feat: Add logprobs support to vLLM backend (#4683) (#4697)

Signed-off-by: Aryan Bagade <aryan@aryanbagade.com>
Signed-off-by: Aryan Bagade <73382554+AryanBagade@users.noreply.github.com>
Co-authored-by: Ryan McCormick <rmccormick@nvidia.com>

components/src/dynamo/vllm/handlers.py

@@ -73,7 +73,8 @@ def build_sampling_params(
     Build SamplingParams from a PreprocessedRequest.
 
     Args:
-        request: The PreprocessedRequest dict with 'sampling_options' and 'stop_conditions'
+        request: The PreprocessedRequest dict with 'sampling_options', 'stop_conditions',
+                 and 'output_options'
         default_sampling_params: Default sampling parameters to initialize with
 
     Returns:
@@ -116,6 +117,41 @@ def build_sampling_params(
             existing = sampling_params.stop_token_ids or []
             sampling_params.stop_token_ids = list(set(existing).union(value))
 
+    # Apply output_options (logprobs, prompt_logprobs, etc.)
+    output_options = request.get("output_options", {})
+    if output_options:
+        # Handle logprobs - vLLM expects this as an integer or None
+        logprobs_value = output_options.get("logprobs")
+        if logprobs_value is not None and logprobs_value != "":
+            try:
+                parsed_logprobs = int(logprobs_value)
+                if parsed_logprobs < 0:
+                    logger.warning(
+                        f"Invalid logprobs value: {logprobs_value} (must be non-negative), ignoring"
+                    )
+                else:
+                    sampling_params.logprobs = parsed_logprobs
+            except (ValueError, TypeError):
+                logger.warning(
+                    f"Invalid logprobs value: {logprobs_value} (must be integer), ignoring"
+                )
+
+        # Handle prompt_logprobs - vLLM expects this as an integer or None
+        prompt_logprobs_value = output_options.get("prompt_logprobs")
+        if prompt_logprobs_value is not None and prompt_logprobs_value != "":
+            try:
+                parsed_prompt_logprobs = int(prompt_logprobs_value)
+                if parsed_prompt_logprobs < 0:
+                    logger.warning(
+                        f"Invalid prompt_logprobs value: {prompt_logprobs_value} (must be non-negative), ignoring"
+                    )
+                else:
+                    sampling_params.prompt_logprobs = parsed_prompt_logprobs
+            except (ValueError, TypeError):
+                logger.warning(
+                    f"Invalid prompt_logprobs value: {prompt_logprobs_value} (must be integer), ignoring"
+                )
+
     # If max_tokens wasn't provided (None or missing), compute a dynamic default
     provided_max_tokens = request.get("stop_conditions", {}).get("max_tokens", None)
     token_ids = request.get("token_ids", [])
@@ -577,6 +613,66 @@ class BaseWorkerHandler(ABC):
             ),
         }
 
