Merge pull request #301 from ndif-team/vllm-testing

NNsight.VLLM tests

conftest.py

@@ -3,7 +3,13 @@
 
 def pytest_addoption(parser):
     parser.addoption("--device", action="store", default="cuda:0")
-
+    parser.addoption(
+        "--tp", 
+        action="store", 
+        type=int, 
+        default="1", 
+        help="An argument for specifying the number of gpus to be used by VLLM"
+    )
 
 def pytest_generate_tests(metafunc):
     # This is called for every test. Only get/set command line arguments

src/nnsight/intervention/graph/graph.py

@@ -370,7 +370,7 @@ def copy(
         new_graph.compiled = self.compiled
 
         for key, value in self.call_counter.items():
-            self.call_counter[memo[key]] = value
+            new_graph.call_counter[memo[key]] = value
 
         if new_graph.compiled:
 
@@ -382,7 +382,7 @@ def copy(
 
             for key, values in self.deferred.items():
 
-                new_graph[memo[key]] = [memo[index] for index in values]
+                new_graph.deferred[memo[key]] = [memo[index] for index in values]
 
             new_graph.grad_subgraph = [memo[index] for index in self.grad_subgraph]
 

src/nnsight/modeling/vllm/sampling.py

@@ -138,10 +138,11 @@ def prepare(
             intervention_graph = None
         elif n_graphs == 1:
             intervention_graph =intervention_graphs[0]
-        else:
+
+        """ else:
             intervention_graph = MultiGraph(intervention_graphs.values())
             
-            InterventionProtocol.shift(intervention_graph)
+            InterventionProtocol.shift(intervention_graph) """
 
         ###########################################
 

src/nnsight/modeling/vllm/vllm.py

@@ -82,13 +82,8 @@ def _load_meta(self, repo_id: str, **kwargs) -> "Module":
         )
 
         # creating the vllm engine configuration
-        """ dict(
-            (field.name, getattr(self, field.name)) for field in fields(self)) """
         vllm_config = engine_args.create_engine_config()
         vllm_config_dict = {field.name: getattr(vllm_config, field.name) for field in fields(type(vllm_config))}
-
-        #dict((field.name, getattr(self, field.name)) for field in fields(engine_args.create_engine_config()))
-        #vllm_config_dict = engine_args.create_engine_config().to_dict()
 
         # starting the distributed environment
         init_distributed_environment(
@@ -103,13 +98,6 @@ def _load_meta(self, repo_id: str, **kwargs) -> "Module":
         initialize_model_parallel(backend="gloo")
 
         # initialize the model
-        """ model = _initialize_model(
-            model_config=vllm_config_dict["model_config"],
-            load_config=vllm_config_dict["load_config"],
-            lora_config=None,
-            cache_config=vllm_config_dict["cache_config"],
-            scheduler_config=vllm_config_dict["scheduler_config"],
-        ) """
 
         model = _initialize_model(vllm_config)
 

