Enable gfx1250 ROCm architecture (#46516)
Signed-off-by: Douglas Lehr <Doug.Lehr@amd.com> Signed-off-by: jpvillam <juan.villamizar@amd.com> Signed-off-by: Daniel <danichan@amd.com> Signed-off-by: Jaden Mathias <jaden.mathias@amd.com> Signed-off-by: jpvillam <jpvillam@amd.com> Signed-off-by: Juan Villamizar <100237675+jpvillam-amd@users.noreply.github.com> Signed-off-by: Douglas Lehr <91553416+dllehr-amd@users.noreply.github.com> Signed-off-by: Jaden Mathias <Jaden.Mathias@amd.com> Co-authored-by: Douglas Lehr <Doug.Lehr@amd.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Co-authored-by: Daniel <danichan@amd.com> Co-authored-by: Jaden Mathias <jaden.mathias@amd.com> Co-authored-by: jpvillam <jpvillam@amd.com> Co-authored-by: root <root@ctheliosr-1b114-f01-2.mnb.dcgpu> Co-authored-by: Douglas Lehr <91553416+dllehr-amd@users.noreply.github.com> Co-authored-by: dllehr <douglehr@amd.com> Co-authored-by: Daniel Chan <Daniel.Chan@amd.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
co-authored by
Douglas Lehr
Claude Opus 4.7
Daniel
Jaden Mathias
jpvillam
root
Douglas Lehr
dllehr
Daniel Chan
Andreas Karatzas
parent
553fcb82d5
commit
4f1da84eb5
@@ -122,7 +122,7 @@ def use_aiter_triton_gemm(n, m, k, dtype):
|
||||
def rocm_unquantized_gemm_impl(
|
||||
x: torch.Tensor, weight: torch.Tensor, bias: torch.Tensor | None = None
|
||||
) -> torch.Tensor:
|
||||
from vllm.platforms.rocm import on_gfx1x, on_gfx9, on_gfx950
|
||||
from vllm.platforms.rocm import on_gfx1x, on_gfx9, on_gfx950, on_gfx1250
|
||||
|
||||
n = x.numel() // x.size(-1)
|
||||
m = weight.shape[0]
|
||||
@@ -164,7 +164,11 @@ def rocm_unquantized_gemm_impl(
|
||||
if use_skinny_reduce_counting:
|
||||
return ops.wvSplitKrc(x, weight, cu_count, bias)
|
||||
|
||||
if use_aiter_triton_gemm(n, m, k, x.dtype):
|
||||
# gfx1250's aiter gemm_a16w16 uses the gluon backend, which requires
|
||||
# K % 256 == 0 (it walks K with fixed-size descriptors and won't pad a
|
||||
# partial last tile). Some whitelisted shapes have K=2880 (e.g. gpt-oss-120b
|
||||
# hidden), so skip aiter there and fall back to the torch GEMM path below.
|
||||
if use_aiter_triton_gemm(n, m, k, x.dtype) and not (on_gfx1250() and k % 256 != 0):
|
||||
from aiter.ops.triton.gemm_a16w16 import gemm_a16w16
|
||||
|
||||
return gemm_a16w16(x, weight, bias)
|
||||
@@ -172,6 +176,8 @@ def rocm_unquantized_gemm_impl(
|
||||
use_skinny = (
|
||||
envs.VLLM_ROCM_USE_SKINNY_GEMM
|
||||
and (on_gfx9() or on_gfx1x())
|
||||
# build (gfx9/gfx11 ISA); fall back to torch GEMM there.
|
||||
# TODO GFX1250: Include once skinny GEMM is supported on gfx1250
|
||||
and x.dtype in [torch.float16, torch.bfloat16]
|
||||
and k % 8 == 0
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user