Files
Mukkesh Ganesh 58b3ad7ba7 feat(gemma4): speculative decoding with the assistant draft (#895)
* fix(scheduler): return the sliding-window slots when a verify step rolls back

* feat(gemma4): add the assistant draft model and its loader

* feat(spec): serve sliding-window targets with a draft that keeps no KV

* fix(spec): budget verify rows in the swa pool and drop the post-verify sync
2026-11-12 20:34:28 -08:00

135 lines
6.0 KiB
Python

"""SpecRunner: the target side of speculative decoding and the order of one step; the Proposer it holds is the draft side."""
from __future__ import annotations
from typing import TYPE_CHECKING, List
import torch
from freetoken.engine.sample import make_device_tensor
from freetoken.engine.spec_protocol import SpecForwardOutput
from .proposer import DraftContext, Proposer, SpecProposal
from .rejection import reject, reject_with_draft_probs
if TYPE_CHECKING:
from freetoken.core import Batch, Req
from freetoken.engine.sample import BatchSamplingArgs
from freetoken.models.blocks import FeatureSink
from freetoken.models.config import SlotStateSpec
def proposer_class(config) -> type[Proposer]:
from freetoken.models.register import _load_attr
from freetoken.spec import DRAFT_PROPOSERS, PROPOSERS
algorithm = config.speculative_algorithm
if algorithm == "mtp":
# a family's draft method picks the proposer behind the mtp CLI name; one with no draft of the
# right method falls through to MTPProposer, whose adjust_config refuses draft-less families
drafts = dict(config.model_spec.spec_drafts)
for method, ref in DRAFT_PROPOSERS.items():
if method in drafts:
return _load_attr(*ref.split(":"))
module, attr = PROPOSERS[algorithm].split(":")
return _load_attr(module, attr)
class SpecRunner:
"""Implements engine.spec_protocol.SpecRuntime; the engine loads it by name and never imports freetoken.spec."""
def __init__(self, config, engine) -> None:
self.config = config
self.device = engine.device
self.k = config.speculative_num_steps
self.proposer = proposer_class(config)(config, engine.model)
self.radix_lookahead = self.proposer.radix_lookahead
self.engine = None
# the step's drafts, from propose_into to the verify_step of the same batch
self._proposal: SpecProposal | None = None
@classmethod
def adjust_config(cls, config, override) -> None:
proposer_cls = proposer_class(config)
proposer_cls.adjust_config(config, override)
default_k, max_k = proposer_cls.num_steps(config)
if config.speculative_num_steps == 0:
override("speculative_num_steps", default_k)
if not 1 <= config.speculative_num_steps <= max_k:
raise ValueError(
f"--speculative-num-steps must be in [1, {max_k}] for "
f"{config.speculative_algorithm}, got {config.speculative_num_steps}"
)
@classmethod
def prepare_target(cls, config, model) -> None:
proposer_class(config).prepare_target(config, model)
@classmethod
def budget(cls, config) -> tuple[int, int]:
return proposer_class(config).budget(config)
@classmethod
def slot_states(cls, config) -> tuple[SlotStateSpec, ...]:
return proposer_class(config).slot_states(config)
@classmethod
def target_may_be_windowed(cls, config) -> bool:
return proposer_class(config).target_may_be_windowed(config)
def attach(self, engine) -> None:
"""Second construction phase, after the pools and before graph capture."""
self.engine = engine
self._host = torch.zeros(self.config.max_running_req + 1, self.k + 2, dtype=torch.int64, pin_memory=True)
pool = engine.linear_state_pool
states = {s.name: pool.slot_state(s.name) for s in self.config.spec_slot_states} if pool is not None else {}
self.proposer.attach(DraftContext(
page_table=engine.page_table, num_pages=engine.num_pages, device=engine.device,
stream=engine.stream, dummy_req=engine.dummy_req, slot_states=states, kv_cache=engine.kv_cache,
))
def detach(self, *, keep_kv: bool) -> None:
self.proposer.detach(keep_kv=keep_kv)
def capture_graphs(self, bs_list: List[int], pool) -> None:
self.proposer.capture_graphs(bs_list, pool)
def feature_sink(self, batch: Batch) -> FeatureSink | None:
return self.proposer.feature_sink(batch)
def on_prefill(self, batch: Batch, next_tokens_gpu: torch.Tensor) -> None:
self.proposer.on_prefill(batch, next_tokens_gpu)
def propose_into(self, reqs: List[Req], token_pool: torch.Tensor) -> None:
"""Draft k tokens per request into token_pool after its pending token at cached_len."""
rows = make_device_tensor([r.table_idx for r in reqs], torch.int64, self.device)
cols = make_device_tensor([r.cached_len for r in reqs], torch.int64, self.device)
self._proposal = self.proposer.propose(reqs, rows, token_pool[rows, cols])
cols = cols.unsqueeze(1) + 1 + torch.arange(self.k, device=self.device)
token_pool[rows.unsqueeze(1), cols] = self._proposal.draft_token_ids.to(token_pool.dtype)
def verify_step(self, batch: Batch, sample_args: BatchSamplingArgs) -> SpecForwardOutput:
engine = self.engine
bs, width = batch.size, self.k + 1
# each request's rows read its pending token then its drafts
drafts = batch.input_ids[: bs * width].view(bs, width)[:, 1:].to(torch.int64)
batch.features = self.feature_sink(batch)
logits = engine.forward_logits(batch)[: bs * width]
proposal, self._proposal = self._proposal, None
if proposal.draft_support_probs is not None:
out, emitted = reject_with_draft_probs(
logits, drafts, proposal.draft_support_ids, proposal.draft_support_probs, sample_args, engine.sampler,
)
else:
out, emitted = reject(logits, drafts, sample_args, engine.sampler)
host = self._host[:bs]
host.copy_(torch.cat([out, emitted.unsqueeze(1)], dim=1), non_blocking=True)
copy_done = torch.cuda.Event()
copy_done.record(engine.stream)
# after the D2H is queued: the drain waits for the copy only, while the draft's work overlaps it
self.proposer.post_verify(batch, out, emitted)
return SpecForwardOutput(out, host[:, :width], copy_done, host[:, width])
def release(self, req: Req) -> None:
self.proposer.release(req)