mcore-migrate-gpt-to-hybrid

por nvidia

Guia de migração para mover checkpoints do GPTModel do Megatron Core, provedores de modelo, comandos de treinamento e mapeamentos de camadas para o HybridModel.

npx skills add https://github.com/nvidia/megatron-lm --skill mcore-migrate-gpt-to-hybrid

GPTModel to HybridModel Migration

Answer-First Migration Guidance

  • The canonical source is docs/user-guide/hybrid-model-migration.md.
  • Read the canonical document completely before answering, planning, reviewing, editing, converting, or training.
  • Keep migration behavior, commands, mappings, prerequisites, limitations, and validation in the canonical document only. Do not duplicate them in this skill.
  • This skill adds only what the canonical document does not cover: the mechanical procedure for editing an existing launch script, and the shell hazards that procedure runs into.

Workflow

  1. Pull the task artifact first: checkpoint metadata, model provider or config, training command, conversion log, diff, or failure output.
  2. Read the canonical migration document completely.
  3. Follow only the relevant document sections. Do not invent an unsupported migration path or silently change the target architecture.
  4. Validate the result proportionately, invoking the relevant repository build and testing skills when applicable.
  5. Report the outcome and link the canonical document for human readers.

Transferring an Existing Launch Script

The canonical document specifies what the migrated command must contain. This section covers how to edit a working script into it without silent breakage. Apply the edits in order.

1. Entrypoint. pretrain_gpt.pypretrain_hybrid.py. When the entrypoint comes from a shell variable or a wrapper, follow it to the real invocation.

2. Generate the pattern instead of typing it. A 96-layer model needs a 192-character pattern; hand-typing invites a silent off-by-one.

n=32                                   # source GPT layer count
blk='*-'                               # '*-' dense, '*E' every-layer MoE
pat=$(printf "$blk%.0s" $(seq $n))

With pipeline segments (seg must divide n, and the segment count must be divisible by --pipeline-model-parallel-size):

n=32; seg=4; per=$((n/seg))
b=$(printf "$blk%.0s" $(seq $per)); pat=$b
for ((i=1;i<seg;i++)); do pat="$pat|$b"; done

3. Replace --num-layers N with --hybrid-layer-pattern. Deleting --num-layers matters: leaving a stale value is only a warning, so it looks healthy while being silently overridden by the pattern-derived count.

4. Add the stack spec, replacing any GPT --spec rather than adding a second one.

5. Delete the pipeline-layout arguments the parser rejects, and repoint --save at a new directory. See the canonical document for both lists.

Bash-array scripts: quoting at the definition site is not enough

Most scripts under examples/ collect arguments in arrays and expand them unquoted:

torchrun ${DISTRIBUTED_ARGS[@]} pretrain_gpt.py ${MODEL_ARGS[@]}

Unquoted ${ARR[@]} re-runs word-splitting and pathname expansion on every element, so the pattern is globbed against the launch directory at expansion time — single-quoting it where the array is defined does not protect it:

touch 'a-b-'; ARGS=(--hybrid-layer-pattern '*-*-')
printf '[%s]\n' ${ARGS[@]}     # -> [--hybrid-layer-pattern] [a-b-]   silently corrupted
printf '[%s]\n' "${ARGS[@]}"   # -> [--hybrid-layer-pattern] [*-*-]   correct

An unmatched glob survives intact, so this passes by luck in most working directories and fails only when some file happens to match. Store the pattern in a variable and quote that array's expansion:

HYBRID_PATTERN=$(printf '*-%.0s' $(seq $NUM_LAYERS))
MODEL_ARGS=( ... --hybrid-layer-pattern "$HYBRID_PATTERN" ... )
torchrun "${DISTRIBUTED_ARGS[@]}" pretrain_hybrid.py "${MODEL_ARGS[@]}" ...

