Xqa decode kernels (#43232)

Signed-off-by: Dan Blanaru <48605845+DanBlanaru@users.noreply.github.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Dan Blanaru
2026-07-02 12:32:05 -07:00
committed by GitHub
co-authored by Matthew Bonanni
parent d715b3aa1e
commit d29125c085
7 changed files with 538 additions and 143 deletions
@@ -78,14 +78,29 @@ def test_supports_sm100_with_artifactory(_art, _cap):
@patch("vllm.envs.VLLM_BATCH_INVARIANT", False)
@patch(
"vllm.utils.flashinfer.current_platform.is_device_capability", return_value=False
)
@patch(
"vllm.utils.flashinfer.current_platform.is_device_capability_family",
return_value=False,
)
def test_supports_non_sm100_platform(_cap):
def test_supports_unsupported_platform(_family, _cap):
assert supports_trtllm_attention() is False
@patch("vllm.envs.VLLM_BATCH_INVARIANT", False)
@patch("vllm.utils.flashinfer.current_platform.is_device_capability", return_value=True)
@patch(
"vllm.utils.flashinfer.current_platform.is_device_capability_family",
return_value=False,
)
@patch("vllm.utils.flashinfer.has_nvidia_artifactory", return_value=True)
def test_supports_sm90_decode_only(_art, _family, _cap):
assert supports_trtllm_attention(is_prefill=False) is True
assert supports_trtllm_attention(is_prefill=True) is False
@patch("vllm.envs.VLLM_BATCH_INVARIANT", False)
@patch(
"vllm.utils.flashinfer.current_platform.is_device_capability_family",