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 15, 12:15 UTC
FIxing the AMD test failures caused by PR#16457 (#17511)
Signed-off-by: Alexei V. Ivanov <alexei.ivanov@amd.com>
Alexei-V-Ivanov-AMD committed
May 1, 2025 at 01:23 UTC
7ab643e425c6c894e37b8bb25b987185dc5c0e9e
2 files changed
+1
-2
docker/Dockerfile.rocm
-1
@@ -118,4 +118,3 @@ ENV TOKENIZERS_PARALLELISM=false
118
ENV HIP_FORCE_DEV_KERNARG=1
119
120
CMD ["/bin/bash"]
121
-
vllm/model_executor/layers/rotary_embedding.py
+1
-1
@@ -80,7 +80,7 @@ def _apply_rotary_emb(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor,
80
is_neox_style: Whether to use the Neox-style or GPT-J-style rotary
81
positional embeddings.
82
"""
83
- if current_platform.is_cuda_alike():
83
+ if current_platform.is_cuda():
84
return apply_rotary_emb(x.unsqueeze(0), cos, sin,
85
not is_neox_style).squeeze(0)
86
else: