Add ROCm 6.4 to `--torch-backend=auto` (#16919)

## Summary

Closes https://github.com/astral-sh/uv/issues/16917.

crates/uv-torch/src/backend.rs

@@ -113,6 +113,10 @@ pub enum TorchMode {
     Cu90,
     /// Use the PyTorch index for CUDA 8.0.
     Cu80,
+    /// Use the PyTorch index for ROCm 6.4.
+    #[serde(rename = "rocm6.4")]
+    #[cfg_attr(feature = "clap", clap(name = "rocm6.4"))]
+    Rocm64,
     /// Use the PyTorch index for ROCm 6.3.
     #[serde(rename = "rocm6.3")]
     #[cfg_attr(feature = "clap", clap(name = "rocm6.3"))]
@@ -272,6 +276,7 @@ impl TorchStrategy {
             TorchMode::Cu91 => TorchBackend::Cu91,
             TorchMode::Cu90 => TorchBackend::Cu90,
             TorchMode::Cu80 => TorchBackend::Cu80,
+            TorchMode::Rocm64 => TorchBackend::Rocm64,
             TorchMode::Rocm63 => TorchBackend::Rocm63,
             TorchMode::Rocm624 => TorchBackend::Rocm624,
             TorchMode::Rocm62 => TorchBackend::Rocm62,
@@ -516,6 +521,7 @@ pub enum TorchBackend {
     Cu91,
     Cu90,
     Cu80,
+    Rocm64,
     Rocm63,
     Rocm624,
     Rocm62,
@@ -647,6 +653,10 @@ impl TorchBackend {
                 TorchSource::PyTorch => &PYTORCH_CU80_INDEX_URL,
                 TorchSource::Pyx => &PYX_CU80_INDEX_URL,
             },
+            Self::Rocm64 => match source {
+                TorchSource::PyTorch => &PYTORCH_ROCM64_INDEX_URL,
+                TorchSource::Pyx => &PYX_ROCM64_INDEX_URL,
+            },
             Self::Rocm63 => match source {
                 TorchSource::PyTorch => &PYTORCH_ROCM63_INDEX_URL,
                 TorchSource::Pyx => &PYX_ROCM63_INDEX_URL,
@@ -774,6 +784,7 @@ impl TorchBackend {
             Self::Cu91 => Some(Version::new([9, 1])),
             Self::Cu90 => Some(Version::new([9, 0])),
             Self::Cu80 => Some(Version::new([8, 0])),
+            Self::Rocm64 => None,
             Self::Rocm63 => None,
             Self::Rocm624 => None,
             Self::Rocm62 => None,
@@ -824,6 +835,7 @@ impl TorchBackend {
             Self::Cu91 => None,
             Self::Cu90 => None,
             Self::Cu80 => None,
+            Self::Rocm64 => Some(Version::new([6, 4])),
             Self::Rocm63 => Some(Version::new([6, 3])),
             Self::Rocm624 => Some(Version::new([6, 2, 4])),
             Self::Rocm62 => Some(Version::new([6, 2])),
@@ -877,6 +889,7 @@ impl FromStr for TorchBackend {
             "cu91" => Ok(Self::Cu91),
             "cu90" => Ok(Self::Cu90),
             "cu80" => Ok(Self::Cu80),
+            "rocm6.4" => Ok(Self::Rocm64),
             "rocm6.3" => Ok(Self::Rocm63),
             "rocm6.2.4" => Ok(Self::Rocm624),
             "rocm6.2" => Ok(Self::Rocm62),
@@ -991,9 +1004,21 @@ static WINDOWS_CUDA_VERSIONS: LazyLock<[(TorchBackend, Version); 26]> = LazyLock
 ///
 /// 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]> =
+static LINUX_AMD_GPU_DRIVERS: LazyLock<[(TorchBackend, AmdGpuArchitecture); 55]> =
     LazyLock::new(|| {
         [
+            // ROCm 6.4
+            (TorchBackend::Rocm64, AmdGpuArchitecture::Gfx900),
+            (TorchBackend::Rocm64, AmdGpuArchitecture::Gfx906),
+            (TorchBackend::Rocm64, AmdGpuArchitecture::Gfx908),
+            (TorchBackend::Rocm64, AmdGpuArchitecture::Gfx90a),
+            (TorchBackend::Rocm64, AmdGpuArchitecture::Gfx942),
+            (TorchBackend::Rocm64, AmdGpuArchitecture::Gfx1030),
+            (TorchBackend::Rocm64, AmdGpuArchitecture::Gfx1100),
+            (TorchBackend::Rocm64, AmdGpuArchitecture::Gfx1101),
+            (TorchBackend::Rocm64, AmdGpuArchitecture::Gfx1102),
+            (TorchBackend::Rocm64, AmdGpuArchitecture::Gfx1200),
+            (TorchBackend::Rocm64, AmdGpuArchitecture::Gfx1201),
             // ROCm 6.3
             (TorchBackend::Rocm63, AmdGpuArchitecture::Gfx900),
             (TorchBackend::Rocm63, AmdGpuArchitecture::Gfx906),
@@ -1100,6 +1125,8 @@ static PYTORCH_CU90_INDEX_URL: LazyLock<IndexUrl> =
     LazyLock::new(|| IndexUrl::from_str("https://download.pytorch.org/whl/cu90").unwrap());
 static PYTORCH_CU80_INDEX_URL: LazyLock<IndexUrl> =
     LazyLock::new(|| IndexUrl::from_str("https://download.pytorch.org/whl/cu80").unwrap());
+static PYTORCH_ROCM64_INDEX_URL: LazyLock<IndexUrl> =
+    LazyLock::new(|| IndexUrl::from_str("https://download.pytorch.org/whl/rocm6.4").unwrap());
 static PYTORCH_ROCM63_INDEX_URL: LazyLock<IndexUrl> =
     LazyLock::new(|| IndexUrl::from_str("https://download.pytorch.org/whl/rocm6.3").unwrap());
 static PYTORCH_ROCM624_INDEX_URL: LazyLock<IndexUrl> =
@@ -1248,6 +1275,10 @@ static PYX_CU80_INDEX_URL: LazyLock<IndexUrl> = LazyLock::new(|| {
     let api_base_url = &*PYX_API_BASE_URL;
     IndexUrl::from_str(&format!("{api_base_url}/simple/astral-sh/cu80")).unwrap()
 });
+static PYX_ROCM64_INDEX_URL: LazyLock<IndexUrl> = LazyLock::new(|| {
+    let api_base_url = &*PYX_API_BASE_URL;
+    IndexUrl::from_str(&format!("{api_base_url}/simple/astral-sh/rocm6.4")).unwrap()
+});
 static PYX_ROCM63_INDEX_URL: LazyLock<IndexUrl> = LazyLock::new(|| {
     let api_base_url = &*PYX_API_BASE_URL;
     IndexUrl::from_str(&format!("{api_base_url}/simple/astral-sh/rocm6.3")).unwrap()

