diff --git a/src/mintpy/gpu/__init__.py b/src/mintpy/gpu/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/src/mintpy/ifgram_inversion_gpu.py b/src/mintpy/gpu/ifgram_inversion.py
similarity index 99%
rename from src/mintpy/ifgram_inversion_gpu.py
rename to src/mintpy/gpu/ifgram_inversion.py
index 3da7cf80f..9342ce0d2 100644
--- a/src/mintpy/ifgram_inversion_gpu.py
+++ b/src/mintpy/gpu/ifgram_inversion.py
@@ -4,7 +4,7 @@
 # GPU-accelerated network inversion                        #
 ############################################################
 # Recommend import:
-#     from mintpy import ifgram_inversion_gpu as ifginv_gpu
+#     from mintpy.gpu import ifgram_inversion as gpu_ifginv
 
 
 """GPU-batched solver for the SBAS network inversion.
diff --git a/src/mintpy/ifgram_inversion.py b/src/mintpy/ifgram_inversion.py
index 52f66afea..7d66c5dd2 100644
--- a/src/mintpy/ifgram_inversion.py
+++ b/src/mintpy/ifgram_inversion.py
@@ -816,7 +816,7 @@ def run_ifgram_inversion_patch(ifgram_file, box=None, ref_phase=None, obs_ds_nam
     # for the full-rank case. Rank-deficient pixels (rare on real SBAS networks)
     # are not handled here; if encountered, NaN/Inf will propagate downstream.
     if solver != 'cpu':
-        from mintpy.ifgram_inversion_gpu import estimate_timeseries_batch
+        from mintpy.gpu.ifgram_inversion import estimate_timeseries_batch
         print(f'estimating time-series via {solver} solver (batched, GPU)')
         ts_sub, q_sub, n_sub = estimate_timeseries_batch(
             A=A, B=B,
diff --git a/tests/test_ifgram_inversion_gpu.py b/tests/test_gpu_ifgram_inversion.py
similarity index 97%
rename from tests/test_ifgram_inversion_gpu.py
rename to tests/test_gpu_ifgram_inversion.py
index 9b8a45ff7..ec656f01a 100644
--- a/tests/test_ifgram_inversion_gpu.py
+++ b/tests/test_gpu_ifgram_inversion.py
@@ -1,4 +1,4 @@
-"""Numerical-equivalence tests for src/mintpy/ifgram_inversion_gpu.py.
+"""Numerical-equivalence tests for src/mintpy/gpu/ifgram_inversion.py.
 
 Compare the GPU-batched solver against the per-pixel CPU reference
 (scipy.linalg.lstsq via mintpy.ifgram_inversion.estimate_timeseries) on
@@ -11,12 +11,12 @@
 
 torch = pytest.importorskip("torch")
 
+from mintpy.gpu.ifgram_inversion import estimate_timeseries_batch
 from mintpy.ifgram_inversion import estimate_timeseries
-from mintpy.ifgram_inversion_gpu import estimate_timeseries_batch
 
 requires_cuda = pytest.mark.skipif(
     not torch.cuda.is_available(),
-    reason="CUDA-capable GPU required for ifgram_inversion_gpu tests",
+    reason="CUDA-capable GPU required for mintpy.gpu.ifgram_inversion tests",
 )