[refactor]: remove training semantics from v2 inference contracts

v2/__init__.py

@@ -1,15 +1,14 @@
-"""v2 — a scoped, CPU-testable realization of the model-native runtime (see v2/README.md).
+"""v2 - the FastVideo inference runtime (see v2/README.md).
 
 > A model card is a (recipe, runtime) pair with a parity obligation.
 > The model owns loop semantics; the runtime owns loop lifecycle.
 > One resident instance runs many loops; one scheduler runs their steps in one currency.
-> Caches are correct by key; parity is correct by test; the interleave gate is non-negotiable.
-> Training records behavior on the same loops it serves.
+> Caches are correct by key; parity is correct by test.
 
-Phase 1 supports Wan2.1-1.3B (T2V) and LTX2.3 (2-stage distilled), plus four training
-methods on Wan2.1-1.3B (finetuning, DMD2, DiffusionNFT, self-forcing). The spine is
-omni-ready (multi-loop ModelInstance, ar_decode/chunk_rollout loop kinds) for the phase-2
-Cosmos3 + vllm-omni omni ports.
+v2 is inference-only. Training, finetuning, distillation, RL, and optimizer
+loops belong to ``fastvideo/train`` or ``fastvideo/training``. v2 only records
+checkpoint provenance in recipe metadata so inference can bind weights to the
+right loop and precision policy.
 
 The core is numpy-only and CPU-testable; heavy Wan/LTX neural forwards become lazy torch
 adapters (see ``v2/platform/backends/``) that are off the test path.

v2/core/card/specs.py

@@ -21,8 +21,6 @@ if TYPE_CHECKING:  # avoid runtime card -> parallel coupling
     pass
 
 
-
-
 # --------------------------------------------------------------------------- #
 # Precision, parallelism, parity, cache, checkpoint contracts                 #
 # --------------------------------------------------------------------------- #
@@ -31,7 +29,6 @@ class PrecisionContract:
     default_dtype: str = "float32"
     component_overrides: dict[str, str] = field(default_factory=dict)
     quantization_scheme: str | None = None  # "nvfp4" | "int8" | None
-    training_precision: str = "float32"  # distinct from serving precision
 
     def dtype_for(self, component_id: str) -> str:
         return self.component_overrides.get(component_id, self.default_dtype)
@@ -52,7 +49,7 @@ class CacheContract:
     eviction: str = "lru"  # "lru" | "fifo" | "none"
     reuse_across_requests: bool = True  # paged/feature=True; slab depends on mode
     per_component: dict[str, int] = field(default_factory=dict)
-    training_mode_disables_recycle: bool = False  # chunk-KV training mode
+    disable_recycle: bool = False  # keep full causal/world context instead of sliding-window recycle
 
 
 @dataclass
@@ -133,7 +130,7 @@ class LoopSpec:
     loop_id: str
     kind: LoopKind
     work_unit_kind: WorkUnitKind
-    behavior_schema: type | None = None  # what to capture for RL (None if not training-relevant)
+    behavior_schema: type | None = None  # optional inference trajectory/debug payload schema
     extension_schema: type | None = None  # per-model LoopState extension (Cosmos3PackedSeq, etc.)
     cache_policy: list[str] = field(default_factory=list)  # cache class names this loop draws from
     graph_capture: str = "eager"  # eager | breakable_cudagraph

v2/core/enums.py

@@ -17,7 +17,6 @@ class LoopKind(str, Enum):
     VAE_TILE = "vae_tile"  # tiled VAE encode/decode
     ENCODER = "encoder"  # one-shot encoder (text/vision) — degenerate single-step loop
     AUDIO_DECODE = "audio_decode"  # vocoder / codec decode
-    TRAIN_FORWARD = "train_forward"  # grad-enabled forward for a training method
 
 
 class WorkUnitKind(str, Enum):
@@ -34,18 +33,16 @@ class WorkUnitKind(str, Enum):
     ENCODER_CHUNK = "encoder_chunk"
     VAE_TILE = "vae_tile"
     AUDIO_CHUNK = "audio_chunk"
-    REWARD_BATCH = "reward_batch"  # RL
-    LOGPROB_BATCH = "logprob_batch"  # RL (likelihood-based)
     TRANSFER = "transfer"
     CACHE_IO = "cache_io"
     GRAPH_CAPTURE = "graph_capture"
 
 
 class ConsistencyLevel(str, Enum):
-    """The consistency ladder. RL methods declare their required rung."""
+    """The inference parity ladder."""
     C0 = "C0"  # component parity (VAE/encoder/block/scheduler-step in isolation)
     C1 = "C1"  # loop parity (full denoise trajectory / AR logits, fixed seed)
-    C2 = "C2"  # behavioral identity (train-forward ≡ serve-forward on the RL objective's quantity)
+    C2 = "C2"  # behavioral identity for inference-visible trajectories
     C3 = "C3"  # distribution parity (rollout distribution under allowed nondeterminism)
     C4 = "C4"  # artifact quality (SSIM / reward agreement / human preference)
 
@@ -55,10 +52,9 @@ class ConsistencyLevel(str, Enum):
 
 
 class ExecutionProfile(str, Enum):
-    """Three forwards, one loop definition — differ only in grad mode + capture."""
+    """Inference execution profiles."""
     SERVE = "serve"  # no-grad, graphed, cached, possibly quantized
-    ROLLOUT = "rollout"  # serve profile + behavior capture
-    TRAIN = "train"  # grad, checkpointed, FSDP-gathered
+    ROLLOUT = "rollout"  # serve profile + trajectory capture / stochastic rollout
 
 
 class Capability(str, Enum):
@@ -75,5 +71,3 @@ class Capability(str, Enum):
     STREAMING_VIDEO_CONTINUATION = "streaming_video_continuation"
     VAE_ENCODE = "vae_encode"
     VAE_DECODE = "vae_decode"
-    POLICY_ROLLOUT = "policy_rollout"  # can serve as an RL rollout engine
-    LOGPROB_RECOMPUTE = "logprob_recompute"

v2/core/loop/sampler.py

@@ -47,13 +47,11 @@ def flow_sde_step_with_logprob(x_t,
                                noise=None,
                                prev_sample=None,
                                noise_scale: float = 0.7):
-    """FlowGRPO-style stochastic step + Gaussian log-prob.
+    """Stochastic diffusion step + Gaussian log-prob metadata.
 
