diff --git a/docs/options/backend.md b/docs/options/backend.md
index 41d6a40da..0ce3ade46 100644
--- a/docs/options/backend.md
+++ b/docs/options/backend.md
@@ -9,16 +9,19 @@ GPU backend to use for inference (default: auto).
 Available backends depend on the detected GPU hardware.
 
 **auto** (default): Automatically selects the preferred backend based on your GPU:
-- **AMD GPUs**: vulkan (Linux/macOS) or rocm (Windows)
+- **AMD GPUs**: vulkan (Linux/macOS) or rocm (WSL2)
 - **NVIDIA GPUs**: cuda; vulkan available as explicit option
-- **Intel GPUs**: vulkan (Linux/macOS) or sycl (Windows); openvino available as explicit option
+- **Intel GPUs**: vulkan (Linux/macOS) or sycl (WSL2); openvino available as explicit option
 - **Ascend NPUs**: cann
 - **MUSA GPUs**: musa
 - **No GPU**: vulkan (CPU fallback)
 
 **Platform-specific behavior**:
 - On **Linux/macOS**, Vulkan provides broad compatibility and is preferred for AMD and Intel GPUs
-- On **Windows**, vulkan is not supported on WSL2, so vendor-specific backends (rocm, sycl) are preferred
+- On **WSL2**, Vulkan is a poor default, so vendor-specific backends (rocm, sycl) are preferred. It
+  remains available, and `--backend=vulkan` still selects it. This covers both Windows, where
+  containers run in the WSL2-backed machine, and ramalama running inside a WSL2 distro, which
+  otherwise looks like Linux
 
 **Explicit backend selection**:
 - **vulkan**: Use Vulkan-based inference (compatible with AMD, NVIDIA, Intel, and CPU)
diff --git a/docs/ramalama-bench.1.md b/docs/ramalama-bench.1.md
index f82372170..c7bd8d3fe 100644
--- a/docs/ramalama-bench.1.md
+++ b/docs/ramalama-bench.1.md
@@ -45,16 +45,19 @@ GPU backend to use for inference (default: auto).
 Available backends depend on the detected GPU hardware.
 
 **auto** (default): Automatically selects the preferred backend based on your GPU:
-- **AMD GPUs**: vulkan (Linux/macOS) or rocm (Windows)
+- **AMD GPUs**: vulkan (Linux/macOS) or rocm (WSL2)
 - **NVIDIA GPUs**: cuda; vulkan available as explicit option
-- **Intel GPUs**: vulkan (Linux/macOS) or sycl (Windows); openvino available as explicit option
+- **Intel GPUs**: vulkan (Linux/macOS) or sycl (WSL2); openvino available as explicit option
 - **Ascend NPUs**: cann
 - **MUSA GPUs**: musa
 - **No GPU**: vulkan (CPU fallback)
 
 **Platform-specific behavior**:
 - On **Linux/macOS**, Vulkan provides broad compatibility and is preferred for AMD and Intel GPUs
-- On **Windows**, vulkan is not supported on WSL2, so vendor-specific backends (rocm, sycl) are preferred
+- On **WSL2**, Vulkan is a poor default, so vendor-specific backends (rocm, sycl) are preferred. It
+  remains available, and `--backend=vulkan` still selects it. This covers both Windows, where
+  containers run in the WSL2-backed machine, and ramalama running inside a WSL2 distro, which
+  otherwise looks like Linux
 
 **Explicit backend selection**:
 - **vulkan**: Use Vulkan-based inference (compatible with AMD, NVIDIA, Intel, and CPU)
diff --git a/docs/ramalama-cuda.7.md b/docs/ramalama-cuda.7.md
index b804dfb82..bee71c0ec 100644
--- a/docs/ramalama-cuda.7.md
+++ b/docs/ramalama-cuda.7.md
@@ -138,6 +138,8 @@ ramalama run granite
 
 This is particularly useful in multi-GPU systems where you want to dedicate specific GPUs to different workloads.
 
+Where the CDI configuration has an entry for each GPU, only the selected ones are passed into the container by the NVIDIA container toolkit; where it only defines `all`, every detected GPU is. Either way the host's other GPU devices, such as `/dev/dri` for an integrated GPU, are left out, since the Vulkan backend offloads onto every device it can enumerate and would otherwise use them. Pass `--device /dev/dri` to `ramalama run` or `ramalama serve` to add them back.
+
 If `CUDA_VISIBLE_DEVICES` is set to an empty string, RamaLama treats it as unset and follows the default GPU-selection behavior.
 
 ```bash
diff --git a/docs/ramalama-perplexity.1.md b/docs/ramalama-perplexity.1.md
index 6cff430ba..c89c3708d 100644
--- a/docs/ramalama-perplexity.1.md
+++ b/docs/ramalama-perplexity.1.md
@@ -45,16 +45,19 @@ GPU backend to use for inference (default: auto).
 Available backends depend on the detected GPU hardware.
 
 **auto** (default): Automatically selects the preferred backend based on your GPU:
-- **AMD GPUs**: vulkan (Linux/macOS) or rocm (Windows)
+- **AMD GPUs**: vulkan (Linux/macOS) or rocm (WSL2)
 - **NVIDIA GPUs**: cuda; vulkan available as explicit option
-- **Intel GPUs**: vulkan (Linux/macOS) or sycl (Windows); openvino available as explicit option
+- **Intel GPUs**: vulkan (Linux/macOS) or sycl (WSL2); openvino available as explicit option
 - **Ascend NPUs**: cann
 - **MUSA GPUs**: musa
 - **No GPU**: vulkan (CPU fallback)
 
 **Platform-specific behavior**:
 - On **Linux/macOS**, Vulkan provides broad compatibility and is preferred for AMD and Intel GPUs
-- On **Windows**, vulkan is not supported on WSL2, so vendor-specific backends (rocm, sycl) are preferred
+- On **WSL2**, Vulkan is a poor default, so vendor-specific backends (rocm, sycl) are preferred. It
+  remains available, and `--backend=vulkan` still selects it. This covers both Windows, where
+  containers run in the WSL2-backed machine, and ramalama running inside a WSL2 distro, which
+  otherwise looks like Linux
 
 **Explicit backend selection**:
 - **vulkan**: Use Vulkan-based inference (compatible with AMD, NVIDIA, Intel, and CPU)
diff --git a/docs/ramalama-run.1.md b/docs/ramalama-run.1.md
index 5ed45f709..1022faf66 100644
--- a/docs/ramalama-run.1.md
+++ b/docs/ramalama-run.1.md
@@ -57,16 +57,19 @@ GPU backend to use for inference (default: auto).
 Available backends depend on the detected GPU hardware.
 
 **auto** (default): Automatically selects the preferred backend based on your GPU:
-- **AMD GPUs**: vulkan (Linux/macOS) or rocm (Windows)
+- **AMD GPUs**: vulkan (Linux/macOS) or rocm (WSL2)
 - **NVIDIA GPUs**: cuda; vulkan available as explicit option
-- **Intel GPUs**: vulkan (Linux/macOS) or sycl (Windows); openvino available as explicit option
+- **Intel GPUs**: vulkan (Linux/macOS) or sycl (WSL2); openvino available as explicit option
 - **Ascend NPUs**: cann
 - **MUSA GPUs**: musa
 - **No GPU**: vulkan (CPU fallback)
 
 **Platform-specific behavior**:
 - On **Linux/macOS**, Vulkan provides broad compatibility and is preferred for AMD and Intel GPUs
-- On **Windows**, vulkan is not supported on WSL2, so vendor-specific backends (rocm, sycl) are preferred
+- On **WSL2**, Vulkan is a poor default, so vendor-specific backends (rocm, sycl) are preferred. It
+  remains available, and `--backend=vulkan` still selects it. This covers both Windows, where
+  containers run in the WSL2-backed machine, and ramalama running inside a WSL2 distro, which
+  otherwise looks like Linux
 
 **Explicit backend selection**:
 - **vulkan**: Use Vulkan-based inference (compatible with AMD, NVIDIA, Intel, and CPU)
diff --git a/docs/ramalama-sandbox-goose.1.md b/docs/ramalama-sandbox-goose.1.md
index 35c1fd3af..f7e52a404 100644
--- a/docs/ramalama-sandbox-goose.1.md
+++ b/docs/ramalama-sandbox-goose.1.md
@@ -53,16 +53,19 @@ GPU backend to use for inference (default: auto).
