[Spec Decode] Enable Gemma4 DSpark adaptive verification with FlashInfer (#57263)

Signed-off-by: zixi-qi <zixi@inferact.ai>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
qizixi
2026-09-26 17:01:17 +08:00
committed by GitHub
co-authored by OpenAI Codex Claude Opus 5.5 Nick Hill
parent 3b4566c5cf
commit a4eb3f25d6
24 changed files with 894 additions and 135 deletions
@@ -455,13 +455,29 @@ def test_mla_dcp_manager_selects_fallback_backends(monkeypatch):
)
def test_dcp_workspace_covers_parallel_drafting():
@pytest.mark.parametrize(
("parallel_drafting", "use_v2_model_runner", "expected"),
[
# Model runner V1's parallel drafter widens requests to 1 + 2 * 3.
(True, False, 28),
(True, True, 16),
# A diffusion canvas of 3 without a speculative config.
(None, True, 16),
],
)
def test_dcp_workspace_covers_decode_width(
parallel_drafting, use_v2_model_runner, expected
):
config = _manager_config()
config.scheduler_config.max_num_batched_tokens = 128
config.num_speculative_tokens = 3
config.speculative_config = MagicMock(parallel_drafting=True)
config.use_v2_model_runner = use_v2_model_runner
if parallel_drafting is not None:
config.speculative_config = MagicMock(
num_speculative_tokens=3, parallel_drafting=parallel_drafting
)
assert dcp.get_dcp_workspace_max_num_tokens(config) == 28
assert dcp.get_dcp_workspace_max_num_tokens(config) == expected
def test_mla_dcp_manager_selects_pcp_combine(monkeypatch):