-    The RL *rollout* sampler (vs the deterministic ODE ``flow_match_euler_step`` used at serve time).
-    Injects noise for exploration and returns the log-prob of the realized sample under the step's
-    Gaussian — the quantity the FlowGRPO PPO ratio uses. ``prev_sample=None`` samples a new step;
-    passing ``prev_sample`` (the rollout sample) recomputes its log-prob under the *current* velocity
-    (the ratio's numerator at update time). Returns ``(prev_sample, log_prob, mean, eff_std)``.
+    This is the stochastic counterpart to deterministic ``flow_match_euler_step``. ``prev_sample=None``
+    samples a new step; passing ``prev_sample`` recomputes its log-prob under the current velocity.
+    Returns ``(prev_sample, log_prob, mean, eff_std)``.
     """
     x_t = np.asarray(x_t, dtype=np.float64)
     velocity = np.asarray(velocity, dtype=np.float64)

v2/core/parity/ladder.py

@@ -1,10 +1,9 @@
 """The consistency ladder + numeric comparison helpers.
 
-C0 component · C1 loop · C2 behavioral · C3 distribution · C4 artifact.
+C0 component · C1 loop · C2 trajectory behavior · C3 distribution · C4 artifact.
 
-The C2 split is load-bearing: likelihood-based methods compare per-step log-probs;
-likelihood-free methods (DiffusionNFT) compare seeded final-sample + prediction-space
-identity (old_deviate / ref-MSE) — there are NO log-probs to match.
+The ladder is for inference parity. C0/C1 are component and loop parity, C2 can compare
+captured trajectory records, C3 covers stochastic distribution parity, and C4 covers artifact quality.
 """
 from __future__ import annotations
 

v2/core/program/workflow.py

@@ -132,8 +132,7 @@ class ParallelWorkflow(Workflow):
 
 class BestOfNWorkflow(Workflow):
     """Inference-time scaling / rejection sampling: generate N candidates (varying seed), score each with
-    a reward scorer, return the best. A feedback loop across models — generator + reward (e.g. the served
-    REWARD_BATCH card) — the other non-linear shape."""
+    a scorer, return the best. This is a feedback loop across servable models, not a training loop."""
 
     def __init__(self, workflow_id, generator_stage, *, scorer, n: int = 4, score_key: str = "latents"):
         super().__init__(workflow_id, [generator_stage])

v2/core/request/params.py

@@ -35,8 +35,8 @@ class DiffusionParams:
     num_frames: int = 81
     fps: int = 16
     seed: int | None = None
-    # FlowGRPO-style RL rollout: use the SDE sampler (with per-step log-prob capture) instead of the
-    # deterministic ODE serve sampler. Default False ⇒ the serve path is unchanged.
+    # Stochastic sampling: use the SDE sampler instead of the deterministic ODE serve sampler.
+    # Default False means the normal serve path is unchanged.
     sde_rollout: bool = False
     sde_noise_scale: float = 0.7
     # Active adapters for this request (LoRA / ControlNet ids on the base).
@@ -47,7 +47,7 @@ class DiffusionParams:
 class CaptureMode(str, Enum):
     """What the loop captures while running."""
     NONE = "none"  # serve forward: no-grad, no capture
-    BEHAVIOR = "behavior"  # rollout forward: emit BehaviorRecord slices for RL/distill
+    BEHAVIOR = "behavior"  # trajectory/debug capture for inference diagnostics
 
 
 @dataclass(frozen=True)

v2/core/types.py

@@ -1,6 +1,6 @@
 """Shared lightweight type aliases for v2.
 
-The core (card/loop/runtime/cache/parity/program/request/training) is numpy-only
+The core (card/loop/runtime/cache/parity/program/request) is numpy-only
 and CPU-testable, so tensors are typed structurally as ``TensorLike``: a
 ``numpy.ndarray`` on the CPU test path, a ``torch.Tensor`` on GPU. The core never
 imports torch — only model-component adapters do, lazily (see

v2/examples.py

@@ -11,7 +11,6 @@ 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
-from v2.training import build_diffusion_nft
 
 
 def _t2v(mid: str, prompt: str, seed: int, steps: int = 4, **kw: Any) -> Request:
@@ -40,17 +39,6 @@ def example_c_causal_streaming(eng) -> None:
           f"streamed_chunks={out.metrics.get('stream_chunks', 0)}")
 
 
-def example_d_rl_rollout() -> None:
-    print("\n(d) RL rollout (DiffusionNFT): the SAME denoise loop + behavior capture (train ≡ serve)")
-    nft = build_diffusion_nft(build_wan21_card(), num_video_per_prompt=4, num_inner_timesteps=2)
-    loss, m = nft.managed_train_step({"prompts": ["a red car", "a blue boat"], "seeds": [1, 2]}, 0)
-    fc = nft.old.caches.stats()["feature"]
-    print(f"    policy_loss={loss['policy_loss']:.3f} kl={loss['kl_div_loss']:.5f} "
-          f"reward_mean={m['reward_mean']:.3f}  consistency={nft.consistency_level().value} (likelihood-free)")
-    print(f"    shared-prompt feature-cache reuse: {fc['hits']} hits / {fc['misses']} misses "
-          f"(K samples encode the prompt once — the 24× reduction)")
-
-
 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()
@@ -78,7 +66,7 @@ async def _serving_demo() -> None:
     import asyncio
 
     from v2.serving.deploy import DynamoWorkerAdapter, FakeDynamoRuntime, LocalFleet, build_deployment_card
-    from v2.recipes.wan21 import build_wan21_card, build_wan_t2v_program
+    from v2.recipes.wan21 import build_wan_t2v_program
     from v2.runtime import AsyncEngine, PoolSet, wan_t2v_disaggregated
     from v2.serving import OmniOpenAIServer
 
@@ -143,12 +131,11 @@ def main() -> None:
     example_a_text_to_video(eng)
     example_b_ltx2_two_stage(eng)
     example_c_causal_streaming(eng)
-    example_d_rl_rollout()
     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, training-on-shared-loops) is real; the neural")
+    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)
 

v2/platform/__init__.py

@@ -3,7 +3,7 @@
 Two tuple-keyed registries (``COMPONENTS``, ``KERNELS``) + a ``Platform`` that detects ``(device,
 arch)`` and resolves both, with the numpy reference as the terminal fallback rung and parity oracle.
 Lets CPU, GPU, and other backends coexist: a model's loops, policies, scheduler, caches, and
-training code never name a device; they call through the component/kernel seams and the resolved
+inference code never names a device; it calls through the component/kernel seams and the resolved
 ``Platform`` decides the implementation.
 