tests/test_vllm.py

@@ -0,0 +1,213 @@
+import pytest
+import nnsight
+import torch
+from typing import TYPE_CHECKING
+
+from nnsight.tracing.backends import Backend
+from nnsight.tracing.protocols import StopProtocol
+
+if TYPE_CHECKING:
+    from nnsight.tracing.graph import Graph
+
+try:
+    from nnsight.modeling.vllm import VLLM
+except:
+    pytest.skip("Skipping VLLM tests", allow_module_level=True)
+
+
+class AssertSavedLenBackend(Backend):
+
+    def __init__(self, len:int) -> None:
+        self.len = len
+
+    def __call__(self, graph: "Graph") -> None:
+
+        try:
+
+            graph.nodes[-1].execute()
+
+        except StopProtocol.StopException:
+
+            pass
+
+        finally:
+
+            assert self.len == len([node for node in graph.nodes if node.done])
+
+            graph.nodes.clear()
+            graph.stack.clear()
+
+
+@pytest.fixture(scope="module")
+def tp(request):
+    tp = request.config.getoption("--tp")
+    if tp > torch.cuda.device_count() or tp < 1:
+        pytest.exit("--tp can't be higher than the number of availale GPUs.")
+    return tp
+
+@pytest.fixture(scope="module")
+def vllm_gpt2(tp: int):
+    return VLLM("gpt2", tensor_parallel_size=tp, dispatch=True)
+
+@pytest.fixture
+def ET_prompt():
+    return "The Eiffel Tower is located in the city of"
+
+@pytest.fixture
+def MSG_prompt():
+    return "Madison Square Garden is located in the city of"
+
+
+def test_single_logit(vllm_gpt2, ET_prompt: str):
+    with vllm_gpt2.trace(ET_prompt, temperature=0.0, top_p=1, backend=AssertSavedLenBackend(1)):
+        logits = vllm_gpt2.logits.output.save()
+
+    next_token = vllm_gpt2.tokenizer.decode(logits.argmax(dim=-1))
+    assert next_token == " Paris"
+
+
+def test_multi_token_generation(vllm_gpt2, MSG_prompt: str):
+    with vllm_gpt2.trace(MSG_prompt, temperature=0.0, top_p=1.0, max_tokens=3):
+        logits = nnsight.list().save()
+        for ii in range(3):
+            logits.append(vllm_gpt2.logits.output)
+            vllm_gpt2.logits.next()
+
+    assert vllm_gpt2.tokenizer.batch_decode([logit.argmax(dim=-1) for logit in logits.value]) == [" New", " York", " City"]
+
+
+""" def test_max_token_generation(vllm_gpt2, ET_prompt: str):
+    with vllm_gpt2.trace(ET_prompt, max_tokens=10):
+        logits = nnsight.list().save()
+        with vllm_gpt2.logits.all():
+            logits.append(vllm_gpt2.logits.output)
+
+    assert len(logits) == 10 """
+
+
+""" def test_sampling(vllm_gpt2, ET_prompt:str):
+    with vllm_gpt2.trace(ET_prompt, temperature=0.8, top_p=0.95, max_tokens=3):
+        samples = nnsight.list().save()
+        with vllm_gpt2.sample.all():
+            li.append(vllm_gpt2.sample.output)
+
+    samples = vllm_gpt2.batch_decode([sample.argmax(dim=-1) for sample in samples])
+    assert samples == [' Canary', ' Wh', 'arf'] """
+
+
+def test_intervention(vllm_gpt2, ET_prompt: str):
+    with vllm_gpt2.trace(ET_prompt, temperature=0.0, top_p=1, backend=AssertSavedLenBackend(2)) as tracer:
+        vllm_gpt2.transformer.h[-2].mlp.output[:] = 0
+        hs = vllm_gpt2.transformer.h[-2].mlp.output.save()
+        logits = vllm_gpt2.logits.output.save()
+
+    next_token = vllm_gpt2.tokenizer.decode(logits.argmax(dim=-1))
+    assert next_token == " London"
+    assert torch.all(hs == 0)
+
+
+def test_swap_intervention(vllm_gpt2, ET_prompt: str):
+    with vllm_gpt2.trace(ET_prompt, temperature=0.0, top_p=1, backend=AssertSavedLenBackend(2)) as tracer:
+        vllm_gpt2.transformer.h[-2].mlp.output = torch.zeros_like(vllm_gpt2.transformer.h[-2].mlp.output)
+        hs = vllm_gpt2.transformer.h[-2].mlp.output.save()
+        logits = vllm_gpt2.logits.output.save()
+
+    next_token = vllm_gpt2.tokenizer.decode(logits.argmax(dim=-1))
+    assert next_token == " London"
+    assert torch.all(hs == 0)
+
+
+def test_batched_intervention(vllm_gpt2, ET_prompt: str,):
