[Bugfix] fix qwen3.5 ep weight loading (#45002)

Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
This commit is contained in:
Jiangyun Zhu
2026-06-09 13:55:30 -07:00
committed by GitHub
parent c9c1540e61
commit ca4cfd8731
3 changed files with 32 additions and 14 deletions
@@ -7,3 +7,4 @@ server_args: >-
--max-model-len 4096
--data-parallel-size 2
--enable-expert-parallel
--no-enable-flashinfer-autotune
+17 -7
View File
@@ -36,6 +36,9 @@ from vllm.distributed import (
get_pp_group,
)
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import (
fused_moe_make_expert_params_mapping,
)
from vllm.model_executor.layers.layernorm import (
GemmaRMSNorm as Qwen3_5RMSNorm,
)
@@ -294,13 +297,20 @@ class Qwen3_5Model(Qwen3NextModel):
loaded_params: set[str] = set()
expert_params_mapping = self.get_expert_mapping()
is_fused_expert = False
base_layer = (
"base_layer." if any(".base_layer." in name for name in params_dict) else ""
)
fused_expert_params_mapping = [
(f"experts.{base_layer}w13_weight", "experts.gate_up_proj", 0, "w1"),
(f"experts.{base_layer}w2_weight", "experts.down_proj", 0, "w2"),
]
fused_expert_params_mapping: list[tuple[str, str, int, str]] = []
for param_name, ckpt_name, _, shard_id in fused_moe_make_expert_params_mapping(
self,
ckpt_gate_proj_name="gate_up_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="gate_up_proj",
num_experts=1,
):
if shard_id == "w3":
continue
parts = ckpt_name.split(".")
fused_expert_params_mapping.append(
(f"{param_name}weight", f"{parts[0]}.{parts[2]}", 0, shard_id)
)
num_experts = (
self.config.num_experts if hasattr(self.config, "num_experts") else 0
)
+14 -7
View File
@@ -209,13 +209,20 @@ class Qwen3_5MultiTokenPredictor(nn.Module):
params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
is_fused_expert = False
base_layer = (
"base_layer." if any(".base_layer." in name for name in params_dict) else ""
)
fused_expert_params_mapping = [
(f"experts.{base_layer}w13_weight", "experts.gate_up_proj", 0, "w1"),
(f"experts.{base_layer}w2_weight", "experts.down_proj", 0, "w2"),
]
fused_expert_params_mapping: list[tuple[str, str, int, str]] = []
for param_name, ckpt_name, _, shard_id in fused_moe_make_expert_params_mapping(
self,
ckpt_gate_proj_name="gate_up_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="gate_up_proj",
num_experts=1,
):
if shard_id == "w3":
continue
parts = ckpt_name.split(".")
fused_expert_params_mapping.append(
(f"{param_name}weight", f"{parts[0]}.{parts[2]}", 0, shard_id)
)
num_experts = (
self.config.num_experts if hasattr(self.config, "num_experts") else 0
)