 Importing this package is cheap and cycle-free (registry + platform classes only); the concrete

v2/platform/backends/accel.py

@@ -45,7 +45,7 @@ class AccelDiT:
         return self._inner(*args, **kwargs)
 
     def __getattr__(self, name: str):
-        # delegate the trainable surface (clone/copy_from/blend_from/mse_grad_step) to the inner toy
+        # delegate the weight/version surface (clone/copy_from/blend_from) to the inner toy
         return getattr(self._inner, name)
 
 

v2/platform/backends/toy.py

@@ -2,8 +2,8 @@
 
 There is no GPU/torch/weights in this environment, so the heavy Wan/LTX neural forwards are
 represented by small, deterministic numpy components. They exercise the *real* loop control
-flow, CFG/flow-shift policies, scheduler steps, cache reuse, parity gates, and training-method
-math with real numbers — just with a toy network instead of a 1.3B DiT.
+flow, CFG/flow-shift policies, scheduler steps, cache reuse, and parity gates with real numbers -
+just with a toy network instead of a 1.3B DiT.
 
 The contract these toys honor is what matters:
   * deterministic given (weights, inputs) → bit-reproducible (the interleave/parity gates rely on this);
@@ -12,8 +12,8 @@ The contract these toys honor is what matters:
 
 On a GPU box, ``ComponentSpec.factory`` swaps these for lazy torch adapters wrapping the real
 ``fastvideo.models`` modules + weights (see each model's ``components.py``); the loops, policies,
-scheduler, caches, parity, and training code are unchanged. That is the whole point of the
-(recipe, runtime) separation.
+scheduler, caches, and parity are unchanged. That is the whole point of the (recipe, runtime)
+separation.
 """
 from __future__ import annotations
 
@@ -105,7 +105,7 @@ class ToyDiT:
         audio = np.tanh(0.9 * au + 0.1 * float(sigma) + 0.3 * a_cond).astype("float32")
         return video, audio
 
-    # --- minimal trainable surface (so training methods do real optimizer steps) --------- #
+    # --- weight/version surface for serving hot-reload tests ----------------------------- #
     def clone(self) -> ToyDiT:
         c = ToyDiT.__new__(ToyDiT)
         c.C, c.s_text, c.s_ctx = self.C, self.s_text, self.s_ctx
@@ -113,40 +113,13 @@ class ToyDiT:
         return c
 
     def blend_from(self, other: ToyDiT, decay: float) -> None:
-        """EMA / decay-blended-old-policy update: self ← decay·self + (1-decay)·other."""
+        """Blend resident weights toward another component version."""
         self.w_x = (decay * self.w_x + (1.0 - decay) * other.w_x).astype("float32")
         self.w_t = (decay * self.w_t + (1.0 - decay) * other.w_t).astype("float32")
 
     def copy_from(self, other: ToyDiT) -> None:
         self.w_x, self.w_t = other.w_x.copy(), other.w_t.copy()
 
-    def mse_grad_step(self,
-                      latent: np.ndarray,
-                      text_embed,
-                      sigma: float,
-                      target: np.ndarray,
-                      lr: float,
-                      context=None,
-                      weight: float = 1.0) -> tuple[float, float]:
-        """One exact SGD step minimizing ``weight·MSE(forward(latent), target)`` w.r.t. ``w_x``.
-
-        Returns (loss, grad_norm). The grad_norm is the #1396-style per-method regression signal.
-        This is a *real* gradient (chain rule through tanh), so the toy student actually learns —
-        the structural stand-in for the FSDP/optimizer step a GPU trainer would run."""
-        latent = np.asarray(latent, dtype=np.float32)
-        target = np.asarray(target, dtype=np.float32)
-        z = self._pre_tanh(latent, text_embed, sigma, context)
-        pred = np.tanh(z)
-        err = pred - target
-        loss = float(weight * np.mean(err**2))
-        n = err.size
-        g = weight * (2.0 * err / n) * (1.0 - pred**2)
-        axes = (list(range(1, g.ndim)), list(range(1, latent.ndim)))
-        grad_wx = np.tensordot(g, latent, axes=axes)  # [C_out, C_in]
-        grad_norm = float(np.linalg.norm(grad_wx))
-        self.w_x = (self.w_x - lr * grad_wx).astype("float32")
-        return loss, grad_norm
-
 
 class ToyTokenizer:
     """Deterministic byte-ish tokenizer for the omni AR pathway (phase 2)."""
@@ -185,82 +158,6 @@ class ToyMoTDiT(ToyDiT):
         return (rng.standard_normal((4, 8)) * 0.1).astype("float32")
 
