Shared planning state; never owns another weight or cache representation.
Source code in vllm/models/qwen3_5/xpu/specialized_decode/planning.py
| class DecodeRuntime:
"""Shared planning state; never owns another weight or cache representation."""
def __init__(self):
self.executors: list[FusedMoeDecode] = []
self.step: DecodeStep | None = None
self.steps: Counter[str] = Counter()
self.history: deque[tuple[tuple[str, ...], DecodeStep]] | None = None
def prepare_step(self, batch: "InputBatch") -> DecodeStep:
reason = None
if batch.num_draft_tokens:
reason = "draft_tokens"
elif (
not 0
< batch.num_reqs
<= batch.num_tokens
<= batch.num_tokens_after_padding
<= MAX_DECODE_ROWS
):
reason = "token_rows"
elif any(
batch.num_tokens_after_padding * e.top_k > MAX_ROUTED_ROWS
for e in self.executors
):
reason = "routed_rows"
self.step = DecodeStep(
batch.num_reqs, batch.num_tokens, batch.num_tokens_after_padding, reason
)
self.steps[reason or "eligible"] += 1
if self.history is not None:
self.history.append((tuple(batch.req_ids), self.step))
return self.step
def prepare_capture(self, num_reqs: int, num_tokens: int) -> None:
reason = (
None if 0 < num_reqs == num_tokens <= MAX_DECODE_ROWS else "capture_shape"
)
self.step = DecodeStep(num_reqs, num_tokens, num_tokens, reason)
def snapshot(self) -> dict:
return {
"layers": len(self.executors),
"planned_steps": dict(self.steps),
"persistent_scratch_bytes": sum(e.scratch_bytes for e in self.executors),
"specialized_weight_bytes": (
0 if all(e.bindings_match() for e in self.executors) else None
),
"native_launches_including_capture": sum(
e.native_launches for e in self.executors
),
"bindings_shared": all(e.bindings_match() for e in self.executors),
}
|