Will Lin
1fe50c0092
The v2 top level had grown to ~27 dirs + 13 loose files — half of them
single-concept abstraction shells (memory/ 86 LOC, transport/ 222, parity/ 178,
extend/ 226) and half vendored fastvideo code sitting as peers to the actual v2
design. Regroup to mirror README section 3 "Planes & dependency order":
core/ enums+types, card, loop, program, parity, request, parallel
(the model-native contracts; no kernels)
runtime/ + folded-in substrate: cache, memory, transport, extend
(the import graph shows only runtime consumes them; compile/cudagraph
already lived here)
serving/ + deploy/ (products / fleet)
_vendor/ all copied fastvideo: models, layers, attention, configs, distributed,
platforms, api, hooks, logging_utils, third_party + fastvideo_args/
utils/logger/envs/forward_context — internal layout unchanged, still
mirrors upstream for diffing
28 dirs + 13 root files -> 8 dirs + 6 root files. Pure mechanical move: 947
absolute-import paths rewritten (v2.X -> v2.{core,runtime,serving,_vendor}.X),
boundary-anchored so platform/ (native dispatch) and platforms/ (vendored CUDA
detect) no longer collide and neither does hooks/ vs extend/. Deleted the empty
v2/loader/. README section 3 + 16 updated; stale test count corrected
(34 files/216 tests -> 22 files/143 tests).
Validated: `import v2` stays torch-free; `v2/run_tests.py` and `pytest v2/tests/`
-> 143 passed, 2 skipped on the numpy toy backend. (On a GPU box force the toy
backend with CUDA_VISIBLE_DEVICES="" or detect() picks cuda.) No external importer
changed — examples use the re-exported `from v2 import VideoGenerator`.
103 lines
3.6 KiB
Python
103 lines
3.6 KiB
Python
"""Request + Session — the only currency crossing the product boundary.
|
|
|
|
A typed ``Request`` carries ``task`` (declared, never inferred), ``inputs``
|
|
(ModalParts), AR ``sampling`` vs ``diffusion`` params, an ``OutputSpec``, and
|
|
per-node overrides. Products construct Requests; they never reach into the model.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import itertools
|
|
from dataclasses import dataclass, field
|
|
|
|
from v2.core.request.cancel import CancelKind, CancelScope
|
|
from v2.core.request.modalpart import ImagePart, ModalPart, TextPart
|
|
from v2.core.request.params import DiffusionParams, OutputSpec, SamplingParams
|
|
from v2.core.request.streams import Stream
|
|
from v2.core.request.tasks import TaskType
|
|
|
|
_counter = itertools.count(1)
|
|
|
|
|
|
def new_request_id() -> str:
|
|
return f"req-{next(_counter)}"
|
|
|
|
|
|
@dataclass(frozen=True)
|
|
class Request:
|
|
request_id: str
|
|
task: TaskType
|
|
model_id: str # which ModelCard to run
|
|
inputs: tuple[ModalPart, ...] = ()
|
|
sampling: SamplingParams = field(default_factory=SamplingParams)
|
|
diffusion: DiffusionParams = field(default_factory=DiffusionParams)
|
|
outputs: OutputSpec = field(default_factory=OutputSpec)
|
|
node_params: dict[str, dict] = field(default_factory=dict) # per graph-node overrides
|
|
priority: int = 0
|
|
|
|
# --- convenience accessors (do not infer task from these) ---
|
|
def prompt(self) -> str:
|
|
for p in self.inputs:
|
|
if isinstance(p, TextPart):
|
|
return p.text
|
|
return ""
|
|
|
|
def image(self):
|
|
for p in self.inputs:
|
|
if isinstance(p, ImagePart):
|
|
return p
|
|
return None
|
|
|
|
def node_override(self, node_id: str) -> dict:
|
|
return self.node_params.get(node_id, {})
|
|
|
|
|
|
def make_request(task: TaskType,
|
|
model_id: str,
|
|
prompt: str = "",
|
|
*,
|
|
inputs: tuple[ModalPart, ...] | None = None,
|
|
sampling: SamplingParams | None = None,
|
|
diffusion: DiffusionParams | None = None,
|
|
outputs: OutputSpec | None = None,
|
|
node_params: dict[str, dict] | None = None,
|
|
priority: int = 0,
|
|
request_id: str | None = None) -> Request:
|
|
"""Ergonomic constructor used by the offline API and tests."""
|
|
parts: tuple[ModalPart, ...] = inputs if inputs is not None else ((TextPart(prompt), ) if prompt else ())
|
|
return Request(
|
|
request_id=request_id or new_request_id(),
|
|
task=task,
|
|
model_id=model_id,
|
|
inputs=parts,
|
|
sampling=sampling or SamplingParams(),
|
|
diffusion=diffusion or DiffusionParams(),
|
|
outputs=outputs or OutputSpec(),
|
|
node_params=node_params or {},
|
|
priority=priority,
|
|
)
|
|
|
|
|
|
@dataclass
|
|
class Session:
|
|
"""A long-lived interactive context.
|
|
|
|
Holds prompt memory, media streams, a cancel scope, and a handle to
|
|
cross-request chunk-KV that persists for a game/scene session (self-forcing /
|
|
world-model rollout). Mutable by design.
|
|
"""
|
|
session_id: str
|
|
prompt_memory: list[str] = field(default_factory=list)
|
|
streams: dict[str, Stream] = field(default_factory=dict)
|
|
cancel_scope: CancelScope | None = None
|
|
kv_handle: object | None = None # persistent chunk-KV CacheHandle (see cache/)
|
|
|
|
def __post_init__(self):
|
|
if self.cancel_scope is None:
|
|
self.cancel_scope = CancelScope(kind=CancelKind.SESSION, target_id=self.session_id)
|
|
|
|
def push_text(self, text: str) -> None:
|
|
self.prompt_memory.append(text)
|
|
|
|
def stream(self, name: str) -> Stream:
|
|
return self.streams.setdefault(name, Stream(f"{self.session_id}:{name}"))
|