Add DCP + Eagle support for Tokenspeed MLA backends (#48180)

Signed-off-by: Pavani Majety <pmajety@nvidia.com>
Signed-off-by: Jingyi Yang <girasoleyang@gmail.com>
Co-authored-by: Jingyi Yang <girasoleyang@gmail.com>
This commit is contained in:
Pavani Majety
2026-07-13 11:46:02 -07:00
committed by GitHub
co-authored by Jingyi Yang
parent 18c4067a54
commit 7fc97042c3
8 changed files with 313 additions and 23 deletions
@@ -154,6 +154,17 @@ def test_use_dcp_fallback(_mock):
assert _call(dcp_world_size=2) is False
@patch("vllm.utils.flashinfer.supports_trtllm_attention", return_value=True)
def test_use_dcp_fallback_prefill(_mock):
# TRTLLM prefill has no cross-rank LSE combine for DCP-sharded KV.
assert _call(dcp_world_size=2, is_prefill=True) is False
@patch("vllm.utils.flashinfer.supports_trtllm_attention", return_value=True)
def test_use_dcp_fallback_prefill_force_on_still_false(_mock):
assert _call(dcp_world_size=2, is_prefill=True, force_use_trtllm=True) is False
@patch("vllm.utils.flashinfer.supports_trtllm_attention", return_value=False)
def test_use_platform_unsupported(_mock):
assert _call() is False