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

[Bugfix] Follow vLLM Qwen-Moe/VL and KV Connector change to fix broken CI (#2181)

Browse Source 
### What this PR does / why we need it?
This pr fix broken CI:
1. Fix the
ee2eb6ecd8
changes, in this commit, they fused the gate and up projections in the
vision MLP, This can improve performance by reducing one matrix
multiplication. so, this pr do the following things:
- Specify that the two linear layers are fused as `mlp.gate_up_proj`
when loading the weights.
    - Use a SiluAndMul activation function.
2. Fix
aefeea0fde,
Update ModelRunnerOutput parameters to adapt to its changes
3. Fix
[vllm-commit](https://github.com/vllm-project/vllm/pull/20815/files#diff-3ffb829a39ab2b3e4706aa28f5e476815f36c3a87b98d6a66514ebedc8f3ffb4R354-R356),
fix qwen moe
### Does this PR introduce _any_ user-facing change?

### How was this patch tested?


- vLLM version: v0.10.0
- vLLM main:
fed5849d3f

---------

Signed-off-by: wangli <wangli858794774@gmail.com>
This commit is contained in:
Li Wang
2025-08-04 21:37:50 +08:00
committed by GitHub
parent e38fab011d
commit ad366bf908
8 changed files with 137 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
vllm_ascend/models/qwen2_5_vl.py
View File

@@ -18,7 +18,7 @@
# limitations under the License.
from functools import partial
from typing import Callable, Iterable, Optional, Set, Tuple
from typing import Callable, Iterable, Optional, Set, Tuple, Union
import torch
import torch.nn as nn
@@ -30,7 +30,8 @@ from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import (
from vllm.config import VllmConfig
from vllm.distributed import parallel_state
from vllm.distributed import utils as dist_utils
from vllm.model_executor.layers.activation import _ACTIVATION_REGISTRY
from vllm.model_executor.layers.activation import (_ACTIVATION_REGISTRY,
get_act_and_mul_fn)
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
@@ -42,6 +43,8 @@ from vllm.model_executor.models.qwen2_5_vl import (
from vllm.model_executor.models.utils import maybe_prefix
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm_ascend.utils import vllm_version_is
MIN_PAD_SIZE = 64 # min_size to pad weight
MAX_PAD_SIZE = 128 # max_size to pad weight
@@ -197,12 +200,16 @@ class AscendQwen2_5_VisionTransformer(Qwen2_5_VisionTransformer):
in_channels=vision_config.in_channels,
hidden_size=self.hidden_size,
)
act_fn = get_act_and_mul_fn(vision_config.hidden_act)
if vllm_version_is("0.10.0"):
act_fn = _ACTIVATION_REGISTRY[vision_config.hidden_act]
self.blocks = nn.ModuleList([
AscendQwen2_5_VisionBlock(
dim=self.hidden_size,
num_heads=self.num_heads,
mlp_hidden_dim=vision_config.intermediate_size,
act_fn=_ACTIVATION_REGISTRY[vision_config.hidden_act],
act_fn=act_fn,
norm_layer=norm_layer,
quant_config=quant_config,
prefix=f"{prefix}.blocks.{layer_idx}")
@@ -291,12 +298,17 @@ class AscendQwen2_5_VisionTransformer(Qwen2_5_VisionTransformer):
def load_weights(self, weights: Iterable[Tuple[str,
torch.Tensor]]) -> Set[str]:
stacked_params_mapping = [
stacked_params_mapping: list[tuple[str, str, Union[str, int]]] = [
# (param_name, shard_name, shard_id)
("qkv_proj", "q_proj", "q"),
("qkv_proj", "k_proj", "k"),
("qkv_proj", "v_proj", "v"),
]
if not vllm_version_is("0.10.0"):
stacked_params_mapping.extend([
("mlp.gate_up_proj.", "mlp.gate_proj.", 0),
("mlp.gate_up_proj.", "mlp.up_proj.", 1),
])
params_dict = dict(self.named_parameters(remove_duplicate=False))
loaded_params: Set[str] = set()
for name, loaded_weight in weights:

10
vllm_ascend/models/qwen2_5_vl_without_padding.py
View File

@@ -30,7 +30,8 @@ from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import (
from vllm.config import VllmConfig
from vllm.distributed import parallel_state
from vllm.distributed import utils as dist_utils
from vllm.model_executor.layers.activation import _ACTIVATION_REGISTRY
from vllm.model_executor.layers.activation import (_ACTIVATION_REGISTRY,
get_act_and_mul_fn)
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.models.qwen2_5_vl import (
@@ -42,6 +43,7 @@ from vllm.model_executor.models.utils import maybe_prefix
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm_ascend.models.qwen2_5_vl import AscendQwen2_5_VisionRotaryEmbedding
from vllm_ascend.utils import vllm_version_is
class AscendQwen2_5_VisionAttention_Without_Padding(Qwen2_5_VisionAttention):
@@ -171,12 +173,16 @@ class AscendQwen2_5_VisionTransformer_Without_Padding(Qwen2_5_VisionTransformer
in_channels=vision_config.in_channels,
hidden_size=self.hidden_size,
)
act_fn = get_act_and_mul_fn(vision_config.hidden_act)
if vllm_version_is("0.10.0"):
act_fn = _ACTIVATION_REGISTRY[vision_config.hidden_act]
self.blocks = nn.ModuleList([
AscendQwen2_5_VisionBlock_Without_Padding(
dim=self.hidden_size,
num_heads=self.num_heads,
mlp_hidden_dim=vision_config.intermediate_size,
act_fn=_ACTIVATION_REGISTRY[vision_config.hidden_act],
act_fn=act_fn,
norm_layer=norm_layer,
quant_config=quant_config,
prefix=f"{prefix}.blocks.{layer_idx}")

2
vllm_ascend/models/qwen3_moe.py
View File

@@ -100,6 +100,8 @@ class CustomQwen3MoeModel(Qwen3MoeModel):
cache_config = vllm_config.cache_config
quant_config = vllm_config.quant_config
parallel_config = vllm_config.parallel_config
self.num_redundant_experts = parallel_config.num_redundant_experts
self.padding_idx = config.pad_token_id
self.vocab_size = config.vocab_size
self.config = config