@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 0bfa1c4f13
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 17, 00:30 UTC

[Misc] Improve error message when LoRA parsing fails (#5194)

Cyrus Leung committed Jun 10, 2024 at 19:38 UTC 0bfa1c4f133737a59bcb94e85ca80f2f4cd68038
2 files changed +20 -9
tests/lora/test_utils.py
+13 -1
@@ -1,12 +1,13 @@
1 from collections import OrderedDict
2
3 +import pytest
4 from torch import nn
5
6 from vllm.lora.utils import parse_fine_tuned_lora_name, replace_submodule
7 from vllm.utils import LRUCache
8
9
9 -def test_parse_fine_tuned_lora_name():
10 +def test_parse_fine_tuned_lora_name_valid():
11 fixture = {
12 ("base_model.model.lm_head.lora_A.weight", "lm_head", True),
13 ("base_model.model.lm_head.lora_B.weight", "lm_head", False),
@@ -35,6 +36,17 @@ def test_parse_fine_tuned_lora_name():
36 assert (module_name, is_lora_a) == parse_fine_tuned_lora_name(name)
37
38
39 +def test_parse_fine_tuned_lora_name_invalid():
40 + fixture = {
41 + "weight",
42 + "base_model.weight",
43 + "base_model.model.weight",
44 + }
45 + for name in fixture:
46 + with pytest.raises(ValueError, match="unsupported LoRA weight"):
47 + parse_fine_tuned_lora_name(name)
48 +
49 +
50 def test_replace_submodule():
51 model = nn.Sequential(
52 OrderedDict([
vllm/lora/utils.py
+7 -8
@@ -94,13 +94,12 @@ def parse_fine_tuned_lora_name(name: str) -> Tuple[str, bool]:
94 is_lora_a whether the tensor is lora_a or lora_b.
95 """
96 parts = name.split(".")
97 - assert parts[0] == "base_model"
98 - assert parts[1] == "model"
99 - if parts[-1] == "weight":
100 - assert parts[-2] == "lora_A" or parts[-2] == "lora_B"
101 - return ".".join(parts[2:-2]), parts[-2] == "lora_A"
97
103 - if parts[-1] == "lora_embedding_A" or parts[-1] == "lora_embedding_B":
104 - return ".".join(parts[2:-1]), parts[-1] == "lora_embedding_A"
98 + if len(parts) >= 2 and parts[0] == "base_model" and parts[1] == "model":
99 + if parts[-1] == "weight":
100 + if parts[-2] == "lora_A" or parts[-2] == "lora_B":
101 + return ".".join(parts[2:-2]), parts[-2] == "lora_A"
102 + elif parts[-1] == "lora_embedding_A" or parts[-1] == "lora_embedding_B":
103 + return ".".join(parts[2:-1]), parts[-1] == "lora_embedding_A"
104
106 - raise ValueError(f"{name} is unsupported format")
105 + raise ValueError(f"{name} is unsupported LoRA weight")