Fixed SEGFAULT when empty encoded inputs are passed to LLM (#1513)

src/cpp/src/sequence_group.hpp

@@ -243,9 +243,12 @@ class SequenceGroup  : public std::enable_shared_from_this<SequenceGroup> {
 
     SequenceGroup(uint64_t request_id, const ov::Tensor input_ids, const ov::genai::GenerationConfig& sampling_params, std::size_t block_size)
         : SequenceGroup(request_id, sampling_params, block_size) {
-        m_prompt_ids.resize(input_ids.get_size());
-        std::copy_n(input_ids.data<int64_t>(), input_ids.get_size(), m_prompt_ids.begin());
-        m_prompt_log_probs.reserve(m_prompt_ids.size());
+        size_t prompt_len = input_ids.get_size();
+        OPENVINO_ASSERT(prompt_len > 0, "Prompt length cannot be 0");
+
+        m_prompt_ids.resize(prompt_len);
+        std::copy_n(input_ids.data<int64_t>(), prompt_len, m_prompt_ids.begin());
+        m_prompt_log_probs.reserve(prompt_len);
 
         // create a single sequence
         add_sequence(Sequence::create(m_next_sequence_id++));

tests/cpp/block_manager.cpp

@@ -10,7 +10,7 @@
 
 TEST(TestBlockManager, general_test) {
     ov::genai::BlockManager bm = ov::genai::BlockManager(6, false, 4);
-    ov::genai::TokenIds prompt_ids;
+    ov::genai::TokenIds prompt_ids = {10, 0};
 
     ov::genai::SequenceGroup::Ptr sequence_group = std::make_shared<ov::genai::SequenceGroup>(
         0,

tests/python_tests/test_llm_pipeline.py

@@ -97,6 +97,14 @@ def test_batch_size_switch():
     ov_pipe.generate(["1", "2"], max_new_tokens=2)
     ov_pipe.generate(["a"], max_new_tokens=2)
 
+
+@pytest.mark.precommit
+@pytest.mark.nightly
+def test_empty_encoded_inputs_throw():
+    ov_pipe = read_model(('katuni4ka/tiny-random-phi3', Path('tiny-random-phi3')))[4]
+    with pytest.raises(RuntimeError):
+        ov_pipe.generate(ov.Tensor(np.array([[]], dtype=np.int64)), max_new_tokens=2)
+
 #
 # Chat scenario
 #