A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 22:15 UTC
[BugFix] Fix false assertion with spec-decode=[2,4,..] and TP>2 (#29036)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Lucas Wilkinson committed
Nov 19, 2025 at 16:43 UTC
8f4f77a7275ecac594f84bdb41b67c95cf3eb26d
1 file changed
+1
-1
vllm/config/compilation.py
+1
-1
@@ -921,7 +921,7 @@ class CompilationConfig:
921
self, uniform_decode_query_len: int, tensor_parallel_size: int
922
):
923
multiple_of = uniform_decode_query_len
924
- if tensor_parallel_size > 1:
924
+ if tensor_parallel_size > 1 and self.pass_config.enable_sequence_parallelism:
925
multiple_of = max(uniform_decode_query_len, tensor_parallel_size)
926
if (
927
multiple_of % uniform_decode_query_len != 0