+    with vllm_gpt2.trace(temperature=0.0, top_p=1, backend=AssertSavedLenBackend(4)) as tracer:
+
+        with tracer.invoke(ET_prompt):
+            clean_hs = vllm_gpt2.transformer.h[-2].mlp.output.save()
+            clean_logits = vllm_gpt2.logits.output.save()
+        with tracer.invoke(ET_prompt):
+            vllm_gpt2.transformer.h[-2].mlp.output[:] = 0
+            corrupted_hs = vllm_gpt2.transformer.h[-2].mlp.output.save()
+            corrupted_logits = vllm_gpt2.logits.output.save()
+
+    clean_token = vllm_gpt2.tokenizer.decode(clean_logits.argmax(dim=-1))
+    corrupted_token = vllm_gpt2.tokenizer.decode(corrupted_logits.argmax(dim=-1))
+
+    assert clean_token == " Paris"
+    assert corrupted_token == " London"
+    assert not torch.all(clean_hs == 0)
+    assert torch.all(corrupted_hs == 0)
+
+
+def test_batched_multi_token_generation(vllm_gpt2, ET_prompt: str, MSG_prompt: str):
+    with vllm_gpt2.trace() as tracer:
+        with tracer.invoke(ET_prompt, max_tokens=3):
+            ET_logits = nnsight.list().save()
+            for ii in range(3):
+                ET_logits.append(vllm_gpt2.logits.output)
+                vllm_gpt2.logits.next()
+        with tracer.invoke(MSG_prompt, max_tokens=5):
+            MSG_logits = nnsight.list().save()
+            for ii in range(5):
+                MSG_logits.append(vllm_gpt2.logits.output)
+                vllm_gpt2.logits.next()
+
+    assert len(ET_logits) == 3
+    assert len(MSG_logits) == 5
+
+
+""" def test_batched_multi_token_generation_with_iter(vllm_gpt2, ET_prompt: str, MSG_prompt: str):
+    with vllm_gpt2.trace(max_tokens=10) as tracer:
+        with tracer.invoke(ET_prompt):
+            ET_logits = nnsight.list().save()
+            with vllm_gpt2.logits.iter[:3]:
+                ET_logits.append(vllm_gpt2.logits.output)
+            #vllm_gpt2.output.save()
+        with tracer.invoke(MSG_prompt, max_tokens=5):
+            MSG_logits = nnsight.list().save()
+            with vllm_gpt2.logits.iter[:5]:
+                MSG_logits.append(vllm_gpt2.logits.output)
+
+    assert len(ET_logits.value) == 3
+    assert len(MSG_logits.value) == 5 """
+
+
+def test_mutli_token_generation_with_intervention(tp, vllm_gpt2, MSG_prompt: str):
+    with vllm_gpt2.trace(MSG_prompt, temperature=0.0, top_p=1.0, max_tokens=5) as tracer:
+        logits = nnsight.list().save()
+        hs_list = nnsight.list().save()
+        for ii in range(5):
+            if ii == 2:
+                vllm_gpt2.transformer.h[-2].output[0][:] = 0
+            hs_list.append(vllm_gpt2.transformer.h[-2].output[0])
+            vllm_gpt2.transformer.h[-2].next()
+            logits.append(vllm_gpt2.logits.output)
+            vllm_gpt2.logits.next()
+
+    assert [torch.all(hs == 0) for hs in hs_list.value] == [False, False, True, False, False]
+
+    if tp == 1:
+        assert vllm_gpt2.tokenizer.batch_decode([logit.argmax(dim=-1) for logit in logits.value]) == [' New', ' York', '\n', '\n', 'The']
+
+
+""" def test_multi_referenced_module(vllm_gpt2, ET_prompt: str):
+    with vllm_gpt2.trace(ET_prompt):
+        act_in = vllm_gpt2.transformer.h[0].mlp.act.input.save()
+        vllm_gpt2.transformer.h[0].mlp.act.next()
+        act_in_other = vllm_gpt2.transformer.h[1].mlp.act.input.save()
+
+    assert not torch.equal(act_in, act_in_other) """
+
+
+def test_tensor_parallelism(tp, vllm_gpt2, ET_prompt: str):
+    if tp < 2:
+        pytest.skip("Skipping test for tp>1!")
+
+    with vllm_gpt2.trace(ET_prompt, temperature=0.0, top_p=1.0):
+        vllm_gpt2.transformer.h[5].mlp.c_fc.output[0][:, 2000:] = 0
+        hs = vllm_gpt2.transformer.h[5].mlp.c_fc.output[0].save()
+        logit = vllm_gpt2.logits.output.save()
+
+    next_token = vllm_gpt2.tokenizer.decode(logit.argmax(dim=-1))
+
+    #assert next_token != " Paris"
+    assert hs.shape == torch.Size([11, 3072])
+    assert torch.all(hs[:, 2000:] == 0)