Commit
·
de5d1e6
1
Parent(s):
87147f5
env update
Browse files- Dockerfile +6 -0
Dockerfile
CHANGED
|
@@ -29,6 +29,12 @@ RUN set -eux; \
|
|
| 29 |
# (optional) preload workaround if still needed
|
| 30 |
ENV LD_PRELOAD=/usr/local/cuda/lib64/libcusparse.so.12:/usr/local/cuda/lib64/libcublas.so.12:/usr/local/cuda/lib64/libcublasLt.so.12:/usr/local/cuda/lib64/libcufft.so.11:/usr/local/cuda/lib64/libcusolver.so.11
|
| 31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
ENV DEBIAN_FRONTEND=noninteractive \
|
| 33 |
PYTHONUNBUFFERED=1 \
|
| 34 |
PIP_NO_CACHE_DIR=1 \
|
|
|
|
| 29 |
# (optional) preload workaround if still needed
|
| 30 |
ENV LD_PRELOAD=/usr/local/cuda/lib64/libcusparse.so.12:/usr/local/cuda/lib64/libcublas.so.12:/usr/local/cuda/lib64/libcublasLt.so.12:/usr/local/cuda/lib64/libcufft.so.11:/usr/local/cuda/lib64/libcusolver.so.11
|
| 31 |
|
| 32 |
+
# Better allocator (less fragmentation than BFC during XLA autotune)
|
| 33 |
+
ENV TF_GPU_ALLOCATOR=cuda_malloc_async
|
| 34 |
+
|
| 35 |
+
# Let cuBLAS use TF32 fast path on Ada (L40S) for big GEMMs
|
| 36 |
+
ENV TF_ENABLE_CUBLAS_TF32=1 NVIDIA_TF32_OVERRIDE=1
|
| 37 |
+
|
| 38 |
ENV DEBIAN_FRONTEND=noninteractive \
|
| 39 |
PYTHONUNBUFFERED=1 \
|
| 40 |
PIP_NO_CACHE_DIR=1 \
|