EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

port deepseekv2 and mtp to main branch (#429)

Browse Source 
### What this PR does / why we need it?
This PR ports all the deepseek graph mode code and mtp code from v0.7.3
to the main branch
---------

Signed-off-by: SidaoY <1024863041@qq.com>
Signed-off-by: linfeng-yuan <1102311262@qq.com>
Signed-off-by: Yizhou Liu <liuyizhou5@h-partners.com>
Signed-off-by: mengwei805 <mengwei25@huawei.com>
Signed-off-by: libaokui <libaokui@huawei.com>
Signed-off-by: q00832892 <qiaoyang19@huawei.com>
Signed-off-by: ganyi <pleaplusone.gy@gmail.com>
Co-authored-by: SidaoY <1024863041@qq.com>
Co-authored-by: linfeng-yuan <1102311262@qq.com>
Co-authored-by: Yizhou Liu <liuyizhou5@h-partners.com>
Co-authored-by: mengwei805 <mengwei25@huawei.com>
Co-authored-by: libaokui <libaokui@huawei.com>
This commit is contained in:
Pleaplusone
2025-04-19 17:38:18 +08:00
committed by GitHub
parent 086423dc35
commit 1a1f9a6d89
33 changed files with 3361 additions and 315 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
vllm_ascend/platform.py
View File

@@ -160,6 +160,8 @@ class NPUPlatform(Platform):
@classmethod
def get_attn_backend_cls(cls, selected_backend, head_size, dtype,
kv_cache_dtype, block_size, use_v1, use_mla):
if use_v1 and use_mla:
return "vllm_ascend.attention.mla_v1.AscendMLABackend"
if use_v1:
return "vllm_ascend.attention.attention_v1.AscendAttentionBackend"
if use_mla:
@@ -191,3 +193,30 @@ class NPUPlatform(Platform):
model configuration.
"""
return True
@classmethod
def destroy_platform_model_parallel(cls) -> None:
from vllm_ascend.distributed.parallel_state import \
destory_ascend_model_parallel
destory_ascend_model_parallel()
@classmethod
def platform_has_backend_register(cls) -> bool:
return True
@classmethod
def platform_register_backend(cls, pg, prefix_store, group_rank,
group_size, backend_options,
timeout) -> None:
from torch.distributed import ProcessGroup, is_hccl_available
assert is_hccl_available()
import torch_npu # noqa
from torch_npu._C._distributed_c10d import ProcessGroupHCCL
backend_options = ProcessGroupHCCL.Options()
backend_options._timeout = timeout
backend_class = ProcessGroupHCCL(prefix_store, group_rank, group_size,
backend_options)
device = torch.device("npu")
backend_class._set_sequence_number_for_group()
backend_type = ProcessGroup.BackendType.CUSTOM
pg._register_backend(device, backend_type, backend_class)