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

[CI] Fix FusedMoEConfig and input batch failure to recover CI (#1602)

Browse Source 
Make CI happy

1.
c1909e7e8c
changed moeConfig init way
2.
48fb076cbc
changed input batch logic.

This PR address these change to vllm-ascend.

Closes: https://github.com/vllm-project/vllm-ascend/issues/1600

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-07-03 18:36:17 +08:00
committed by GitHub
parent d96da1f00c
commit a45dfde283
11 changed files with 173 additions and 134 deletions
Download Patch File Download Diff File Expand all files Collapse all files

62
vllm_ascend/ops/fused_moe.py
View File

@@ -26,11 +26,11 @@ from vllm.config import get_current_vllm_config
from vllm.distributed import (GroupCoordinator, get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_reduce)
from vllm.distributed.parallel_state import get_dp_group, get_tp_group
from vllm.distributed.parallel_state import (get_dp_group, get_tp_group,
get_world_group)
from vllm.forward_context import get_forward_context
from vllm.model_executor.layers.fused_moe.layer import (
FusedMoE, FusedMoEParallelConfig, MoEConfig, UnquantizedFusedMoEMethod,
determine_expert_map)
FusedMoE, UnquantizedFusedMoEMethod, determine_expert_map)
from vllm.model_executor.layers.quantization.base_config import \
QuantizationConfig
@@ -40,7 +40,16 @@ from vllm_ascend.distributed.parallel_state import get_ep_group, get_etp_group
from vllm_ascend.ops.expert_load_balancer import ExpertLoadBalancer
from vllm_ascend.utils import (FusedMoEState, dispose_tensor,
get_fused_moe_state, is_310p, npu_stream_switch,
npu_wait_tensor)
npu_wait_tensor, vllm_version_is)
if vllm_version_is("0.9.1"):
from vllm.model_executor.layers.fused_moe.layer import \
FusedMoEParallelConfig
from vllm.model_executor.layers.fused_moe.layer import \
MoEConfig as FusedMoEConfig
else:
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig, FusedMoEParallelConfig)
MOE_ALL2ALL_BUFFER: bool = envs_ascend.MOE_ALL2ALL_BUFFER
@@ -933,7 +942,7 @@ def select_experts(
class AscendUnquantizedFusedMoEMethod(UnquantizedFusedMoEMethod):
def __init__(self, moe: MoEConfig = None):
def __init__(self, moe: FusedMoEConfig = None):
super().__init__(moe=moe)
vllm_config = get_current_vllm_config()
@@ -1110,13 +1119,21 @@ class AscendFusedMoE(FusedMoE):
vllm_config = get_current_vllm_config()
self.moe_parallel_config: FusedMoEParallelConfig = (
FusedMoEParallelConfig.make(
if vllm_version_is("0.9.1"):
self.moe_parallel_config = FusedMoEParallelConfig.make(
tp_size_=(tp_size if tp_size is not None else
get_tensor_model_parallel_world_size()),
dp_size_=(dp_size if dp_size is not None else
get_dp_group().world_size),
vllm_parallel_config=vllm_config.parallel_config))
vllm_parallel_config=vllm_config.parallel_config)
else:
self.moe_parallel_config = FusedMoEParallelConfig.make(
tp_size_=(tp_size if tp_size is not None else
get_tensor_model_parallel_world_size()),
dp_size_=(dp_size if dp_size is not None else
get_dp_group().world_size),
world_size_=get_world_group().world_size,
vllm_parallel_config=vllm_config.parallel_config)
self.top_k = top_k
self.num_experts = num_experts
@@ -1167,15 +1184,26 @@ class AscendFusedMoE(FusedMoE):
raise ValueError("Only softmax scoring function is supported for "
"non-grouped topk.")
moe = MoEConfig(
num_experts=self.global_num_experts,
experts_per_token=top_k,
hidden_dim=hidden_size,
num_local_experts=self.local_num_experts,
moe_parallel_config=self.moe_parallel_config,
# TODO (bnell): this needs to be fixed for quantized types.
in_dtype=params_dtype,
)
if vllm_version_is("0.9.1"):
moe = FusedMoEConfig(
num_experts=self.global_num_experts,
experts_per_token=top_k,
hidden_dim=hidden_size,
num_local_experts=self.local_num_experts,
moe_parallel_config=self.moe_parallel_config,
# TODO (bnell): this needs to be fixed for quantized types.
in_dtype=params_dtype,
)
else:
moe = FusedMoEConfig.make(
num_experts=self.global_num_experts,
experts_per_token=top_k,
hidden_dim=hidden_size,
num_local_experts=self.local_num_experts,
moe_parallel_config=self.moe_parallel_config,
# TODO (bnell): this needs to be fixed for quantized types.
in_dtype=params_dtype,
quant_config=quant_config)
if quant_config is None:
self.quant_method = AscendUnquantizedFusedMoEMethod(moe)