[V1] Logprobs and prompt logprobs support (vllm-project#9880)

This PR is adding support for sample logprobs & prompt logprobs to vLLM v1.

New behavior:

- During model execution, model runner computes sample logprobs (if user-provided logprobs setting is not None) and prompt logprobs (if user-provided prompt_logprobs setting is not None). For both sample and prompt logprobs, the engine core returns 3 vectors: token ids, token logprob values, token ranks. Ranks reflect tokens' 1-indexed positions in the vocabulary vector after sorting the vocabulary by log probability in descending order.
- In scheduler.update_from_output(), sample and prompt logprobs are incorporated into the EngineCoreOutput data structure which is transferred to the engine client. If multiprocessing is enabled, then sample and prompt logprobs will be (de)serialized when the EngineCoreOutput data structure is (de)serialized.
- During output processing, the LogprobsProcessor transforms the triplet of token ids, token logprobs values, and token ranks into the OpenAI-compatible List[Dict[token id,Logprob]] format (for sample and prompt logprobs respectively.)
- Each Logprob instance (whether sample- or prompt-) consists of a token's log-probability, rank, and detokenized string representation. Note that logprob detokenization is handled by the LogprobsProcessor not the detokenizer.

Signed-off-by: Andrew Feldman <[email protected]>
Signed-off-by: Nick Hill <[email protected]>
Signed-off-by: [email protected] <[email protected]>

Co-authored-by: [email protected] <[email protected]>
Co-authored-by: Nick Hill <[email protected]>
Signed-off-by: Louis Ulmer <[email protected]>

Author: 3 people

tests/v1/core/test_scheduler.py

@@ -195,8 +195,8 @@ def test_schedule_partial_requests():
         req_ids=[request.request_id for request in requests],
         req_id_to_index=req_to_index,
         sampled_token_ids=[0] * len(requests),
-        logprob_token_ids_cpu=None,
-        logprobs_cpu=None,
+        logprobs=None,
+        prompt_logprobs_dict={},
     )
     scheduler.update_from_output(output, model_runner_output)
 

tests/v1/engine/conftest.py

@@ -0,0 +1,90 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import List, Tuple
+
+import pytest
+import torch
+from transformers import AutoTokenizer
+
+from tests.v1.engine.utils import (NUM_PROMPT_LOGPROBS_UNDER_TEST,
+                                   NUM_SAMPLE_LOGPROBS_UNDER_TEST, PROMPT_LEN,
+                                   TOKENIZER_NAME,
+                                   DummyOutputProcessorTestVectors,
+                                   generate_dummy_prompt_logprobs_tensors,
+                                   generate_dummy_sample_logprobs)
+from vllm.engine.arg_utils import EngineArgs
+from vllm.transformers_utils.tokenizer_group import init_tokenizer_from_configs
+
+from tests.v1.engine.utils import FULL_STRINGS  # isort: skip
+
+EngineCoreSampleLogprobsType = List[Tuple[torch.Tensor, torch.Tensor]]
+EngineCorePromptLogprobsType = Tuple[torch.Tensor, torch.Tensor]
+
+
+def _build_test_vectors_no_logprobs() -> DummyOutputProcessorTestVectors:
+    """Generate output processor dummy test vectors, without logprobs
+    
+    Returns:
+      DummyOutputProcessorTestVectors instance with no logprobs
+    """
+
+    tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_NAME)
+    vllm_config = EngineArgs(model=TOKENIZER_NAME).create_engine_config()
+    # Tokenize prompts under test & create dummy generated tokens
+    prompt_tokens = [
+        tokenizer(text).input_ids[:PROMPT_LEN] for text in FULL_STRINGS
+    ]
+    generation_tokens = [
+        tokenizer(text).input_ids[PROMPT_LEN:] for text in FULL_STRINGS
+    ]
+    # Generate prompt strings
+    prompt_strings = [
+        tokenizer.decode(prompt_tokens, skip_special_tokens=True)
+        for prompt_tokens in prompt_tokens
+    ]
+    prompt_strings_len = [
+        len(prompt_string) for prompt_string in prompt_strings
+    ]
+    return DummyOutputProcessorTestVectors(
+        tokenizer=tokenizer,
+        tokenizer_group=init_tokenizer_from_configs(
+            vllm_config.model_config, vllm_config.scheduler_config,
+            vllm_config.parallel_config, vllm_config.lora_config),
+        vllm_config=vllm_config,
+        full_tokens=[tokenizer(text).input_ids for text in FULL_STRINGS],
+        prompt_tokens=prompt_tokens,
+        generation_tokens=generation_tokens,
+        prompt_strings=prompt_strings,
+        prompt_strings_len=prompt_strings_len,
+        generation_strings=[
+            text[prompt_len:]
+            for text, prompt_len in zip(FULL_STRINGS, prompt_strings_len)
+        ],
+        prompt_logprobs=[],
+        generation_logprobs=[])
+
+
+@pytest.fixture
+def dummy_test_vectors() -> DummyOutputProcessorTestVectors:
+    """Generate output processor dummy test vectors, with logprobs
+    
+    Returns:
+      DummyOutputProcessorTestVectors instance with logprobs
+    """
+    # Build dummy test vectors without logprobs
+    dtv = _build_test_vectors_no_logprobs()
+    # Inject logprobs into dummy test vectors
+    # data structure
+    dtv.generation_logprobs = [
+        generate_dummy_sample_logprobs(
+            sampled_tokens_list=tokens_list,
+            num_logprobs=NUM_SAMPLE_LOGPROBS_UNDER_TEST,
+            tokenizer=dtv.tokenizer) for tokens_list in dtv.generation_tokens
+    ]
+    dtv.prompt_logprobs = [
+        generate_dummy_prompt_logprobs_tensors(
+            prompt_tokens_list=tokens_list,
+            num_logprobs=NUM_PROMPT_LOGPROBS_UNDER_TEST,
+            tokenizer=dtv.tokenizer) for tokens_list in dtv.prompt_tokens
+    ]
+    return dtv

