Add auto-detection for AMD GPUs (#14176)

## Summary

Allows `--torch-backend=auto` to detect AMD GPUs. The approach is fairly
well-documented inline, but I opted for `rocm_agent_enumerator` over
(e.g.) `rocminfo` since it seems to be the recommended approach for
scripting:
https://rocm.docs.amd.com/projects/rocminfo/en/latest/how-to/use-rocm-agent-enumerator.html.

Closes https://github.com/astral-sh/uv/issues/14086.

## Test Plan

```
root@rocm-jupyter-gpu-mi300x1-192gb-devcloud-atl1:~# ./uv-linux-libc-11fb582c5c046bae09766ceddd276dcc5bb41218/uv pip install torch --torch-backend=auto
Resolved 11 packages in 251ms
Prepared 2 packages in 6ms
Installed 11 packages in 257ms
 + filelock==3.18.0
 + fsspec==2025.5.1
 + jinja2==3.1.6
 + markupsafe==3.0.2
 + mpmath==1.3.0
 + networkx==3.5
 + pytorch-triton-rocm==3.3.1
 + setuptools==80.9.0
 + sympy==1.14.0
 + torch==2.7.1+rocm6.3
 + typing-extensions==4.14.0
```

---------

Co-authored-by: Zanie Blue <contact@zanie.dev>

crates/uv-resolver/src/resolver/mod.rs

@@ -1814,7 +1814,7 @@ impl<InstalledPackages: InstalledPackagesProvider> ResolverState<InstalledPackag
                     .options
                     .torch_backend
                     .as_ref()
-                    .filter(|torch_backend| matches!(torch_backend, TorchStrategy::Auto { .. }))
+                    .filter(|torch_backend| matches!(torch_backend, TorchStrategy::Cuda { .. }))
                     .and_then(|_| pins.get(name, version).and_then(ResolvedDist::index))
                     .map(IndexUrl::url)
                     .and_then(SystemDependency::from_index)

crates/uv-static/src/env_vars.rs

@@ -718,10 +718,14 @@ impl EnvVars {
     /// This is a quasi-standard variable, described, e.g., in `ncurses(3x)`.
     pub const COLUMNS: &'static str = "COLUMNS";
 
-    /// The CUDA driver version to assume when inferring the PyTorch backend.
+    /// The CUDA driver version to assume when inferring the PyTorch backend (e.g., `550.144.03`).
     #[attr_hidden]
     pub const UV_CUDA_DRIVER_VERSION: &'static str = "UV_CUDA_DRIVER_VERSION";
 
+    /// The AMD GPU architecture to assume when inferring the PyTorch backend (e.g., `gfx1100`).
+    #[attr_hidden]
+    pub const UV_AMD_GPU_ARCHITECTURE: &'static str = "UV_AMD_GPU_ARCHITECTURE";
+
     /// Equivalent to the `--torch-backend` command-line argument (e.g., `cpu`, `cu126`, or `auto`).
     pub const UV_TORCH_BACKEND: &'static str = "UV_TORCH_BACKEND";
 

crates/uv-torch/src/accelerator.rs

@@ -13,17 +13,30 @@ pub enum AcceleratorError {
     Version(#[from] uv_pep440::VersionParseError),
     #[error(transparent)]
     Utf8(#[from] std::string::FromUtf8Error),
+    #[error("Unknown AMD GPU architecture: {0}")]
+    UnknownAmdGpuArchitecture(String),
 }
 
