Configuration

Environment variables and runtime configuration objects you'll encounter.

Build-time

Variable	Effect
USE_CUDA=0	Skip CUDA build
USE_ROCM=1	ROCm build
USE_MPS=0	Skip MPS
USE_DISTRIBUTED=0	Skip distributed
USE_NCCL=0	Skip NCCL
USE_GLOO=0	Skip Gloo
USE_FBGEMM=0	Skip FBGEMM
USE_KINETO=0	Skip Kineto profiler
USE_NNPACK=0	Skip NNPACK
USE_XNNPACK=0	Skip XNNPACK
USE_MKLDNN=0	Skip oneDNN/MKLDNN
USE_OPENMP=0	Skip OpenMP
BUILD_TEST=0	Skip C++ test binaries
MAX_JOBS=N	Limit parallel build jobs
CMAKE_BUILD_PARALLEL_LEVEL	CMake parallelism
DEBUG=1	-O0 -g
REL_WITH_DEB_INFO=1	-O2 -g
TORCH_CUDA_ARCH_LIST	e.g., 7.0;8.0;9.0
CUDA_HOME, CUDNN_HOME, MKL_ROOT	Toolkit paths
BUILD_CAFFE2=0	Skip residual Caffe2 (default off)
BUILD_BINARY=0	Skip binary build

CUDA runtime

Variable	Effect
CUDA_LAUNCH_BLOCKING=1	Serialize CUDA launches
CUBLAS_WORKSPACE_CONFIG=:16:8 or :4096:8	Make cuBLAS deterministic
PYTORCH_CUDA_ALLOC_CONF	Caching allocator knobs
PYTORCH_NO_CUDA_MEMORY_CACHING=1	Disable allocator caching (debug)
CUDA_VISIBLE_DEVICES	Restrict visible GPUs
TORCH_USE_CUDA_DSA	Enable device-side asserts
TORCH_CUDNN_V8_API_ENABLED=1	Use cuDNN v8 frontend

PYTORCH_CUDA_ALLOC_CONF accepts comma-separated key:value pairs:

Key	Default	Notes
max_split_size_mb	unlimited	Don't split blocks larger than this
garbage_collection_threshold	0	Aggressively collect to keep below threshold
expandable_segments	False	Use cuMemMap segments
roundup_power2_divisions	0	More size classes per power-of-2 bucket
pinned_use_cuda_host_register	False	Pin host memory via cudaHostRegister
pinned_num_register_threads	1	Worker threads for the above

Compile stack

Variable	Effect
TORCH_LOGS	Comma-separated logger names; +/- prefix for level adjust
TORCHDYNAMO_VERBOSE=1	Verbose dynamo
TORCH_COMPILE_DEBUG=1	Dump every IR / kernel artefact to torch_compile_debug/
TORCHINDUCTOR_CACHE_DIR	Where Inductor caches generated kernels
TORCHINDUCTOR_FX_GRAPH_CACHE=1	Enable FX graph cache (default 1)
TORCHINDUCTOR_AUTOTUNE_LOCAL_CACHE=1	Enable autotune local cache
TORCHINDUCTOR_MAX_AUTOTUNE=1	Try every choice, pick fastest (slow but maximum perf)
TORCHINDUCTOR_TRITON_DEBUG_SYNC_KERNEL	Synchronize after every kernel for debugging
TORCH_INDUCTOR_LOG_KERNELS=1	Log every kernel emitted
TORCH_TRACE	Path for structured tracing dump

The Python configs are objects: torch._dynamo.config, torch._inductor.config, torch.compiler.config. Use torch._dynamo.config.patch(...) as a contextmanager / decorator to change values temporarily — per CLAUDE.md, do not manually save/restore.

Distributed

Variable	Effect
MASTER_ADDR, MASTER_PORT	Default rendezvous endpoint
WORLD_SIZE, RANK, LOCAL_RANK	Set by torchrun / elastic launcher
NCCL_DEBUG=INFO	NCCL chatter
NCCL_SOCKET_IFNAME=eth0	Restrict NCCL to specific NIC
NCCL_IB_HCA=mlx5_0	InfiniBand device
TORCH_NCCL_TRACE_BUFFER_SIZE	Flight-recorder buffer entries
TORCH_NCCL_DUMP_ON_TIMEOUT=1	Dump flight recorder on watchdog timeout
TORCH_NCCL_DESYNC_DEBUG=1	Detect mismatched collectives
TORCH_DISTRIBUTED_DEBUG=DETAIL	Per-rank diagnostic logging
TORCH_NCCL_BLOCKING_WAIT=1	Synchronize after every collective
TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC=600	Heartbeat watchdog period

Autograd / numerics

Variable	Effect
TORCH_USE_DETERMINISTIC_ALGORITHMS=1	Equivalent of torch.use_deterministic_algorithms(True)
TORCH_WARN_ALWAYS=1	Don't deduplicate warnings
TORCH_LOGS=structured	Structured trace logging
TORCH_DISABLE_AUTOGRAD_FALLBACK=1	Don't auto-fallback for missing autograd kernels (dev only)

Profiling

Variable	Effect
KINETO_LOG_LEVEL=INFO	Kineto profiler logging
KINETO_DAEMON_INIT_DELAY_S	Delay before daemon connects

Where to look

File	Purpose
torch/_dynamo/config.py	Dynamo configs
torch/_inductor/config.py	Inductor configs
torch/_logging/	TORCH_LOGS parser
c10/cuda/CUDAAllocatorConfig.cpp	PYTORCH_CUDA_ALLOC_CONF
torch/distributed/elastic/	Distributed launcher env vars