 Available backends depend on the detected GPU hardware.
 
 **auto** (default): Automatically selects the preferred backend based on your GPU:
-- **AMD GPUs**: vulkan (Linux/macOS) or rocm (Windows)
+- **AMD GPUs**: vulkan (Linux/macOS) or rocm (WSL2)
 - **NVIDIA GPUs**: cuda; vulkan available as explicit option
-- **Intel GPUs**: vulkan (Linux/macOS) or sycl (Windows); openvino available as explicit option
+- **Intel GPUs**: vulkan (Linux/macOS) or sycl (WSL2); openvino available as explicit option
 - **Ascend NPUs**: cann
 - **MUSA GPUs**: musa
 - **No GPU**: vulkan (CPU fallback)
 
 **Platform-specific behavior**:
 - On **Linux/macOS**, Vulkan provides broad compatibility and is preferred for AMD and Intel GPUs
-- On **Windows**, vulkan is not supported on WSL2, so vendor-specific backends (rocm, sycl) are preferred
+- On **WSL2**, Vulkan is a poor default, so vendor-specific backends (rocm, sycl) are preferred. It
+  remains available, and `--backend=vulkan` still selects it. This covers both Windows, where
+  containers run in the WSL2-backed machine, and ramalama running inside a WSL2 distro, which
+  otherwise looks like Linux
 
 **Explicit backend selection**:
 - **vulkan**: Use Vulkan-based inference (compatible with AMD, NVIDIA, Intel, and CPU)
diff --git a/docs/ramalama-sandbox-opencode.1.md b/docs/ramalama-sandbox-opencode.1.md
index 5af6bf194..b067e8679 100644
--- a/docs/ramalama-sandbox-opencode.1.md
+++ b/docs/ramalama-sandbox-opencode.1.md
@@ -53,16 +53,19 @@ GPU backend to use for inference (default: auto).
 Available backends depend on the detected GPU hardware.
 
 **auto** (default): Automatically selects the preferred backend based on your GPU:
-- **AMD GPUs**: vulkan (Linux/macOS) or rocm (Windows)
+- **AMD GPUs**: vulkan (Linux/macOS) or rocm (WSL2)
 - **NVIDIA GPUs**: cuda; vulkan available as explicit option
-- **Intel GPUs**: vulkan (Linux/macOS) or sycl (Windows); openvino available as explicit option
+- **Intel GPUs**: vulkan (Linux/macOS) or sycl (WSL2); openvino available as explicit option
 - **Ascend NPUs**: cann
 - **MUSA GPUs**: musa
 - **No GPU**: vulkan (CPU fallback)
 
 **Platform-specific behavior**:
 - On **Linux/macOS**, Vulkan provides broad compatibility and is preferred for AMD and Intel GPUs
-- On **Windows**, vulkan is not supported on WSL2, so vendor-specific backends (rocm, sycl) are preferred
+- On **WSL2**, Vulkan is a poor default, so vendor-specific backends (rocm, sycl) are preferred. It
+  remains available, and `--backend=vulkan` still selects it. This covers both Windows, where
+  containers run in the WSL2-backed machine, and ramalama running inside a WSL2 distro, which
+  otherwise looks like Linux
 
 **Explicit backend selection**:
 - **vulkan**: Use Vulkan-based inference (compatible with AMD, NVIDIA, Intel, and CPU)
diff --git a/docs/ramalama-sandbox-pi.1.md b/docs/ramalama-sandbox-pi.1.md
index 868af0d3e..44129449f 100644
--- a/docs/ramalama-sandbox-pi.1.md
+++ b/docs/ramalama-sandbox-pi.1.md
@@ -57,16 +57,19 @@ GPU backend to use for inference (default: auto).
 Available backends depend on the detected GPU hardware.
 
 **auto** (default): Automatically selects the preferred backend based on your GPU:
-- **AMD GPUs**: vulkan (Linux/macOS) or rocm (Windows)
+- **AMD GPUs**: vulkan (Linux/macOS) or rocm (WSL2)
 - **NVIDIA GPUs**: cuda; vulkan available as explicit option
-- **Intel GPUs**: vulkan (Linux/macOS) or sycl (Windows); openvino available as explicit option
+- **Intel GPUs**: vulkan (Linux/macOS) or sycl (WSL2); openvino available as explicit option
 - **Ascend NPUs**: cann
 - **MUSA GPUs**: musa
 - **No GPU**: vulkan (CPU fallback)
 
 **Platform-specific behavior**:
 - On **Linux/macOS**, Vulkan provides broad compatibility and is preferred for AMD and Intel GPUs
-- On **Windows**, vulkan is not supported on WSL2, so vendor-specific backends (rocm, sycl) are preferred
+- On **WSL2**, Vulkan is a poor default, so vendor-specific backends (rocm, sycl) are preferred. It
+  remains available, and `--backend=vulkan` still selects it. This covers both Windows, where
+  containers run in the WSL2-backed machine, and ramalama running inside a WSL2 distro, which
+  otherwise looks like Linux
 
 **Explicit backend selection**:
 - **vulkan**: Use Vulkan-based inference (compatible with AMD, NVIDIA, Intel, and CPU)
diff --git a/docs/ramalama-serve.1.md b/docs/ramalama-serve.1.md
index 83400ba02..20e85951c 100644
--- a/docs/ramalama-serve.1.md
+++ b/docs/ramalama-serve.1.md
@@ -86,16 +86,19 @@ GPU backend to use for inference (default: auto).
 Available backends depend on the detected GPU hardware.
 
 **auto** (default): Automatically selects the preferred backend based on your GPU:
-- **AMD GPUs**: vulkan (Linux/macOS) or rocm (Windows)
+- **AMD GPUs**: vulkan (Linux/macOS) or rocm (WSL2)
 - **NVIDIA GPUs**: cuda; vulkan available as explicit option
-- **Intel GPUs**: vulkan (Linux/macOS) or sycl (Windows); openvino available as explicit option
+- **Intel GPUs**: vulkan (Linux/macOS) or sycl (WSL2); openvino available as explicit option
 - **Ascend NPUs**: cann
 - **MUSA GPUs**: musa
 - **No GPU**: vulkan (CPU fallback)
 
 **Platform-specific behavior**:
 - On **Linux/macOS**, Vulkan provides broad compatibility and is preferred for AMD and Intel GPUs
-- On **Windows**, vulkan is not supported on WSL2, so vendor-specific backends (rocm, sycl) are preferred
+- On **WSL2**, Vulkan is a poor default, so vendor-specific backends (rocm, sycl) are preferred. It
+  remains available, and `--backend=vulkan` still selects it. This covers both Windows, where
+  containers run in the WSL2-backed machine, and ramalama running inside a WSL2 distro, which
+  otherwise looks like Linux
 
 **Explicit backend selection**:
 - **vulkan**: Use Vulkan-based inference (compatible with AMD, NVIDIA, Intel, and CPU)
diff --git a/docs/ramalama.conf b/docs/ramalama.conf
index cc54ea239..4b24d6d00 100644
--- a/docs/ramalama.conf
+++ b/docs/ramalama.conf
@@ -240,9 +240,9 @@
 #
 # Valid options: auto, vulkan, rocm, cuda, sycl, openvino, cann, musa
 # - auto (default): Automatically selects the preferred backend based on detected GPU
-#   - AMD GPUs: vulkan (Linux/macOS) or rocm (Windows)
+#   - AMD GPUs: vulkan (Linux/macOS) or rocm (WSL2)
 #   - NVIDIA GPUs: cuda; vulkan available as explicit option
-#   - Intel GPUs: vulkan (Linux/macOS) or sycl (Windows); openvino available as explicit option
+#   - Intel GPUs: vulkan (Linux/macOS) or sycl (WSL2); openvino available as explicit option
 #   - Ascend NPUs: cann
 #   - MUSA GPUs: musa
 #   - No GPU: vulkan (CPU fallback)
@@ -254,9 +254,10 @@
 # - cann: Use Huawei CANN backend (Ascend NPUs only); uses quay.io/ramalama/cann
 # - musa: Use Moore Threads MUSA backend (MUSA GPUs only); uses quay.io/ramalama/musa
 #
-# Platform-specific behavior: On Windows, vulkan is not supported on WSL2, so
-# vendor-specific backends (rocm for AMD, sycl for Intel) are automatically
-# preferred when using backend="auto".
