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:
co-authored by
Jingyi Yang
parent
18c4067a54
commit
7fc97042c3
@@ -439,11 +439,13 @@ def use_trtllm_attention(
|
||||
if force_use_trtllm is not None and not force_use_trtllm:
|
||||
return False
|
||||
|
||||
# Decode context parallel is not supported
|
||||
# TRTLLM prefill attends only the DCP-local KV shard and has no
|
||||
# cross-rank LSE combine, so it cannot be used with DCP; fall back to
|
||||
# FlashInfer's DCP prefill path. TRTLLM decode under DCP is selected
|
||||
# separately (all-gathered query heads + LSE combine in forward).
|
||||
if dcp_world_size > 1:
|
||||
logger.warning_once(
|
||||
"Trtllm does not support returning LSE and as a result "
|
||||
"does not support DCP, reverting to FlashInfer"
|
||||
"TRTLLM prefill does not support DCP, reverting to FlashInfer"
|
||||
)
|
||||
return False
|
||||
|
||||
|
||||
Reference in New Issue
Block a user