[diffusion] fix: restore nested condition rows in FlowGRPO actor replay - #534

What does this PR do?

Fixes a RuntimeError: slice() not supported for NestedTensor on dim=0 that blocks
MiniMax H3 Ref2VA FlowGRPO training on current main. T2VA and FL2VA are
unaffected.

Ref2VA reference condition rows (condition_video_rows / condition_audio_rows)
are padded to a global length and turned into jagged nested tensors by
embeds_padding_2_no_padding. The FlowGRPO diffusion engine never restored them
to dense tensors, so the MiniMax H3 Ref2VA adapter's [:, :count] slice in
diffusers_training_adapter.py crashed before the first training step.

Checklist Before Starting

• Search for similar PRs. There is no open PR fixing this; the only related
  open work is [3/N][diffusion, rollout, tests] refactor: unify diffusion rollout request contract #480 (rollout request refactor) and [doc, recipe] chore: document MiniMax H3 training support #529 (docs), which are
  unrelated to nested condition rows.
• Format the PR title as [diffusion] fix: ... (validated with
  tests/special_sanity/check_pr_title.py).

Test

CPU regression coverage in
tests/pipelines/test_minimax_h3_ref2va_flow_grpo_on_cpu.py:

• test_ref2va_actor_replays_full_layout_and_scores_only_targets: builds the
  complete Ref2VA replay micro-batch with nested condition rows and drives it
  through PPODiffusersFSDPEngine.prepare_model_inputs, verifying the engine
  restores dense rows before the adapter performs [:, :count] slicing.
• test_ref2va_engine_unpads_nested_condition_rows: directly covers dense
  restoration and row-count handling for heterogeneous nested rows.
• test_ref2va_engine_rejects_mismatched_nested_mask: verifies nested rows
  without a nested mask fail closed.

Commands run (project venv, CPU):

PYTHONPATH=$PWD python -m pytest \
  tests/pipelines/test_minimax_h3_ref2va_flow_grpo_on_cpu.py::test_ref2va_actor_replays_full_layout_and_scores_only_targets -q
# 1 passed

PYTHONPATH=$PWD python -m pytest \
  tests/pipelines/test_minimax_h3_flow_grpo_on_cpu.py \
  tests/pipelines/test_minimax_h3_diffusion_nft_on_cpu.py \
  tests/pipelines/test_minimax_h3_ref2va_flow_grpo_on_cpu.py \
  tests/pipelines/test_minimax_h3_ref2va_nft_on_cpu.py \
  tests/workers/test_diffusers_dtype_on_cpu.py -q
# 93 passed

pre-commit run --files \
  verl_omni/workers/engine/fsdp/diffusers_impl.py \
  tests/pipelines/test_minimax_h3_ref2va_flow_grpo_on_cpu.py
# Passed

Negative-control check: temporarily replacing
PPODiffusersFSDPEngine._unpad_condition_rows with a no-op makes the full
actor-replay test fail with the original
RuntimeError: slice(): not supported for NestedTensor on dim=0.

A stacked eight-GPU TinyRandom Ref2VA FlowGRPO smoke also completed one rollout
and actor-update step (training/global_step: 1) and produced eight MP4s plus
one JSONL rollout record.

Design & Code Changes

• Hoist _unpad_condition_rows from NFTDiffusersFSDPEngine into the shared
  DiffusersFSDPEngine base so it is available to all diffusion engines.
• Call self._unpad_condition_rows(micro_batch) in
  PPODiffusersFSDPEngine.prepare_model_inputs before delegating to the adapter, so
  condition rows reach the MiniMax H3 Ref2VA adapter as dense padded tensors.
• Remove the now-duplicated _unpad_condition_rows definition from the NFT engine
  (it inherits the base version; its existing call site still resolves correctly).
• Leave DPODiffusersFSDPEngine unchanged because there is no Ref2VA DPO pipeline.

Checklist Before Submitting

• Read the Contribute Guide.
• Apply pre-commit checks (see Test section).
• Added engine-entry regression coverage for the restored-rows and
  mismatched-mask paths.

AI assistance

AI assistance (pi coding agent) was used for this change.

A human submitter (NancyFyong) reviewed every changed line and verified the
commands/results above.