[ROCm][Kernel] Add HybridW4A16LinearKernel: Triton prefill + HIP skinny decode (#40977)

Signed-off-by: Matthias Gehre <matthias.gehre@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Matthias Gehre
2026-07-13 20:22:00 -07:00
committed by GitHub
co-authored by Andreas Karatzas
parent b50ef9c6ed
commit 382bbd5144
11 changed files with 2165 additions and 5 deletions
@@ -16,9 +16,28 @@ from vllm.model_executor.kernels.linear import (
from vllm.platforms import current_platform
from vllm.scalar_type import scalar_types
if current_platform.is_rocm():
from vllm.platforms.rocm import on_gfx1x, on_gfx1100
else:
on_gfx1100 = on_gfx1x = lambda: False # noqa: E731
# Group sizes the HIP skinny path of RDNAHybridW4A16LinearKernel instantiates.
_HYBRID_GROUP_SIZES = (32, 64, 128)
def _expected_rocm_kernel(weight_type, group_size: int) -> str:
"""Mirror the ROCm priority order in ``_POSSIBLE_KERNELS[ROCM]``:
RDNA3 (gfx1100, symmetric uint4b8) -> Hybrid (gfx11/gfx12) -> Triton.
"""
if on_gfx1100() and weight_type == scalar_types.uint4b8:
return "RDNA3W4A16LinearKernel"
if on_gfx1x() and group_size in _HYBRID_GROUP_SIZES:
return "RDNAHybridW4A16LinearKernel"
return "TritonW4A16LinearKernel"
@pytest.mark.skipif(not current_platform.is_rocm(), reason="ROCm only")
def test_choose_mp_linear_kernel_picks_triton_w4a16_for_uint4b8():
def test_choose_mp_linear_kernel_uint4b8():
# int4 weights, 16-bit activations (CT W4A16 typical config).
K, N = 1024, 256
config = MPLinearLayerConfig(
@@ -32,12 +51,12 @@ def test_choose_mp_linear_kernel_picks_triton_w4a16_for_uint4b8():
)
kernel_type = choose_mp_linear_kernel(config)
assert kernel_type.__name__ == "TritonW4A16LinearKernel"
assert kernel_type.__name__ == _expected_rocm_kernel(scalar_types.uint4b8, 128)
@pytest.mark.skipif(not current_platform.is_rocm(), reason="ROCm only")
def test_choose_mp_linear_kernel_picks_triton_w4a16_for_uint4_asymmetric():
# Asymmetric int4 weights should also be supported (explicit zero points).
def test_choose_mp_linear_kernel_uint4_asymmetric():
# Asymmetric int4 weights (explicit zero points).
K, N = 512, 512
config = MPLinearLayerConfig(
full_weight_shape=(K, N),
@@ -50,4 +69,4 @@ def test_choose_mp_linear_kernel_picks_triton_w4a16_for_uint4_asymmetric():
)
kernel_type = choose_mp_linear_kernel(config)
assert kernel_type.__name__ == "TritonW4A16LinearKernel"
assert kernel_type.__name__ == _expected_rocm_kernel(scalar_types.uint4, 64)