yindo/rocm-automated
1
0
Fork 0
mirror of https://github.com/BillyOutlast/rocm-automated.git synced 2026-02-04 03:51:19 +01:00
Code Issues Packages Projects Releases Wiki Activity

reneabling extra docker compoients after good run

Browse Source
This commit is contained in:
John Doe
2025-11-24 00:24:56 -05:00
parent e9a99d0ae5
commit 49073253d9
1 changed files with 21 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

42
docker-compose.yaml
View File

@@ -107,32 +107,32 @@ services:
- MIOPEN_USER_DB_PATH=/tmp/.miopen
- MIOPEN_CUSTOM_CACHE_DIR=/tmp/.miopen
# === GPU targeting ===
# - HCC_AMDGPU_TARGET="gfx1100" # Change for your GPU
# - PYTORCH_ROCM_ARCH="gfx1100" # e.g., gfx1030 for RX 6800/6900
- HCC_AMDGPU_TARGET="gfx1100" # Change for your GPU
- PYTORCH_ROCM_ARCH="gfx1100" # e.g., gfx1030 for RX 6800/6900
# === Memory allocator tuning ===
# - PYTORCH_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:6144
- PYTORCH_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:6144
# === Precision and performance ===
# - TORCH_BLAS_PREFER_HIPBLASLT=0
# - TORCHINDUCTOR_MAX_AUTOTUNE_GEMM_BACKENDS="CK,TRITON,ROCBLAS"
# - TORCHINDUCTOR_MAX_AUTOTUNE_GEMM_SEARCH_SPACE="BEST"
# - TORCHINDUCTOR_FORCE_FALLBACK=0
- TORCH_BLAS_PREFER_HIPBLASLT=0
- TORCHINDUCTOR_MAX_AUTOTUNE_GEMM_BACKENDS="CK,TRITON,ROCBLAS"
- TORCHINDUCTOR_MAX_AUTOTUNE_GEMM_SEARCH_SPACE="BEST"
- TORCHINDUCTOR_FORCE_FALLBACK=0
# === Flash Attention ===
# - FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE"
# - FLASH_ATTENTION_BACKEND="flash_attn_triton_amd"
# - FLASH_ATTENTION_TRITON_AMD_SEQ_LEN=4096
# - USE_CK=ON
# - TRANSFORMERS_USE_FLASH_ATTENTION=1
# - TRITON_USE_ROCM=ON
# - TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
- FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE"
- FLASH_ATTENTION_BACKEND="flash_attn_triton_amd"
- FLASH_ATTENTION_TRITON_AMD_SEQ_LEN=4096
- USE_CK=ON
- TRANSFORMERS_USE_FLASH_ATTENTION=1
- TRITON_USE_ROCM=ON
- TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
# === CPU threading ===
# - OMP_NUM_THREADS=8
# - MKL_NUM_THREADS=8
# - NUMEXPR_NUM_THREADS=8
- OMP_NUM_THREADS=8
- MKL_NUM_THREADS=8
- NUMEXPR_NUM_THREADS=8
# === Experimental ROCm flags ===
# - HSA_ENABLE_ASYNC_COPY=1
# - HSA_ENABLE_SDMA=1
# - MIOPEN_FIND_MODE=2
# - MIOPEN_ENABLE_CACHE=1
- HSA_ENABLE_ASYNC_COPY=1
- HSA_ENABLE_SDMA=1
- MIOPEN_FIND_MODE=2
- MIOPEN_ENABLE_CACHE=1
# - MIGRAPHX_MLIR_USE_SPECIFIC_OPS="attention"
ports:
- "8188:8188"