 
-class ToyPromptRefiner:
-    """Toy LM prompt-refiner — UniRL/PromptRL's Qwen role (a *separate expert*, not MoT-shared).
-
-    A categorical policy over a small set of refinement "actions" (stand-in for the LM emitting a
-    refined prompt). Sampling an action picks an embedding offset that shifts the diffusion
-    conditioning; the realized reward then drives the policy by REINFORCE. This is the *second
-    trainable expert* in the unified-RL stress test: one reward → a token-policy-gradient update
-    here AND a FlowGRPO update on the DiT, under one advantage (PromptRL §6).
-
-    Real gradient: loss = −A·logπ(a); ∇_logits = −A·(onehot(a) − softmax). So the toy LM actually
-    learns to prefer the reward-favored action — the LM-side analogue of ToyDiT.mse_grad_step.
-    """
-
-    def __init__(self, n_actions: int = 8, dim: int = TEXT_DIM, seq: int = TEXT_SEQ, seed: int = 0):
-        self.n = int(n_actions)
-        self.dim, self.seq = int(dim), int(seq)
-        self.logits: np.ndarray = np.zeros(self.n, dtype="float64")  # the trainable policy params
-        rng = np.random.default_rng(seed)
-        # each action ⇒ a fixed conditioning offset (the "content" of that refined prompt)
-        self._emb = (rng.standard_normal((self.n, self.seq, self.dim)) * 0.1).astype("float32")
-
-    def _probs(self) -> np.ndarray:
-        z = self.logits - self.logits.max()
-        e = np.exp(z)
-        return e / e.sum()
-
-    def sample_refinement(self, rng) -> tuple[int, float]:
-        """Sample an action (exploration) and return (action, logπ(action))."""
-        p = self._probs()
-        a = int(rng.choice(self.n, p=p))
-        return a, float(np.log(p[a] + 1e-12))
-
-    def argmax_refinement(self) -> tuple[int, float]:
-        a = int(np.argmax(self.logits))
-        return a, float(np.log(self._probs()[a] + 1e-12))
-
-    def action_logprob(self, action: int) -> float:
-        return float(np.log(self._probs()[int(action)] + 1e-12))
-
-    def ar_forward(self, tokens) -> int:
-        """und-pathway hook so the declared ``ar_decode`` loop is runnable (greedy ⇒ interleave-safe)."""
-        return int(np.argmax(self.logits))
-
-    def refined_embed(self, base_embed, action: int) -> np.ndarray:
-        """Apply the chosen refinement to the base text embedding → the diffusion conditioning."""
-        base = np.zeros((self.seq, self.dim), dtype="float32") if base_embed is None \
-            else np.asarray(base_embed, dtype="float32")
-        return (base + self._emb[int(action)]).astype("float32")
-
-    def pg_step(self, action: int, advantage: float, lr: float) -> float:
-        """REINFORCE on the categorical policy. Returns grad_norm (the per-method regression signal)."""
-        p = self._probs()
-        onehot: np.ndarray = np.zeros(self.n, dtype="float64")
-        onehot[int(action)] = 1.0
-        grad = -float(advantage) * (onehot - p)  # ∇_logits (−A·logπ(a))
-        self.logits = self.logits - float(lr) * grad  # descent on −A·logπ ⇒ ascent on A·logπ
-        return float(np.linalg.norm(grad))
-
-    def kl_to(self, other: ToyPromptRefiner) -> float:
-        """KL(π_self ‖ π_other) over the categorical — the LM-PG reference-KL penalty term."""
-        p, q = self._probs(), other._probs()
-        return float(np.sum(p * (np.log(p + 1e-12) - np.log(q + 1e-12))))
-
-    def clone(self) -> ToyPromptRefiner:
-        c = ToyPromptRefiner.__new__(ToyPromptRefiner)
-        c.n, c.dim, c.seq = self.n, self.dim, self.seq
-        c.logits, c._emb = self.logits.copy(), self._emb.copy()
-        return c
-
-    def copy_from(self, other: ToyPromptRefiner) -> None:
-        self.logits = other.logits.copy()
-
-    def blend_from(self, other: ToyPromptRefiner, decay: float) -> None:
-        self.logits = decay * self.logits + (1.0 - decay) * other.logits
-
-
 class ToyTalker(ToyMoTDiT):
     """Toy Talker — Qwen-Omni stage 1 (AR over a speech-codec vocab, conditioned on the Thinker).
 

v2/platform/registry.py

@@ -40,7 +40,7 @@ from collections.abc import Callable, Iterable
 # Op names — the stable identifiers a loop dispatches on (not free functions). #
 # --------------------------------------------------------------------------- #
 FLOW_MATCH_STEP = "flow_match_step"  # deterministic flow-match Euler solver step (ODE serve)
-FLOW_SDE_STEP = "flow_sde_step"  # FlowGRPO stochastic step + log-prob (RL rollout)
+FLOW_SDE_STEP = "flow_sde_step"  # stochastic diffusion step + log-prob metadata
 
 
 def _always_available() -> bool:

v2/recipes/bagel/card.py

@@ -38,12 +38,13 @@ def build_bagel_card(model_id: str = "bagel-mot") -> ModelCard:
         return ARDecodeLoop(loop_id="generate_text", transformer_id="transformer", max_tokens=6)
 
     def image_factory():
-        return WanDenoiseLoop(loop_id="generate_image",
-                              cfg=cfg,
-                              flow_shift=flow,
-                              precision=precision,
-                              expert=expert,
-                              )
+        return WanDenoiseLoop(
+            loop_id="generate_image",
+            cfg=cfg,
+            flow_shift=flow,
+            precision=precision,
+            expert=expert,
+        )
 
     components = {
         "tokenizer":
@@ -93,7 +94,7 @@ def build_bagel_card(model_id: str = "bagel-mot") -> ModelCard:
             "feature": CacheContract("feature", max_bytes=1 << 24, reuse_across_requests=True),
             "paged_kv": CacheContract("paged_kv", max_bytes=1 << 24, block_bytes=1 << 12, reuse_across_requests=False),
         },
-        precision=PrecisionContract(default_dtype="float32", training_precision="float32"),
+        precision=PrecisionContract(default_dtype="float32"),
         parallelism=ParallelismContract(valid_plans=[ParallelPlan.single()], default_plan=ParallelPlan.single()),
     )
     return card.validate()

v2/recipes/common.py

@@ -1,7 +1,6 @@
 """Shared component-node helpers (text encode with feature cache, etc.).
 
-The content-hash feature cache lets a K-sample RL group encode its shared prompt once instead of
-K times.
+The content-hash feature cache lets repeated requests reuse prompt embeddings safely.
 """
 from __future__ import annotations
 

v2/recipes/cosmos3/card.py

@@ -37,12 +37,13 @@ def build_cosmos3_card(model_id: str = "cosmos3-vfm") -> ModelCard:
         return ARDecodeLoop(loop_id="ar_decode", transformer_id="transformer", max_tokens=6)
 
     def dn_factory():
-        return WanDenoiseLoop(loop_id="diffusion_denoise",
-                              cfg=cfg,
-                              flow_shift=flow,
-                              precision=precision,
-                              expert=expert,
-                              )
+        return WanDenoiseLoop(
+            loop_id="diffusion_denoise",
+            cfg=cfg,
+            flow_shift=flow,
+            precision=precision,
+            expert=expert,
+        )
 
     components = {
         "tokenizer":
@@ -89,8 +90,7 @@ def build_cosmos3_card(model_id: str = "cosmos3-vfm") -> ModelCard:
         components=components,
         loops=loops,
         capabilities=CapabilityMatrix.of(Capability.TEXT_TO_VIDEO, Capability.REASONING_TEXT,
-                                         Capability.TEXT_TO_VIDEO_SOUND, Capability.VAE_DECODE,
-                                         Capability.POLICY_ROLLOUT),
+                                         Capability.TEXT_TO_VIDEO_SOUND, Capability.VAE_DECODE),
         recipe=RecipeSpec(method="base",
                           assumes_loop="diffusion_denoise",
                           assumes_precision="float32",
@@ -100,7 +100,7 @@ def build_cosmos3_card(model_id: str = "cosmos3-vfm") -> ModelCard:
             "feature": CacheContract("feature", max_bytes=1 << 24, reuse_across_requests=True),
             "paged_kv": CacheContract("paged_kv", max_bytes=1 << 24, block_bytes=1 << 12, reuse_across_requests=False),
         },
-        precision=PrecisionContract(default_dtype="float32", training_precision="float32"),
+        precision=PrecisionContract(default_dtype="float32"),
         parallelism=ParallelismContract(valid_plans=[ParallelPlan.single()], default_plan=ParallelPlan.single()),
     )
     return card.validate()