 #[derive(Debug, Clone, Eq, PartialEq)]
 pub enum Accelerator {
+    /// The CUDA driver version (e.g., `550.144.03`).
+    ///
+    /// This is in contrast to the CUDA toolkit version (e.g., `12.8.0`).
     Cuda { driver_version: Version },
+    /// The AMD GPU architecture (e.g., `gfx906`).
+    ///
+    /// This is in contrast to the user-space ROCm version (e.g., `6.4.0-47`) or the kernel-mode
+    /// driver version (e.g., `6.12.12`).
+    Amd {
+        gpu_architecture: AmdGpuArchitecture,
+    },
 }
 
 impl std::fmt::Display for Accelerator {
     fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
         match self {
             Self::Cuda { driver_version } => write!(f, "CUDA {driver_version}"),
+            Self::Amd { gpu_architecture } => write!(f, "AMD {gpu_architecture}"),
         }
     }
 }
@@ -33,9 +46,11 @@ impl Accelerator {
     ///
     /// Query, in order:
     /// 1. The `UV_CUDA_DRIVER_VERSION` environment variable.
+    /// 2. The `UV_AMD_GPU_ARCHITECTURE` environment variable.
     /// 2. `/sys/module/nvidia/version`, which contains the driver version (e.g., `550.144.03`).
     /// 3. `/proc/driver/nvidia/version`, which contains the driver version among other information.
     /// 4. `nvidia-smi --query-gpu=driver_version --format=csv,noheader`.
+    /// 5. `rocm_agent_enumerator`, which lists the AMD GPU architectures.
     pub fn detect() -> Result<Option<Self>, AcceleratorError> {
         // Read from `UV_CUDA_DRIVER_VERSION`.
         if let Ok(driver_version) = std::env::var(EnvVars::UV_CUDA_DRIVER_VERSION) {
@@ -44,6 +59,15 @@ impl Accelerator {
             return Ok(Some(Self::Cuda { driver_version }));
         }
 
+        // Read from `UV_AMD_GPU_ARCHITECTURE`.
+        if let Ok(gpu_architecture) = std::env::var(EnvVars::UV_AMD_GPU_ARCHITECTURE) {
+            let gpu_architecture = AmdGpuArchitecture::from_str(&gpu_architecture)?;
+            debug!(
+                "Detected AMD GPU architecture from `UV_AMD_GPU_ARCHITECTURE`: {gpu_architecture}"
+            );
+            return Ok(Some(Self::Amd { gpu_architecture }));
+        }
+
         // Read from `/sys/module/nvidia/version`.
         match fs_err::read_to_string("/sys/module/nvidia/version") {
             Ok(content) => {
@@ -100,7 +124,34 @@ impl Accelerator {
             );
         }
 
-        debug!("Failed to detect CUDA driver version");
+        // Query `rocm_agent_enumerator` to detect the AMD GPU architecture.
+        //
+        // See: https://rocm.docs.amd.com/projects/rocminfo/en/latest/how-to/use-rocm-agent-enumerator.html
+        if let Ok(output) = std::process::Command::new("rocm_agent_enumerator").output() {
+            if output.status.success() {
+                let stdout = String::from_utf8(output.stdout)?;
+                if let Some(gpu_architecture) = stdout
+                    .lines()
+                    .map(str::trim)
+                    .filter_map(|line| AmdGpuArchitecture::from_str(line).ok())
+                    .min()
+                {
+                    debug!(
+                        "Detected AMD GPU architecture from `rocm_agent_enumerator`: {gpu_architecture}"
+                    );
+                    return Ok(Some(Self::Amd { gpu_architecture }));
+                }
+            } else {
+                debug!(
+                    "Failed to query AMD GPU architecture with `rocm_agent_enumerator` with status `{}`: {}",
+                    output.status,
+                    String::from_utf8_lossy(&output.stderr)
+                );
+            }
+        }
+
+        debug!("Failed to detect GPU driver version");
+
         Ok(None)
     }
 }
@@ -129,6 +180,63 @@ fn parse_proc_driver_nvidia_version(content: &str) -> Result<Option<Version>, Ac
     Ok(Some(driver_version))
 }
 
