[Bugfix][Model] Fix Aria expert weight names and layout (#57487)

Signed-off-by: Jiang Wu <245355730+JiangLLM@users.noreply.github.com>
Co-authored-by: Jiang Wu <245355730+JiangLLM@users.noreply.github.com>
This commit is contained in:
Jiang Wu
2026-09-18 21:06:30 -07:00
committed by GitHub
co-authored by Jiang Wu
parent 9d75dd4544
commit 01c7bf8813
2 changed files with 109 additions and 4 deletions
+100
View File
@@ -0,0 +1,100 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
import torch
from transformers import AriaTextConfig
from transformers.models.aria.modeling_aria import AriaTextMoELayer as HFMoE
from vllm.config import VllmConfig
from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import ( # noqa: E501
CompressedTensorsConfig,
)
from vllm.model_executor.models.aria import AriaTextModel, AriaTextMoELayer
from vllm.model_executor.models.utils import AutoWeightsLoader
@pytest.mark.cpu_test
@pytest.mark.usefixtures("dist_init")
@pytest.mark.parametrize(
"checkpoint_name,loaded_name,param_name",
[
(
"experts.fc1.weight",
"experts.w13_weight",
"experts.routed_experts.w13_weight",
),
(
"experts.fc2.weight",
"experts.w2_weight",
"experts.routed_experts.w2_weight",
),
(
"shared_experts.down_proj.weight",
"shared_experts.down_proj.weight",
"shared_experts.down_proj.weight",
),
],
)
def test_aria_expert_weights_load_with_checkpoint_layout(
checkpoint_name: str, loaded_name: str, param_name: str
):
"""Real Aria checkpoint weights must reach the right parameter and layout."""
config = AriaTextConfig(
hidden_size=32,
intermediate_size=64,
moe_num_experts=2,
moe_topk=1,
moe_num_shared_experts=1,
)
checkpoint_weight = HFMoE(config).state_dict()[checkpoint_name]
checkpoint_weight.copy_(
torch.arange(
checkpoint_weight.numel(), dtype=checkpoint_weight.dtype
).reshape_as(checkpoint_weight)
)
expected = checkpoint_weight.clone()
if checkpoint_name.startswith("experts."):
expected = expected.transpose(-1, -2)
layer = AriaTextMoELayer(config, quant_config=None, prefix="mlp")
loaded = AutoWeightsLoader(layer).load_weights(
[(checkpoint_name, checkpoint_weight)], mapper=AriaTextModel.hf_to_vllm_mapper
)
assert loaded == {loaded_name}
torch.testing.assert_close(
dict(layer.named_parameters())[param_name], expected, rtol=0, atol=0
)
@pytest.mark.cpu_test
def test_aria_quant_config_renames_expert_modules():
quant_config = CompressedTensorsConfig(
target_scheme_map={},
ignore=[
"model.layers.0.mlp.experts.fc1",
"model.layers.0.mlp.experts.fc2",
],
quant_format="pack-quantized",
)
AriaTextModel.__new__(
AriaTextModel, vllm_config=VllmConfig(quant_config=quant_config)
)
assert quant_config.ignore == [
"model.layers.0.mlp.experts.gate_up_proj",
"model.layers.0.mlp.experts.down_proj",
]
@pytest.mark.cpu_test
@pytest.mark.parametrize(
"name,expected", [("fc1", "gate_up_proj"), ("fc2", "down_proj")]
)
def test_aria_expert_scale_names(name: str, expected: str):
suffixes = ["weight_scale", "input_scale"]
names = [f"model.layers.0.mlp.experts.{name}.{suffix}" for suffix in suffixes]
assert AriaTextModel.hf_to_vllm_mapper.apply_list(names) == [
f"model.layers.0.mlp.experts.{expected}.{suffix}" for suffix in suffixes
]
+9 -4
View File
@@ -3,6 +3,7 @@
from collections.abc import Iterable, Mapping, Sequence
from typing import Annotated, Literal
import regex as re
import torch
import torch.nn as nn
from transformers import AriaConfig, AriaTextConfig, BatchFeature
@@ -222,6 +223,7 @@ class AriaTextMoELayer(nn.Module):
intermediate_size=config.intermediate_size,
quant_config=quant_config,
prefix=f"{prefix}.experts",
is_fused_checkpoint_transposed=True,
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
@@ -267,14 +269,17 @@ class AriaTextModel(LlamaModel, SupportsQuant):
"gate_up_proj": ["gate_proj", "up_proj"],
}
# Aria packs all experts into single (transposed) fc1/fc2 tensors, which is
# exactly the pre-fused checkpoint layout FusedMoE self-loads once fc1/fc2
# are renamed to the fused gate_up_proj/down_proj names.
# The fused expert loader expects names without the .weight suffix.
hf_to_vllm_mapper = LlamaModel.hf_to_vllm_mapper | WeightsMapper(
orig_to_new_regex={
re.compile(r"experts\.fc1\.weight$"): "experts.gate_up_proj",
re.compile(r"experts\.fc2\.weight$"): "experts.down_proj",
},
# Quantization configs also use the expert module names.
orig_to_new_substr={
"experts.fc1": "experts.gate_up_proj",
"experts.fc2": "experts.down_proj",
}
},
)
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):