diff --git a/.vscode/settings.json b/.vscode/settings.json
index 6abf0d3d6d..62b4261af3 100644
--- a/.vscode/settings.json
+++ b/.vscode/settings.json
@@ -11,4 +11,7 @@
     "rust-analyzer.cargo.features": [
         "cuda",
     ],
+    "files.associations": {
+        "cmath": "cpp"
+    },
 }
\ No newline at end of file
diff --git a/candle-kernels/src/cuda_utils.cuh b/candle-kernels/src/cuda_utils.cuh
index da8a1fe1c1..f3af71b18c 100644
--- a/candle-kernels/src/cuda_utils.cuh
+++ b/candle-kernels/src/cuda_utils.cuh
@@ -235,7 +235,7 @@ __device__ __forceinline__ __nv_bfloat16 copysigng(__nv_bfloat16 a, __nv_bfloat1
 #define F8E4M3_TO_FLOAT(x) __half2float(__nv_cvt_fp8_to_halfraw(x.__x, __NV_E4M3))
 
 __device__ __forceinline__ __nv_fp8_e4m3 powg(__nv_fp8_e4m3 a, __nv_fp8_e4m3 b) { return __nv_fp8_e4m3(powf(F8E4M3_TO_FLOAT(a), F8E4M3_TO_FLOAT(b))); }
-__device__ __forceinline__ bool isnang(__nv_fp8_e4m3 a) { return isnanf(F8E4M3_TO_FLOAT(a)); }
+__device__ __forceinline__ bool isnang(__nv_fp8_e4m3 a) { return isnan(F8E4M3_TO_FLOAT(a)); }
 __device__ __forceinline__ __nv_fp8_e4m3 sqrtg(__nv_fp8_e4m3 a) { return __nv_fp8_e4m3(sqrtf(F8E4M3_TO_FLOAT(a))); }
 __device__ __forceinline__ __nv_fp8_e4m3 cosg(__nv_fp8_e4m3 a) { return __nv_fp8_e4m3(cosf(F8E4M3_TO_FLOAT(a))); }
 __device__ __forceinline__ __nv_fp8_e4m3 sing(__nv_fp8_e4m3 a) { return __nv_fp8_e4m3(sinf(F8E4M3_TO_FLOAT(a))); }