+/// A GPU architecture for AMD GPUs.
+///
+/// See: <https://rocm.docs.amd.com/projects/install-on-linux/en/latest/reference/system-requirements.html>
+#[derive(Debug, Copy, Clone, Eq, PartialEq, Ord, PartialOrd)]
+pub enum AmdGpuArchitecture {
+    Gfx900,
+    Gfx906,
+    Gfx908,
+    Gfx90a,
+    Gfx942,
+    Gfx1030,
+    Gfx1100,
+    Gfx1101,
+    Gfx1102,
+    Gfx1200,
+    Gfx1201,
+}
+
+impl FromStr for AmdGpuArchitecture {
+    type Err = AcceleratorError;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s {
+            "gfx900" => Ok(Self::Gfx900),
+            "gfx906" => Ok(Self::Gfx906),
+            "gfx908" => Ok(Self::Gfx908),
+            "gfx90a" => Ok(Self::Gfx90a),
+            "gfx942" => Ok(Self::Gfx942),
+            "gfx1030" => Ok(Self::Gfx1030),
+            "gfx1100" => Ok(Self::Gfx1100),
+            "gfx1101" => Ok(Self::Gfx1101),
+            "gfx1102" => Ok(Self::Gfx1102),
+            "gfx1200" => Ok(Self::Gfx1200),
+            "gfx1201" => Ok(Self::Gfx1201),
+            _ => Err(AcceleratorError::UnknownAmdGpuArchitecture(s.to_string())),
+        }
+    }
+}
+
+impl std::fmt::Display for AmdGpuArchitecture {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            Self::Gfx900 => write!(f, "gfx900"),
+            Self::Gfx906 => write!(f, "gfx906"),
+            Self::Gfx908 => write!(f, "gfx908"),
+            Self::Gfx90a => write!(f, "gfx90a"),
+            Self::Gfx942 => write!(f, "gfx942"),
+            Self::Gfx1030 => write!(f, "gfx1030"),
+            Self::Gfx1100 => write!(f, "gfx1100"),
+            Self::Gfx1101 => write!(f, "gfx1101"),
+            Self::Gfx1102 => write!(f, "gfx1102"),
+            Self::Gfx1200 => write!(f, "gfx1200"),
+            Self::Gfx1201 => write!(f, "gfx1201"),
+        }
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

crates/uv-torch/src/backend.rs

@@ -47,7 +47,7 @@ use uv_normalize::PackageName;
 use uv_pep440::Version;
 use uv_platform_tags::Os;
 
-use crate::{Accelerator, AcceleratorError};
+use crate::{Accelerator, AcceleratorError, AmdGpuArchitecture};
 