+    @staticmethod
+    def _extract_logprobs(
+        output, num_output_tokens_so_far: int
+    ) -> tuple[list[float] | None, list[list[dict]] | None]:
+        """
+        Extract logprobs from vLLM CompletionOutput for new tokens.
+
+        Args:
+            output: vLLM CompletionOutput object
+            num_output_tokens_so_far: Number of tokens already processed
+
+        Returns:
+            Tuple of (log_probs, top_logprobs) in Dynamo's expected format:
+            - log_probs: List of log probabilities for each new token
+            - top_logprobs: List of top logprobs dicts for each new token
+        """
+        if output.logprobs is None:
+            return None, None
+
+        # Get logprobs for new tokens only
+        new_logprobs = output.logprobs[num_output_tokens_so_far:]
+        if not new_logprobs:
+            return None, None
+
+        log_probs = []
+        top_logprobs = []
+
+        for token_idx, token_logprobs_dict in enumerate(new_logprobs):
+            if token_logprobs_dict is None:
+                continue
+
+            # Get the actual token_id that was generated at this position
+            actual_token_id = output.token_ids[num_output_tokens_so_far + token_idx]
+
+            # Extract log probability for the selected token
+            # vLLM guarantees the selected token is always in the logprobs dict
+            selected_logprob = token_logprobs_dict[actual_token_id]
+            log_probs.append(float(selected_logprob.logprob))
+
+            # Build top_logprobs list for this token position
+            token_top_logprobs = []
+            for tok_id, logprob_info in token_logprobs_dict.items():
+                token_top_logprobs.append(
+                    {
+                        "rank": (
+                            logprob_info.rank if hasattr(logprob_info, "rank") else 0
+                        ),
+                        "token_id": tok_id,
+                        "token": (
+                            logprob_info.decoded_token
+                            if hasattr(logprob_info, "decoded_token")
+                            else None
+                        ),
+                        "logprob": float(logprob_info.logprob),
+                    }
+                )
+            top_logprobs.append(token_top_logprobs)
+
+        return log_probs if log_probs else None, top_logprobs if top_logprobs else None
+
     async def generate_tokens(
         self,
         prompt,
@@ -622,6 +718,16 @@ class BaseWorkerHandler(ABC):
                     output = res.outputs[0]
                     next_total_toks = len(output.token_ids)
                     out = {"token_ids": output.token_ids[num_output_tokens_so_far:]}
+
+                    # Extract logprobs for new tokens if available
+                    log_probs, top_logprobs = self._extract_logprobs(
+                        output, num_output_tokens_so_far
+                    )
+                    if log_probs is not None:
+                        out["log_probs"] = log_probs
+                    if top_logprobs is not None:
+                        out["top_logprobs"] = top_logprobs
+
                     if output.finish_reason:
                         out["finish_reason"] = output.finish_reason
                         out[

tests/serve/test_vllm.py

@@ -19,7 +19,9 @@ from tests.utils.engine_process import EngineConfig
 from tests.utils.payload_builder import (
     chat_payload,
     chat_payload_default,
+    chat_payload_with_logprobs,
     completion_payload_default,
+    completion_payload_with_logprobs,
     metric_payload_default,
 )
 from tests.utils.payloads import ToolCallingChatPayload
@@ -59,6 +61,29 @@ vllm_configs = {
             metric_payload_default(min_num_requests=6, backend="vllm"),
         ],
     ),
+    "aggregated_logprobs": VLLMConfig(
+        name="aggregated_logprobs",
+        directory=vllm_dir,
+        script_name="agg.sh",
+        marks=[pytest.mark.gpu_1],
+        model="Qwen/Qwen3-0.6B",
+        request_payloads=[
+            chat_payload_with_logprobs(
+                repeat_count=2,
+                expected_response=["AI", "knock", "joke"],
+                max_tokens=30,
+                temperature=0.0,
+                top_logprobs=3,
+            ),
+            completion_payload_with_logprobs(
+                repeat_count=2,
+                expected_response=["AI", "knock", "joke"],
+                max_tokens=30,
+                temperature=0.0,
+                logprobs=5,
+            ),
+        ],
+    ),
     "aggregated_lmcache": VLLMConfig(
         name="aggregated_lmcache",
         directory=vllm_dir,

tests/utils/payload_builder.py

@@ -153,6 +153,10 @@ def chat_payload(
     }
     if temperature is not None:
         body["temperature"] = temperature
+    if logprobs is not None:
+        body["logprobs"] = logprobs
+    if top_logprobs is not None:
+        body["top_logprobs"] = top_logprobs
 
     if top_logprobs is not None:
         body["top_logprobs"] = top_logprobs
@@ -307,3 +311,83 @@ def make_completions_health_check(port: int, model: str):
             return False
 
