[Feature] Add SM103 (Blackwell Ultra) Support to vLLM (#30484)

Signed-off-by: LopezCastroRoberto <robertol.c510@gmail.com>
Signed-off-by: Roberto L. Castro <38211239+LopezCastroRoberto@users.noreply.github.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
Roberto L. Castro
2025-12-12 19:34:23 -08:00
committed by GitHub
co-authored by youkaichao
parent 57e9bf1864
commit 4fa7ce46f3
21 changed files with 53 additions and 33 deletions
+2 -2
View File
@@ -38,7 +38,7 @@ class DeepGemmQuantScaleFMT(Enum):
return DeepGemmQuantScaleFMT.FLOAT32
return (
DeepGemmQuantScaleFMT.UE8M0
if current_platform.is_device_capability(100)
if current_platform.is_device_capability_family(100)
else DeepGemmQuantScaleFMT.FLOAT32_CEIL_UE8M0
)
@@ -50,7 +50,7 @@ def is_deep_gemm_supported() -> bool:
"""
is_supported_arch = current_platform.is_cuda() and (
current_platform.is_device_capability(90)
or current_platform.is_device_capability(100)
or current_platform.is_device_capability_family(100)
)
return envs.VLLM_USE_DEEP_GEMM and has_deep_gemm() and is_supported_arch
+3 -1
View File
@@ -264,7 +264,9 @@ def supports_trtllm_attention() -> bool:
return False
# Requires SM100 and NVIDIA artifactory to be accessible to download cubins
return current_platform.is_device_capability(100) and has_nvidia_artifactory()
return (
current_platform.is_device_capability_family(100) and has_nvidia_artifactory()
)
def force_use_trtllm_attention() -> bool | None: