From f2906b492e9173ce39920a3927699f202b2cc2e5 Mon Sep 17 00:00:00 2001 From: aoshen02 Date: Wed, 12 Aug 2026 13:53:50 +0800 Subject: [PATCH] [Bugfix] Avoid repeated layerwise reload warning scans (#49505) Signed-off-by: aoshen02 Co-authored-by: Codex --- .../model_loader/test_reload.py | 31 +++++++++++++++++++ .../model_loader/reload/layerwise.py | 4 +-- 2 files changed, 33 insertions(+), 2 deletions(-) diff --git a/tests/model_executor/model_loader/test_reload.py b/tests/model_executor/model_loader/test_reload.py index 4963f3ae5b..f32ccb118f 100644 --- a/tests/model_executor/model_loader/test_reload.py +++ b/tests/model_executor/model_loader/test_reload.py @@ -2,12 +2,14 @@ # SPDX-FileCopyrightText: Copyright contributors to the vLLM project import gc import inspect +from unittest.mock import Mock from weakref import WeakKeyDictionary, ref import pytest import torch from torch.nn.parameter import UninitializedParameter +import vllm.model_executor.model_loader.reload.layerwise as reload_layerwise import vllm.model_executor.model_loader.reload.meta as reload_meta from vllm.config import ModelConfig from vllm.model_executor.layers.attention import MMEncoderAttention @@ -571,6 +573,35 @@ def test_get_numel_loaded_caps_at_param_size(): assert num_loaded == 10 +def test_layerwise_loading_warning_only_checks_new_layers(monkeypatch): + layers = [torch.nn.Linear(16, 1, bias=False) for _ in range(2)] + + def partial_weight_loader(param, loaded_weight): + param.view(-1)[: loaded_weight.numel()].copy_(loaded_weight) + + for layer in layers: + layer.weight.requires_grad_(False) + layer.weight.weight_loader = partial_weight_loader + reload_layerwise.initialize_online_processing(layer) + + monkeypatch.setattr(reload_layerwise, "has_device_tensors", lambda _: True) + get_info_size = Mock(return_value=0) + warning_once = Mock() + monkeypatch.setattr(reload_layerwise, "get_info_size", get_info_size) + monkeypatch.setattr(reload_layerwise.logger, "warning_once", warning_once) + + reload_layerwise.LOADING_LAYERS.clear() + try: + for layer in layers: + for _ in range(3): + layer.weight.weight_loader(layer.weight, torch.ones(1)) + finally: + reload_layerwise.LOADING_LAYERS.clear() + + assert get_info_size.call_count == 2 + warning_once.assert_called_once() + + class _ComposedLoaderLayer(torch.nn.Module): """Mimics a Mamba2 mixer's equal-numel direct params (A, D, dt_bias). diff --git a/vllm/model_executor/model_loader/reload/layerwise.py b/vllm/model_executor/model_loader/reload/layerwise.py index 5aa510a624..5a73738e15 100644 --- a/vllm/model_executor/model_loader/reload/layerwise.py +++ b/vllm/model_executor/model_loader/reload/layerwise.py @@ -200,9 +200,9 @@ def make_online_process_loader(layer: torch.nn.Module, param_name: str) -> Calla return ret # Log warnings allocating excessive buffers on device - if has_device_tensors(bound_args): + if has_device_tensors(bound_args) and layer not in LOADING_LAYERS: LOADING_LAYERS.add(layer) - if len(LOADING_LAYERS) >= 2: + if len(LOADING_LAYERS) == 2: names = sorted([layer.__class__.__name__ for layer in LOADING_LAYERS]) mem_used = sum( get_info_size(LAYERWISE_INFO[layer]) for layer in LOADING_LAYERS