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, 03:00 UTC
[ROCm][Bugfix] Fixed several bugs related to rccl path and attention selector logic (#3699)
Hongxia Yang committed
Mar 29, 2024 at 17:52 UTC
9765b5c4061b37c0db417aa6c638b0ca90183be1
4 files changed
+5
-5
Dockerfile.rocm
+1
-1
@@ -90,6 +90,6 @@ RUN cd /app \
90
&& cd ..
91
92
RUN python3 -m pip install --upgrade pip
93
-RUN python3 -m pip install --no-cache-dir ray[all]
93
+RUN python3 -m pip install --no-cache-dir ray[all]==2.9.3
94
95
CMD ["/bin/bash"]
requirements-rocm.txt
+1
-1
@@ -5,7 +5,7 @@ starlette
5
requests
6
py-cpuinfo
7
psutil
8
-ray >= 2.9
8
+ray == 2.9.3
9
sentencepiece # Required for LLaMA tokenizer.
10
numpy
11
tokenizers>=0.15.0
vllm/attention/backends/xformers.py
+2
-2
@@ -405,8 +405,8 @@ def _check_use_naive_attention() -> bool:
405
if not is_hip():
406
return False
407
# For ROCm, check whether flash attention is installed or not.
408
- has_flash_attn = importlib.util.find_spec("flash_attn") is None
409
- if not has_flash_attn:
408
+ use_naive_attention = importlib.util.find_spec("flash_attn") is None
409
+ if use_naive_attention:
410
logger.warning("flash_attn is not installed. Using naive attention. "
411
"This will take significantly more GPU memory.")
412
return True
vllm/model_executor/parallel_utils/pynccl.py
+1
-1
@@ -41,7 +41,7 @@ else:
41
if torch.version.cuda is not None:
42
so_file = "libnccl.so.2"
43
elif torch.version.hip is not None:
44
- so_file = "librccl.so.2"
44
+ so_file = "librccl.so.1"
45
else:
46
raise ValueError("NCCL only supports CUDA and ROCm backends.")
47
logger.debug(f"Loading nccl from library {so_file}")