SolitaryThinker
9b8838834c
Wire the capture/replay lifecycle into the driven-loop step boundary — the other half of Path A (hand-fused kernels behind the registry + piecewise cudagraphs, no compiler). Models and tests the correctness-critical control logic; replay re-runs the current step thunk (CPU models the lifecycle, not the GPU speedup). - GraphCapturer on the instance (cross-request cache), wired into RuntimeLoopContext.execute and gated by LoopSpec.graph_capture == "breakable_cudagraph". Capture key = (device, arch, loop, shape_sig, resident-weight-versions, graph_key). Eager-break for non-capturable (SDE) and interceptor-overridden steps. Never executes a stored thunk, so interleaved requests can't smear state. - WorkPlan.capturable / graph_key. wan21 sets capturable=not sde and folds compute dtype (shape_sig.dtype) + CFG branch set + expert + scheduler-precision into the key — closing a cross-precision key-collision corruption path a non-fp32 build would otherwise hit on a real GPU (audit finding). - Version-in-key auto-invalidation + real eviction: set_weights_version evicts the synced component's graphs (duck-typed, so card/ imports no runtime), preventing a GPU graph leak across FlowGRPO's per-iteration syncs. - register_kernel gains the workspace_bytes capture-safety contract (declared in the matrix; cuda cells declare real scratch, numpy reference is 0). - 11 tests: capture-once/replay-many, eager-break (SDE + override), capture ≡ pure-eager bit-identical, recapture on shape + weight-version change with eviction, eager-loop gating, accel-backend capture, and capturer unit tests (key discrimination, eager-break, workspace-collision safety net, eviction). Honestly deferred (bite a real GPU, not the CPU tests): the static-buffer refactor of the step body, admission budgeting of capture cost (GRAPH_CAPTURE), and per-card opt-in beyond wan21 — all documented in cudagraph.py + README. 195 tests pass.
95 lines
4.4 KiB
Python
95 lines
4.4 KiB
Python
"""RuntimeLoopContext — the concrete inversion point (design_v3 §5.1, §11).
|
|
|
|
This is the single seam: ``execute(plan)`` runs the kernel thunk built by the model's
|
|
``next()``, wrapped by interceptors (compute-altering) and observers (read-only). The model
|
|
never sees the scheduler; the scheduler never sees the model's math. Per-request loop state
|
|
is bound here so interceptors read state from ``LoopState`` (never module globals).
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
from time import perf_counter
|
|
from typing import Any
|
|
|
|
from .._enums import ExecutionProfile
|
|
from ..loop.contracts import StepResult, WorkPlan
|
|
from ..request.streams import OmniEvent, Stream, StreamChunk
|
|
|
|
|
|
class RuntimeLoopContext:
|
|
def __init__(self, instance: Any, *, observers, interceptors, slots: dict[str, Any],
|
|
stream: Stream, cancel_scope, profile: ExecutionProfile, metrics: dict,
|
|
request_id: str):
|
|
self.instance = instance
|
|
self.observers = observers
|
|
self.interceptors = interceptors
|
|
self.slots = slots # the program's typed named edges
|
|
self.stream = stream
|
|
self.cancel_scope = cancel_scope
|
|
self.profile = profile
|
|
self.metrics = metrics
|
|
self.request_id = request_id
|
|
self.state = None # bound per step by the LoopRunner
|
|
|
|
def bind_state(self, state) -> None:
|
|
self.state = state
|
|
|
|
def execute(self, plan: WorkPlan) -> StepResult:
|
|
self.observers.emit("step_start", plan=plan)
|
|
# Interceptors may supply an override for the *expensive forward* (e.g. a cached
|
|
# prediction). The step body still runs the cheap scheduler step (combine + solver),
|
|
# so the override changes compute, not the loop's state-transition shape.
|
|
override = None
|
|
if self.interceptors.active and self.state is not None:
|
|
override = self.interceptors.before(plan, self.state)
|
|
# only count a skip the step body will actually honor (must carry a forward result)
|
|
if override is not None and "noise_pred" in override:
|
|
self.metrics["skipped_steps"] = self.metrics.get("skipped_steps", 0) + 1
|
|
|
|
def _eager(ov):
|
|
return plan.run(self.instance, ov) if plan.run is not None else {}
|
|
|
|
t0 = perf_counter()
|
|
out = self._run_step(plan, override, _eager)
|
|
dt = perf_counter() - t0
|
|
if isinstance(out, StepResult):
|
|
result = out
|
|
if result.actual_seconds == 0.0:
|
|
result.actual_seconds = dt
|
|
else:
|
|
result = StepResult(output=dict(out), actual_seconds=dt)
|
|
|
|
if self.interceptors.active and self.state is not None:
|
|
self.interceptors.after(plan, self.state, result)
|
|
|
|
self.observers.emit("step_complete", plan=plan, result=result)
|
|
self.metrics["steps"] = self.metrics.get("steps", 0) + 1
|
|
self.metrics["gpu_seconds"] = self.metrics.get("gpu_seconds", 0.0) + result.actual_seconds
|
|
return result
|
|
|
|
def _run_step(self, plan: WorkPlan, override, eager):
|
|
"""Execute the step body, under piecewise CUDA-graph capture/replay when the loop declares
|
|
graph_capture="breakable_cudagraph" (design_v3 §6.2; Path A). Otherwise plain eager — the
|
|
unchanged path. Replay still runs the current thunk (CPU models the lifecycle, not the
|
|
speedup); see runtime/cudagraph.py."""
|
|
if plan.run is None:
|
|
return {}
|
|
loop = self.instance.card.loops.get(plan.loop_id) if self.instance.card else None
|
|
if getattr(loop, "graph_capture", "eager") != "breakable_cudagraph":
|
|
return eager(override)
|
|
if self.instance.graphs is None:
|
|
from .cudagraph import GraphCapturer
|
|
self.instance.graphs = GraphCapturer()
|
|
return self.instance.graphs.dispatch(plan, self.instance, override, eager)
|
|
|
|
def emit(self, chunk: StreamChunk) -> None:
|
|
self.stream.emit(OmniEvent(type="media.chunk", request_id=self.request_id,
|
|
seq=chunk.seq, payload={"chunk": chunk}))
|
|
self.metrics["stream_chunks"] = self.metrics.get("stream_chunks", 0) + 1
|
|
|
|
def check_cancel(self) -> None:
|
|
if self.cancel_scope is not None:
|
|
self.cancel_scope.check() # raises request.Cancelled at the step boundary
|
|
|
|
def observe(self, event: str, **kw) -> None:
|
|
self.observers.emit(event, **kw)
|