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 5, 02:15 UTC
Fix #33773: Replace unconditional pandas import with PlaceholderModule (#39990)
Signed-off-by: Netanel Haber <58652339+netanel-haber@users.noreply.github.com>
Netanel Haber committed
Apr 16, 2026 at 12:06 UTC
98700c6105b0313d924126cba428219111ee8f6f
1 file changed
+6
-1
vllm/_aiter_ops.py
+6
-1
@@ -3,18 +3,23 @@
3
import functools
4
from collections.abc import Callable
5
6
-import pandas as pd
6
import torch
7
from torch._ops import OpOverload
8
9
import vllm.envs as envs
10
from vllm.platforms import current_platform
11
+from vllm.utils.import_utils import PlaceholderModule
12
from vllm.utils.torch_utils import direct_register_custom_op
13
from vllm.v1.attention.ops.rocm_aiter_mla_sparse import (
14
rocm_aiter_sparse_attn_indexer,
15
rocm_aiter_sparse_attn_indexer_fake,
16
)
17
18
+try:
19
+ import pandas as pd
20
+except ImportError:
21
+ pd = PlaceholderModule("pandas")
22
+
23
# fp8_dtype is not cached.
24
# on ROCm the fp8_dtype always calls is_fp8_fnuz
25
# which is a host op, so we cache it once here.