     return _check_completions_endpoint
+
+
+def chat_payload_with_logprobs(
+    content: Union[str, List[Dict[str, Any]]] = TEXT_PROMPT,
+    repeat_count: int = 1,
+    expected_response: Optional[List[str]] = None,
+    max_tokens: int = 50,
+    temperature: float = 0.0,
+    top_logprobs: int = 3,
+) -> ChatPayloadWithLogprobs:
+    """
+    Create a chat payload that requests and validates logprobs in the response.
+
+    Args:
+        content: Message content (text or structured content list)
+        repeat_count: Number of times to repeat the request
+        expected_response: List of strings expected in the response text
+        max_tokens: Maximum tokens to generate
+        temperature: Sampling temperature
+        top_logprobs: Number of top logprobs to return per token
+
+    Returns:
+        ChatPayloadWithLogprobs that validates logprobs in response
+    """
+    body: Dict[str, Any] = {
+        "messages": [
+            {
+                "role": "user",
+                "content": content,
+            }
+        ],
+        "max_tokens": max_tokens,
+        "temperature": temperature,
+        "logprobs": True,
+        "top_logprobs": top_logprobs,
+    }
+
+    return ChatPayloadWithLogprobs(
+        body=body,
+        repeat_count=repeat_count,
+        expected_log=[],
+        expected_response=expected_response or ["AI", "knock", "joke"],
+    )
+
+
+def completion_payload_with_logprobs(
+    prompt: str = TEXT_PROMPT,
+    repeat_count: int = 1,
+    expected_response: Optional[List[str]] = None,
+    max_tokens: int = 50,
+    temperature: float = 0.0,
+    logprobs: int = 5,
+) -> CompletionPayloadWithLogprobs:
+    """
+    Create a completion payload that requests and validates logprobs in the response.
+
+    Args:
+        prompt: Text prompt
+        repeat_count: Number of times to repeat the request
+        expected_response: List of strings expected in the response text
+        max_tokens: Maximum tokens to generate
+        temperature: Sampling temperature
+        logprobs: Number of logprobs to return per token
+
+    Returns:
+        CompletionPayloadWithLogprobs that validates logprobs in response
+    """
+    body: Dict[str, Any] = {
+        "prompt": prompt,
+        "max_tokens": max_tokens,
+        "temperature": temperature,
+        "logprobs": logprobs,
+    }
+
+    return CompletionPayloadWithLogprobs(
+        body=body,
+        repeat_count=repeat_count,
+        expected_log=[],
+        expected_response=expected_response or ["AI", "knock", "joke"],
+    )

tests/utils/payloads.py

@@ -14,6 +14,7 @@
 # limitations under the License.
 
 import logging
+import math
 import re
 import time
 from copy import deepcopy
@@ -183,6 +184,14 @@ class ChatPayloadWithLogprobs(ChatPayload):
                         "top_logprobs" in item
                     ), "Missing 'top_logprobs' in logprobs content"
 
+                    # Sanity check: logprob should be valid (not nan/inf/positive)
+                    logprob_val = item["logprob"]
+                    assert not math.isnan(logprob_val), "logprob is NaN"
+                    assert not math.isinf(logprob_val), "logprob is infinite"
+                    assert (
+                        logprob_val <= 0
+                    ), f"logprob should be <= 0, got {logprob_val}"
+
                 logger.info(
                     f"✓ Logprobs validation passed: found {len(content_logprobs)} tokens with logprobs"
                 )
@@ -281,6 +290,20 @@ class CompletionPayloadWithLogprobs(CompletionPayload):
                 assert len(token_logprobs) == len(
                     tokens
                 ), "Mismatch between token_logprobs and tokens length"
+
+                # Sanity check: each logprob should be valid (not nan/inf/positive)
+                for i, logprob_val in enumerate(token_logprobs):
+                    if logprob_val is not None:  # First token can be None
+                        assert not math.isnan(
+                            logprob_val
+                        ), f"logprob at index {i} is NaN"
+                        assert not math.isinf(
+                            logprob_val
+                        ), f"logprob at index {i} is infinite"
+                        assert (
+                            logprob_val <= 0
+                        ), f"logprob at index {i} should be <= 0, got {logprob_val}"
+
                 logger.info(
                     f"✓ Logprobs validation passed: found {len(token_logprobs)} tokens with logprobs"
                 )