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 14, 05:30 UTC
[CI Failure] Disable FlashInfer RoPE to unblock CI (#25299)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Sep 20, 2025 at 04:16 UTC
e08a3a3fdbdb5408f904a237b31ff2447a336b2f
1 file changed
+7
-7
vllm/model_executor/layers/rotary_embedding/base.py
+7
-7
@@ -6,8 +6,6 @@ from typing import Optional
6
import torch
7
8
from vllm.model_executor.custom_op import CustomOp
9
-from vllm.platforms import current_platform
10
-from vllm.utils.flashinfer import has_flashinfer
9
10
from .common import apply_rotary_emb_torch
11
@@ -32,13 +30,15 @@ class RotaryEmbedding(CustomOp):
30
self.base = base
31
self.is_neox_style = is_neox_style
32
self.dtype = dtype
33
+ # TODO(mgoin): disabled for now due to failures
34
# Flashinfer only supports head_size=64, 128, 256, 512.
35
# https://github.com/flashinfer-ai/flashinfer/blob/ebfd655efe830048dba5d582aaa61d61d1cf9a87/include/flashinfer/utils.cuh#L174-L202
37
- self.use_flashinfer = (self.enabled()
38
- and dtype in (torch.float16, torch.bfloat16)
39
- and current_platform.is_cuda()
40
- and has_flashinfer()
41
- and self.head_size in [64, 128, 256, 512])
36
+ # self.use_flashinfer = (self.enabled()
37
+ # and dtype in (torch.float16, torch.bfloat16)
38
+ # and current_platform.is_cuda()
39
+ # and has_flashinfer()
40
+ # and self.head_size in [64, 128, 256, 512])
41
+ self.use_flashinfer = False
42
43
cache = self._compute_cos_sin_cache()
44
if not self.use_flashinfer: