Integration SM100 FlashInfer fused allreduce RMSNorm (#20691)

Signed-off-by: ilmarkov <imarkov@redhat.com> Co-authored-by: ilmarkov <imarkov@redhat.com>
2025-07-12 03:58:15 +02:00
parent 7b828e30d5
commit fc0f41d10a
4 changed files with 514 additions and 6 deletions
--- a/vllm/config.py
+++ b/vllm/config.py
@@ -3962,6 +3962,10 @@ class PassConfig:
    """Whether to enable sequence parallelism."""
    enable_async_tp: bool = False
    """Whether to enable async TP."""
+    enable_fi_allreduce_fusion: bool = False
+    """Whether to enable flashinfer allreduce fusion."""
+    fi_allreduce_fusion_max_token_num: int = 1024
+    """Max number of tokens to used in flashinfer allreduce fusion."""

    # TODO(luka) better pass enabling system.