-
Notifications
You must be signed in to change notification settings - Fork 2.9k
Expand file tree
/
Copy pathlogits_processor.py
More file actions
75 lines (67 loc) · 3.13 KB
/
Copy pathlogits_processor.py
File metadata and controls
75 lines (67 loc) · 3.13 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
# ===----------------------------------------------------------------------=== #
# Copyright (c) 2026, Modular Inc. All rights reserved.
#
# Licensed under the Apache License v2.0 with LLVM Exceptions:
# https://llvm.org/LICENSE.txt
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ===----------------------------------------------------------------------=== #
"""Provides functions for applying logits processors to model output batches before sampling."""
from __future__ import annotations
from max.driver import CPU, Buffer
from max.pipelines.context import TextGenerationContextType
from max.pipelines.context.logit_processors_type import (
BatchLogitsProcessor,
BatchProcessorInputs,
ProcessorInputs,
)
def apply_logits_processors(
context_batch: list[TextGenerationContextType],
batch_logits: Buffer,
batch_logit_offsets: Buffer | None,
batch_processors: list[BatchLogitsProcessor] | None = None,
) -> None:
"""Applies logits processors to a batch of logits.
Args:
context_batch: The batch of contexts containing the inputs to the model.
batch_logits: The model logits, a float32 tensor with shape `(N_batch, vocab_size)`.
batch_logit_offsets: If the model returns multiple logits, this is a tensor with
shape `(batch_size + 1, 1)` that contains the offsets of each sequence in
the batch. Otherwise, this is `None`.
logits_processors: List of logits processors to apply to the logits for
each context in the batch. The length of this list must match the
number of contexts in the batch.
batch_processors: List of batch processors to apply to the batch logits.
These are applied in order after the individual context-level
processors.
"""
batch_logit_offsets_cpu: Buffer | None = None
for i, context in enumerate(context_batch):
processors = context.sampling_params.logits_processors
if processors is None:
continue
if batch_logit_offsets_cpu is None and batch_logit_offsets is not None:
batch_logit_offsets_cpu = batch_logit_offsets.to(CPU())
if batch_logit_offsets_cpu is not None:
start_idx = batch_logit_offsets_cpu[i].item()
end_idx = batch_logit_offsets_cpu[i + 1].item()
else:
start_idx = i
end_idx = i + 1
logits = batch_logits[start_idx:end_idx, :]
assert isinstance(logits, Buffer)
for processor in processors:
processor(ProcessorInputs(logits=logits, context=context))
if batch_processors is not None:
for batch_processor in batch_processors:
batch_processor(
BatchProcessorInputs(
logits=batch_logits,
logit_offsets=batch_logit_offsets,
context_batch=context_batch,
)
)