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:
Juan Villamizar
2026-07-31 01:41:12 +00:00
committed by GitHub
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
33 changed files with 1850 additions and 104 deletions
@@ -374,9 +374,14 @@ class AiterFp8BlockScaledMMKernel(Fp8BlockScaledMMLinearKernel):
super().__init__(config)
n, k = config.weight_shape
self.use_triton = (
not current_platform.is_fp8_fnuz()
and rocm_aiter_ops.is_triton_gemm_w8a8_tuned(n, k)
_on_gfx1250 = False
if current_platform.is_rocm():
from vllm.platforms.rocm import on_gfx1250
_on_gfx1250 = on_gfx1250()
self.use_triton = not current_platform.is_fp8_fnuz() and (
rocm_aiter_ops.is_triton_gemm_w8a8_tuned(n, k) or _on_gfx1250
)
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
@@ -104,10 +104,10 @@ class RowWiseTorchFP8ScaledMMLinearKernel(TorchFP8ScaledMMLinearKernel):
if not current_platform.is_rocm():
return False, "requires ROCm."
from vllm.platforms.rocm import on_mi3xx
from vllm.platforms.rocm import get_cdna_version
if not on_mi3xx():
return False, "requires MI3xx."
if get_cdna_version() <= 2:
return False, "requires CDNA3+"
if compute_capability is not None and compute_capability < 94:
return False, "requires compute capability 94 and above."
@@ -79,10 +79,15 @@ class ROCmFP8ScaledMMLinearKernel(FP8ScaledMMLinearKernel):
if not current_platform.is_rocm():
return False, "requires ROCm."
from vllm.platforms.rocm import on_gfx12x, on_mi3xx
from vllm.platforms.rocm import get_cdna_version, on_gfx12x, on_gfx1250
if not (on_mi3xx() or on_gfx12x()):
return False, "requires MI3xx or gfx12x"
# wvSplitKQ (skinny GEMM) is excluded from the gfx1250 build.
if on_gfx1250():
return False, "wvSplitKQ (skinny GEMM) is not built on gfx1250"
# Restore RDNA4 (gfx12x) dropped by the get_cdna_version()>2 refactor.
if get_cdna_version() <= 2 and not on_gfx12x():
return False, "requires CDNA3+ (gfx942/gfx950) or RDNA4 (gfx12x)"
if not envs.VLLM_ROCM_USE_SKINNY_GEMM:
return False, "requires VLLM_ROCM_USE_SKINNY_GEMM to be enabled."