 /// The strategy to use when determining the appropriate PyTorch index.
 #[derive(Debug, Copy, Clone, Eq, PartialEq, serde::Deserialize, serde::Serialize)]
@@ -178,8 +178,13 @@ pub enum TorchMode {
 /// The strategy to use when determining the appropriate PyTorch index.
 #[derive(Debug, Clone, Eq, PartialEq)]
 pub enum TorchStrategy {
-    /// Select the appropriate PyTorch index based on the operating system and CUDA driver version.
-    Auto { os: Os, driver_version: Version },
+    /// Select the appropriate PyTorch index based on the operating system and CUDA driver version (e.g., `550.144.03`).
+    Cuda { os: Os, driver_version: Version },
+    /// Select the appropriate PyTorch index based on the operating system and AMD GPU architecture (e.g., `gfx1100`).
+    Amd {
+        os: Os,
+        gpu_architecture: AmdGpuArchitecture,
+    },
     /// Use the specified PyTorch index.
     Backend(TorchBackend),
 }
@@ -188,16 +193,17 @@ impl TorchStrategy {
     /// Determine the [`TorchStrategy`] from the given [`TorchMode`], [`Os`], and [`Accelerator`].
     pub fn from_mode(mode: TorchMode, os: &Os) -> Result<Self, AcceleratorError> {
         match mode {
-            TorchMode::Auto => {
-                if let Some(Accelerator::Cuda { driver_version }) = Accelerator::detect()? {
-                    Ok(Self::Auto {
-                        os: os.clone(),
-                        driver_version: driver_version.clone(),
-                    })
-                } else {
-                    Ok(Self::Backend(TorchBackend::Cpu))
-                }
-            }
+            TorchMode::Auto => match Accelerator::detect()? {
+                Some(Accelerator::Cuda { driver_version }) => Ok(Self::Cuda {
+                    os: os.clone(),
+                    driver_version: driver_version.clone(),
+                }),
+                Some(Accelerator::Amd { gpu_architecture }) => Ok(Self::Amd {
+                    os: os.clone(),
+                    gpu_architecture,
+                }),
+                None => Ok(Self::Backend(TorchBackend::Cpu)),
+            },
             TorchMode::Cpu => Ok(Self::Backend(TorchBackend::Cpu)),
             TorchMode::Cu128 => Ok(Self::Backend(TorchBackend::Cu128)),
             TorchMode::Cu126 => Ok(Self::Backend(TorchBackend::Cu126)),
@@ -267,25 +273,27 @@ impl TorchStrategy {
     /// Return the appropriate index URLs for the given [`TorchStrategy`].
     pub fn index_urls(&self) -> impl Iterator<Item = &IndexUrl> {
         match self {
-            TorchStrategy::Auto { os, driver_version } => {
+            TorchStrategy::Cuda { os, driver_version } => {
                 // If this is a GPU-enabled package, and CUDA drivers are installed, use PyTorch's CUDA
                 // indexes.
                 //
                 // See: https://github.com/pmeier/light-the-torch/blob/33397cbe45d07b51ad8ee76b004571a4c236e37f/light_the_torch/_patch.py#L36-L49
                 match os {
-                    Os::Manylinux { .. } | Os::Musllinux { .. } => Either::Left(Either::Left(
-                        LINUX_DRIVERS
-                            .iter()
-                            .filter_map(move |(backend, version)| {
-                                if driver_version >= version {
-                                    Some(backend.index_url())
-                                } else {
-                                    None
-                                }
-                            })
-                            .chain(std::iter::once(TorchBackend::Cpu.index_url())),
-                    )),
-                    Os::Windows => Either::Left(Either::Right(
+                    Os::Manylinux { .. } | Os::Musllinux { .. } => {
+                        Either::Left(Either::Left(Either::Left(
+                            LINUX_CUDA_DRIVERS
+                                .iter()
+                                .filter_map(move |(backend, version)| {
+                                    if driver_version >= version {
+                                        Some(backend.index_url())
+                                    } else {
+                                        None
+                                    }
+                                })
+                                .chain(std::iter::once(TorchBackend::Cpu.index_url())),
+                        )))
+                    }
+                    Os::Windows => Either::Left(Either::Left(Either::Right(
                         WINDOWS_CUDA_VERSIONS
                             .iter()
                             .filter_map(move |(backend, version)| {
@@ -296,7 +304,7 @@ impl TorchStrategy {
                                 }
                             })
                             .chain(std::iter::once(TorchBackend::Cpu.index_url())),
-                    )),
+                    ))),
                     Os::Macos { .. }
                     | Os::FreeBsd { .. }
                     | Os::NetBsd { .. }
@@ -306,11 +314,42 @@ impl TorchStrategy {
                     | Os::Haiku { .. }
                     | Os::Android { .. }
                     | Os::Pyodide { .. } => {
-                        Either::Right(std::iter::once(TorchBackend::Cpu.index_url()))
+                        Either::Right(Either::Left(std::iter::once(TorchBackend::Cpu.index_url())))
                     }
                 }
             }
-            TorchStrategy::Backend(backend) => Either::Right(std::iter::once(backend.index_url())),
+            TorchStrategy::Amd {
+                os,
+                gpu_architecture,
+            } => match os {
+                Os::Manylinux { .. } | Os::Musllinux { .. } => Either::Left(Either::Right(
+                    LINUX_AMD_GPU_DRIVERS
+                        .iter()
+                        .filter_map(move |(backend, architecture)| {
+                            if gpu_architecture == architecture {
+                                Some(backend.index_url())
+                            } else {
+                                None
+                            }
+                        })
+                        .chain(std::iter::once(TorchBackend::Cpu.index_url())),
+                )),
+                Os::Windows
+                | Os::Macos { .. }
+                | Os::FreeBsd { .. }
+                | Os::NetBsd { .. }
+                | Os::OpenBsd { .. }
+                | Os::Dragonfly { .. }
+                | Os::Illumos { .. }
+                | Os::Haiku { .. }
+                | Os::Android { .. }
+                | Os::Pyodide { .. } => {
+                    Either::Right(Either::Left(std::iter::once(TorchBackend::Cpu.index_url())))
+                }
+            },
+            TorchStrategy::Backend(backend) => {
+                Either::Right(Either::Right(std::iter::once(backend.index_url())))
+            }
         }
     }
 }
@@ -578,7 +617,7 @@ impl FromStr for TorchBackend {
 /// Linux CUDA driver versions and the corresponding CUDA versions.
 ///
 /// See: <https://github.com/pmeier/light-the-torch/blob/33397cbe45d07b51ad8ee76b004571a4c236e37f/light_the_torch/_cb.py#L150-L213>
-static LINUX_DRIVERS: LazyLock<[(TorchBackend, Version); 24]> = LazyLock::new(|| {
+static LINUX_CUDA_DRIVERS: LazyLock<[(TorchBackend, Version); 24]> = LazyLock::new(|| {
     [
         // Table 2 from
         // https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html
@@ -651,6 +690,73 @@ static WINDOWS_CUDA_VERSIONS: LazyLock<[(TorchBackend, Version); 24]> = LazyLock
     ]
 });
 
