opencl: fix crash when warming up MoE on Adreno

[lhez]

Overview

When warming up MoE models on Adreno (in this case, gpt-oss-20b-mxfp4), it crashes with invalid workgroup size.

This is because the warmup run ne20 = 128 (use all experts) and the workgroup size ends up exceeding the max workgroup size of 1024. During a normal run, ne20 is the number of used experts and the workgroup size does not exceed the max workgroup size.

llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp

Lines 12788 to 12790 in 1e5ad35

	size_t histogram_global_size[] = {(size_t)(((ne21 + 63) / 64) * 64), static_cast<size_t>(ne20), 1};
	size_t histogram_local_size[] = {64, static_cast<size_t>(ne20), 1};
	backend_ctx->enqueue_ndrange_kernel(kernel, 3, histogram_global_size, histogram_local_size, src);

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Asked AI to investigate

[lhez]

@ggml-org/maintainers Can I get another approval?