wip

tests/engine/test_args.py

@@ -0,0 +1,24 @@
+import pytest
+
+from aphrodite.common.utils import FlexibleArgumentParser
+from aphrodite.engine.args_tools import EngineArgs
+
+
+@pytest.mark.parametrize(("arg", "expected"), [
+    (None, None),
+    ("image=16", {
+        "image": 16
+    }),
+    ("image=16,video=2", {
+        "image": 16,
+        "video": 2
+    }),
+])
+def test_limit_mm_per_prompt_parser(arg, expected):
+    parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
+    if arg is None:
+        args = parser.parse_args([])
+    else:
+        args = parser.parse_args(["--limit-mm-per-prompt", arg])
+
+    assert args.limit_mm_per_prompt == expected

tests/engine/test_computed_prefix_block.py

@@ -0,0 +1,34 @@
+import pytest
+
+from aphrodite.engine.arg_utils import EngineArgs
+from vllm.engine.llm_engine import LLMEngine
+from vllm.sampling_params import SamplingParams
+
+
+@pytest.mark.parametrize("model", ["facebook/opt-125m"])
+@pytest.mark.parametrize("block_size", [16])
+def test_computed_prefix_blocks(model: str, block_size: int):
+    # This test checks if we are able to run the engine to completion
+    # without triggering asserts.
+    # We are in a scenario where all blocks from the second request's prompt
+    # are full and already computed when the second request arrives.
+    prompt = (
+        "You are a helpful assistant. How do I build a car from cardboard and "
+        "paper clips? Is there an easy to follow video tutorial available "
+        "online for free?")
+    prompt2 = (
+        " Please recommend to me some resources where I can learn not only to "
+        "handle technical difficulties of building a car, but also "
+        "decoration.")
+
+    engine_args = EngineArgs(model=model,
+                             block_size=block_size,
+                             enable_prefix_caching=True)
+
+    engine = LLMEngine.from_engine_args(engine_args)
+    sampling_params = SamplingParams()
+
+    engine.add_request("0", prompt + prompt2, sampling_params)
+    engine.step()
+    engine.add_request("1", prompt, sampling_params)
+    engine.step()

tests/engine/test_computed_prefix_blocks.py

@@ -0,0 +1,34 @@
+import pytest
+
+from aphrodite.common.sampling_params import SamplingParams
+from aphrodite.engine.aphrodite_engine import AphroditeEngine
+from aphrodite.engine.args_tools import EngineArgs
+
+
+@pytest.mark.parametrize("model", ["facebook/opt-125m"])
+@pytest.mark.parametrize("block_size", [16])
+def test_computed_prefix_blocks(model: str, block_size: int):
+    # This test checks if we are able to run the engine to completion
+    # without triggering asserts.
+    # We are in a scenario where all blocks from the second request's prompt
+    # are full and already computed when the second request arrives.
+    prompt = (
+        "You are a helpful assistant. How do I build a car from cardboard and "
+        "paper clips? Is there an easy to follow video tutorial available "
+        "online for free?")
+    prompt2 = (
+        " Please recommend to me some resources where I can learn not only to "
+        "handle technical difficulties of building a car, but also "
+        "decoration.")
+
+    engine_args = EngineArgs(model=model,
+                             block_size=block_size,
+                             enable_prefix_caching=True)
+
+    engine = AphroditeEngine.from_engine_args(engine_args)
+    sampling_params = SamplingParams()
+
+    engine.add_request("0", prompt + prompt2, sampling_params)
+    engine.step()
+    engine.add_request("1", prompt, sampling_params)
+    engine.step()