+/// Linux AMD GPU architectures and the corresponding PyTorch backends.
+///
+/// These were inferred by running the following snippet for each ROCm version:
+///
+/// ```python
+/// import torch
+///
+/// print(torch.cuda.get_arch_list())
+/// ```
+///
+/// AMD also provides a compatibility matrix: <https://rocm.docs.amd.com/en/latest/compatibility/compatibility-matrix.html>;
+/// however, this list includes a broader array of GPUs than those in the matrix.
+static LINUX_AMD_GPU_DRIVERS: LazyLock<[(TorchBackend, AmdGpuArchitecture); 44]> =
+    LazyLock::new(|| {
+        [
+            // ROCm 6.3
+            (TorchBackend::Rocm63, AmdGpuArchitecture::Gfx900),
+            (TorchBackend::Rocm63, AmdGpuArchitecture::Gfx906),
+            (TorchBackend::Rocm63, AmdGpuArchitecture::Gfx908),
+            (TorchBackend::Rocm63, AmdGpuArchitecture::Gfx90a),
+            (TorchBackend::Rocm63, AmdGpuArchitecture::Gfx942),
+            (TorchBackend::Rocm63, AmdGpuArchitecture::Gfx1030),
+            (TorchBackend::Rocm63, AmdGpuArchitecture::Gfx1100),
+            (TorchBackend::Rocm63, AmdGpuArchitecture::Gfx1101),
+            (TorchBackend::Rocm63, AmdGpuArchitecture::Gfx1102),
+            (TorchBackend::Rocm63, AmdGpuArchitecture::Gfx1200),
+            (TorchBackend::Rocm63, AmdGpuArchitecture::Gfx1201),
+            // ROCm 6.2.4
+            (TorchBackend::Rocm624, AmdGpuArchitecture::Gfx900),
+            (TorchBackend::Rocm624, AmdGpuArchitecture::Gfx906),
+            (TorchBackend::Rocm624, AmdGpuArchitecture::Gfx908),
+            (TorchBackend::Rocm624, AmdGpuArchitecture::Gfx90a),
+            (TorchBackend::Rocm624, AmdGpuArchitecture::Gfx942),
+            (TorchBackend::Rocm624, AmdGpuArchitecture::Gfx1030),
+            (TorchBackend::Rocm624, AmdGpuArchitecture::Gfx1100),
+            (TorchBackend::Rocm624, AmdGpuArchitecture::Gfx1101),
+            (TorchBackend::Rocm624, AmdGpuArchitecture::Gfx1102),
+            (TorchBackend::Rocm624, AmdGpuArchitecture::Gfx1200),
+            (TorchBackend::Rocm624, AmdGpuArchitecture::Gfx1201),
+            // ROCm 6.2
+            (TorchBackend::Rocm62, AmdGpuArchitecture::Gfx900),
+            (TorchBackend::Rocm62, AmdGpuArchitecture::Gfx906),
+            (TorchBackend::Rocm62, AmdGpuArchitecture::Gfx908),
+            (TorchBackend::Rocm62, AmdGpuArchitecture::Gfx90a),
+            (TorchBackend::Rocm62, AmdGpuArchitecture::Gfx1030),
+            (TorchBackend::Rocm62, AmdGpuArchitecture::Gfx1100),
+            (TorchBackend::Rocm62, AmdGpuArchitecture::Gfx942),
+            // ROCm 6.1
+            (TorchBackend::Rocm61, AmdGpuArchitecture::Gfx900),
+            (TorchBackend::Rocm61, AmdGpuArchitecture::Gfx906),
+            (TorchBackend::Rocm61, AmdGpuArchitecture::Gfx908),
+            (TorchBackend::Rocm61, AmdGpuArchitecture::Gfx90a),
+            (TorchBackend::Rocm61, AmdGpuArchitecture::Gfx942),
+            (TorchBackend::Rocm61, AmdGpuArchitecture::Gfx1030),
+            (TorchBackend::Rocm61, AmdGpuArchitecture::Gfx1100),
+            (TorchBackend::Rocm61, AmdGpuArchitecture::Gfx1101),
+            // ROCm 6.0
+            (TorchBackend::Rocm60, AmdGpuArchitecture::Gfx900),
+            (TorchBackend::Rocm60, AmdGpuArchitecture::Gfx906),
+            (TorchBackend::Rocm60, AmdGpuArchitecture::Gfx908),
+            (TorchBackend::Rocm60, AmdGpuArchitecture::Gfx90a),
+            (TorchBackend::Rocm60, AmdGpuArchitecture::Gfx1030),
+            (TorchBackend::Rocm60, AmdGpuArchitecture::Gfx1100),
+            (TorchBackend::Rocm60, AmdGpuArchitecture::Gfx942),
+        ]
+    });
+
 static CPU_INDEX_URL: LazyLock<IndexUrl> =
     LazyLock::new(|| IndexUrl::from_str("https://download.pytorch.org/whl/cpu").unwrap());
 static CU128_INDEX_URL: LazyLock<IndexUrl> =

