[V1] LogitsProcessor programming model (#16728)

Signed-off-by: Nick Hill <nhill@redhat.com>
Signed-off-by: Andrew Feldman <afeldman@neuralmagic.com>
Signed-off-by: Andrew Feldman <afeldman@redhat.com>
Co-authored-by: Nick Hill <nhill@redhat.com>
This commit is contained in:
afeldman-nm
2025-07-02 09:10:42 -07:00
committed by GitHub
co-authored by Nick Hill
parent c1909e7e8c
commit 48fb076cbc
13 changed files with 1401 additions and 393 deletions
-16
View File
@@ -7,22 +7,6 @@ from vllm.model_executor.layers.utils import apply_penalties
from vllm.utils import is_pin_memory_available, make_tensor_with_pad
def apply_min_token_penalties(
logits: torch.Tensor, output_token_ids: list[list[int]],
min_tokens: dict[int, tuple[int, set[int]]]) -> None:
"""
Applies minimum token penalty by setting the logits of the stop tokens
to -inf.
"""
min_tokens_logits_to_penalize: list[tuple[int, int]] = []
for index, (min_token, stop_token_ids) in min_tokens.items():
if len(output_token_ids[index]) < min_token:
for stop_token_id in stop_token_ids:
min_tokens_logits_to_penalize.append((index, stop_token_id))
if min_tokens_logits_to_penalize:
logits[tuple(zip(*min_tokens_logits_to_penalize))] = -float("inf")
def apply_all_penalties(
logits: torch.Tensor,
prompt_token_ids: torch.Tensor,