[1/N][warmup][DSv4] Migrate generic MLA metadata and indexing kernels (#50175)

Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>
Co-authored-by: Codex <codex@openai.com>
This commit is contained in:
Roberto L. Castro
2026-08-31 23:26:05 -04:00
committed by GitHub
co-authored by Codex
parent 188716ace7
commit e16b5e518d
17 changed files with 1261 additions and 617 deletions
@@ -604,6 +604,33 @@ class MLAAttention(nn.Module, AttentionLayerBase):
self.use_sparse = use_sparse
if vllm_config.kernel_config.enable_jit_warmup:
backend_name = self.attn_backend.get_name()
if backend_name in (
"FLASHMLA_SPARSE",
"FLASHINFER_MLA_SPARSE",
"FLASHINFER_MLA_SPARSE_SM120",
"DEEPSEEK_V32_INDEXER",
):
from vllm.v1.attention.backends.mla.compressor_utils import (
_COMPRESSED_SLOT_MAPPING_KERNEL,
)
from vllm.v1.attention.backends.mla.indexer import (
_BUILD_PREFILL_CHUNK_METADATA_KERNEL,
_PREPARE_UNIFORM_DECODE_KERNEL,
)
_COMPRESSED_SLOT_MAPPING_KERNEL.register_warmup()
_PREPARE_UNIFORM_DECODE_KERNEL.register_warmup()
_BUILD_PREFILL_CHUNK_METADATA_KERNEL.register_warmup()
if backend_name != "DEEPSEEK_V32_INDEXER":
from vllm.v1.attention.backends.mla.sparse_swa import (
_COMPUTE_PREFILL_METADATA_KERNEL,
)
_COMPUTE_PREFILL_METADATA_KERNEL.register_warmup()
self.dcp_manager: MLADCPManager | None = None
if self.impl.dcp_world_size > 1:
query_dtype = (
@@ -652,6 +679,26 @@ class MLAAttention(nn.Module, AttentionLayerBase):
def bind_kv_cache(self, kv_cache: torch.Tensor) -> None:
# [B, H=1, N, C] -> [B, N, C]
self.kv_cache = kv_cache.squeeze(1)
if (
self._vllm_config.kernel_config.enable_jit_warmup
and self.attn_backend.get_name()
in (
"FLASHMLA_SPARSE",
"FLASHINFER_MLA_SPARSE",
"FLASHINFER_MLA_SPARSE_SM120",
"DEEPSEEK_V32_INDEXER",
)
):
from vllm.v1.attention.backends.mla.sparse_utils import (
_CONVERT_REQ_INDEX_TO_GLOBAL_INDEX_KERNEL,
)
row_width = self.kv_cache.shape[-1]
assert self.kv_cache.stride(0) % row_width == 0
_CONVERT_REQ_INDEX_TO_GLOBAL_INDEX_KERNEL.register_warmup(
self._vllm_config,
block_stride_rows=self.kv_cache.stride(0) // row_width,
)
@property
def chunked_prefill_workspace_size(self) -> int: