ggml-webgpu: Enable NVIDIA self-hosted CI (#22976)

* Enabel nvidia ci for webgpu

* Address precision issues

* fix placement

* Relax more set_rows and div

* Try relaxing all f16

* formatting and naming

* Add comment explaining max_nmse_err logic

Added comment referencing pull request for clarification.

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);
         }
     }