SolitaryThinker
270fae959d
A scoped, CPU-testable realization of design_v3.md — the architecture where the atomic unit is a typed (recipe, runtime) ModelCard, the model owns loop semantics while the runtime owns loop lifecycle, one resident instance runs many loops, and training records behavior on the same loops it serves. Implements: - card/ loop/ runtime/ cache/ memory/ parallel/ parity/ extend/ program/ request/ training/ spanning design_v3 §4-§13: ModelCard + validate(); driven loops (init/next/advance/finalize); step-interleaving Engine with reservation-before-admission + per-class caches keyed by CacheKey; the C0-C4 consistency ladder + the non-negotiable batch-of-N interleave parity gate. - Inference: Wan2.1-1.3B (T2V), LTX2.3 (two-stage distilled, shared transformer), Wan-causal (chunk rollout + slab-KV streaming). - Training (Wan2.1-1.3B), each driving the SAME loops the engine serves: finetune (flow-match), DMD2 (teacher/critic distribution matching), DiffusionNFT (likelihood-free C2, samples from the decay-blended old policy, group-relative advantages, shared-prompt cache reuse), self-forcing (causal chunk loop). The engine never imports training (the §10 dependency rule, grep-verified). numpy-only core (no torch/GPU here); heavy Wan/LTX forwards are deterministic toy stand-ins with lazy torch-adapter seams (ComponentSpec.load_id/factory) for a GPU box. 49 tests pass via pytest and a zero-dependency runner; interleave parity verified (and a buggy module-global interceptor provably breaks it). Omni-ready spine (ar_decode/chunk_step loop kinds, multi-loop instances, LoopState.extension) for the phase-2 Cosmos3 + vllm-omni omni ports. Run: python3 -m pytest mini_fastvideo/tests/ -q ; python3 -m mini_fastvideo.examples
61 lines
1.7 KiB
Python
61 lines
1.7 KiB
Python
"""Artifact — a named, typed output with provenance (design_v3 §12).
|
|
|
|
"This kills the ``extra['audio']`` pattern — audio carries its sample rate as a
|
|
first-class artifact, not a dict passenger." Every artifact records which program
|
|
node produced it (``producer``), so multi-modal fan-out outputs are traceable.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
from dataclasses import dataclass, field
|
|
|
|
from .._types import TensorLike
|
|
from .modalpart import Modality
|
|
|
|
|
|
@dataclass
|
|
class Artifact:
|
|
name: str # output slot name, e.g. "video", "audio", "text"
|
|
producer: str = "" # program node id that produced it (provenance)
|
|
meta: dict[str, object] = field(default_factory=dict)
|
|
|
|
|
|
@dataclass
|
|
class VideoArtifact(Artifact):
|
|
frames: TensorLike = None # T x H x W x C
|
|
fps: int = 16
|
|
|
|
|
|
@dataclass
|
|
class AudioArtifact(Artifact):
|
|
samples: TensorLike = None
|
|
sample_rate: int = 44100 # first-class, not an extra-dict passenger
|
|
|
|
|
|
@dataclass
|
|
class TextArtifact(Artifact):
|
|
token_ids: tuple[int, ...] = ()
|
|
text: str = ""
|
|
|
|
|
|
@dataclass
|
|
class TensorArtifact(Artifact):
|
|
tensor: TensorLike = None # generic (e.g. action stream)
|
|
|
|
|
|
@dataclass
|
|
class LatentArtifact(Artifact):
|
|
latent: TensorLike = None
|
|
of_modality: Modality = Modality.VIDEO
|
|
|
|
|
|
@dataclass
|
|
class Output:
|
|
"""The collected result of running a program for one request."""
|
|
request_id: str
|
|
artifacts: dict[str, Artifact] = field(default_factory=dict)
|
|
metrics: dict[str, float] = field(default_factory=dict)
|
|
error: str | None = None # set on partial/aborted delivery (failure isolation)
|
|
|
|
def get(self, name: str) -> Artifact | None:
|
|
return self.artifacts.get(name)
|