Nemo Mbridge Perf Sequence Packing

Sequence Packing Skill

For stable background and recommendation level, see:

• @docs/training/packed-sequences.md
• @skills/nemo-mbridge-perf-sequence-packing/card.yaml

Enablement

Offline packed SFT for LLM finetuning:

import math

from megatron.bridge.data.datasets.packed_sequence import PackedSequenceSpecs

cfg.train.micro_batch_size = 1
cfg.train.global_batch_size = 8
cfg.dataset.seq_length = 8192
cfg.model.seq_length = 8192
cfg.dataset.enable_offline_packing = True

cp_size = cfg.model.context_parallel_size
tp_size = cfg.model.tensor_model_parallel_size
cp_multiple = 2 * cp_size if cp_size > 1 else 1
sp_multiple = cp_size * tp_size if cfg.model.sequence_parallel and tp_size > 1 else 1
cfg.dataset.offline_packing_specs = PackedSequenceSpecs(
    packed_sequence_size=8192,
    pad_seq_to_mult=math.lcm(cp_multiple, sp_multiple),
)

Choose the offline pack length

For text-only LLM SFT and PEFT verification, start with an 8192-token offline pack when the model context limit, memory, and model-family support allow it. Benchmark pack lengths at equal token slots per optimizer step:

token_slots_per_step = packed_sequence_size * global_batch_size

For example, 2K/GBS32, 4K/GBS16, and 8K/GBS8 each expose 65,536 token slots per step. Longer packs aggregate more source examples into each physical MBS1 row and can reduce gradient accumulation and per-step overhead. They also increase activation memory and may expose kernel-width constraints, so select the largest measured configuration that fits rather than assuming longer is always faster.

Offline packing requires MBS1. Require global_batch_size % data_parallel_size == 0 and global_batch_size >= data_parallel_size; an 8K/GBS8 workload therefore needs DP no larger than 8. Keep model.seq_length, dataset.seq_length, and packed_sequence_size equal, use a fresh packed-data output root after changing any of them, and inspect the resolved post-setup configuration.

Equal token slots do not make different pack lengths numerically identical: the longer target changes truncation and pack membership. Rerun finite-loss, no-skip/NaN, and convergence sentinels before replacing verified evidence.

For finetuning with CP enabled:

cfg.model.context_parallel_size = 2
cfg.model.calculate_per_token_loss = True
cfg.ddp.average_in_collective = False

Use the same alignment formula for SFT and PEFT. It produces 1 for TP1/CP1 with SP disabled and 4 for TP4/CP1 with SP enabled. Offline packing does not derive the value automatically, so pin it explicitly and rebuild packed data after a topology change.

If a dispatcher or kernel requires a fixed final token width:

cfg.dataset.dataset_kwargs = {
    **(cfg.dataset.dataset_kwargs or {}),
    "pad_to_max_length": True,
}

Choose packed_sequence_size to satisfy the kernel multiple. For example, HybridEP with a 128-token combine chunk requires a width divisible by 128. This is separate from pad_seq_to_mult, which aligns each constituent sequence for CP/SP.

If CUDA graphs are enabled for this packed path, fixed token width is required and packed metadata must also have a static shape:

cfg.dataset.offline_packing_specs.pad_cu_seqlens = True
cfg.dataset.dataset_kwargs["pad_to_max_length"] = True

Note: pad_cu_seqlens = True also requires a metadata JSON file alongside the packed dataset (asserted in src/megatron/bridge/data/datasets/sft.py). Custom packed datasets that omit the metadata file will hit an assertion at dataset initialization.

In-batch packing for GPT SFT and supported VLM finetuning:

cfg.dataset.enable_in_batch_packing = True
cfg.dataset.dataloader_type = "single"
cfg.train.micro_batch_size = 4

For local or materialized GPT-SFT JSONL, this keeps the existing mmap-backed dataset and performs tokenization lazily. Both prompt/completion (GPTSFTDataset) and chat (GPTSFTChatDataset) preserve their loss-mask semantics. Use dataloader_type="single" or "cyclic" so every DataLoader yield is one logical microbatch; GPT-SFT in-batch packing does not support the global-batch "batch" dataloader.

Energon online packing for Qwen-VL uses Energon's per-worker candidate buffer instead of limiting selection to one collator micro batch:

cfg.dataset.packing_buffer_size = 16
cfg.dataset.micro_batch_size = 1
cfg.train.micro_batch_size = 1
cfg.model.calculate_per_token_loss = True
cfg.ddp.average_in_collective = False

packing_buffer_size is the sole native-packing selector; leave the legacy collator- and step-owned packing flags at their defaults. Use vlm_step. The buffer size counts prepared candidate samples per worker, not bytes or packed tokens. Since prepared image/video patch tensors remain in host memory until selection, start at 8-16 for high-resolution or video data and measure worker RSS, first-batch latency, and bin fill before increasing it. This path does not write offline packs; the source WebDataset shards remain unchanged. It supports eager Qwen-VL with MBS1 and rejects MTP, CUDA graphs, Qwen3-VL DistTrain, and PP. Requested MoE expert-parallel communication overlap is disabled with a warning. Standard eager alltoall EP has functional coverage for Qwen3.6-35B-A3B at TP1/PP1/EP8 with overlap disabled; this is not performance evidence. Other EP dispatchers are accepted with fixed-width native packs but do not yet have equivalent runtime evidence. The Qwen-VL model derives a MoE padding mask from logical and physical THD boundaries so fixed-width gaps do not enter auxiliary-loss, z-loss, or expert-bias statistics. Current MCore may still dispatch padded positions; expert-capacity/token-dropping configurations lack native-packing runtime coverage.

Long-context baseline:

cfg.model.seq_length = 16384
cfg.dataset.seq_length = 16384
cfg.model.context_parallel_size = 2