Environment Variables¶

RBLN PyTorch reads its configuration from environment variables. TORCH_RBLN_* variables belong to torch-rbln; RBLN_* variables in the device-mapping group belong to the RBLN runtime and are documented here because torch-rbln interprets them. Set variables before the process starts unless a row says the variable is read at runtime.

When a variable is read matters:

Logging and diagnostics¶

1

export TORCH_RBLN_LOG_LEVEL=INFO

Execution and fallback¶

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

# Disable every configurable fallback category; the paths listed above still run on CPU
export TORCH_RBLN_DISABLE_FALLBACK=all

# Only surface unsupported operators
export TORCH_RBLN_DISABLE_FALLBACK=unsupported_op

# Two categories
export TORCH_RBLN_DISABLE_FALLBACK=compile_error,strided_copy_error

# Back to the default
unset TORCH_RBLN_DISABLE_FALLBACK

Device mapping¶

By default each physical NPU maps 1:1 to a logical device rbln:N. On ATOM™, to group several NPUs into one logical device for Rebellions Scalable Design (RSD), set one of the grouping variables. On Rebel100™ leave them unset: a card is a fixed RSD-4 group of its four chiplets, and several cards are combined with rbln-ccl (Tutorial: Collectives with rbln-ccl). All ids below are indices into the pool that RBLN_VISIBLE_DEVICES leaves visible.

Priority: RBLN_DEVICE_MAP > RBLN_NPUS_PER_DEVICE > default 1:1.

Examples with four visible NPUs:

With six visible NPUs and RBLN_NPUS_PER_DEVICE=4, rbln:0 = NPUs [0, 1, 2, 3] and NPUs 4, 5 stay unused with a warning.

[RBLN] Device Topology Initialized:
+----------------+------------------+---------------------+
| Logical Device | Physical NPU IDs | Status              |
+----------------+------------------+---------------------+
| rbln:0         | [ 0, 1 ]         | Active (Aggregated) |
| rbln:1         | [ 2, 3 ]         | Active (Aggregated) |
+----------------+------------------+---------------------+

[RBLN] Device Topology Initialized:
+----------------+------------------+---------------------+
| Logical Device | Physical NPU IDs | Status              |
+----------------+------------------+---------------------+
| rbln:0         | [ 0, 1 ]         | Active (Aggregated) |
| -              | [ 2 ]            | Unused              |
+----------------+------------------+---------------------+
[Warning] 1 physical NPU(s) are unused due to grouping constraints.

Distributed¶

The control-plane variables are read when torch.distributed.init_process_group(backend="rbln-ccl") creates the process group, not at import.

On one host the defaults are enough: launch one process per NPU with torchrun, give each process its own NPU through RBLN_VISIBLE_DEVICES before the first device use, and call init_process_group(backend="rbln-ccl"). Tutorial: Collectives with rbln-ccl shows the launch. Across hosts, set RBLN_ROOT_IP and RBLN_LOCAL_IP on every rank before init_process_group, and pin the data-plane NIC with RBLN_RDMA_HCA or RBLN_RDMA_IP when the host has more than one RDMA interface.

Profiling¶

With RBLN_PROFILER=1 set for the process, a torch.profiler.profile(activities=[ProfilerActivity.CPU, ProfilerActivity.PrivateUse1]) region records RBLN device events next to the ATen calls, and the runtime writes its own profile file (rbln_<timestamp>.pb) for the RBLN Profiler tools when the process exits. Tutorial: Profiling on Rebel100™ walks through a run.

ABI check¶

import torch verifies that this torch-rbln build and the librbln.so it loads agree on the rebel ABI contract. See Troubleshoot for the messages.

Memory¶

Development-only options¶

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

# Benchmark without the NaN/Inf scan
export TORCH_RBLN_DEV_DISABLE_OP_CPU_FALLBACK=nan_inf

# Run under pdb without CPU fallback
export TORCH_RBLN_DEV_DISABLE_OP_CPU_FALLBACK=trace

# Several checks
export TORCH_RBLN_DEV_DISABLE_OP_CPU_FALLBACK=dtype,scalar

# Back to the default
unset TORCH_RBLN_DEV_DISABLE_OP_CPU_FALLBACK

See also¶

• Execution Semantics - the behaviors these variables change
• Troubleshoot - which variable to reach for on each symptom
• RBLN Compiler › Troubleshoot - device selection with RBLN_VISIBLE_DEVICES from the runtime's side
• RBLN Profiler - RBLN_PROFILER and the profiler output