tests/v1/engine/test_async_llm.py

@@ -2,10 +2,11 @@
 
 import asyncio
 from contextlib import ExitStack
-from typing import List, Tuple
+from typing import List, Optional, Tuple
 
 import pytest
 
+from tests.v1.engine.utils import PLP_APC_UNSUPPORTED_MSG
 from vllm import SamplingParams
 from vllm.engine.arg_utils import AsyncEngineArgs
 from vllm.platforms import current_platform
@@ -21,13 +22,19 @@
                               disable_log_requests=True)
 
 
-async def generate(engine: AsyncLLM, request_id: str,
+async def generate(engine: AsyncLLM,
+                   request_id: str,
                    output_kind: RequestOutputKind,
-                   max_tokens: int) -> Tuple[int, str]:
+                   max_tokens: int,
+                   prompt_logprobs: Optional[int] = None) -> Tuple[int, str]:
+    # Ensure generate doesn't complete too fast for cancellation test.
+    await asyncio.sleep(0.2)
+
     count = 0
     sampling_params = SamplingParams(max_tokens=max_tokens,
                                      output_kind=output_kind,
-                                     temperature=0)
+                                     temperature=0,
+                                     prompt_logprobs=prompt_logprobs)
     async for out in engine.generate(request_id=request_id,
                                      prompt="Hello my name is Robert and",
                                      sampling_params=sampling_params):
@@ -43,6 +50,40 @@ async def generate(engine: AsyncLLM, request_id: str,
     return count, request_id
 
 
+@pytest.mark.parametrize(
+    "output_kind", [RequestOutputKind.DELTA, RequestOutputKind.FINAL_ONLY])
+@pytest.mark.asyncio
+async def test_async_llm_refuses_prompt_logprobs_with_apc(
+        monkeypatch, output_kind: RequestOutputKind):
+    """Test passes if AsyncLLM raises an exception when it is configured
+    for automatic prefix caching and it receives a request with
+    prompt_logprobs enabled, which is incompatible."""
+    # TODO(rickyx): Remove monkeypatch VLLM_USE_V1 setting once we have a
+    # better way to test V1 so that in the future when we switch, we don't
+    # have to change all the tests.
+    monkeypatch.setenv("VLLM_USE_V1", "1")
+    # Create AsyncLLM engine with APC
+    apc_engine_args = AsyncEngineArgs(model="facebook/opt-125m",
+                                      enable_prefix_caching=True,
+                                      gpu_memory_utilization=0.8,
+                                      disable_log_requests=True)
+    engine = AsyncLLM.from_engine_args(apc_engine_args)
+    try:
+        with pytest.raises(ValueError) as excinfo:
+            # Issue a request with prompt logprobs enabled, which should fail
+            await asyncio.create_task(
+                generate(engine,
+                         "request-0",
+                         output_kind,
+                         10,
+                         prompt_logprobs=5))
+        # Validate exception string is correct
+        assert str(excinfo.value) == PLP_APC_UNSUPPORTED_MSG
+    finally:
+        # Shut down engine
+        engine.shutdown()
+
+
 @pytest.mark.parametrize(
     "output_kind", [RequestOutputKind.DELTA, RequestOutputKind.FINAL_ONLY])
 @pytest.mark.asyncio

tests/v1/engine/test_llm_engine.py

@@ -0,0 +1,23 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import pytest
+
+from tests.v1.engine.utils import PLP_APC_UNSUPPORTED_MSG
+from vllm import LLM, SamplingParams
+
+
+def test_llm_engine_refuses_prompt_logprobs_with_apc(monkeypatch):
+    """Test passes if LLMEngine raises an exception when it is configured
+    for automatic prefix caching and it receives a request with
+    prompt_logprobs enabled, which is incompatible."""
+
+    monkeypatch.setenv("VLLM_USE_V1", "1")
+    # TODO(nick): Single-proc to work around a ZMQ shutdown hang for now.
+    monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
+    with pytest.raises(ValueError) as excinfo:
+        LLM(model="facebook/opt-125m", enable_prefix_caching=True).generate(
+            "Hello, my name is",
+            SamplingParams(temperature=0.8, top_p=0.95, prompt_logprobs=5))
+
+    # Validate exception string is correct
+    assert str(excinfo.value) == PLP_APC_UNSUPPORTED_MSG