[V1] Logprobs and prompt logprobs support (#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 <afeldman@neuralmagic.com>
Signed-off-by: Nick Hill <nhill@redhat.com>
Signed-off-by: rshaw@neuralmagic.com <rshaw@neuralmagic.com>


Co-authored-by: rshaw@neuralmagic.com <rshaw@neuralmagic.com>
Co-authored-by: Nick Hill <nhill@redhat.com>
This commit is contained in:
afeldman-nm
2025-02-07 07:26:20 -08:00
committed by GitHub
co-authored by rshaw@neuralmagic.com <rshaw@neuralmagic.com> Nick Hill
parent 538fab93cd
commit 0630d4537a
30 changed files with 2865 additions and 283 deletions
+12 -15
View File
@@ -176,7 +176,9 @@ class InputBatch:
self.generators: Dict[int, torch.Generator] = {}
self.num_logprobs: Dict[str, int] = {}
self.prompt_logprob_reqs: Set[str] = set()
# NOTE(rob): num_prompt_logprobs only includes reqs
# that are currently in the prefill phase.
self.num_prompt_logprobs: Dict[str, int] = {}
def add_request(
self,
@@ -238,11 +240,10 @@ class InputBatch:
if request.generator is not None:
self.generators[req_index] = request.generator
num_logprobs = sampling_params.logprobs
if num_logprobs is not None and num_logprobs > 0:
self.num_logprobs[req_id] = num_logprobs
if sampling_params.prompt_logprobs:
self.prompt_logprob_reqs.add(req_id)
if sampling_params.logprobs is not None:
self.num_logprobs[req_id] = sampling_params.logprobs
if sampling_params.prompt_logprobs is not None:
self.num_prompt_logprobs[req_id] = sampling_params.prompt_logprobs
# Add request lora ID
if request.lora_request:
@@ -272,7 +273,7 @@ class InputBatch:
self.repetition_penalties_reqs.discard(req_id)
self.generators.pop(req_index, None)
self.num_logprobs.pop(req_id, None)
self.prompt_logprob_reqs.discard(req_id)
self.num_prompt_logprobs.pop(req_id, None)
# LoRA
lora_id = self.request_lora_mapping[req_index]
@@ -297,7 +298,7 @@ class InputBatch:
self.repetition_penalties_reqs.clear()
self.generators.clear()
self.num_logprobs.clear()
self.prompt_logprob_reqs.clear()
self.num_prompt_logprobs.clear()
self.request_lora_mapping.fill(0)
self.lora_id_to_lora_request.clear()
self.lora_id_to_request_ids.clear()
@@ -489,13 +490,9 @@ class InputBatch:
and len(self.repetition_penalties_reqs) == 0)
@property
def max_num_logprobs(self) -> int:
return max(self.num_logprobs.values()) if self.num_logprobs else 0
@property
def no_logprob(self) -> bool:
return len(self.num_logprobs) == 0
def max_num_logprobs(self) -> Optional[int]:
return max(self.num_logprobs.values()) if self.num_logprobs else None
@property
def no_prompt_logprob(self) -> bool:
return len(self.prompt_logprob_reqs) == 0
return not self.num_prompt_logprobs