From 17b5961dbc36c9d4b33f06c88a8af0edebddfa1b Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=D0=9C=D0=B0=D0=BA=D1=81?= <maxteneff@gmail.com>
Date: Sat, 17 Oct 2026 20:06:59 +0300
Subject: [PATCH 1/2] Load transformer models saved on GPU on CPU-only machines

Map the fitted checkpoint to CPU when unpickling transformer models, so a
model trained on GPU can be loaded with `load_model`, `load` or `loads` on a
machine without CUDA. Weights are moved to the target device before inference
and by the trainer when fitting, so nothing changes on GPU machines.

Fixes #286
---
 CHANGELOG.md                              |  5 +++++
 rectools/models/nn/transformers/base.py   |  4 +++-
 tests/models/nn/transformers/test_base.py | 20 ++++++++++++++++++++
 3 files changed, 28 insertions(+), 1 deletion(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 7bb8cd98..c17af062 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -6,6 +6,11 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 
+## Unreleased
+
+### Fixed
+- Loading fitted transformer models (`SASRecModel`, `BERT4RecModel`) that were saved on GPU with `load_model`, `load` and `loads` on a machine without GPU ([#286](https://github.com/MTSWebServices/RecTools/issues/286))
+
 ## [0.19.0] - 11.06.2026
 
 ### Added
diff --git a/rectools/models/nn/transformers/base.py b/rectools/models/nn/transformers/base.py
index 2183005c..8df88de5 100644
--- a/rectools/models/nn/transformers/base.py
+++ b/rectools/models/nn/transformers/base.py
@@ -668,7 +668,9 @@ def __getstate__(self) -> object:
 
     def __setstate__(self, state: tp.Dict[str, tp.Any]) -> None:
         if "fitted_checkpoint" in state:
-            checkpoint = torch.load(io.BytesIO(state["fitted_checkpoint"]), weights_only=False)
+            # Load to CPU so that a model fitted on GPU can be loaded on a machine without one.
+            # Weights are moved to the right device before inference and by the trainer when fitting.
+            checkpoint = torch.load(io.BytesIO(state["fitted_checkpoint"]), map_location="cpu", weights_only=False)
             loaded = self._model_from_checkpoint(checkpoint)
         else:
             loaded = self.from_config(state["model_config"])
diff --git a/tests/models/nn/transformers/test_base.py b/tests/models/nn/transformers/test_base.py
index 217a89a7..57cb80cc 100644
--- a/tests/models/nn/transformers/test_base.py
+++ b/tests/models/nn/transformers/test_base.py
@@ -184,6 +184,26 @@ def test_save_load_for_fitted_model(
         model.fit(dataset_item_features)
         assert_save_load_do_not_change_model(model, dataset_item_features)
 
+    @pytest.mark.parametrize("model_cls", (SASRecModel, BERT4RecModel))
+    def test_load_fitted_model_saved_on_gpu_without_gpu(
+        self,
+        model_cls: tp.Type[TransformerModelBase],
+        dataset: Dataset,
+        monkeypatch: pytest.MonkeyPatch,
+    ) -> None:
+        model = model_cls.from_config({"deterministic": True})
+        model.fit(dataset)
+
+        # Save the model as if its tensors were on a GPU
+        with monkeypatch.context() as m:
+            m.setattr(torch.serialization, "location_tag", lambda storage: "cuda:0")
+            data = model.dumps()
+
+        monkeypatch.setattr(torch.cuda, "is_available", lambda: False)
+        recovered_model = model_cls.loads(data)
+
+        self._assert_same_reco(model, recovered_model, dataset)
+
     @pytest.mark.parametrize("test_dataset", ("dataset", "dataset_item_features"))
     @pytest.mark.parametrize("model_cls", (SASRecModel, BERT4RecModel))
     @pytest.mark.parametrize(

From 2117a2b758d27dd46b45578a419cdbfb44b8a649 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=D0=9C=D0=B0=D0=BA=D1=81?= <maxteneff@gmail.com>
Date: Sat, 17 Oct 2026 20:08:30 +0300
Subject: [PATCH 2/2] Link the changelog entry to the pull request

---
 CHANGELOG.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index c17af062..162600fb 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -9,7 +9,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ## Unreleased
 
 ### Fixed
-- Loading fitted transformer models (`SASRecModel`, `BERT4RecModel`) that were saved on GPU with `load_model`, `load` and `loads` on a machine without GPU ([#286](https://github.com/MTSWebServices/RecTools/issues/286))
+- Loading fitted transformer models (`SASRecModel`, `BERT4RecModel`) that were saved on GPU with `load_model`, `load` and `loads` on a machine without GPU ([#327](https://github.com/MTSWebServices/RecTools/pull/327))
 
 ## [0.19.0] - 11.06.2026