Verify the rewrite

Both checks are cheap and catch the common slips:

# 1. No rejected or stale arguments survived -- must print nothing.
grep -nE -- '--(num-layers|num-layers-per-virtual-pipeline-stage|num-virtual-stages-per-pipeline-rank|pipeline-model-parallel-layout|account-for-embedding-in-pipeline-split|account-for-loss-in-pipeline-split|hybrid-override-pattern|fim-data)\b' train_hybrid.sh

# 2. Pattern shape -- attn and mlp must each equal the source GPT layer count.
p='*-*-|*-*-'
main=${p%%/*}; main=${main//|/}
attn=${main//[^\*]/}; mlp=${main//[^-E]/}; segs=${p%%/*}; segs=${segs//[^|]/}
echo "layers=${#main} attn=${#attn} mlp=${#mlp} segments=$(( ${#segs} + 1 ))"

Then diff the migrated script against the original: it should contain the edits above and nothing else.

Expected result of an architecture-preserving transfer

A *- or *E transfer changes the layer indexing, not the model. On a measured 8-block dense run (2 GPUs, bf16, seq 4096, 100 iterations, identical seed and data), pretrain_gpt.py --num-layers 8 and pretrain_hybrid.py --hybrid-layer-pattern '*-*-*-*-*-*-*-*-' produced:

  • identical parameter counts (2,818,641,920 on both);
  • HybridModel: ... layers='*-*-*-*-*-*-*-*-' (16 layers) from the allocator;
  • steady-state throughput within 0.1% (490.7 vs 491.2 ms/iter);
  • identical loss for the first two iterations, then a zero-mean drift of |Δ| ≤ 0.08 attributable to kernel/reduction ordering.

Treat a systematic loss offset, a parameter-count difference, or a throughput gap beyond noise as a migration bug, not as expected behavior. Note that per-iteration wall clock early in a run is dominated by dataset-cache warmup, so compare steady-state iterations only.


Documentation Drift

If the implementation and migration guide disagree:

  1. Report the discrepancy before continuing.
  2. If the task authorizes a correction, update the canonical document first.
  3. Do not add a competing migration rule to this skill.

Mais skills de nvidia

compileiq-debug
nvidia
Use quando algo está errado: Search() trava, todas as avaliações retornam INVALID_SCORE, as pontuações não estão melhorando, toda configuração retorna o mesmo número, erros de ptxas…
create-github-pr
nvidia
Crie pull requests do GitHub usando a CLI gh. Use quando o usuário quiser criar um novo PR, enviar código para revisão ou abrir um pull request. Palavras-chave de acionamento -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Escaneia outras issues abertas para encontrar aquelas que um determinado PR pode também corrigir ou quebrar acidentalmente. Gera oportunidades de correção adjacentes e riscos de contradição com arquivo:linha…
fhir-basics
nvidia
Ensina aos agentes como funcionam as APIs FHIR R4, quais recursos estão disponíveis, como consultá-los com parâmetros de busca e como analisar corretamente todos os formatos de resposta…
compileiq-validate-result
nvidia
Use APÓS a conclusão de uma Pesquisa e ANTES de reivindicar qualquer aceleração ou enviar um ACF. Carrega o CSV dump_results, extrai os K melhores candidatos (objetivo único)…
changelog-audit
nvidia
Auditar o CHANGELOG.md do Warp antes de um lançamento: recuperar entradas perdidas, ordenar por impacto ao usuário, refinar a linguagem das entradas, ajustar quebras de linha e (no modo de branch de lançamento) incrementar comparação…
maintain-dynamic-plugins
nvidia
Manter carregadores de plugins dinâmicos do NeMo Relay, manifestos, SDKs nativos em Rust, protocolo de worker gRPC, SDK de worker Python, documentação, testes e cobertura do fluxo de lançamento
dgx-diagnose
nvidia
Diagnostique problemas comuns do DGX Station GB300 — falhas de CUDA, direcionamento incorreto de GPU, bugs de contêiner vLLM/SGLang, problemas de estado MIG, erros de NVLink/Fabric Manager,…