v2/recipes/flux2/card.py

@@ -107,7 +107,7 @@ def _build_flux2_card(model_id: str, *, text_encoder_load_id: str, text_encoder_
         parity=ParitySpec(consistency_levels=[ConsistencyLevel.C1],
                           tests=[ParityTestSpec(name="denoise_trajectory", level=ConsistencyLevel.C1, tap="latents")]),
         caches={"feature": CacheContract(cache_class="feature", max_bytes=1 << 24, reuse_across_requests=True)},
-        precision=PrecisionContract(default_dtype="float32", training_precision="float32"),
+        precision=PrecisionContract(default_dtype="float32"),
         parallelism=ParallelismContract(valid_plans=[ParallelPlan.single()], default_plan=ParallelPlan.single()),
         sampling_defaults=sampling_defaults,
     )

v2/recipes/ltx2/card.py

@@ -135,7 +135,7 @@ def build_ltx2_card(model_id: str = "ltx2-2stage-distilled") -> ModelCard:
             negative_prompt=""),
         parity=ParitySpec(consistency_levels=[ConsistencyLevel.C1]),
         caches={"feature": CacheContract(cache_class="feature", max_bytes=1 << 24, reuse_across_requests=True)},
-        precision=PrecisionContract(default_dtype="float32", training_precision="float32"),
+        precision=PrecisionContract(default_dtype="float32"),
         parallelism=ParallelismContract(valid_plans=[ParallelPlan.single()], default_plan=ParallelPlan.single()),
     )
     return card.validate()
@@ -211,7 +211,7 @@ def build_ltx2_base_card(model_id: str = "ltx2-single-stage") -> ModelCard:
                                            negative_prompt=LTX2_NEG),
         parity=ParitySpec(consistency_levels=[ConsistencyLevel.C1]),
         caches={"feature": CacheContract(cache_class="feature", max_bytes=1 << 24, reuse_across_requests=True)},
-        precision=PrecisionContract(default_dtype="float32", training_precision="float32"),
+        precision=PrecisionContract(default_dtype="float32"),
         parallelism=ParallelismContract(valid_plans=[ParallelPlan.single()], default_plan=ParallelPlan.single()),
     )
     return card.validate()
@@ -301,7 +301,7 @@ def build_ltx2_3_card(model_id: str = "ltx2.3-distilled") -> ModelCard:
             }),
         parity=ParitySpec(consistency_levels=[ConsistencyLevel.C1]),
         caches={"feature": CacheContract(cache_class="feature", max_bytes=1 << 24, reuse_across_requests=True)},
-        precision=PrecisionContract(default_dtype="float32", training_precision="float32"),
+        precision=PrecisionContract(default_dtype="float32"),
         parallelism=ParallelismContract(valid_plans=[ParallelPlan.single()], default_plan=ParallelPlan.single()),
     )
     return card.validate()

v2/recipes/matrixgame2/card.py

@@ -127,7 +127,7 @@ def build_matrixgame2_card(model_id: str = "matrix-game-2.0-base-distilled",
         parity=ParitySpec(consistency_levels=[ConsistencyLevel.C1],
                           tests=[ParityTestSpec(name="denoise_trajectory", level=ConsistencyLevel.C1, tap="latents")]),
         caches={"feature": CacheContract(cache_class="feature", max_bytes=1 << 24, reuse_across_requests=True)},
-        precision=PrecisionContract(default_dtype="float32", training_precision="float32"),
+        precision=PrecisionContract(default_dtype="float32"),
         parallelism=ParallelismContract(valid_plans=[ParallelPlan.single()], default_plan=ParallelPlan.single()),
         sampling_defaults=sampling_defaults or SamplingDefaults(num_steps=3,
                                                                 guidance_scale=1.0,

v2/recipes/qwen_omni/card.py

@@ -8,7 +8,7 @@ talker token, and vocoder chunk is a runtime-visible WorkUnit.
 A third weight-sharing topology, in the same Card/Loop/Program vocabulary:
 
   * Cosmos3/BAGEL — ONE shared MoT module, two loops (weight sharing within a request).
-  * unified/UniRL — TWO disjoint experts, two loops, jointly RL-trained.
+  * multi-expert inference — TWO disjoint experts, two loops, one request.
   * Qwen-Omni    — THREE disjoint experts (thinker, talker, vocoder), three loops, cascaded:
                    the talker conditions on the thinker's tokens+hidden state, the vocoder on the
                    talker's speech tokens; streaming codec→waveform.
@@ -51,10 +51,7 @@ def build_qwen_omni_card(model_id: str = "qwen-omni-tts") -> ModelCard:
                             prompt_slot="talker_prompt_tokens")
 
     def vocoder_factory():
-        return VocoderLoop(loop_id="vocoder",
-                           vocoder_id="vocoder",
-                           chunk_tokens=2,
-                           speech_slot="speech_tokens")
+        return VocoderLoop(loop_id="vocoder", vocoder_id="vocoder", chunk_tokens=2, speech_slot="speech_tokens")
 
     components = {
         "tokenizer":
@@ -115,8 +112,7 @@ def build_qwen_omni_card(model_id: str = "qwen-omni-tts") -> ModelCard:
         family="qwen_omni",
         components=components,
         loops=loops,
-        capabilities=CapabilityMatrix.of(Capability.REASONING_TEXT, Capability.TEXT_TO_SPEECH,
-                                         Capability.POLICY_ROLLOUT),
+        capabilities=CapabilityMatrix.of(Capability.REASONING_TEXT, Capability.TEXT_TO_SPEECH),
         recipe=RecipeSpec(method="base",
                           assumes_loop="vocoder",
                           assumes_precision="float32",
@@ -125,7 +121,7 @@ def build_qwen_omni_card(model_id: str = "qwen-omni-tts") -> ModelCard:
         caches={
             "paged_kv": CacheContract("paged_kv", max_bytes=1 << 24, block_bytes=1 << 12, reuse_across_requests=False),
         },
-        precision=PrecisionContract(default_dtype="float32", training_precision="float32"),
+        precision=PrecisionContract(default_dtype="float32"),
         parallelism=ParallelismContract(valid_plans=[ParallelPlan.single()], default_plan=ParallelPlan.single()),
     )
     return card.validate()

