[Bugfix] Wait for the linear bias before layerwise online processing (#49805)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Harry Mellor
2026-07-25 17:56:32 +00:00
committed by GitHub
co-authored by Claude Opus 5
parent 26d725c334
commit 9321aff536
4 changed files with 67 additions and 10 deletions
@@ -10,9 +10,11 @@ from torch.nn.parameter import UninitializedParameter
import vllm.model_executor.model_loader.reload.meta as reload_meta
from vllm.model_executor.layers.linear import QKVParallelLinear
from vllm.model_executor.layers.quantization.base_config import QuantizeMethodBase
from vllm.model_executor.model_loader.reload.layerwise import (
finalize_layerwise_reload,
initialize_layerwise_reload,
initialize_online_processing,
record_metadata_for_reloading,
)
from vllm.model_executor.model_loader.reload.meta import (
@@ -278,6 +280,58 @@ def test_layerwise_reload_composed_loader_does_not_drop_params(monkeypatch):
assert torch.equal(layer.D, loaded["D"])
class _RecordingQuantMethod(QuantizeMethodBase):
"""Records the layer's bias at the moment processing runs."""
uses_meta_device = True
def __init__(self):
self.bias_at_process = None
def create_weights(self, layer, *weight_args, **extra_weight_attrs):
pass
def apply(self, layer, *args, **kwargs):
raise NotImplementedError
def process_weights_after_loading(self, layer):
self.bias_at_process = layer.bias.detach().clone()
class _LateBiasLayer(torch.nn.Module):
"""Mimics an online-quantized linear: `weight` is created on meta by
`create_weights()`, which wraps the loaders, and the linear base registers
`bias` afterwards."""
def __init__(self, quant_method):
super().__init__()
self.quant_method = quant_method
weight = torch.nn.Parameter(torch.empty(4, 2, device="meta"))
weight.weight_loader = default_weight_loader
self.register_parameter("weight", weight)
initialize_online_processing(self)
bias = torch.nn.Parameter(torch.zeros(4))
bias.weight_loader = default_weight_loader
self.register_parameter("bias", bias)
def test_online_processing_waits_for_late_registered_bias():
# Regression test: `bias` is skipped by the meta device paths, but it is
# still loaded by a weight loader. Excluding it from the processing trigger
# finalized the layer one load early, so the trailing bias was written into
# an already-processed layer (e.g. over FP8 Marlin's permuted bias).
quant_method = _RecordingQuantMethod()
layer = _LateBiasLayer(quant_method)
loaded_bias = torch.full((4,), 3.0)
layer.weight.weight_loader(layer.weight, torch.full((4, 2), 2.0))
assert quant_method.bias_at_process is None
layer.bias.weight_loader(layer.bias, loaded_bias)
assert quant_method.bias_at_process is not None
assert torch.equal(quant_method.bias_at_process, loaded_bias)
def test_layerwise_reload_skips_non_persistent_parameter_alias_buffers(monkeypatch):
layer = _AliasedBufferLayer()
model = torch.nn.Sequential(layer)
@@ -14,7 +14,7 @@ from vllm.model_executor.layers.quantization.base_config import QuantizeMethodBa
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from .meta import (
SKIP_TENSORS,
SKIP_LOAD_TENSORS,
capture_layer_to_meta,
get_numel_loaded,
materialize_layer,
@@ -140,7 +140,7 @@ def _wrap_parameters_weight_loader(layer: torch.nn.Module) -> None:
"""Wrap each parameter's weight loader."""
# Note that nested wrapping will occur for shared tensors
for name, tensor in get_layer_tensors(layer).items():
if name in SKIP_TENSORS:
if name in SKIP_LOAD_TENSORS:
continue
if _get_weight_loader(tensor).__name__ != "online_process_loader":
tensor.weight_loader = make_online_process_loader(layer, name)
@@ -20,20 +20,23 @@ __all__ = [
"get_numel_loaded",
]
# Modules whose tensors are never moved to, or materialized from, the meta device.
SKIP_MODULES: set[str] = {"HadamardTransform"}
SKIP_TENSORS: set[str] = {
# Tensors never loaded by a weight loader, so the layerwise trigger ignores them.
SKIP_LOAD_TENSORS: set[str] = {
"_expert_map",
"expert_mask",
"expert_global_to_physical",
"expert_physical_to_global",
"expert_local_to_global",
"e_score_correction_bias",
# Built after create_weights(), so it is not tracked by the layerwise-reload
# trigger and would be re-materialized into uninitialized memory. Skip it.
"bias",
}
# Tensors which are never moved to, or materialized from, the meta device.
# `bias` is built after create_weights(), so it is never on meta to begin with.
SKIP_TENSORS: set[str] = SKIP_LOAD_TENSORS | {"bias"}
def to_meta_tensor(tensor: torch.Tensor) -> torch.Tensor:
"""Convert a tensor to a meta tensor while preserving class and attributes."""
@@ -33,15 +33,15 @@ def get_layer_params_buffers(layer: torch.nn.Module) -> LayerTensors:
def get_layer_size(layer: torch.nn.Module) -> int:
"""Calculate total number of elements across loadable tensors in a layer.
Excludes SKIP_TENSORS (e.g. _expert_map) which are never moved to meta
device and never loaded via weight_loader during layerwise reload.
Excludes SKIP_LOAD_TENSORS (e.g. _expert_map) which are never loaded via
weight_loader during layerwise reload.
"""
from .meta import SKIP_TENSORS
from .meta import SKIP_LOAD_TENSORS
return sum(
tensor.numel()
for name, tensor in get_layer_tensors(layer).items()
if name not in SKIP_TENSORS
if name not in SKIP_LOAD_TENSORS
)