+# Platform-specific behavior: vulkan is not supported on WSL2, so vendor-specific
+# backends (rocm for AMD, sycl for Intel) are automatically preferred when using
+# backend="auto". This covers both Windows and ramalama running inside a WSL2
+# distro, which otherwise looks like Linux.
 #
 #backend = "auto"
 
diff --git a/docs/ramalama.conf.5.md b/docs/ramalama.conf.5.md
index 529913a09..261d76eb6 100644
--- a/docs/ramalama.conf.5.md
+++ b/docs/ramalama.conf.5.md
@@ -239,9 +239,9 @@ This setting affects which container image is selected and how GPU resources are
 Valid options: `auto`, `vulkan`, `rocm`, `cuda`, `sycl`, `openvino`, `cann`, `musa`.
 
 - **auto** (default): Automatically selects the preferred backend based on detected GPU:
-  - AMD GPUs: vulkan (Linux/macOS) or rocm (Windows)
+  - AMD GPUs: vulkan (Linux/macOS) or rocm (WSL2)
   - NVIDIA GPUs: cuda; vulkan available as explicit option
-  - Intel GPUs: vulkan (Linux/macOS) or sycl (Windows); openvino available as explicit option
+  - Intel GPUs: vulkan (Linux/macOS) or sycl (WSL2); openvino available as explicit option
   - Ascend NPUs: cann
   - MUSA GPUs: musa
   - No GPU: vulkan (CPU fallback)
@@ -254,7 +254,7 @@ Valid options: `auto`, `vulkan`, `rocm`, `cuda`, `sycl`, `openvino`, `cann`, `mu
 - **cann**: Use Huawei CANN backend (Ascend NPUs only); uses `quay.io/ramalama/cann`
 - **musa**: Use Moore Threads MUSA backend (MUSA GPUs only); uses `quay.io/ramalama/musa`
 
-**Platform-specific behavior**: On Windows, vulkan is not supported on WSL2, so vendor-specific backends (rocm for AMD, sycl for Intel) are automatically preferred when using `backend="auto"`.
+**Platform-specific behavior**: vulkan is not supported on WSL2, so vendor-specific backends (rocm for AMD, sycl for Intel) are automatically preferred when using `backend="auto"`. This covers both Windows and ramalama running inside a WSL2 distro, which otherwise looks like Linux.
 
 Example configuration:
 
diff --git a/ramalama/common.py b/ramalama/common.py
index fa38f08bf..de5e690ea 100644
--- a/ramalama/common.py
+++ b/ramalama/common.py
@@ -11,7 +11,7 @@
 import string
 import subprocess
 import sys
-from collections.abc import Callable, Sequence
+from collections.abc import Callable, Collection, Sequence
 from dataclasses import dataclass
 from functools import lru_cache
 from pathlib import Path
@@ -488,6 +488,33 @@ def check_metal(args: ContainerArgType) -> bool:
     return platform.system() == "Darwin"
 
 
+@lru_cache(maxsize=1)
+def in_wsl() -> bool:
+    """True when the interpreter itself is running inside a WSL distro.
+
+    False on native Windows, where ramalama drives a podman machine instead.
+    """
+    try:
+        with open("/proc/sys/kernel/osrelease") as f:
+            return "microsoft" in f.read().lower()
+    except OSError:
+        return False
+
+
+def is_windows_or_wsl() -> bool:
+    """True where containers reach GPUs through WSL rather than native devices.
+
+    Covers both a native Windows interpreter, which runs containers in the
+    WSL2-backed podman machine, and ramalama running inside a WSL distro
+    itself, which platform.system() reports as "Linux".
+
+    WSL exposes GPUs through /dev/dxg, so Vulkan there means mesa's dzn driver
+    translating to D3D12 (no cooperative matrix support, no compute tuning) or
+    a silent llvmpipe fallback.
+    """
+    return platform.system() == "Windows" or in_wsl()
+
+
 @lru_cache(maxsize=1)
 def has_nvidia_vulkan_icd() -> bool:
     """True when NVIDIA's Vulkan ICD manifest is installed on the host.
@@ -722,7 +749,19 @@ def set_gpu_type_env_vars():
 ]
 
 
-def get_gpu_devices():
+def get_gpu_devices(accel_env_vars: Optional[Collection[str]] = None) -> dict[str, str]:
+    """The host GPU devices to hand to the container, given the accelerator in play.
+
+    An NVIDIA GPU does not come in this way: the container toolkit injects the
+    device nodes of the GPUs that were asked for, DRM nodes included. Mapping
+    the host's GPU devices in as well can then only add ones that are not the
+    accelerator in play - an iGPU on a hybrid host, or a GPU left out of a
+    narrowed selection - and llama.cpp's Vulkan backend offloads onto every
+    device it can enumerate. "--device" remains for anyone who wants them.
+    """
+    if "CUDA_VISIBLE_DEVICES" in (get_gpu_type_env_vars() if accel_env_vars is None else accel_env_vars):
+        return {}
+
     devices = {}
     for dev in ["dri", "kfd", "accel"]:
         path = "/dev/" + dev
diff --git a/ramalama/compose.py b/ramalama/compose.py
index ab53e62c5..449b3b132 100644
--- a/ramalama/compose.py
+++ b/ramalama/compose.py
@@ -98,7 +98,7 @@ def _gen_mmproj_volume(self) -> str:
         return f'\n      - "{self.src_mmproj_path}:{self.dest_mmproj_path}:ro"'
 
     def _gen_devices(self) -> str:
-        devices = get_gpu_devices()
+        devices = get_gpu_devices(get_accel_env_vars())
 
         if not devices:
             return ""
diff --git a/ramalama/engine.py b/ramalama/engine.py
index 01ffdadc1..a6918a19a 100644
--- a/ramalama/engine.py
+++ b/ramalama/engine.py
@@ -3,7 +3,6 @@
 import glob
 import json
 import os
-import platform
 import subprocess
 import sys
 import time
@@ -22,7 +21,9 @@
     exec_cmd,
     genname,
     get_accel_env_vars,
+    get_gpu_devices,
     host_path,
+    is_windows_or_wsl,
     perror,
     run_cmd,
 )
@@ -120,12 +121,14 @@ def add_device_options(self):
         if ramalama.common.podman_machine_accel:
             self.exec_args += ["--device", "/dev/dri"]
 
-        for path in ["/dev/dri", "/dev/kfd", "/dev/accel", "/dev/davinci*", "/dev/devmm_svm", "/dev/hisi_hdc"]:
+        env_vars = get_accel_env_vars()
+        gpu_devices = list(get_gpu_devices(env_vars).values())
+        for path in [*gpu_devices, "/dev/davinci*", "/dev/devmm_svm", "/dev/hisi_hdc"]:
             for dev in glob.glob(path):
                 self.exec_args += ["--device", dev]
 
-        intel_windows_added = False
-        for k, v in get_accel_env_vars().items():
+        wsl_devices_added = False
+        for k, v in env_vars.items():
             # Special case for Cuda
             if k == "CUDA_VISIBLE_DEVICES":
                 # Pass in only the GPUs the user selected rather than all of
@@ -147,11 +150,17 @@ def add_device_options(self):
                 v = container_cuda_visible_devices(v)
             elif k == "MUSA_VISIBLE_DEVICES":
                 self.exec_args += ["--env", "MTHREADS_VISIBLE_DEVICES=all"]
-            elif k == "INTEL_VISIBLE_DEVICES":
-                if platform.system() == "Windows" and not intel_windows_added:
+            elif k in ("HIP_VISIBLE_DEVICES", "INTEL_VISIBLE_DEVICES"):
+                # WSL exposes the GPU as /dev/dxg with its driver libraries in
+                # /usr/lib/wsl, whether ramalama runs on native Windows against
+                # the podman machine or inside the distro itself, where
+                # platform.system() reports "Linux". That is how both the AMD
+                # and the Intel GPU come in, so the rocm backend needs it as
+                # much as sycl does.
