Add FLASHINFER_MLA to test_mla_backends and add B200 CI run (#27663)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Matthew Bonanni
2025-10-31 11:12:19 -07:00
committed by GitHub
parent 5e8862e9e0
commit f29aeb5a25
4 changed files with 208 additions and 64 deletions
@@ -6,7 +6,7 @@ from typing import ClassVar
import torch
from flashinfer.decode import trtllm_batch_decode_with_kv_cache_mla
from vllm.attention.backends.abstract import AttentionLayer, AttentionType
from vllm.attention.backends.abstract import AttentionLayer, AttentionType, MultipleOf
from vllm.logger import init_logger
from vllm.v1.attention.backends.mla.common import (
MLACommonBackend,
@@ -40,6 +40,10 @@ class FlashInferMLABackend(MLACommonBackend):
def get_builder_cls() -> type["FlashInferMLAMetadataBuilder"]:
return FlashInferMLAMetadataBuilder
@classmethod
def get_supported_kernel_block_size(cls) -> list[int | MultipleOf]:
return [32, 64]
g_fi_workspace = torch.zeros(
FLASHINFER_MLA_WORKSPACE_BUFFER_SIZE,