docs/guides/integration/pytorch.md

@@ -444,10 +444,10 @@ $ # With an environment variable.
 $ UV_TORCH_BACKEND=auto uv pip install torch
 ```
 
-When enabled, uv will query for the installed CUDA driver version and use the most-compatible
-PyTorch index for all relevant packages (e.g., `torch`, `torchvision`, etc.). If no such CUDA driver
-is found, uv will fall back to the CPU-only index. uv will continue to respect existing index
-configuration for any packages outside the PyTorch ecosystem.
+When enabled, uv will query for the installed CUDA driver and AMD GPU versions then use the
+most-compatible PyTorch index for all relevant packages (e.g., `torch`, `torchvision`, etc.). If no
+such GPU is found, uv will fall back to the CPU-only index. uv will continue to respect existing
+index configuration for any packages outside the PyTorch ecosystem.
 
 You can also select a specific backend (e.g., CUDA 12.6) with `--torch-backend=cu126` (or
 `UV_TORCH_BACKEND=cu126`):
@@ -460,5 +460,4 @@ $ # With an environment variable.
 $ UV_TORCH_BACKEND=cu126 uv pip install torch torchvision
 ```
 
-At present, `--torch-backend` is only available in the `uv pip` interface, and only supports
-detection of CUDA drivers (as opposed to other accelerators like ROCm or Intel GPUs).
+At present, `--torch-backend` is only available in the `uv pip` interface.