+                if is_windows_or_wsl() and not wsl_devices_added:
                     self.exec_args += ["--device", "/dev/dxg"]
                     self.exec_args += ["--mount", "type=bind,src=/usr/lib/wsl,dst=/usr/lib/wsl"]
-                    intel_windows_added = True
+                    wsl_devices_added = True
 
             self.exec_args += ["-e", f"{k}={v}"]
 
diff --git a/ramalama/kube.py b/ramalama/kube.py
index a8c4ca3be..443d5a78f 100644
--- a/ramalama/kube.py
+++ b/ramalama/kube.py
@@ -85,7 +85,7 @@ def _gen_volumes(self) -> Tuple[str, str]:
     def _gen_devices(self) -> Tuple[str, str]:
         mounts = ""
         volumes = ""
-        for name, path in get_gpu_devices().items():
+        for name, path in get_gpu_devices(get_accel_env_vars()).items():
             mounts += f"""
         - mountPath: {path}
           name: {name}"""
@@ -219,7 +219,7 @@ def __gen_resources(self) -> str:
           limits:
              'nvidia.com/gpu=all': 1"""
 
-        devices = get_gpu_devices()
+        devices = get_gpu_devices(get_accel_env_vars())
         if devices:
             limits = "".join(f"\n             'podman.io/device={path}': 1" for path in devices.values())
             return f"""
diff --git a/ramalama/plugins/runtimes/inference/llama_cpp.py b/ramalama/plugins/runtimes/inference/llama_cpp.py
index 0a4a537e6..4ddc598a3 100644
--- a/ramalama/plugins/runtimes/inference/llama_cpp.py
+++ b/ramalama/plugins/runtimes/inference/llama_cpp.py
@@ -4,7 +4,6 @@
 import copy
 import json
 import os
-import platform
 import shutil
 import subprocess
 import sys
@@ -42,6 +41,7 @@
     genname,
     get_gpu_type_env_vars,
     has_nvidia_vulkan_icd,
+    is_windows_or_wsl,
     run_cmd,
     set_accel_env_vars,
     set_gpu_type_env_vars,
@@ -141,9 +141,7 @@ def __call__(self, parser, namespace, values, option_string=None):
 
 def get_gpu_backend_preferences(gpu_type: str) -> list[str]:
     """Returns preferred backends for a given GPU type in order of preference.
-    On Windows, vulkan is not supported on WSL2, so vendor backends are preferred."""
-    is_windows = platform.system() == "Windows"
-
+    Vulkan is a poor default on WSL2, so vendor backends are preferred there."""
     preferences = {
         "HIP_VISIBLE_DEVICES": ["vulkan", "rocm"],  # AMD: Vulkan preferred
         "CUDA_VISIBLE_DEVICES": ["cuda", "vulkan"],  # NVIDIA: CUDA preferred
@@ -154,7 +152,7 @@ def get_gpu_backend_preferences(gpu_type: str) -> list[str]:
         "GGML_VK_VISIBLE_DEVICES": ["vulkan"],  # Vulkan: Vulkan only
     }
 
-    if is_windows:
+    if is_windows_or_wsl():
         preferences["HIP_VISIBLE_DEVICES"] = ["rocm", "vulkan"]
         preferences["INTEL_VISIBLE_DEVICES"] = ["sycl", "vulkan", "openvino"]
 
@@ -206,11 +204,13 @@ def get_available_backends() -> list[str]:
     return ["auto", "vulkan"]
 
 
+# GGML_VK_VISIBLE_DEVICES, the vulkan backend, is deliberately absent: it runs
+# in the default image, and hardcoding the published tag here would ignore
+# default_image / RAMALAMA_DEFAULT_IMAGE whenever a GPU is detected.
 _LLAMA_CPP_IMAGES: dict[str, str] = {
     "ASAHI_VISIBLE_DEVICES": version_tagged_image("quay.io/ramalama/asahi"),
     "ASCEND_VISIBLE_DEVICES": version_tagged_image("quay.io/ramalama/cann"),
     "CUDA_VISIBLE_DEVICES": version_tagged_image("quay.io/ramalama/cuda"),
-    "GGML_VK_VISIBLE_DEVICES": version_tagged_image("quay.io/ramalama/ramalama"),
     "HIP_VISIBLE_DEVICES": version_tagged_image("quay.io/ramalama/rocm"),
     "INTEL_VISIBLE_DEVICES": version_tagged_image("quay.io/ramalama/intel-gpu"),
     "OPENVINO_VISIBLE_DEVICES": version_tagged_image("quay.io/ramalama/openvino"),
@@ -374,8 +374,15 @@ def get_container_image(self, config: Any, detected_gpu_type: str) -> Optional[s
         if gpu_type == "GGML_VK_VISIBLE_DEVICES" and detected_gpu_type == "CUDA_VISIBLE_DEVICES":
             warn_without_nvidia_vulkan_icd()
 
-        override = config.images.get(gpu_type) if gpu_type else None
-        return override if override else _LLAMA_CPP_IMAGES.get(gpu_type, config.default_image)
+        # In auto mode an image configured for the detected GPU is the user
+        # pinning an image for their hardware, so it takes precedence over one
+        # configured for the backend that was resolved from that hardware.
+        override_keys = [detected_gpu_type, gpu_type] if backend == "auto" else [gpu_type]
+        for key in override_keys:
+            if key and (override := config.images.get(key)):
+                return override
+
+        return _LLAMA_CPP_IMAGES.get(gpu_type, config.default_image)
 
     def _container_image_is_ggml(self, args: argparse.Namespace) -> bool:
         if not args.container or args.dryrun:
diff --git a/ramalama/quadlet.py b/ramalama/quadlet.py
index 88564a347..a5b3e5382 100644
--- a/ramalama/quadlet.py
+++ b/ramalama/quadlet.py
@@ -4,7 +4,16 @@
 import shlex
 from typing import Optional, Tuple
 
-from ramalama.common import MNT_DIR, RAG_DIR, ContainerEntryPoint, get_accel, get_accel_env_vars
+# Live reference for checking global vars
+import ramalama.common
+from ramalama.common import (
+    MNT_DIR,
+    RAG_DIR,
+    ContainerEntryPoint,
+    container_cuda_visible_devices,
+    get_accel,
+    get_accel_env_vars,
+)
 from ramalama.file import UnitFile
 from ramalama.host_utils import format_bind_host_publish_prefix, is_loopback_bind_host
 
@@ -68,11 +77,16 @@ def generate(self) -> list[UnitFile]:
         quadlet_file = UnitFile(container_file_name)
         quadlet_file.add("Unit", "Description", f"RamaLama {self.name} AI Model Service")
         quadlet_file.add("Unit", "After", "local-fs.target")
-        quadlet_file.add("Container", "AddDevice", "-/dev/accel")
-        quadlet_file.add("Container", "AddDevice", "-/dev/dri")
-        quadlet_file.add("Container", "AddDevice", "-/dev/kfd")
         if get_accel() == "cuda":
-            quadlet_file.add("Container", "AddDevice", "nvidia.com/gpu=all")
+            # The container toolkit brings in the NVIDIA device nodes itself, so
+            # the host's GPU devices would only add ones that are not the GPU in
+            # play, such as an iGPU. See get_gpu_devices().
