Fix cu128 index selection for pinned Torch versions

tests/test_platform_detection.py

@@ -121,6 +121,23 @@ def test_nvidia_gpu_linux(monkeypatch):
     assert get_torch_platform(gpu_infos) == expected
 
 
+def test_nvidia_gpu_demotes_to_cu124_for_pinned_torch_below_2_7(monkeypatch):
+    monkeypatch.setattr("torchruntime.platform_detection.os_name", "Windows")
+    monkeypatch.setattr("torchruntime.platform_detection.arch", "amd64")
+    monkeypatch.setattr("torchruntime.platform_detection.py_version", (3, 11))
+    monkeypatch.setattr("torchruntime.platform_detection.get_nvidia_arch", lambda device_names: 8.6)
+
+    gpu_infos = [GPU(NVIDIA, "NVIDIA", 0x1234, "GeForce", True)]
+
+    assert get_torch_platform(gpu_infos) == "cu128"
+    assert get_torch_platform(gpu_infos, packages=["torch==2.6.0"]) == "cu124"
+    assert get_torch_platform(gpu_infos, packages=["torch<2.7.0"]) == "cu124"
+    assert get_torch_platform(gpu_infos, packages=["torch<=2.7.0"]) == "cu128"
+    assert get_torch_platform(gpu_infos, packages=["torch!=2.7.0"]) == "cu128"
+    assert get_torch_platform(gpu_infos, packages=["torch>=2.7.0,!=2.7.0,!=2.7.1,<2.8.0"]) == "cu128"
+    assert get_torch_platform(gpu_infos, packages=["torchvision==0.21.0"]) == "cu124"
+
+
 def test_nvidia_gpu_mac(monkeypatch):
     monkeypatch.setattr("torchruntime.platform_detection.os_name", "Darwin")
     monkeypatch.setattr("torchruntime.platform_detection.arch", "arm64")

torchruntime/installer.py

@@ -98,7 +98,7 @@ def install(packages=[], use_uv=False):
     """
 
     gpu_infos = get_gpus()
-    torch_platform = get_torch_platform(gpu_infos)
+    torch_platform = get_torch_platform(gpu_infos, packages=packages)
     cmds = get_install_commands(torch_platform, packages)
     cmds = get_pip_commands(cmds, use_uv=use_uv)
     run_commands(cmds)

torchruntime/platform_detection.py

@@ -1,21 +1,63 @@
-import re
 import sys
 import platform
 
+from packaging.requirements import Requirement
+from packaging.version import Version
+
 from .gpu_db import get_nvidia_arch, get_amd_gfx_info
 from .consts import AMD, INTEL, NVIDIA, CONTACT_LINK
 
 os_name = platform.system()
 arch = platform.machine().lower()
 py_version = sys.version_info
 
+_CUDA_12_8_MIN_VERSIONS = {
+    "torch": Version("2.7.0"),
+    "torchaudio": Version("2.7.0"),
+    "torchvision": Version("0.22.0"),
+}
+
+
+def _packages_require_cuda_12_4(packages):
+    if not packages:
+        return False
+
+    for package in packages:
+        try:
+            requirement = Requirement(package)
+        except Exception:
+            continue
+
+        name = requirement.name.lower().replace("_", "-")
+        threshold = _CUDA_12_8_MIN_VERSIONS.get(name)
+        if not threshold or not requirement.specifier:
+            continue
+
+        test_versions = [
+            threshold,
+            Version(f"{threshold.major}.{threshold.minor}.{threshold.micro + 1}"),
+            Version(f"{threshold.major}.{threshold.minor}.{threshold.micro + 2}"),
+            Version(f"{threshold.major}.{threshold.minor}.{threshold.micro + 3}"),
+            Version(f"{threshold.major}.{threshold.minor + 1}.0"),
+            Version(f"{threshold.major + 1}.0.0"),
+        ]
+
+        allows_threshold_or_higher = any(
+            requirement.specifier.contains(str(version), prereleases=True) for version in test_versions
+        )
+        if not allows_threshold_or_higher:
+            return True
+
+    return False
 
-def get_torch_platform(gpu_infos):
+
+def get_torch_platform(gpu_infos, packages=[]):
     """
     Determine the appropriate PyTorch platform to use based on the system architecture, OS, and GPU information.
 
     Args:
         gpu_infos (list of `torchruntime.device_db.GPU` instances)
+        packages (list of str): Optional list of torch/torchvision/torchaudio requirement strings.
 
     Returns:
         str: A string representing the platform to use. Possible values:
@@ -53,12 +95,12 @@ def get_torch_platform(gpu_infos):
             integrated_devices.append(device)
 
     if discrete_devices:
-        return _get_platform_for_discrete(discrete_devices)
+        return _get_platform_for_discrete(discrete_devices, packages=packages)
 
     return _get_platform_for_integrated(integrated_devices)
 
 
-def _get_platform_for_discrete(gpu_infos):
+def _get_platform_for_discrete(gpu_infos, packages=None):
     vendor_ids = set(gpu.vendor_id for gpu in gpu_infos)
 
     if len(vendor_ids) > 1:
@@ -126,6 +168,9 @@ def _get_platform_for_discrete(gpu_infos):
             if (arch_version > 3.7 and arch_version < 7.5) or py_version < (3, 9):
                 return "cu124"
 
+            if _packages_require_cuda_12_4(packages):
+                return "cu124"
+
             return "cu128"
         elif os_name == "Darwin":
             raise NotImplementedError(