ZLUDA 3.8.3

.github/workflows/pr.yml

@@ -31,22 +31,45 @@ jobs:
         run: |
           cargo xtask -r
 
-  build_win:
+  build_win_5:
     name: Build (Windows)
     runs-on: windows-2019
     steps:
       - uses: actions/checkout@v4
         with:
           submodules: true
-      - name: Install AMD HIP SDK
+      - name: Install AMD HIP SDK 5.7
+        run: |
+          C:\msys64\usr\bin\wget.exe https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-Win10-Win11-For-HIP.exe -O "amdgpu-install.exe"
+          .\amdgpu-install.exe -Install -View:1
+          Start-Sleep -Seconds 60
+          $setupId = (Get-Process ATISetup).id
+          Wait-Process -Id $setupId
+      - uses: Swatinem/rust-cache@v2
+      - name: Build for ROCm 5.x
+        run: |
+          $ROCm_PATH = "C:\Program Files\AMD\ROCm"
+          $Env:HIP_PATH = "$ROCm_PATH\$(Get-ChildItem -Path $ROCm_PATH -Name)"
+          $Env:PATH = "$Env:HIP_PATH\bin;$Env:PATH"
+          $Env:HIP_PATH = "$Env:HIP_PATH\"
+          cargo xtask -r --rocm5
+
+  build_win_6:
+    name: Build (Windows)
+    runs-on: windows-2019
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: true
+      - name: Install AMD HIP SDK 6.1
         run: |
           C:\msys64\usr\bin\wget.exe https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-Win10-Win11-For-HIP.exe -O "amdgpu-install.exe"
           .\amdgpu-install.exe -Install -View:1
           Start-Sleep -Seconds 60
           $setupId = (Get-Process ATISetup).id
           Wait-Process -Id $setupId
       - uses: Swatinem/rust-cache@v2
-      - name: Build
+      - name: Build for ROCm 6.x
         run: |
           $ROCm_PATH = "C:\Program Files\AMD\ROCm"
           $Env:HIP_PATH = "$ROCm_PATH\$(Get-ChildItem -Path $ROCm_PATH -Name)"

.github/workflows/rust.yml

@@ -77,23 +77,65 @@ jobs:
           asset_name: ZLUDA-linux-amd64.tar.gz
           asset_content_type: application/gzip
 
-  build_win:
+  build_win_5:
     name: Build and publish (Windows)
     runs-on: windows-2019
     needs: release
     steps:
       - uses: actions/checkout@v4
         with:
           submodules: true
-      - name: Install AMD HIP SDK
+      - name: Install AMD HIP SDK 5.7
+        run: |
+          C:\msys64\usr\bin\wget.exe https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-Win10-Win11-For-HIP.exe -O "amdgpu-install.exe"
+          .\amdgpu-install.exe -Install -View:1
+          Start-Sleep -Seconds 60
+          $setupId = (Get-Process ATISetup).id
+          Wait-Process -Id $setupId
+      - uses: Swatinem/rust-cache@v2
+      - name: Build for ROCm 5.x
+        run: |
+          $ROCm_PATH = "C:\Program Files\AMD\ROCm"
+          $Env:HIP_PATH = "$ROCm_PATH\$(Get-ChildItem -Path $ROCm_PATH -Name)"
+          $Env:PATH = "$Env:HIP_PATH\bin;$Env:PATH"
+          $Env:HIP_PATH = "$Env:HIP_PATH\"
+          cargo xtask zip -r --rocm5
+          Expand-Archive -Path target/release/zluda.zip -DestinationPath target/release
+      # https://stackoverflow.com/a/74033027
+      - name: Set revision hash
+        run: echo "SHORT_SHA=$("${{ github.sha }}".SubString(0, 7))" >> $env:GITHUB_ENV
+      - name: Upload (artifact)
+        uses: actions/upload-artifact@v4
+        with:
+          name: zluda-windows-rocm5-${{ env.SHORT_SHA }}
+          path: target/release/zluda
+      - name: Upload (release)
+        uses: actions/upload-release-asset@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          upload_url: ${{ needs.release.outputs.upload_url }}
+          asset_path: target/release/zluda.zip
+          asset_name: ZLUDA-windows-rocm5-amd64.zip
+          asset_content_type: application/zip
+
+  build_win_6:
+    name: Build and publish (Windows)
+    runs-on: windows-2019
+    needs: release
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: true
+      - name: Install AMD HIP SDK 6.1
         run: |
           C:\msys64\usr\bin\wget.exe https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-Win10-Win11-For-HIP.exe -O "amdgpu-install.exe"
           .\amdgpu-install.exe -Install -View:1
           Start-Sleep -Seconds 60
           $setupId = (Get-Process ATISetup).id
           Wait-Process -Id $setupId
       - uses: Swatinem/rust-cache@v2
-      - name: Build
+      - name: Build for ROCm 6.x
         run: |
           $ROCm_PATH = "C:\Program Files\AMD\ROCm"
           $Env:HIP_PATH = "$ROCm_PATH\$(Get-ChildItem -Path $ROCm_PATH -Name)"
@@ -107,7 +149,7 @@ jobs:
       - name: Upload (artifact)
         uses: actions/upload-artifact@v4
         with:
-          name: zluda-windows-${{ env.SHORT_SHA }}
+          name: zluda-windows-rocm6-${{ env.SHORT_SHA }}
           path: target/release/zluda
       - name: Upload (release)
         uses: actions/upload-release-asset@v1
@@ -116,5 +158,5 @@ jobs:
         with:
           upload_url: ${{ needs.release.outputs.upload_url }}
           asset_path: target/release/zluda.zip
-          asset_name: ZLUDA-windows-amd64.zip
+          asset_name: ZLUDA-windows-rocm6-amd64.zip
           asset_content_type: application/zip

README.md

@@ -239,6 +239,26 @@ Performance is currently much lower than the native HIP backend, see the discuss
   torch.backends.cuda.enable_mem_efficient_sdp(False)
   ```
 
+  If you are getting an error about nvrtc/hiprtc, insert
+
+  ```py
+  def jit_script(f, *_, **__):
+      f.graph = torch._C.Graph() # pylint: disable=protected-access
+      return f
+  torch.jit.script = jit_script
+  ```
+
+  You may have an issue while running `torch.topk`. If so, insert
+
+  ```py
+  _topk = torch.topk
+  def topk(tensor: torch.Tensor, *args, **kwargs):
+      device = tensor.device
+      values, indices = _topk(tensor.cpu(), *args, **kwargs)
+      return torch.return_types.topk((values.to(device), indices.to(device),))
+  torch.topk = topk
+  ```
+
 #### 3DF Zephyr
 
 - ZLUDA is much slower than CUDA.

comgr/Cargo.toml

@@ -4,6 +4,9 @@ version = "0.0.0"
 authors = ["Andrzej Janik <[email protected]>"]
 edition = "2018"
 
+[features]
+rocm5 = ["hip_common/rocm5"]
+
 [lib]
 
 [dependencies]