+            for name in ramalama.common.nvidia_selected_devices or ["all"]:
+                quadlet_file.add("Container", "AddDevice", f"nvidia.com/gpu={name}")
+        else:
+            quadlet_file.add("Container", "AddDevice", "-/dev/accel")
+            quadlet_file.add("Container", "AddDevice", "-/dev/dri")
+            quadlet_file.add("Container", "AddDevice", "-/dev/kfd")
         quadlet_file.add("Container", "Image", f"{self.image}")
         quadlet_file.add("Container", "RunInit", "true")
         quadlet_file.add("Container", "Environment", "HOME=/tmp")
@@ -132,6 +146,10 @@ def _gen_mmproj_volume(self, quadlet_file: UnitFile):
     def _gen_env(self, quadlet_file: UnitFile):
         env_var_string = ""
         for k, v in get_accel_env_vars().items():
+            if k == "CUDA_VISIBLE_DEVICES":
+                # AddDevice above passes in just the selected GPUs, so the
+                # container renumbers them, exactly as it does for "ramalama run".
+                v = container_cuda_visible_devices(v)
             quadlet_file.add("Container", "Environment", f"{k}={v}")
         for e in self.args.env:
             quadlet_file.add("Container", "Environment", f"{e}")
diff --git a/test/unit/conftest.py b/test/unit/conftest.py
index 50d6bb225..49c5b7faa 100644
--- a/test/unit/conftest.py
+++ b/test/unit/conftest.py
@@ -45,6 +45,17 @@ def _isolate_from_toolbox():
     in_toolbox.cache_clear()
 
 
+@pytest.fixture(autouse=True)
+def _clear_wsl_cache():
+    """in_wsl() caches its /proc read, so tests patching it must not leak into
+    each other or inherit whatever the host answered first."""
+    from ramalama.common import in_wsl
+
+    in_wsl.cache_clear()
+    yield
+    in_wsl.cache_clear()
+
+
 @pytest.fixture(autouse=True)
 def _clear_vulkan_icd_cache():
     """The ICD probe and the warning it drives are cached for the process, so
diff --git a/test/unit/data/test_compose/with_nvidia_gpu.yaml b/test/unit/data/test_compose/with_nvidia_gpu.yaml
index d0768ce29..36d06108f 100644
--- a/test/unit/data/test_compose/with_nvidia_gpu.yaml
+++ b/test/unit/data/test_compose/with_nvidia_gpu.yaml
@@ -11,10 +11,6 @@ services:
       - "8080:8080"
     environment:
       - CUDA_VISIBLE_DEVICES=0
-    devices:
-      - "/dev/accel:/dev/accel"
-      - "/dev/dri:/dev/dri"
-      - "/dev/kfd:/dev/kfd"
     deploy:
       resources:
         reservations:
diff --git a/test/unit/data/test_compose/with_nvidia_gpu_selection.yaml b/test/unit/data/test_compose/with_nvidia_gpu_selection.yaml
index e9905e863..c1972f9e8 100644
--- a/test/unit/data/test_compose/with_nvidia_gpu_selection.yaml
+++ b/test/unit/data/test_compose/with_nvidia_gpu_selection.yaml
@@ -11,10 +11,6 @@ services:
       - "8080:8080"
     environment:
       - CUDA_VISIBLE_DEVICES=0
-    devices:
-      - "/dev/accel:/dev/accel"
-      - "/dev/dri:/dev/dri"
-      - "/dev/kfd:/dev/kfd"
     deploy:
       resources:
         reservations:
diff --git a/test/unit/data/test_compose/with_nvidia_gpu_vulkan_image.yaml b/test/unit/data/test_compose/with_nvidia_gpu_vulkan_image.yaml
index 8c7aed896..eae7d6d11 100644
--- a/test/unit/data/test_compose/with_nvidia_gpu_vulkan_image.yaml
+++ b/test/unit/data/test_compose/with_nvidia_gpu_vulkan_image.yaml
@@ -11,10 +11,6 @@ services:
       - "8080:8080"
     environment:
       - CUDA_VISIBLE_DEVICES=0
-    devices:
-      - "/dev/accel:/dev/accel"
-      - "/dev/dri:/dev/dri"
-      - "/dev/kfd:/dev/kfd"
     deploy:
       resources:
         reservations:
diff --git a/test/unit/data/test_quadlet/basic/tinyllama.container b/test/unit/data/test_quadlet/basic/tinyllama.container
index 8fab939a1..161bd3baa 100644
--- a/test/unit/data/test_quadlet/basic/tinyllama.container
+++ b/test/unit/data/test_quadlet/basic/tinyllama.container
@@ -3,9 +3,6 @@ Description=RamaLama tinyllama AI Model Service
 After=local-fs.target
 
 [Container]
-AddDevice=-/dev/accel
-AddDevice=-/dev/dri
-AddDevice=-/dev/kfd
 AddDevice=nvidia.com/gpu=all
 Image=testimage
 RunInit=true
diff --git a/test/unit/data/test_quadlet/draft_model/tinyllama.container b/test/unit/data/test_quadlet/draft_model/tinyllama.container
index cd2c41d7f..95e8a30f5 100644
--- a/test/unit/data/test_quadlet/draft_model/tinyllama.container
+++ b/test/unit/data/test_quadlet/draft_model/tinyllama.container
@@ -3,9 +3,6 @@ Description=RamaLama tinyllama AI Model Service
 After=local-fs.target
 
 [Container]
-AddDevice=-/dev/accel
-AddDevice=-/dev/dri
-AddDevice=-/dev/kfd
 AddDevice=nvidia.com/gpu=all
 Image=testimage
 RunInit=true
diff --git a/test/unit/data/test_quadlet/empty/tinyllama.container b/test/unit/data/test_quadlet/empty/tinyllama.container
index 8fab939a1..161bd3baa 100644
--- a/test/unit/data/test_quadlet/empty/tinyllama.container
+++ b/test/unit/data/test_quadlet/empty/tinyllama.container
@@ -3,9 +3,6 @@ Description=RamaLama tinyllama AI Model Service
 After=local-fs.target
 
 [Container]
-AddDevice=-/dev/accel
-AddDevice=-/dev/dri
-AddDevice=-/dev/kfd
 AddDevice=nvidia.com/gpu=all
 Image=testimage
 RunInit=true
diff --git a/test/unit/data/test_quadlet/modelfromstore/modelfromstore.container b/test/unit/data/test_quadlet/modelfromstore/modelfromstore.container
index 8716fda61..8eb3f6f8f 100644
--- a/test/unit/data/test_quadlet/modelfromstore/modelfromstore.container
+++ b/test/unit/data/test_quadlet/modelfromstore/modelfromstore.container
@@ -3,9 +3,6 @@ Description=RamaLama modelfromstore AI Model Service
 After=local-fs.target
 
 [Container]
-AddDevice=-/dev/accel
-AddDevice=-/dev/dri
-AddDevice=-/dev/kfd
 AddDevice=nvidia.com/gpu=all
 Image=testimage
 RunInit=true
diff --git a/test/unit/data/test_quadlet/modelfromstore_add_to_unit/modelfromstore_add_to_unit.container b/test/unit/data/test_quadlet/modelfromstore_add_to_unit/modelfromstore_add_to_unit.container
index a9c2b998e..837c0b9d2 100644
