Fix thread safety CI Test

[cavusmustafa]

Fix: test-thread-safety map::at crash on the OpenVINO backend

For: ravi9/llama.cpp dev_backend_openvino
Patch: fix-thread-safety-get_tensor_ov_name.patch (1 file, ggml-decoder.cpp, +20/−4)
Branch: fix_thread_safety_upstream @ b601163ee (based on c6311c391, no gemma4/MoE code)

Symptom

CI test 33 test-thread-safety fails:

GGML OpenVINO backend std::exception: map::at
graph_compute: ggml_backend_sched_graph_compute_async failed with error -1
llama_decode: failed to decode, ret = -3
One or more threads failed.

(stories15M, -ngl 99 -np 4 -t 2, 3 models × 4 contexts.)

Root cause

get_tensor_ov_name() disambiguates COMPUTE/kvcache tensors that share a ggml name by
appending #<ggml_hash_find(visited_hash_set, tensor)>. That hash value is a hash-set
slot derived from the tensor pointer, so it differs between cgraph instances.

The decoder / infer-request cache (ov_runtime_context) is process-global and keyed
by graph_key (node names + count). The 3 same-architecture models in the test produce
the same graph_key and share one cache entry. On a cache hit:

• the OV param names captured at compile time embed one instance's hash slots
  (e.g. cache_k_l0#975);
• they are looked up against the reused decoder's m_inputs, rebuilt by update_io()
  from a different cgraph instance with different hash slots;
• → the name isn't found → std::map::at throws.

Confirmed via GDB: throw is in GgmlOvDecoder::get_input_ggml_tensor(name="cache_k_l0#975")
→ m_inputs.at(name). Not concurrency-specific — also reproduces at -np 1 -t 1.

Bisect

commit	test-thread-safety
parent of 49852ce (4102b17e1)	✅ "All threads finished without errors"
49852ce … c6311c391 (HEAD)	❌ map::at

49852ce ("remove the unique name in llama.cpp; add new ov name in ov bk") introduced
get_tensor_ov_name, i.e. this regression.

Fix

Derive the suffix from the tensor's topological index in cgraph->nodes[] /
cgraph->leafs[] instead of the pointer-hash slot. That index is identical for
corresponding tensors across graphs of the same structure, so cached OV names resolve on
reuse — while still uniquely disambiguating same-named tensors within a graph.

if ((tensor->flags & GGML_TENSOR_FLAG_COMPUTE) || GgmlOvDecoder::is_kvcache(tensor, nullptr)) {
    for (int i = 0; i < cgraph->n_nodes; i++)
        if (cgraph->nodes[i] == tensor) return std::string(tensor->name) + "#" + std::to_string(i);
    for (int i = 0; i < cgraph->n_leafs; i++)
        if (cgraph->leafs[i] == tensor) return std::string(tensor->name) + "#l" + std::to_string(i);
}
return tensor->name;

Verification (OpenVINO CPU, OV 2026.2.1)

• test-thread-safety → ✅ "All threads finished without errors" (exit 0).
• test-backend-ops -b OPENVINO0 → ✅ 2622/2622, 0 fail.
• Also verified on the gemma4-MoE branch (rebased on the same base): 26B MoE CPU still
  greedy-decodes "France"; no op regressions.