biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[ROCm][Kernel] Add gfx950 support for skinny gemms (#18010)

Browse Source 
Signed-off-by: charlifu <charlifu@amd.com>
This commit is contained in:
Charlie Fu
2025-05-31 09:40:05 -05:00
committed by GitHub
parent f2c3f66d59
commit 306d60401d
5 changed files with 91 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
vllm/platforms/rocm.py
View File

@@ -105,9 +105,15 @@ def on_gfx1x() -> bool:
@cache
def on_mi250_mi300() -> bool:
def on_mi3xx() -> bool:
GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
return any(arch in GPU_ARCH for arch in ["gfx90a", "gfx942"])
return any(arch in GPU_ARCH for arch in ["gfx942", "gfx950"])
@cache
def on_gfx9() -> bool:
GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
return any(arch in GPU_ARCH for arch in ["gfx90a", "gfx942", "gfx950"])
@cache