xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-09 21:35:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Hardware][Apple-CPU] Enable native bfloat16 on Apple Silicon (M2 and later) (#24129)

Browse Source 
Signed-off-by: ignaciosica <mignacio.sica@gmail.com>
This commit is contained in:
Ignacio Sica 2025-09-10 00:50:21 -03:00 committed by GitHub
parent 7e7db04310
commit 3c2156b3af
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 6 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/platforms/cpu.py
View File

@ -75,12 +75,12 @@ class CpuPlatform(Platform):
def supported_dtypes(self) -> list[torch.dtype]:
if self.get_cpu_architecture() == CpuArchEnum.POWERPC:
return [torch.bfloat16, torch.float32]
elif sys.platform.startswith(
"darwin") and self.get_cpu_architecture() == CpuArchEnum.ARM:
# TODO: change this condition to check if the platform support bf16
# instead of checking the OS. For instance M2 shall supports bf16
# already. But we need to modify `cpu_extension.cmake` to activate
# the feature in the build.
elif (self.get_cpu_architecture() == CpuArchEnum.ARM
and sys.platform.startswith("darwin")):
if (subprocess.check_output(
["sysctl -n hw.optional.arm.FEAT_BF16"],
shell=True).strip() == b"1"):
return [torch.bfloat16, torch.float16, torch.float32]
return [torch.float16, torch.float32]
# x86/aarch64 CPU has supported both bf16 and fp16 natively.
return [torch.bfloat16, torch.float16, torch.float32]