[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
@@ -24,6 +24,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
)
from vllm.model_executor.warmup.jit_warmup import (
WarmupIntRange,
kernel_launcher,
zip_inputs,
)
from vllm.model_executor.warmup.jit_warmup_triton_helper import (
@@ -31,7 +32,6 @@ from vllm.model_executor.warmup.jit_warmup_triton_helper import (
TritonPointerInputVariant,
TritonWarmupTensor,
VllmTritonJitKernel,
kernel_launcher,
triton_scalar_specialization_rep,
)
from vllm.platforms import current_platform
@@ -521,11 +521,17 @@ def dequantize_and_gather_k_cache(
if has_cutedsl():
# lazily import, otherwise some tests fail due to CUDA driver init failure.
from vllm.models.deepseek_v4.nvidia.ops.dequant_gather_k_cutedsl import (
dequantize_and_gather_k_cache_cutedsl,
_DEQUANT_GATHER_K_CACHE_CUTEDSL_KERNEL,
)
dequantize_and_gather_k_cache_cutedsl(
out, k_cache, seq_lens, gather_lens, block_table, block_size, offset
_DEQUANT_GATHER_K_CACHE_CUTEDSL_KERNEL(
out=out,
k_cache=k_cache,
seq_lens=seq_lens,
gather_lens=gather_lens,
block_table=block_table,
block_size=block_size,
offset=offset,
)
return