[am17an]

Overview

Fix issue when always offloading. Reproduce via:

GGML_OP_OFFLOAD_MIN_BATCH=0 ./build/bin/llama-bench -m ~/workspace/Qwen3.8-Flash-Next-GGUF/UD-IQ1_S/Qwen3.8-Flash-Next-UD-IQ1_S-00001-of-00003.gguf -lm dio -ncmoe 999 -p 256 -ub 64

Zero-sized tensors appear in the last layer of graph for e.g. in qwen4exp:

llama.cpp/src/models/qwen4exp.cpp

Lines 400 to 402 in 3c64099

	if (il == n_layer - 1 && inp_out_ids) {
	// everything below is per token, so drop the rows that produce no output
	cur = ggml_get_rows(ctx0, cur, inp_out_ids);

when always offloading this loop does an out-of-bounds access:

llama.cpp/ggml/src/ggml-backend.cpp

Lines 1767 to 1770 in 3c64099

	int id = 0;
	while (!ggml_bitset_get(used_ids.data(), id)) {
	id++;
	}

This PR exits early when the ids tensors is empty.

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes, for reproducing