Is this page helpful?

Determinism Operation Catalog#

This content is for developers reviewing or extending deterministic-mode coverage. Terms are defined in the glossary.

The catalog has two buckets:

  • Operations with a deterministic code path

  • Operations that deterministic mode cannot support yet

The project goal is to shrink the second bucket and make the first bucket faster.

Most operations in a training step need no entry here. The following are deterministic as-is:

  • Elementwise ops

  • GEMMs under the pinned cuBLAS workspace

  • Rank-indexed collectives (all-gather, all-to-all, broadcast)

  • Stable sorts

  • Unique-index writes

The tables list only the operations where a choice is made.

Deterministic Code Path Operations#

Selected by torch.are_deterministic_algorithms_enabled() or config.deterministic_mode. The default-mode path stays in the other branch.

Operation

Where

Deterministic Path

Default Path

MoE token unpermute (combine)

megatron/core/transformer/moe/moe_utils.py

index_add_ — deterministic under torch deterministic algorithms and CUDA-graph safe

scatter_add_ (atomic accumulation)

MoE routing map and probabilities

megatron/core/transformer/moe/moe_utils.py

index_put_(accumulate=False) row-wise writes

out-of-place scatter

Vocab-parallel embedding

megatron/core/tensor_parallel/layers.py

direct indexing weight[idx] (deterministic backward)

F.embedding (non-deterministic atomic backward)

Gated-delta-net kernel

megatron/core/ssm/gated_delta_net.py

torch chunk_gated_delta_rule

FLA fused kernel

Gated-delta-net causal conv1d

megatron/core/ssm/gated_delta_net/

F.conv1d (plus transposes)

FLA causal_conv1d

Mamba/SSM Triton ops

megatron/core/ssm/ops/common/determinism.py

one fixed autotune config plus a zero-initialized tiled workspace reduced with an ordered sum

timing-based autotune, uninitialized workspace

Mamba/GDP causal conv1d

megatron/core/ssm/causal_conv1d.py

causal_conv1d >= 1.6.0 — per-block workspace for the weight and bias gradients, reduced with an ordered sum

atomicAdd accumulation (order varies per launch)

Transformer Engine attention

megatron/core/extensions/transformer_engine.py

requires NVTE_ALLOW_NONDETERMINISTIC_ALGO=0, under which TE selects only backends that support deterministic execution (including deterministic FlashAttention backward)

TE picks freely, including atomic-accumulation attention backward

Inference DP scheduling and RL rollout order

megatron/core/inference/engines/dynamic_engine.py, megatron/rl/rl_utils.py

sort by stable key

completion order

The two conv rows are different kernels. Mamba and GDP call Dao-AILab’s causal_conv1d, which has its own deterministic reduction; gated-delta-net binds FLA’s, which does not, so it falls back to F.conv1d.

The following environment controls make the rest of the step deterministic. The flag --deterministic-mode validates and defaults these settings. For details, refer to megatron/training/determinism.py.

  • NCCL_ALGO=Ring. The tree is rejected because its reduction order is not user-controllable.

  • CUBLAS_WORKSPACE_CONFIG=:4096:8 (or :16:8).

  • NVTE_ALLOW_NONDETERMINISTIC_ALGO=0.

  • MAMBA_DETERMINISTIC must not be disabled.

  • CAUSAL_CONV1D_DETERMINISTIC must not be disabled. Needs causal_conv1d >= 1.6.0.

Operations Without Determinism Support#

Deterministic mode either rejects these at validation (fails closed) or they are known open gaps.

Operation or Feature

Where Enforced or Observed

Status

Fused cross-entropy loss (--cross-entropy-loss-fusion)

rejected by --deterministic-mode (megatron/training/determinism.py)

The fused kernel is non-deterministic. Whether a deterministic variant is feasible remains an open question. Until then, the framework uses the native vocab-parallel path.

TP communication overlap (--tp-comm-overlap)

rejected by --deterministic-mode

Overlapped collective ordering is not reproducible.

Packed sequence (thd) in gated-delta-net

assertion in megatron/core/ssm/gated_delta_net.py

No deterministic packed-sequence SSM path exists yet.

Cross-allocation floating-point collectives (TP all-reduce, DP grad reduce-scatter)

open gap

NCCL_ALGO=Ring pins the algorithm but not the physical ring an allocation receives. The environment variable alone does not guarantee bit-exactness across different allocations for these reductions. Runs repeated within one allocation, or on allocations with identical topology, remain bit-exact.

Performance Notes#

The deterministic paths above cost roughly 15% of step time compared to default mode, which varies by model. Models that rely heavily on Mixture of Experts (MoE) pay more. Measured examples range from approximately four percent on a large dense model to approximately 17% on a hybrid MoE model. The measured hotspots are:

  • The deterministic MoE scatter and unpermute path

  • The sorted router top-k

  • Attention backward

  • Grouped-GEMM weight gradient (wgrad)

Reducing this cost is a tracked workstream in issue #5785. A change to any row above needs a bit-exact test and a comparison of deterministic and default performance (tests/performance_tests/shell_test_utils/determinism/).