From 889140f8111535d948f7ca8a0885ffe517aa0eab Mon Sep 17 00:00:00 2001 From: Cedar Date: Fri, 21 Feb 2025 13:35:02 -0800 Subject: [PATCH] move test file location --- ...h_size_consistency.py => direct_to_batcher_test.py} | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-) rename app_tests/integration_tests/llm/shortfin/{service_only/test_batch_size_consistency.py => direct_to_batcher_test.py} (92%) diff --git a/app_tests/integration_tests/llm/shortfin/service_only/test_batch_size_consistency.py b/app_tests/integration_tests/llm/shortfin/direct_to_batcher_test.py similarity index 92% rename from app_tests/integration_tests/llm/shortfin/service_only/test_batch_size_consistency.py rename to app_tests/integration_tests/llm/shortfin/direct_to_batcher_test.py index 53e024b32..606e21d32 100644 --- a/app_tests/integration_tests/llm/shortfin/service_only/test_batch_size_consistency.py +++ b/app_tests/integration_tests/llm/shortfin/direct_to_batcher_test.py @@ -98,7 +98,15 @@ async def run(self): def test_batch_and_nobatch_consistency(server_instance): - """Test that requests produce identical results regardless of batch size.""" + """ + Test that requests produce identical results regardless of batch size. + + If this test fails, it means that changing the batch size changes the generation results. + + Look for kvcache corruption due to + - improper seq_len / current_position handling in service.py + - improper masking in sharktank + """ with server_instance.start_service_only() as generate_service: # Create and run the test process test_process = BatchConsistencyTestProcess(