--- a/test/unit/data/test_quadlet/modelfromstore_add_to_unit/modelfromstore_add_to_unit.container
+++ b/test/unit/data/test_quadlet/modelfromstore_add_to_unit/modelfromstore_add_to_unit.container
@@ -3,9 +3,6 @@ Description=RamaLama modelfromstore_add_to_unit AI Model Service
 After=local-fs.target
 
 [Container]
-AddDevice=-/dev/accel
-AddDevice=-/dev/dri
-AddDevice=-/dev/kfd
 AddDevice=nvidia.com/gpu=all
 Image=testimage
 RunInit=true
diff --git a/test/unit/data/test_quadlet/modelfromstore_ct/modelfromstore_ct.container b/test/unit/data/test_quadlet/modelfromstore_ct/modelfromstore_ct.container
index 24b7a1ffb..2b132ee9c 100644
--- a/test/unit/data/test_quadlet/modelfromstore_ct/modelfromstore_ct.container
+++ b/test/unit/data/test_quadlet/modelfromstore_ct/modelfromstore_ct.container
@@ -3,9 +3,6 @@ Description=RamaLama modelfromstore_ct AI Model Service
 After=local-fs.target
 
 [Container]
-AddDevice=-/dev/accel
-AddDevice=-/dev/dri
-AddDevice=-/dev/kfd
 AddDevice=nvidia.com/gpu=all
 Image=testimage
 RunInit=true
diff --git a/test/unit/data/test_quadlet/modelfromstore_mmproj/modelfromstore_mmproj.container b/test/unit/data/test_quadlet/modelfromstore_mmproj/modelfromstore_mmproj.container
index ea5dbfdbe..a8b4e0aae 100644
--- a/test/unit/data/test_quadlet/modelfromstore_mmproj/modelfromstore_mmproj.container
+++ b/test/unit/data/test_quadlet/modelfromstore_mmproj/modelfromstore_mmproj.container
@@ -3,9 +3,6 @@ Description=RamaLama modelfromstore_mmproj AI Model Service
 After=local-fs.target
 
 [Container]
-AddDevice=-/dev/accel
-AddDevice=-/dev/dri
-AddDevice=-/dev/kfd
 AddDevice=nvidia.com/gpu=all
 Image=testimage
 RunInit=true
diff --git a/test/unit/data/test_quadlet/multipart/gpt-oss-120b.container b/test/unit/data/test_quadlet/multipart/gpt-oss-120b.container
index bcb282308..76a0898cf 100644
--- a/test/unit/data/test_quadlet/multipart/gpt-oss-120b.container
+++ b/test/unit/data/test_quadlet/multipart/gpt-oss-120b.container
@@ -3,9 +3,6 @@ Description=RamaLama gpt-oss-120b AI Model Service
 After=local-fs.target
 
 [Container]
-AddDevice=-/dev/accel
-AddDevice=-/dev/dri
-AddDevice=-/dev/kfd
 AddDevice=nvidia.com/gpu=all
 Image=testimage
 RunInit=true
diff --git a/test/unit/data/test_quadlet/oci_basic/oci-model.container b/test/unit/data/test_quadlet/oci_basic/oci-model.container
index 806548842..4485fcd63 100644
--- a/test/unit/data/test_quadlet/oci_basic/oci-model.container
+++ b/test/unit/data/test_quadlet/oci_basic/oci-model.container
@@ -3,9 +3,6 @@ Description=RamaLama oci-model AI Model Service
 After=local-fs.target
 
 [Container]
-AddDevice=-/dev/accel
-AddDevice=-/dev/dri
-AddDevice=-/dev/kfd
 AddDevice=nvidia.com/gpu=all
 Image=testimage
 RunInit=true
diff --git a/test/unit/data/test_quadlet/oci_port/oci-model-port.container b/test/unit/data/test_quadlet/oci_port/oci-model-port.container
index e892c39b8..467280e3c 100644
--- a/test/unit/data/test_quadlet/oci_port/oci-model-port.container
+++ b/test/unit/data/test_quadlet/oci_port/oci-model-port.container
@@ -3,9 +3,6 @@ Description=RamaLama oci-model-port AI Model Service
 After=local-fs.target
 
 [Container]
-AddDevice=-/dev/accel
-AddDevice=-/dev/dri
-AddDevice=-/dev/kfd
 AddDevice=nvidia.com/gpu=all
 Image=testimage
 RunInit=true
diff --git a/test/unit/data/test_quadlet/oci_rag/oci-model-rag.container b/test/unit/data/test_quadlet/oci_rag/oci-model-rag.container
index 3d966f567..9d6cb7fe6 100644
--- a/test/unit/data/test_quadlet/oci_rag/oci-model-rag.container
+++ b/test/unit/data/test_quadlet/oci_rag/oci-model-rag.container
@@ -3,9 +3,6 @@ Description=RamaLama oci-model-rag AI Model Service
 After=local-fs.target
 
 [Container]
-AddDevice=-/dev/accel
-AddDevice=-/dev/dri
-AddDevice=-/dev/kfd
 AddDevice=nvidia.com/gpu=all
 Image=testimage
 RunInit=true
diff --git a/test/unit/data/test_quadlet/portmapping/tinyllama.container b/test/unit/data/test_quadlet/portmapping/tinyllama.container
index 63b5894af..79515b4bb 100644
--- a/test/unit/data/test_quadlet/portmapping/tinyllama.container
+++ b/test/unit/data/test_quadlet/portmapping/tinyllama.container
@@ -3,9 +3,6 @@ Description=RamaLama tinyllama AI Model Service
 After=local-fs.target
 
 [Container]
-AddDevice=-/dev/accel
-AddDevice=-/dev/dri
-AddDevice=-/dev/kfd
 AddDevice=nvidia.com/gpu=all
 Image=testimage
 RunInit=true
diff --git a/test/unit/test_common.py b/test/unit/test_common.py
index cd074e1b5..4f913d0d5 100644
--- a/test/unit/test_common.py
+++ b/test/unit/test_common.py
@@ -26,11 +26,14 @@
     ensure_image,
     find_in_cdi,
     get_accel,
+    get_gpu_devices,
     has_nvidia_vulkan_icd,
     host_available,
     host_cmd,
     host_path,
     in_toolbox,
+    in_wsl,
+    is_windows_or_wsl,
     load_cdi_config,
     populate_volume_from_image,
     rm_until_substring,
@@ -964,3 +967,74 @@ def test_icd_installed(self, icd_dir):
     def test_only_other_vendors(self):
         with patch("ramalama.common.glob.glob", return_value=[]):
             assert not has_nvidia_vulkan_icd()
+
+
+class TestIsWindowsOrWsl:
+    """is_windows_or_wsl() must catch both native Windows and ramalama running inside a WSL2 distro."""
+
+    @pytest.mark.parametrize(
+        "osrelease,expected",
+        [
+            ("5.15.167.4-microsoft-standard-WSL2\n", True),
+            ("6.6.87.2-microsoft-standard-WSL2+\n", True),
+            ("7.1.13-100.fc43.x86_64\n", False),
+        ],
+    )
+    def test_in_wsl_reads_osrelease(self, osrelease, expected):
+        with patch("builtins.open", mock_open(read_data=osrelease)):
+            assert in_wsl() == expected
+
+    def test_in_wsl_missing_osrelease(self):
+        with patch("builtins.open", side_effect=OSError("no /proc")):
+            assert not in_wsl()
+
+    def test_native_windows(self):
+        # No /proc to read on a native Windows interpreter.