docs/reference/cli.md

@@ -4301,6 +4301,7 @@ by <code>--python-version</code>.</p>
 <li><code>cu91</code>:  Use the PyTorch index for CUDA 9.1</li>
 <li><code>cu90</code>:  Use the PyTorch index for CUDA 9.0</li>
 <li><code>cu80</code>:  Use the PyTorch index for CUDA 8.0</li>
+<li><code>rocm6.4</code>:  Use the PyTorch index for ROCm 6.4</li>
 <li><code>rocm6.3</code>:  Use the PyTorch index for ROCm 6.3</li>
 <li><code>rocm6.2.4</code>:  Use the PyTorch index for ROCm 6.2.4</li>
 <li><code>rocm6.2</code>:  Use the PyTorch index for ROCm 6.2</li>
@@ -4586,6 +4587,7 @@ be used with caution, as it can modify the system Python installation.</p>
 <li><code>cu91</code>:  Use the PyTorch index for CUDA 9.1</li>
 <li><code>cu90</code>:  Use the PyTorch index for CUDA 9.0</li>
 <li><code>cu80</code>:  Use the PyTorch index for CUDA 8.0</li>
+<li><code>rocm6.4</code>:  Use the PyTorch index for ROCm 6.4</li>
 <li><code>rocm6.3</code>:  Use the PyTorch index for ROCm 6.3</li>
 <li><code>rocm6.2.4</code>:  Use the PyTorch index for ROCm 6.2.4</li>
 <li><code>rocm6.2</code>:  Use the PyTorch index for ROCm 6.2</li>
@@ -4899,6 +4901,7 @@ should be used with caution, as it can modify the system Python installation.</p
 <li><code>cu91</code>:  Use the PyTorch index for CUDA 9.1</li>
 <li><code>cu90</code>:  Use the PyTorch index for CUDA 9.0</li>
 <li><code>cu80</code>:  Use the PyTorch index for CUDA 8.0</li>
+<li><code>rocm6.4</code>:  Use the PyTorch index for ROCm 6.4</li>
 <li><code>rocm6.3</code>:  Use the PyTorch index for ROCm 6.3</li>
 <li><code>rocm6.2.4</code>:  Use the PyTorch index for ROCm 6.2.4</li>
 <li><code>rocm6.2</code>:  Use the PyTorch index for ROCm 6.2</li>

uv.schema.json

@@ -2647,6 +2647,11 @@
           "type": "string",
           "const": "cu80"
         },
+        {
+          "description": "Use the PyTorch index for ROCm 6.4.",
+          "type": "string",
+          "const": "rocm6.4"
+        },
         {
           "description": "Use the PyTorch index for ROCm 6.3.",
           "type": "string",