v2/recipes/wan21/i2v.py

@@ -53,12 +53,13 @@ def build_wan21_i2v_card(model_id: str = "wan2.1-i2v-1.3b",
     precision, expert = PrecisionPolicy(compute_dtype="float32", scheduler_step_in_fp32=True), NoRouting("transformer")
 
     def loop_factory():
-        return WanDenoiseLoop(loop_id="i2v_denoise",
-                              cfg=cfg,
-                              flow_shift=flow,
-                              precision=precision,
-                              expert=expert,
-                              )
+        return WanDenoiseLoop(
+            loop_id="i2v_denoise",
+            cfg=cfg,
+            flow_shift=flow,
+            precision=precision,
+            expert=expert,
+        )
 
     components = {
         "text_encoder":
@@ -109,7 +110,7 @@ def build_wan21_i2v_card(model_id: str = "wan2.1-i2v-1.3b",
                           consistency_required=ConsistencyLevel.C1),
         parity=ParitySpec(consistency_levels=[ConsistencyLevel.C1]),
         caches={"feature": CacheContract("feature", max_bytes=1 << 24, reuse_across_requests=True)},
-        precision=PrecisionContract(default_dtype="float32", training_precision="float32"),
+        precision=PrecisionContract(default_dtype="float32"),
         parallelism=ParallelismContract(valid_plans=[ParallelPlan.single()], default_plan=ParallelPlan.single()),
         sampling_defaults=SamplingDefaults(num_steps=40,
                                            guidance_scale=5.0,
@@ -138,12 +139,13 @@ def build_wan22_i2v_a14b_card(model_id: str = "wan2.2-i2v-a14b",
     expert = BoundaryTimestepRouting(high_noise="transformer", low_noise="transformer_2", boundary=boundary)
 
     def loop_factory():
-        return WanDenoiseLoop(loop_id="i2v_denoise",
-                              cfg=cfg,
-                              flow_shift=flow,
-                              precision=precision,
-                              expert=expert,
-                              )
+        return WanDenoiseLoop(
+            loop_id="i2v_denoise",
+            cfg=cfg,
+            flow_shift=flow,
+            precision=precision,
+            expert=expert,
+        )
 
     def _dit(cid: str) -> ComponentSpec:
         return ComponentSpec(cid,
@@ -198,7 +200,7 @@ def build_wan22_i2v_a14b_card(model_id: str = "wan2.2-i2v-a14b",
                           consistency_required=ConsistencyLevel.C1),
         parity=ParitySpec(consistency_levels=[ConsistencyLevel.C1]),
         caches={"feature": CacheContract("feature", max_bytes=1 << 24, reuse_across_requests=True)},
-        precision=PrecisionContract(default_dtype="float32", training_precision="float32"),
+        precision=PrecisionContract(default_dtype="float32"),
         parallelism=ParallelismContract(valid_plans=[ParallelPlan.single()], default_plan=ParallelPlan.single()),
         sampling_defaults=SamplingDefaults(num_steps=40,
                                            guidance_scale=3.5,

v2/recipes/wan21/program.py

@@ -1,7 +1,7 @@
-"""Wan2.1 T2V program: text_encode → diffusion_denoise → vae_decode.
+"""Wan2.1 T2V program: text_encode -> diffusion_denoise -> vae_decode.
 
-The text-encode node uses the content-hash feature cache: the same prompt across requests (e.g. a
-K=24 RL group) encodes once and reuses.
+The text-encode node uses the content-hash feature cache: the same prompt across requests encodes
+once and reuses.
 """
 from __future__ import annotations
 

v2/recipes/wan_causal/card.py

@@ -32,7 +32,7 @@ def build_wan_causal_card(model_id: str = "wan-causal-sf-1.3b",
                           num_chunks: int = 7,
                           chunk_size: int = 3,
                           steps_per_chunk: int = 4,
-                          training_mode: bool = False) -> ModelCard:
+                          preserve_full_context: bool = False) -> ModelCard:
     # Self-forcing DMD reference (fastvideo SelfForcingWanT2V480PConfig + causal_denoising.py): the
     # distilled student is CFG-FREE (guidance 1.0, single forward/step) and denoises with the 4-step DMD
     # schedule (dmd_denoising_steps=[1000,750,500,250], warped — reproduced by FlowShiftPolicy(5.0) over 4
@@ -44,14 +44,15 @@ def build_wan_causal_card(model_id: str = "wan-causal-sf-1.3b",
     precision = PrecisionPolicy(compute_dtype="float32", scheduler_step_in_fp32=True)
 
     def loop_factory():
-        return ChunkRolloutLoop(loop_id="chunk_rollout",
-                                num_chunks=num_chunks,
-                                chunk_size=chunk_size,
-                                steps_per_chunk=steps_per_chunk,
-                                cfg=cfg,
-                                flow_shift=flow,
-                                precision=precision,
-                                )
+        return ChunkRolloutLoop(
+            loop_id="chunk_rollout",
+            num_chunks=num_chunks,
+            chunk_size=chunk_size,
+            steps_per_chunk=steps_per_chunk,
+            cfg=cfg,
+            flow_shift=flow,
+            precision=precision,
+        )
 
     components = {
         "text_encoder":
@@ -89,7 +90,7 @@ def build_wan_causal_card(model_id: str = "wan-causal-sf-1.3b",
         components=components,
         loops=loops,
         capabilities=CapabilityMatrix.of(Capability.TEXT_TO_VIDEO, Capability.STREAMING_VIDEO_CONTINUATION,
-                                         Capability.VAE_DECODE, Capability.POLICY_ROLLOUT),
+                                         Capability.VAE_DECODE),
         recipe=RecipeSpec(method="self_forcing",
                           parents=["wan2.1-1.3b"],
                           assumes_loop="chunk_rollout",
@@ -103,10 +104,10 @@ def build_wan_causal_card(model_id: str = "wan-causal-sf-1.3b",
             CacheContract(cache_class="slab_kv",
                           max_bytes=1 << 26,
                           reuse_across_requests=False,
-                          per_component={"window": (num_chunks if training_mode else 2)},
-                          training_mode_disables_recycle=training_mode),
+                          per_component={"window": (num_chunks if preserve_full_context else 2)},
+                          disable_recycle=preserve_full_context),
         },
-        precision=PrecisionContract(default_dtype="float32", training_precision="float32"),
+        precision=PrecisionContract(default_dtype="float32"),
         parallelism=ParallelismContract(valid_plans=[ParallelPlan.single()], default_plan=ParallelPlan.single()),
         sampling_defaults=SamplingDefaults(
             num_steps=4, guidance_scale=1.0, height=480, width=832, num_frames=81, fps=16,

v2/recipes/wan_causal/loop.py

@@ -1,9 +1,9 @@
 """ChunkRolloutLoop — causal/AR video.
 
