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 14, 00:30 UTC
[Model] Use the same fused_moe configs for all H200 devices (#23642)
Signed-off-by: Roger Meier <r.meier@siemens.com>
Roger Meier committed
Oct 31, 2025 at 01:36 UTC
2918c1b49c88c29783c86f78d2c4221cb9622379
1 file changed
+3
vllm/model_executor/layers/fused_moe/fused_moe.py
+3
@@ -818,6 +818,9 @@ def get_config_file_name(
818
E: int, N: int, dtype: str | None, block_shape: list[int] | None = None
819
) -> str:
820
device_name = current_platform.get_device_name().replace(" ", "_")
821
+ # Set device_name to H200 if a device from the H200 family is detected
822
+ if "H200" in device_name:
823
+ device_name = "H200"
824
dtype_selector = "" if not dtype else f",dtype={dtype}"
825
block_shape_selector = (
826
"" if not block_shape or not all(block_shape) else f",block_shape={block_shape}"