+        with (
+            patch("ramalama.common.platform.system", return_value="Windows"),
+            patch("builtins.open", side_effect=OSError("no /proc")),
+        ):
+            assert is_windows_or_wsl()
+
+    def test_inside_wsl_distro(self):
+        # platform.system() reports "Linux" from inside the distro.
+        with (
+            patch("ramalama.common.platform.system", return_value="Linux"),
+            patch("builtins.open", mock_open(read_data="5.15.167.4-microsoft-standard-WSL2\n")),
+        ):
+            assert is_windows_or_wsl()
+
+    def test_native_linux(self):
+        with (
+            patch("ramalama.common.platform.system", return_value="Linux"),
+            patch("builtins.open", mock_open(read_data="7.1.13-100.fc43.x86_64\n")),
+        ):
+            assert not is_windows_or_wsl()
+
+
+class TestGetGpuDevices:
+    devices = {"/dev/accel": "accel", "/dev/dri": "dri", "/dev/kfd": "kfd"}
+
+    def _get_gpu_devices(self, accel_env_vars, environ=None):
+        with (
+            patch("os.path.exists", lambda path: path in self.devices),
+            patch.dict("os.environ", environ or {}, clear=True),
+        ):
+            return get_gpu_devices(accel_env_vars)
+
+    def test_host_devices(self):
+        assert self._get_gpu_devices({"HIP_VISIBLE_DEVICES": "0"}) == {
+            "accel": "/dev/accel",
+            "dri": "/dev/dri",
+            "kfd": "/dev/kfd",
+        }
+
+    def test_nvidia_gets_none(self):
+        # The container toolkit passes the NVIDIA GPUs in itself. Anything the
+        # host's GPU devices would add on top is a GPU that was not asked for,
+        # an iGPU say, and the vulkan backend would offload onto it.
+        assert self._get_gpu_devices({"CUDA_VISIBLE_DEVICES": "0"}) == {}
+
+    def test_defaults_to_the_environment(self):
+        assert self._get_gpu_devices(None, {"CUDA_VISIBLE_DEVICES": "0"}) == {}
+        assert self._get_gpu_devices(None, {"HIP_VISIBLE_DEVICES": "0"}) != {}
diff --git a/test/unit/test_engine.py b/test/unit/test_engine.py
index 7f2b0c78e..69b0b9799 100644
--- a/test/unit/test_engine.py
+++ b/test/unit/test_engine.py
@@ -58,6 +58,80 @@ def test_cuda_device_options_docker_selection(self):
         exec_args = self._cuda_device_args("docker", "1,2", selected=["1", "2"])
         self.assertEqual(exec_args, ["--gpus", '"device=1,2"', "-e", "CUDA_VISIBLE_DEVICES=0,1"])
 
+    def _host_gpu_device_args(self, accel_env_vars):
+        engine = ramalama.engine.Engine(self.base_args)
+        engine.exec_args = []
+        host_devices = ["/dev/accel", "/dev/dri", "/dev/kfd"]
+        with (
+            patch("ramalama.engine.get_accel_env_vars", return_value=dict(accel_env_vars)),
+            # A GPU that WSL would expose as /dev/dxg is a native one here.
+            patch("ramalama.engine.is_windows_or_wsl", return_value=False),
+            patch("os.path.exists", lambda path: path in host_devices),
+            patch("glob.glob", lambda path: [path] if path in host_devices else []),
+            patch.object(ramalama.common, "podman_machine_accel", False),
+            patch.object(ramalama.common, "nvidia_selected_devices", []),
+        ):
+            engine.add_device_options()
+        return engine.exec_args
+
+    def test_host_gpu_devices(self):
+        exec_args = self._host_gpu_device_args({"HIP_VISIBLE_DEVICES": "0"})
+        self.assertEqual(
+            exec_args,
+            [
+                "--device",
+                "/dev/accel",
+                "--device",
+                "/dev/dri",
+                "--device",
+                "/dev/kfd",
+                "-e",
+                "HIP_VISIBLE_DEVICES=0",
+            ],
+        )
+
+    def test_host_gpu_devices_left_out_for_nvidia(self):
+        # The container toolkit passes the NVIDIA GPUs in itself, so /dev/dri
+        # could only add a GPU that was not asked for - an iGPU on a hybrid
+        # host - and the Vulkan backend would offload onto it.
+        exec_args = self._host_gpu_device_args({"CUDA_VISIBLE_DEVICES": "0"})
+        self.assertEqual(exec_args, ["--device", "nvidia.com/gpu=all", "-e", "CUDA_VISIBLE_DEVICES=0"])
+
+    def _wsl_device_args(self, accel_env_var, windows_or_wsl):
+        engine = ramalama.engine.Engine(self.base_args)
+        engine.exec_args = []
+        with (
+            patch("ramalama.engine.get_accel_env_vars", return_value={accel_env_var: "0"}),
+            patch("ramalama.engine.is_windows_or_wsl", return_value=windows_or_wsl),
+            patch("glob.glob", return_value=[]),
+            patch.object(ramalama.common, "podman_machine_accel", False),
+        ):
+            engine.add_device_options()
+        return engine.exec_args
+
+    def test_intel_device_options_native(self):
+        exec_args = self._wsl_device_args("INTEL_VISIBLE_DEVICES", False)
+        self.assertEqual(exec_args, ["-e", "INTEL_VISIBLE_DEVICES=0"])
+
+    def test_wsl_device_options(self):
+        # WSL exposes the AMD GPU the same way as the Intel one. Also covers
+        # ramalama running inside a WSL distro, where platform.system() reports
+        # "Linux" but the GPU is still /dev/dxg.
+        for accel_env_var in ("HIP_VISIBLE_DEVICES", "INTEL_VISIBLE_DEVICES"):
+            with self.subTest(accel_env_var=accel_env_var):
+                exec_args = self._wsl_device_args(accel_env_var, True)
+                self.assertEqual(
+                    exec_args,
+                    [
+                        "--device",
+                        "/dev/dxg",
+                        "--mount",
+                        "type=bind,src=/usr/lib/wsl,dst=/usr/lib/wsl",
+                        "-e",
+                        f"{accel_env_var}=0",
+                    ],
+                )
+
     def test_add_container_labels(self):
         args = Namespace(**vars(self.base_args), MODEL="test-model", port="8080", subcommand="run")
         engine = ramalama.engine.Engine(args)
@@ -249,7 +323,7 @@ def test_remove_network_dryrun_skips_remove(self, mock_run_cmd):
         "none-host",
     ],
 )
-@patch("ramalama.engine.platform.system", return_value="Linux")
+@patch("ramalama.host_utils.platform.system", return_value="Linux")
 def test_add_port_with_host(mock_system, host, port, expected_port_arg):
     base_args = Namespace(
         engine="podman",
@@ -294,7 +368,7 @@ def test_add_port_with_host_on_vm_engine(system, host, port, expected_port_arg):
         host=host,
         port=port,
     )
-    with patch("ramalama.engine.platform.system", return_value=system):
+    with patch("ramalama.host_utils.platform.system", return_value=system):
         engine = ramalama.engine.Engine(base_args)
     p_index = engine.exec_args.index("-p")
     assert engine.exec_args[p_index + 1] == expected_port_arg
diff --git a/test/unit/test_inference_engine_plugins.py b/test/unit/test_inference_engine_plugins.py
index 0a5198a54..80f45ad29 100644
--- a/test/unit/test_inference_engine_plugins.py
+++ b/test/unit/test_inference_engine_plugins.py
@@ -29,6 +29,13 @@
 from ramalama.plugins.runtimes.inference.vllm import VllmPlugin
 
 
+@pytest.fixture(autouse=True)
+def not_windows_or_wsl(monkeypatch):
+    """Backend defaults differ on Windows and WSL, so pin the native path by
+    default. Tests covering that behaviour override this."""
