[5/N][warmup][DSv4] Migrate NVIDIA CuTeDSL attention kernels (#53566)

Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>
Co-authored-by: Codex <codex@openai.com>
This commit is contained in:
Roberto L. Castro
2026-09-12 00:05:59 -04:00
committed by GitHub
co-authored by Codex
parent 485421b1c3
commit 120ec4ebd2
28 changed files with 3124 additions and 2551 deletions
@@ -12,16 +12,15 @@ from cutlass import Float32, Int32, Uint32, Uint64
from quack.compile_utils import make_fake_tensor
from vllm.cute_utils import recast_val
from vllm.model_executor.warmup.jit_warmup import kernel_launcher
from vllm.model_executor.warmup.jit_warmup_cutedsl_helper import (
CuTeDSLLaunchSpec,
VllmCuTeDSLJitKernel,
cutedsl_kernel_launcher,
)
from vllm.model_executor.warmup.jit_warmup_triton_helper import (
LaunchSpec,
TritonWarmupTensor,
VllmTritonJitKernel,
kernel_launcher,
)
from vllm.triton_utils import tl, triton
@@ -569,7 +568,7 @@ class StableTopKFromGatheredCandidatesKernel(
)
return gathered, out
@cutedsl_kernel_launcher
@kernel_launcher
def __call__(
self,
gathered: torch.Tensor,
@@ -577,16 +576,11 @@ class StableTopKFromGatheredCandidatesKernel(
*,
topk: int,
) -> CuTeDSLLaunchSpec[CompileKey]:
compile_key = self.dispatch(topk=topk, num_candidates=gathered.shape[1])
return (
compile_key,
(gathered, out),
{
"gathered_shape": tuple(gathered.shape),
"out_shape": tuple(out.shape),
"topk": topk,
},
compile_key = self.dispatch(
topk=topk,
num_candidates=gathered.shape[1],
)
return compile_key, (gathered, out)
_PACK_DCP_TOPK_CANDIDATES_KERNEL = PackDCPTopkCandidatesKernel()