Support EAGLE speculative decoding in the KV-canary (#26813)
This commit is contained in:
@@ -255,5 +255,40 @@ def _drive_one_cycle(manager, forward_batch) -> None:
|
||||
manager.post_ops_maybe_inside_graph(forward_batch, pre_ops_output)
|
||||
|
||||
|
||||
class TestCanaryManagerActiveSingleForwardManagerDispatch(CanaryManagerTestCase):
|
||||
def test_pre_ops_maybe_inside_graph_dispatches_to_bracketed_sfm(
|
||||
self,
|
||||
) -> None:
|
||||
"""Verify the dispatcher routes phase 2 to the bracketed SingleForwardManager."""
|
||||
manager = make_manager(device=self.device, speculative_num_steps=3)
|
||||
forward_batch = make_forward_batch(self.device)
|
||||
target_sfm = manager._single_forward_managers[1]
|
||||
observed: list[object] = []
|
||||
original_phase_2 = target_sfm.pre_ops_maybe_inside_graph
|
||||
|
||||
def _record(fb):
|
||||
observed.append(fb)
|
||||
return original_phase_2(fb)
|
||||
|
||||
target_sfm.pre_ops_maybe_inside_graph = _record
|
||||
manager._single_forward_managers[0].pre_ops_outside_graph(
|
||||
maybe_inaccurate_forward_batch=forward_batch
|
||||
)
|
||||
manager._single_forward_managers[1].pre_ops_outside_graph(
|
||||
maybe_inaccurate_forward_batch=forward_batch
|
||||
)
|
||||
with manager.with_active_single_forward_manager(1):
|
||||
manager.pre_ops_maybe_inside_graph(forward_batch)
|
||||
self.assertEqual(observed, [forward_batch])
|
||||
|
||||
def test_pre_ops_maybe_inside_graph_asserts_outside_bracket(
|
||||
self,
|
||||
) -> None:
|
||||
manager = make_manager(device=self.device)
|
||||
forward_batch = make_forward_batch(self.device)
|
||||
with self.assertRaises(AssertionError):
|
||||
manager.pre_ops_maybe_inside_graph(forward_batch)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
|
||||
@@ -0,0 +1,34 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import unittest
|
||||
|
||||
from sglang.test.ci.ci_register import register_cuda_ci
|
||||
from sglang.test.mock_model.utils import MOCK_MODEL_PATH, run_mock_model_bench_serving
|
||||
from sglang.test.test_utils import CustomTestCase
|
||||
|
||||
register_cuda_ci(est_time=600, stage="extra-a", runner_config="1-gpu-small")
|
||||
|
||||
|
||||
class TestE2ESpeculativeEagle(CustomTestCase):
|
||||
def test_spec_eagle_no_canary_violation(self) -> None:
|
||||
run_mock_model_bench_serving(
|
||||
extra_server_args=[
|
||||
"--speculative-algorithm",
|
||||
"EAGLE",
|
||||
"--speculative-draft-model-path",
|
||||
MOCK_MODEL_PATH,
|
||||
"--speculative-num-steps",
|
||||
"1",
|
||||
"--speculative-eagle-topk",
|
||||
"1",
|
||||
"--speculative-num-draft-tokens",
|
||||
"2",
|
||||
"--mem-fraction-static",
|
||||
"0.45",
|
||||
],
|
||||
input_check_enabled=False,
|
||||
)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user