Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Bugfix] Fix crash with llama 3.2 vision models and guided decoding #9631

Merged
merged 4 commits into from
Oct 25, 2024
Merged

[Bugfix] Fix crash with llama 3.2 vision models and guided decoding #9631

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
f3ce9b0
fix crash with llama 3.2 vision models and guided decoding
tjohnson31415 Oct 23, 2024
0a58760
vectorize allowed token check
tjohnson31415 Oct 23, 2024
dcefbf0
further optimization
njhill Oct 25, 2024
ce70737
Merge remote-tracking branch 'origin/main' into fix-llama-vision-guid…
njhill Oct 25, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 11 additions & 3 deletions vllm/model_executor/guided_decoding/outlines_logits_processors.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,11 +15,11 @@
# limitations under the License.
import copy
import json
import math
from collections import defaultdict
from functools import lru_cache
from typing import Callable, DefaultDict, Dict, List, Union

import numpy as np
import torch
from lark import Lark
from outlines import grammars
Expand Down Expand Up @@ -77,9 +77,17 @@ def __call__(self, input_ids: List[int],
f"Unsupported instruction type {type(instruction)}")

mask = torch.full((scores.shape[-1], ),
-math.inf,
-torch.inf,
device=scores.device)
mask[allowed_tokens] = 0
# The tokenizer may support more token ids than the model can generate,
# eg. Llama 3.2 Vision models have an `<|image|>` token with id 128256
# but scores.shape == torch.Size([128256])
# Using NumPy is faster for filtering token ids
allowed_tokens = np.array(allowed_tokens, dtype=np.int64)
allowed_tokens = torch.tensor(allowed_tokens, device=scores.device)
allowed_tokens = allowed_tokens.masked_select(
allowed_tokens < scores.shape[-1])
mask.index_fill_(0, allowed_tokens, 0)
scores.add_(mask)
return scores

Expand Down