-Loop shape: outer loop over latent-frame chunks × inner denoise per chunk, carrying a per-chunk
-KV/context via the slab-KV cache (sink/local window in inference; full history in self-forcing
-training mode). Each completed chunk is streamable (``StepResult.emit``), and the rollout profile
-captures per-chunk behavior for the self-forcing distillation method.
+Loop shape: outer loop over latent-frame chunks x inner denoise per chunk, carrying a per-chunk
+KV/context via the slab-KV cache. The cache can use a sink/local window or preserve full request
+history. Each completed chunk is streamable (``StepResult.emit``), and the rollout profile can
+capture per-chunk inference trajectory records.
 
 State is entirely in ``LoopState`` (the slab list + chunk position), so interleaving causal
 rollouts of two sessions cannot smear context — and the slab-KV pool is namespaced per request.
@@ -151,9 +151,7 @@ class ChunkRolloutLoop:
                             seq=st.scratch["chunk_idx"],
                             data=x,
                             preview=False))
-        res = ResourceRequest(
-                              resident_bytes=int(x.nbytes),
-                              peak_activation_bytes=int(x.nbytes * len(branches)))
+        res = ResourceRequest(resident_bytes=int(x.nbytes), peak_activation_bytes=int(x.nbytes * len(branches)))
         return WorkPlan(loop_id=self.loop_id,
                         instance_id=st.instance_id,
                         kind=WorkUnitKind.CHUNK_STEP,

v2/runtime/cache/classes.py

@@ -4,7 +4,7 @@ No single unified block pool: cache classes differ by 150-500x in natural granul
 each gets its own statically budgeted pool behind one ``CacheHandle``. Four classes:
   * ``FeatureCache``   — content-hash keyed, partitioned by adapter+weights (text/vision encoders)
   * ``ResidualCache``  — cache-dit residuals, scoped per request AND per CFG branch
-  * ``SlabKVCache``    — chunk-KV slabs (self-forcing / world models); training mode disables recycle
+  * ``SlabKVCache``    — chunk-KV slabs for causal/world models; can disable recycle
   * ``PagedKVCache``   — paged text-KV stub for ar_decode (phase-2 omni); minority case, lazy
 
 KV is the minority case — a pure bidirectional deployment (Wan/LTX T2V) allocates none of it.
@@ -71,7 +71,7 @@ class FeatureCache(_Pool):
             self.used_bytes -= nb
 
     def invalidate_weights(self, version: str) -> None:
-        """RL update_weights bumps weight epoch → drop entries from older epochs (wholesale)."""
+        """Drop entries from older resident-weight epochs."""
         drop = [h for h, (_v, _nb, k) in self._store.items() if k.weights_version != version]
         for h in drop:
             _v, nb, _k = self._store.pop(h)
@@ -124,21 +124,21 @@ class Slab:
 class SlabKVCache(_Pool):
     """Chunk-KV slabs for causal/world-model rollout.
 
-    ``training_mode`` disables mid-rollout recycling so activation-checkpoint recompute
-    doesn't double-advance the cache (self-forcing).
+    ``disable_recycle`` preserves the full request-local context. Otherwise the
+    cache behaves as a sliding window with the declared ``window`` size.
     """
 
     def __init__(self, policy: CachePolicy):
         super().__init__(policy)
         self._store: dict[str, list[Slab]] = {}
         self.window = max(1, policy.per_component.get("window", 1 << 30))
-        self.training_mode = policy.training_mode_disables_recycle
+        self.disable_recycle = policy.disable_recycle
 
     def append(self, namespace: str, slab: Slab) -> None:
         slabs = self._store.setdefault(namespace, [])
         slabs.append(slab)
         self.used_bytes += _nbytes(slab.k) + _nbytes(slab.v)
-        if not self.training_mode and len(slabs) > self.window:
+        if not self.disable_recycle and len(slabs) > self.window:
             dropped = slabs.pop(0)  # sliding-window recycle (inference only)
             self.used_bytes -= _nbytes(dropped.k) + _nbytes(dropped.v)
 

v2/runtime/cache/keys.py

@@ -2,8 +2,7 @@
 
 If a field can change output semantics, it is in the key (incorrect reuse is worse than no reuse).
 The serving hazard this kills: a request that shares a prompt but differs in te-LoRA stack must not
-serve stale embeddings — so the key is *partitioned* by ``adapter_versions``, not flushed. An RL
-``update_weights`` bumps ``weights_version`` and invalidates wholesale.
+serve stale embeddings, so the key is *partitioned* by ``adapter_versions``.
 """
 from __future__ import annotations
 
@@ -15,7 +14,7 @@ from typing import Any
 def content_hash(obj: Any) -> str:
     """Stable content hash for feature-cache keys (text/vision embeddings).
 
-    Lets K rollout samples of one prompt reuse a single text encode.
+    Lets repeated requests for one prompt reuse a single text encode.
     """
     h = hashlib.sha256()
     if isinstance(obj, str):
@@ -78,4 +77,4 @@ class CachePolicy:
     eviction: str = "lru"  # "lru" | "fifo" | "none"
     reuse_across_requests: bool = True
     per_component: dict[str, int] = field(default_factory=dict)
-    training_mode_disables_recycle: bool = False
+    disable_recycle: bool = False

v2/runtime/cache/manager.py

@@ -2,7 +2,7 @@
 
 Static partitioning makes cross-class fragmentation impossible (jumbo slab traffic cannot
 strand text-KV pages and vice versa). Each class gets a budget carved at init from the card's
-CacheContracts. ``invalidate_weights`` implements the wholesale RL weight-epoch bump.
+CacheContracts.
 """
 from __future__ import annotations
 
