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 6, 07:00 UTC
refactor: abstract deepgemm support into platform (#37519)
Co-authored-by: sherryC41 <sherry.c.c41@gmail.com>
SherryC41 committed
Mar 21, 2026 at 01:54 UTC
6ec5e9fd37efc3634f509bc16e34f0d9a3cce528
3 files changed
+13
-4
vllm/platforms/cuda.py
+5
@@ -511,6 +511,11 @@ class CudaPlatformBase(Platform):
511
def support_static_graph_mode(cls) -> bool:
512
return True
513
514
+ @classmethod
515
+ def support_deep_gemm(cls) -> bool:
516
+ """Currently, only Hopper and Blackwell GPUs are supported."""
517
+ return cls.is_device_capability(90) or cls.is_device_capability_family(100)
518
+
519
@classmethod
520
def num_compute_units(cls, device_id: int = 0) -> int:
521
return torch.cuda.get_device_properties(device_id).multi_processor_count
vllm/platforms/interface.py
+7
@@ -712,6 +712,13 @@ class Platform:
712
"""
713
return False
714
715
+ @classmethod
716
+ def support_deep_gemm(cls) -> bool:
717
+ """
718
+ Returns if DeepGEMM is supported by the current platform.
719
+ """
720
+ return False
721
+
722
@classmethod
723
def use_custom_op_collectives(cls) -> bool:
724
"""
vllm/utils/deep_gemm.py
+1
-4
@@ -70,10 +70,7 @@ def is_deep_gemm_supported() -> bool:
70
"""Return `True` if DeepGEMM is supported on the current platform.
71
Currently, only Hopper and Blackwell GPUs are supported.
72
"""
73
- is_supported_arch = current_platform.is_cuda() and (
74
- current_platform.is_device_capability(90)
75
- or current_platform.is_device_capability_family(100)
76
- )
73
+ is_supported_arch = current_platform.support_deep_gemm()
74
return envs.VLLM_USE_DEEP_GEMM and has_deep_gemm() and is_supported_arch
75
76