[Spec Decode] Enable adaptive DSpark on SM100 sparse MLA (#52783)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Misha Goin
2026-08-25 13:31:15 -04:00
committed by GitHub
parent 70021a0081
commit 7de96050c5
7 changed files with 293 additions and 18 deletions
@@ -5,9 +5,13 @@ from types import SimpleNamespace
import numpy as np
from vllm.v1.attention.backend import AttentionCGSupport
from vllm.v1.worker.gpu.async_utils import StepTimingSample
from vllm.v1.worker.gpu.attn_utils import AttentionCGSupportInfo
from vllm.v1.worker.gpu.spec_decode import adaptive_verification as adaptive_module
from vllm.v1.worker.gpu.spec_decode.adaptive_verification import (
AdaptiveVerificationManager,
maybe_create_adaptive_verification_manager,
)
from vllm.v1.worker.gpu.structured_outputs import _build_grammar_mapping
@@ -31,6 +35,60 @@ def make_manager(
return manager
def test_manager_scopes_varlen_check_without_weakening_runner_cg_mode(monkeypatch):
class Backend:
@classmethod
def supports_device_cpu_query_lens_mismatch(cls):
return True
class Builder:
def __init__(self, support):
self.support = support
def get_cudagraph_support(self, *_args):
return self.support
def group(layer_name, support):
builder = Builder(support)
return SimpleNamespace(
layer_names=[layer_name],
backend=Backend,
kv_cache_spec=None,
get_metadata_builder=lambda _index: builder,
)
groups = [
[
group("target", AttentionCGSupport.ALWAYS),
group("draft", AttentionCGSupport.UNIFORM_BATCH),
]
]
runner_support = AttentionCGSupportInfo(
AttentionCGSupport.UNIFORM_BATCH, "DraftBackend"
)
created = object()
monkeypatch.setattr(
adaptive_module,
"AdaptiveVerificationManager",
lambda *_args, **_kwargs: created,
)
manager = maybe_create_adaptive_verification_manager(
enable_adaptive_verification=True,
attn_groups=groups,
attn_cg_support=runner_support,
req_states=object(),
query_start_loc=object(),
num_bonus_tokens=1,
max_total_logits=1,
vllm_config=None,
target_layer_names={"target"},
)
assert manager is created
assert runner_support.min_cg_support == AttentionCGSupport.UNIFORM_BATCH
def test_budget_stops_where_marginal_drafts_stop_paying_for_themselves():
# Verification is cheap up to two extra tokens, then jumps 100x; only the
# highest-confidence draft is worth the cheap slot.