                     "a teapot",
                     sampling=SamplingParams(max_tokens=6, seed=2),
                     diffusion=DiffusionParams(num_steps=4, seed=2)))
    print(f"    BAGEL (generate_text→generate_image): text={o2.artifacts['text'].text} "
          f"image={o2.artifacts['image'].tensor.shape}")
    print(f"    scheduler priced BOTH WorkUnit kinds (runtime-visible, not one opaque stage): "
          f"{dict(eng.admission.metrics.by_kind)}")


async def _serving_demo() -> None:
    import asyncio

    from v2.serving.deploy import DynamoWorkerAdapter, FakeDynamoRuntime, LocalFleet, build_deployment_card
    from v2.recipes.wan21 import build_wan_t2v_program
    from v2.runtime import AsyncEngine, PoolSet, wan_t2v_disaggregated
    from v2.serving import OmniOpenAIServer

    eng = build_default_engine()
    build_omni_engine(eng)
    ae = AsyncEngine(eng)

    # disaggregated pools: encoder → denoiser → decoder
    card = build_wan21_card()
    pools = PoolSet(wan_t2v_disaggregated(), card)
    pools.warmup()
    ae.register_disaggregated("wan-disagg", pools, build_wan_t2v_program())
    out = await ae.generate(
        make_request(TaskType.T2V, "wan-disagg", "a wave", diffusion=DiffusionParams(num_steps=4, seed=1)))
    print(f"    disaggregated T2V (enc→den→dec): video={out.artifacts['video'].frames.shape} "
          f"cross-pool transfers={out.metrics['transfers']:.0f}")

    # our own OpenAI server over a real socket
    server = OmniOpenAIServer(ae, engine_id="worker-0")
    host, port = await server.serve(port=0)

    async def http(method: str, path: str, body: bytes = b"") -> str:
        r, w = await asyncio.open_connection(host, port)
        w.write(f"{method} {path} HTTP/1.1\r\nHost: x\r\nContent-Length: {len(body)}\r\n\r\n".encode() + body)
        await w.drain()
        data = await r.read()
        w.close()
        return data.decode("utf-8", "replace")

    health = await http("GET", "/health")
    sse = await http("POST", "/v1/chat/completions",
                     b'{"model":"cosmos3-vfm","messages":[{"role":"user","content":"a comet"}],"stream":true}')
    n_chunks = sse.count("data: ")
    print(
        f"    OpenAI server: /health ok={'healthy' in health}; chat SSE streamed {n_chunks} chunks (omni reason→denoise)"
    )
    await server.close()

    # our own fleet router + Dynamo adapter (frontable, not relied upon) — same DeploymentCard
    dcard = build_deployment_card("worker-0", [card])
    fleet = LocalFleet("least_loaded")
    fleet.register("worker-0", ae, dcard)
    routed = fleet.route(make_request(TaskType.T2V, "wan2.1-1.3b", "x", diffusion=DiffusionParams(num_steps=2)))
    dyn = FakeDynamoRuntime()
    dyn.register_worker(DynamoWorkerAdapter(ae, dcard))
    print(f"    LocalFleet routes to '{routed.worker_id}'; Dynamo adapter registered "
          f"({len(dyn.registry)} worker) — both consume the SAME DeploymentCard")


def example_h_serving_and_fleet() -> None:
    import asyncio
    print("\n(h) Serving + fleet (OUR OWN — Dynamo-optional): async engine, role pools, OpenAI server")
    asyncio.run(_serving_demo())


def main() -> None:
    print("=" * 78)
    print("v2 — worked examples. One runtime, many loops.")
    print("=" * 78)
    eng = build_default_engine()
    print(f"registered (recipe, runtime) cards: {list(eng._registry)}")
    example_a_text_to_video(eng)
    example_b_ltx2_two_stage(eng)
    example_c_causal_streaming(eng)
    example_g_omni_mot()
    example_h_serving_and_fleet()
    print("\n" + "=" * 78)
    print("All examples ran on CPU with numpy toy components. The architecture (cards, driven")
    print("loops, scheduler, caches, parity, and serving) is real; the neural")
    print("forwards are toys. On a GPU box, swap ComponentSpec.factory for the torch adapters.")
    print("=" * 78)


if __name__ == "__main__":
    main()