EngineX-Ascend/enginex-ascend-910-vllm
11
0
Fork 1
Code Issues Pull Requests Actions Projects Releases Wiki Activity

init v0.23.0

Browse Source 
Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>
This commit is contained in:
Sun Ruoxi
2026-08-27 15:11:51 +08:00
parent b582a8e7d1
commit 7f8a1b1f7a
2849 changed files with 712887 additions and 22001 deletions
Download Patch File Download Diff File Expand all files Collapse all files

160
vllm_ascend/lora/utils.py
View File

@@ -1,91 +1,25 @@
from typing import Optional
import vllm
from torch import nn
from transformers import PretrainedConfig
from vllm.config import LoRAConfig
from vllm.lora.layers import (ColumnParallelLinearWithLoRA,
MergedColumnParallelLinearWithLoRA,
MergedQKVParallelLinearWithLoRA,
QKVParallelLinearWithLoRA,
RowParallelLinearWithLoRA,
VocabParallelEmbeddingWithLoRA)
from vllm.lora.layers.utils import _not_fully_sharded_can_replace
from vllm.lora.layers import (
MergedQKVParallelLinearWithLoRA,
MergedQKVParallelLinearWithShardedLoRA,
QKVParallelLinearWithLoRA,
QKVParallelLinearWithShardedLoRA,
)
from vllm.lora.layers.utils import _fully_sharded_can_replace, _not_fully_sharded_can_replace
from vllm_ascend.ops.linear import (AscendColumnParallelLinear,
AscendMergedColumnParallelLinear,
AscendQKVParallelLinear,
AscendRowParallelLinear)
from vllm_ascend.ops.vocab_parallel_embedding import \
AscendVocabParallelEmbedding
class AscendColumnParallelLinearWithLoRA(ColumnParallelLinearWithLoRA):
@classmethod
def can_replace_layer(
cls,
source_layer: nn.Module,
lora_config: LoRAConfig,
packed_modules_list: list,
model_config: Optional[PretrainedConfig],
) -> bool:
return type(source_layer) is AscendColumnParallelLinear
class AscendMergedColumnParallelLinearWithLoRA(
MergedColumnParallelLinearWithLoRA):
@classmethod
def can_replace_layer(
cls,
source_layer: nn.Module,
lora_config: LoRAConfig,
packed_modules_list: list,
model_config: Optional[PretrainedConfig],
) -> bool:
return type(source_layer) is AscendMergedColumnParallelLinear
class AscendRowParallelLinearWithLoRA(RowParallelLinearWithLoRA):
@classmethod
def can_replace_layer(
cls,
source_layer: nn.Module,
lora_config: LoRAConfig,
packed_modules_list: list,
model_config: Optional[PretrainedConfig],
) -> bool:
return type(source_layer) is AscendRowParallelLinear
class AscendVocabParallelEmbeddingWithLoRA(VocabParallelEmbeddingWithLoRA):
@classmethod
def can_replace_layer(
cls,
source_layer: nn.Module,
lora_config: LoRAConfig,
packed_modules_list: list,
model_config: Optional[PretrainedConfig],
) -> bool:
return type(source_layer) is AscendVocabParallelEmbedding
from vllm_ascend.lora.fused_moe import (
AscendFusedMoE3DWithLoRA,
AscendFusedMoEWithLoRA,
)
from vllm_ascend.ops.linear import (
AscendQKVParallelLinear,
)
class AscendQKVParallelLinearWithLoRA(QKVParallelLinearWithLoRA):
@classmethod
@_not_fully_sharded_can_replace
def can_replace_layer(cls, source_layer: nn.Module,
lora_config: LoRAConfig, packed_modules_list: list,
model_config: Optional[PretrainedConfig]) -> bool:
return type(source_layer) is AscendQKVParallelLinear and len(
packed_modules_list) == 1
class AscendMergedQKVParallelLinearWithLoRA(MergedQKVParallelLinearWithLoRA):
@classmethod
@_not_fully_sharded_can_replace
def can_replace_layer(
@@ -93,18 +27,62 @@ class AscendMergedQKVParallelLinearWithLoRA(MergedQKVParallelLinearWithLoRA):
source_layer: nn.Module,
lora_config: LoRAConfig,
packed_modules_list: list,
model_config: Optional[PretrainedConfig],
model_config: PretrainedConfig | None,
) -> bool:
return (type(source_layer) is AscendQKVParallelLinear
and len(packed_modules_list) == 3)
return type(source_layer) is AscendQKVParallelLinear and len(packed_modules_list) == 1
class AscendMergedQKVParallelLinearWithLoRA(MergedQKVParallelLinearWithLoRA):
@classmethod
@_not_fully_sharded_can_replace
def can_replace_layer(
cls,
source_layer: nn.Module,
lora_config: LoRAConfig,
packed_modules_list: list,
model_config: PretrainedConfig | None,
) -> bool:
return type(source_layer) is AscendQKVParallelLinear and len(packed_modules_list) == 3
class AscendMergedQKVParallelLinearWithShardedLoRA(MergedQKVParallelLinearWithShardedLoRA):
@classmethod
@_fully_sharded_can_replace
def can_replace_layer(
cls,
source_layer: nn.Module,
lora_config: LoRAConfig,
packed_modules_list: list,
model_config: PretrainedConfig | None = None,
) -> bool:
return type(source_layer) is AscendQKVParallelLinear and len(packed_modules_list) == 3
class AscendQKVParallelLinearWithShardedLoRA(QKVParallelLinearWithShardedLoRA):
@classmethod
@_fully_sharded_can_replace
def can_replace_layer(
cls,
source_layer: nn.Module,
lora_config: LoRAConfig,
packed_modules_list: list,
model_config: PretrainedConfig | None = None,
) -> bool:
return type(source_layer) is AscendQKVParallelLinear and len(packed_modules_list) == 1
def refresh_all_lora_classes():
vllm.lora.utils._all_lora_classes.add(AscendColumnParallelLinearWithLoRA)
vllm.lora.utils._all_lora_classes.add(
AscendMergedColumnParallelLinearWithLoRA)
vllm.lora.utils._all_lora_classes.add(AscendRowParallelLinearWithLoRA)
vllm.lora.utils._all_lora_classes.add(AscendVocabParallelEmbeddingWithLoRA)
vllm.lora.utils._all_lora_classes.add(AscendQKVParallelLinearWithLoRA)
vllm.lora.utils._all_lora_classes.add(
AscendMergedQKVParallelLinearWithLoRA)
ascend_classes = (
AscendQKVParallelLinearWithLoRA,
AscendMergedQKVParallelLinearWithLoRA,
AscendMergedQKVParallelLinearWithShardedLoRA,
AscendQKVParallelLinearWithShardedLoRA,
AscendFusedMoEWithLoRA,
AscendFusedMoE3DWithLoRA,
)
# vLLM #35077 changed _all_lora_classes from set to ordered tuple.
# Append the Ascend classes in a deterministic order.
vllm.lora.utils._all_lora_classes = (
*ascend_classes,
*vllm.lora.utils._all_lora_classes,
)