@@ -32,7 +32,7 @@ class CacheManager:
                     eviction=cc.eviction,
                     reuse_across_requests=cc.reuse_across_requests,
                     per_component=dict(cc.per_component),
-                    training_mode_disables_recycle=cc.training_mode_disables_recycle,
+                    disable_recycle=cc.disable_recycle,
                 ))
         return cls(policies)
 

v2/runtime/cudagraph.py

@@ -19,7 +19,7 @@ correctness-critical and bug-prone regardless of device:
   * the **eager-break** — a step that is model-declared non-capturable (host RNG / data-dependent
     branch, e.g. the FlowGRPO SDE step) or that an interceptor overrode this step never touches the
     graph cache (vLLM's ``@eager_break_during_capture``).
-  * **invalidation** — a resident-weights version bump (RL weight sync) or a shape change changes the
+  * **invalidation** — a resident-weights version bump or a shape change changes the
     key → recapture, so a graph never outlives the weights/shape it was captured for. This is
     AUTOMATIC: the resident-weight version is *in* the key, so a synced component's old graphs simply
     become unreachable and the next step recaptures.
@@ -159,7 +159,7 @@ class GraphCapturer:
         return plan.graph_fn(instance, g.workspace)
 
     def invalidate(self, component_ids: Any = None) -> int:
-        """Evict captured graphs whose resident weights just changed (RL weight sync), mirroring
+        """Evict captured graphs whose resident weights just changed, mirroring
         ``CacheManager.invalidate_components``. Without this, a synced component's old graphs become
         unreachable (the version is in the key) but linger — a real GPU memory leak across FlowGRPO's
         per-iteration syncs. With ``component_ids=None``, clears all. Returns the count dropped."""

v2/runtime/memory/allocator.py

@@ -3,7 +3,7 @@
 Reservation must be pre-flight: a scheduler that admits a diffusion step and then discovers it
 cannot allocate the VAE tile is wrong.
 
-Sleep/wake is component-granular (tags are component names) for RL: drop DiT + caches,
+Sleep/wake is component-granular (tags are component names): drop DiT + caches,
 keep VAE/text-encoder resident (CuMem-style).
 """
 from __future__ import annotations

v2/tests/test_cache.py

@@ -1,4 +1,4 @@
-"""Per-class cache pools: feature partition/invalidation, slab-KV window/training."""
+"""Per-class cache pools: feature partition/invalidation and slab-KV windows."""
 from __future__ import annotations
 
 import numpy as np
@@ -23,7 +23,7 @@ def test_feature_cache_weight_invalidation_is_wholesale():
     fc = FeatureCache(CachePolicy("feature", max_bytes=1 << 20))
     k = CacheKey(model_id="m", component_id="te", weights_version="v0")
     fc.put(k, np.ones((2,), dtype="float32"))
-    fc.invalidate_weights("v1")                # RL update_weights bump
+    fc.invalidate_weights("v1")
     assert fc.get(k) is None
 
 
@@ -35,17 +35,16 @@ def test_feature_cache_budget_eviction():
     assert fc.used_bytes <= 8
 
 
-def test_slab_kv_inference_window_vs_training_mode():
+def test_slab_kv_window_vs_disable_recycle():
     inf = SlabKVCache(CachePolicy("slab_kv", reuse_across_requests=False, per_component={"window": 2}))
     for i in range(4):
         inf.append("r", Slab(i, np.zeros((3,), dtype="float32"), None))
     assert len(inf.get("r")) == 2              # sliding window recycles (inference)
 
-    tr = SlabKVCache(CachePolicy("slab_kv", training_mode_disables_recycle=True,
-                                 per_component={"window": 2}))
+    tr = SlabKVCache(CachePolicy("slab_kv", disable_recycle=True, per_component={"window": 2}))
     for i in range(4):
         tr.append("r", Slab(i, np.zeros((3,), dtype="float32"), None))
-    assert len(tr.get("r")) == 4               # no mid-rollout recycle in training mode (self-forcing)
+    assert len(tr.get("r")) == 4               # no mid-rollout recycle when full context is requested
     assert tr.used_bytes > 0
     tr.clear_namespace("r")
     assert tr.used_bytes == 0

v2/tests/test_cudagraph.py

@@ -89,7 +89,7 @@ def test_weight_version_bump_recaptures_and_evicts_stale():
     inst = _wan()
     _drive(inst, seed=1, steps=4)
     assert inst.graphs.stats["captures"] == 1 and inst.graphs.captured == 1
-    inst.set_weights_version("v1", ["transformer"])    # RL weight sync: bumps version + evicts graph
+    inst.set_weights_version("v1", ["transformer"])    # version bump evicts stale graphs
     assert inst.graphs.captured == 0                    # stale graph dropped (no leak)
     _drive(inst, seed=1, steps=4)
     assert inst.graphs.stats["captures"] == 2           # recaptured under the new version...

v2/tests/test_review_fixes.py

@@ -43,12 +43,12 @@ def test_feature_key_partitions_by_adapter_stack():
     assert inst.caches.stats()["feature"]["misses"] == m + 1
 
 
-def test_transformer_weight_sync_keeps_text_encoder_cache():
+def test_transformer_weight_reload_keeps_text_encoder_cache():
     card = build_wan21_card()
     inst = load_card(card, cache_manager=CacheManager.from_card(card))
     cached_text_encode(inst, "p")
     assert inst.caches.stats()["feature"]["used_bytes"] > 0
-    inst.set_weights_version("w1", components=["transformer"])   # RL-style transformer-only sync
+    inst.set_weights_version("w1", components=["transformer"])   # transformer-only reload
     h = inst.caches.stats()["feature"]["hits"]
     cached_text_encode(inst, "p")                   # text-encoder embed survived ⇒ hit (not flushed)
     assert inst.caches.stats()["feature"]["hits"] == h + 1

v2/tests/test_thinker_talker.py

@@ -1,8 +1,7 @@
 """Qwen-Omni thinker→talker→vocoder — the third weight-sharing topology (vllm-omni
 ``qwen2_5_omni`` pipeline).
 
-Asserts the Card/Loop/Program vocabulary holds for a THREE-expert, THREE-loop cascade — the topology
-neither MoT (one shared module, two loops) nor joint-RL (two disjoint experts, two loops) covers:
+Asserts the Card/Loop/Program vocabulary holds for a THREE-expert, THREE-loop cascade:
 
   * three disjoint experts (thinker, talker, vocoder) on three loop *types* (ar_decode, ar_decode,
     audio_decode) in one request — every token and vocoder chunk a runtime-visible WorkUnit;
@@ -102,4 +101,3 @@ def test_vocoder_streams_audio_chunks():
     # the audio-streaming delta is exactly the vocoder's per-chunk emits (AR text emits are common)
     assert on.metrics["stream_chunks"] - off.metrics["stream_chunks"] == n_chunks
 
-