iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

[Bugfix] Fix weight loading issues caused by #41184 (#45054)

Browse Source 
Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2026-06-10 01:20:13 -04:00
committed by GitHub
parent 2c9c07c85e
commit f4966f8b3d
5 changed files with 55 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files
vllm/model_executor/models/aria.py
+2 -2
View File
@@ -335,8 +335,8 @@ class AriaTextModel(LlamaModel, SupportsQuant):
packed_modules_mapping = {
"qkv_proj": ["q_proj", "k_proj", "v_proj"],
"gate_up_proj": ["gate_proj", "up_proj"],
"experts.w13_weight": ["experts.fc1.weight"],
"experts.w2_weight": ["experts.fc2.weight"],
"experts.routed_experts.w13_weight": ["experts.fc1.weight"],
"experts.routed_experts.w2_weight": ["experts.fc2.weight"],
}
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
vllm/model_executor/models/qwen3_vl_moe.py
+12 -2
View File
@@ -187,8 +187,18 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
"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"),
(
f"experts.routed_experts.{base_layer}w13_weight",
"experts.gate_up_proj",
0,
"w1",
),
(
f"experts.routed_experts.{base_layer}w2_weight",
"experts.down_proj",
0,
"w2",
),
]
num_experts = self.config.num_experts
for name, loaded_weight in weights:
vllm/model_executor/models/step3_text.py
+15 -3
View File
@@ -422,9 +422,21 @@ class Step3TextModel(nn.Module):
)
expert_params_mapping = [
(f".moe.experts.{base_layer}w13_weight", ".moe.gate_proj.weight", "w1"),
(f".moe.experts.{base_layer}w13_weight", ".moe.up_proj.weight", "w3"),
(f".moe.experts.{base_layer}w2_weight", ".moe.down_proj.weight", "w2"),
(
f".moe.experts.routed_experts.{base_layer}w13_weight",
".moe.gate_proj.weight",
"w1",
),
(
f".moe.experts.routed_experts.{base_layer}w13_weight",
".moe.up_proj.weight",
"w3",
),
(
f".moe.experts.routed_experts.{base_layer}w2_weight",
".moe.down_proj.weight",
"w2",
),
]
disable_moe_stacked_params = [data[1] for data in expert_params_mapping]
vllm/model_executor/models/step3p5.py
+24 -12
View File
@@ -635,36 +635,48 @@ class Step3p5Model(nn.Module):
# Old packed 3D format: .moe.gate_proj.weight [num_experts, out, in]
expert_params_mapping = [
(f".moe.experts.{base_layer}w13_weight", ".moe.gate_proj.weight", "w1"),
(f".moe.experts.{base_layer}w13_weight", ".moe.up_proj.weight", "w3"),
(f".moe.experts.{base_layer}w2_weight", ".moe.down_proj.weight", "w2"),
(
f".moe.experts.{base_layer}w13_weight_scale_2",
f".moe.experts.routed_experts.{base_layer}w13_weight",
".moe.gate_proj.weight",
"w1",
),
(
f".moe.experts.routed_experts.{base_layer}w13_weight",
".moe.up_proj.weight",
"w3",
),
(
f".moe.experts.routed_experts.{base_layer}w2_weight",
".moe.down_proj.weight",
"w2",
),
(
f".moe.experts.routed_experts.{base_layer}w13_weight_scale_2",
".moe.gate_proj.weight_scale_2",
"w1",
),
(
f".moe.experts.{base_layer}w13_weight_scale_2",
f".moe.experts.routed_experts.{base_layer}w13_weight_scale_2",
".moe.up_proj.weight_scale_2",
"w3",
),
(
f".moe.experts.{base_layer}w2_weight_scale_2",
f".moe.experts.routed_experts.{base_layer}w2_weight_scale_2",
".moe.down_proj.weight_scale_2",
"w2",
),
(
f".moe.experts.{base_layer}w13_weight_scale",
f".moe.experts.routed_experts.{base_layer}w13_weight_scale",
".moe.gate_proj.weight_scale",
"w1",
),
(
f".moe.experts.{base_layer}w13_weight_scale",
f".moe.experts.routed_experts.{base_layer}w13_weight_scale",
".moe.up_proj.weight_scale",
"w3",
),
(
f".moe.experts.{base_layer}w2_weight_scale",
f".moe.experts.routed_experts.{base_layer}w2_weight_scale",
".moe.down_proj.weight_scale",
"w2",
),
@@ -672,17 +684,17 @@ class Step3p5Model(nn.Module):
# input scales are stored as moe.{gate,up,down}_proj.input_scale
# rather than the standard per-expert format handled generically.
(
f".moe.experts.{base_layer}w13_input_scale",
f".moe.experts.routed_experts.{base_layer}w13_input_scale",
".moe.gate_proj.input_scale",
"w1",
),
(
f".moe.experts.{base_layer}w13_input_scale",
f".moe.experts.routed_experts.{base_layer}w13_input_scale",
".moe.up_proj.input_scale",
"w3",
),
(
f".moe.experts.{base_layer}w2_input_scale",
f".moe.experts.routed_experts.{base_layer}w2_input_scale",
".moe.down_proj.input_scale",
"w2",
),
vllm/models/deepseek_v4/quant_config.py
+2 -6
View File
@@ -8,7 +8,6 @@ from typing import TYPE_CHECKING
from vllm.config import get_current_vllm_config
from vllm.model_executor.layers.fused_moe import (
MoERunner,
RoutedExperts,
UnquantizedFusedMoEMethod,
)
@@ -133,7 +132,7 @@ class DeepseekV4FP8Config(Fp8Config):
return None
def get_quant_method(self, layer, prefix):
if isinstance(layer, (MoERunner, RoutedExperts)):
if isinstance(layer, RoutedExperts):
if is_layer_skipped(
prefix=prefix,
ignored_layers=self.ignored_layers,
@@ -156,9 +155,6 @@ class DeepseekV4FP8Config(Fp8Config):
return super().get_quant_method(layer, prefix)
def is_mxfp4_quant(self, prefix, layer):
if (
not isinstance(layer, (MoERunner, RoutedExperts))
or self.expert_dtype != "fp4"
):
if not isinstance(layer, RoutedExperts) or self.expert_dtype != "fp4":
return False
return self.moe_quant_algo != "NVFP4"