+    monkeypatch.setattr("ramalama.plugins.runtimes.inference.llama_cpp.is_windows_or_wsl", lambda: False)
+
+
 def make_ns(
     container=True,
     file=None,
@@ -752,6 +759,36 @@ def test_get_container_image_user_override(self):
         image = self.plugin.get_container_image(config, "CUDA_VISIBLE_DEVICES")
         assert image == "custom/cuda:v1.0"
 
+    def test_get_container_image_vulkan_follows_default_image(self):
+        # AMD resolves to the vulkan backend, which runs in the default image,
+        # so RAMALAMA_DEFAULT_IMAGE has to be honoured here too.
+        config = MagicMock()
+        config.runtimes = {"llama_cpp": {"backend": "auto"}}
+        config.images.get.return_value = None
+        config.default_image = "my-registry.example/ramalama:snapshot"
+        image = self.plugin.get_container_image(config, "HIP_VISIBLE_DEVICES")
+        assert image == "my-registry.example/ramalama:snapshot"
+
+    def test_get_container_image_user_override_for_detected_gpu(self):
+        # The override is keyed by the detected GPU, while auto resolves to the
+        # vulkan backend; the image the user pinned for their hardware wins.
+        config = MagicMock()
+        config.runtimes = {"llama_cpp": {"backend": "auto"}}
+        config.images.get.side_effect = lambda key, default=None: {
+            "HIP_VISIBLE_DEVICES": "custom/rocm:v1.0",
+        }.get(key, default)
+        image = self.plugin.get_container_image(config, "HIP_VISIBLE_DEVICES")
+        assert image == "custom/rocm:v1.0"
+
+    def test_get_container_image_user_override_for_resolved_backend(self):
+        config = MagicMock()
+        config.runtimes = {"llama_cpp": {"backend": "auto"}}
+        config.images.get.side_effect = lambda key, default=None: {
+            "GGML_VK_VISIBLE_DEVICES": "custom/vulkan:v1.0",
+        }.get(key, default)
+        image = self.plugin.get_container_image(config, "HIP_VISIBLE_DEVICES")
+        assert image == "custom/vulkan:v1.0"
+
     @patch.dict("os.environ", clear=True, INTEL_VISIBLE_DEVICES="1")
     def test_intel_no_openvino(self):
         ns = make_ns()
@@ -1296,8 +1333,8 @@ def test_backend_selection(backend: str, gpu_env: str, expected_result: str, mon
 @pytest.mark.parametrize(
     "backend,gpu_env,expected_result",
     [
-        # Auto mode on Windows: ROCm for AMD, CUDA for NVIDIA, sycl for Intel
-        ("auto", "HIP_VISIBLE_DEVICES", version_tagged_image("quay.io/ramalama/rocm")),  # AMD -> ROCm on Windows
+        # Auto mode on WSL2: ROCm for AMD, CUDA for NVIDIA, sycl for Intel
+        ("auto", "HIP_VISIBLE_DEVICES", version_tagged_image("quay.io/ramalama/rocm")),  # AMD -> ROCm on WSL2
         ("auto", "CUDA_VISIBLE_DEVICES", version_tagged_image("quay.io/ramalama/cuda")),  # NVIDIA -> CUDA
         ("auto", "INTEL_VISIBLE_DEVICES", version_tagged_image("quay.io/ramalama/intel-gpu")),  # Intel -> sycl
         # Explicit backends still work, vulkan included
@@ -1310,10 +1347,10 @@ def test_backend_selection(backend: str, gpu_env: str, expected_result: str, mon
         ("openvino", "INTEL_VISIBLE_DEVICES", version_tagged_image("quay.io/ramalama/openvino")),
     ],
 )
-def test_backend_selection_windows(backend: str, gpu_env: str, expected_result: str, monkeypatch):
-    """Test that Windows defaults to vendor-specific backends for AMD and Intel."""
+def test_backend_selection_wsl(backend: str, gpu_env: str, expected_result: str, monkeypatch):
+    """Test that WSL2 defaults to vendor-specific backends for AMD and Intel."""
     monkeypatch.setattr("ramalama.common.get_accel", lambda: "none")
-    monkeypatch.setattr("ramalama.plugins.runtimes.inference.llama_cpp.platform.system", lambda: "Windows")
+    monkeypatch.setattr("ramalama.plugins.runtimes.inference.llama_cpp.is_windows_or_wsl", lambda: True)
 
     with NamedTemporaryFile('w', delete_on_close=False) as f:
         f.write(f"""\
@@ -1445,16 +1482,16 @@ def test_get_available_backends(gpu_env: Optional[str], expected_backends: list[
 @pytest.mark.parametrize(
     "gpu_env,expected_backends",
     [
-        ("HIP_VISIBLE_DEVICES", ["auto", "rocm", "vulkan"]),  # AMD: ROCm preferred on Windows
+        ("HIP_VISIBLE_DEVICES", ["auto", "rocm", "vulkan"]),  # AMD: ROCm preferred on WSL2
         ("CUDA_VISIBLE_DEVICES", ["auto", "cuda", "vulkan"]),  # NVIDIA: same on all platforms
-        ("INTEL_VISIBLE_DEVICES", ["auto", "sycl", "vulkan", "openvino"]),  # Intel: sycl preferred on Windows
+        ("INTEL_VISIBLE_DEVICES", ["auto", "sycl", "vulkan", "openvino"]),  # Intel: sycl preferred on WSL2
         (None, ["auto", "vulkan"]),  # No GPU: same on all platforms
     ],
 )
-def test_get_available_backends_windows(gpu_env: Optional[str], expected_backends: list[str], monkeypatch):
-    """Test that available backends on Windows prefer vendor-specific backends."""
+def test_get_available_backends_wsl(gpu_env: Optional[str], expected_backends: list[str], monkeypatch):
+    """Test that available backends on WSL2 prefer vendor-specific backends."""
     monkeypatch.setattr("ramalama.common.get_accel", lambda: "none")
-    monkeypatch.setattr("ramalama.plugins.runtimes.inference.llama_cpp.platform.system", lambda: "Windows")
+    monkeypatch.setattr("ramalama.plugins.runtimes.inference.llama_cpp.is_windows_or_wsl", lambda: True)
 
     env = {}
     if gpu_env:
diff --git a/test/unit/test_quadlet.py b/test/unit/test_quadlet.py
index 72a33b0a2..280b6e621 100644
--- a/test/unit/test_quadlet.py
+++ b/test/unit/test_quadlet.py
@@ -325,3 +325,25 @@ def test_quadlet_generate(input: Input, expected_files_path: Path, monkeypatch):
             del expected_files[file.filename]
 
     assert expected_files == dict()
+
+
+def test_quadlet_nvidia_selection(monkeypatch):
+    """A narrowed CUDA_VISIBLE_DEVICES reserves just those GPUs, as it does for
+    "ramalama run": the Vulkan backend cannot be filtered any other way."""
+    monkeypatch.setattr("os.path.exists", lambda path: False)
+    monkeypatch.setattr("ramalama.quadlet.get_accel", lambda: "cuda")
+    monkeypatch.setattr("ramalama.quadlet.get_accel_env_vars", lambda: {"CUDA_VISIBLE_DEVICES": "1,2"})
+    monkeypatch.setattr("ramalama.common.nvidia_selected_devices", ["1", "2"])
+
+    files = Quadlet("tinyllama", ("/blob", "model"), None, None, Args(), [], False, None, None).generate()
+    with io.StringIO() as sio:
+        for file in files:
+            file._write(sio)
+        content = sio.getvalue()
+
+    assert "AddDevice=nvidia.com/gpu=1" in content
+    assert "AddDevice=nvidia.com/gpu=2" in content
+    assert "AddDevice=nvidia.com/gpu=all" not in content
+    # The container sees the two GPUs as 0 and 1, so the host's indices would
+    # name a device that is not there.
+    assert "Environment=CUDA_VISIBLE_DEVICES=0,1" in content