cannot enable cl_khr_fp_16 extension on this platform

[maded2]

get this error when trying to load a model

my GPU: Linux Box (NVIDIA GEFORC/ GTX 1080 TI) 10GB

[david-thrower]

TLDR: This may be a structural issue:

This is a first pass answer. There's no promise this is the reason in your case, but I found this:

The cl_khr_fp16 extension, which enables half-precision (FP16) support in OpenCL, is not fully supported on many NVIDIA GPUs, despite the hardware capabilities. This is apparently a long-standing issue for OpenCL when trying to utilize FP16 operations on NVIDIA.

A few quick points:

1. NVIDIA does not officially report cl_khr_fp16 as a fully supported extension in their OpenCL implementations[1].

2. Even when the extension appears to be enabled, certain FP16 operations may still fail or produce unexpected results[1].

If I had to guess, it is probably Tiny Grad or llama.cpp that as the intermediate component to blame, but I don't know enough details about what you did, what model you used, what options, env vars, flags, etc ... you have set, so this is just a guess.

[1] https://forums.developer.nvidia.com/t/cl-khr-fp16-opencl-support/191106