tests/engine/test_custom_executor.py

@@ -0,0 +1,90 @@
+import asyncio
+import os
+
+import pytest
+
+from aphrodite.common.sampling_params import SamplingParams
+from aphrodite.engine.args_tools import AsyncEngineArgs, EngineArgs
+from aphrodite.engine.async_aphrodite import AphroditeEngine, AsyncAphrodite
+from aphrodite.executor.gpu_executor import GPUExecutor, GPUExecutorAsync
+
+
+class Mock:
+    ...
+
+
+class CustomGPUExecutor(GPUExecutor):
+
+    def execute_model(self, *args, **kwargs):
+        # Drop marker to show that this was ran
+        with open(".marker", "w"):
+            ...
+        return super().execute_model(*args, **kwargs)
+
+
+class CustomGPUExecutorAsync(GPUExecutorAsync):
+
+    async def execute_model_async(self, *args, **kwargs):
+        with open(".marker", "w"):
+            ...
+        return await super().execute_model_async(*args, **kwargs)
+
+
+@pytest.mark.parametrize("model", ["facebook/opt-125m"])
+def test_custom_executor_type_checking(model):
+    with pytest.raises(ValueError):
+        engine_args = EngineArgs(model=model,
+                                 distributed_executor_backend=Mock)
+        AphroditeEngine.from_engine_args(engine_args)
+    with pytest.raises(ValueError):
+        engine_args = AsyncEngineArgs(model=model,
+                                      distributed_executor_backend=Mock)
+        AsyncAphrodite.from_engine_args(engine_args)
+    with pytest.raises(TypeError):
+        engine_args = AsyncEngineArgs(
+            model=model, distributed_executor_backend=CustomGPUExecutor)
+        AsyncAphrodite.from_engine_args(engine_args)
+
+
+@pytest.mark.parametrize("model", ["facebook/opt-125m"])
+def test_custom_executor(model, tmpdir):
+    cwd = os.path.abspath(".")
+    os.chdir(tmpdir)
+    try:
+        assert not os.path.exists(".marker")
+
+        engine_args = EngineArgs(
+            model=model, distributed_executor_backend=CustomGPUExecutor)
+        engine = AphroditeEngine.from_engine_args(engine_args)
+        sampling_params = SamplingParams(max_tokens=1)
+
+        engine.add_request("0", "foo", sampling_params)
+        engine.step()
+
+        assert os.path.exists(".marker")
+    finally:
+        os.chdir(cwd)
+
+
+@pytest.mark.parametrize("model", ["facebook/opt-125m"])
+def test_custom_executor_async(model, tmpdir):
+    cwd = os.path.abspath(".")
+    os.chdir(tmpdir)
+    try:
+        assert not os.path.exists(".marker")
+
+        engine_args = AsyncEngineArgs(
+            model=model, distributed_executor_backend=CustomGPUExecutorAsync)
+        engine = AsyncAphrodite.from_engine_args(engine_args)
+        sampling_params = SamplingParams(max_tokens=1)
+
+        async def t():
+            stream = await engine.add_request("0", "foo", sampling_params)
+            async for x in stream:
+                ...
+
+        asyncio.run(t())
+
+        assert os.path.exists(".marker")
+    finally:
+        os.chdir(cwd)

tests/engine/test_detokenization.py

@@ -0,0 +1,32 @@
+import pytest
+
+from aphrodite.common.sampling_params import SamplingParams
+from aphrodite.endpoints.llm import LLM
+
+
+@pytest.mark.parametrize("model", ["facebook/opt-125m"])
+def test_computed_prefix_blocks(model: str):
+    # This test checks if the engine generates completions both with and
+    # without optional detokenization, that detokenization includes text
+    # and no-detokenization doesn't, and that both completions have the same
+    # token_ids.
+    prompt = (
+        "You are a helpful assistant. How do I build a car from cardboard and "
+        "paper clips? Is there an easy to follow video tutorial available "
+        "online for free?")
+
+    llm = LLM(model=model)
+    sampling_params = SamplingParams(max_tokens=10,
+                                     temperature=0.0,
+                                     detokenize=False)
+
+    outputs_no_detokenization = llm.generate(prompt,
+                                             sampling_params)[0].outputs[0]
+    sampling_params.detokenize = True
+    outputs_with_detokenization = llm.generate(prompt,
+                                               sampling_params)[0].outputs[0]
+
+    assert outputs_no_detokenization.text == ''
+    assert outputs_with_detokenization.text != ''
+    assert outputs_no_detokenization.token_ids == \
+        outputs_with_detokenization.token_ids