perf(data): parallelize NeMo Gym image encoding (cherry-pick #3721)

[DanialTaheri]

What does this PR do ?

Add a one line overview of what this PR aims to accomplish.

Cherry-picks #3721 (perf(data): parallelize NeMo Gym image encoding, 55227be9c88b723ca6301d66330703eb1ee8479c on main) onto super-v3.5-posttraining.

encode_images_in_examples previously resolved and base64-encoded every local image reference serially, once per occurrence. The picked commit collects the references first, encodes each unique source once through a bounded ThreadPoolExecutor (NEMO_GYM_IMAGE_ENCODE_MAX_WORKERS = 8), then mutates the payload on the caller thread.

Issues

List issues that this PR closes (syntax):

None — this is a backport of an already-merged main commit.

Usage

• You can potentially add a usage example below

No API change. The only call site is nemo_rl/environments/nemo_gym.py:848, which is byte-identical on this branch and on main.

from nemo_rl.data.multimodal_utils import encode_images_in_examples

# Same signature and in-place mutation semantics as before; local image paths
# are now resolved concurrently and deduplicated across the batch.
encode_images_in_examples(nemo_gym_examples)

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests? — the four tests added by perf(data): parallelize NeMo Gym image encoding #3721 to tests/unit/data/test_multimodal_image_encoding.py come along with the pick.
• Did you run the unit tests and functional tests locally? Visit our Testing Guide for how to run tests — not run locally; relying on CI here (see below).
• Did you add or update any necessary documentation? — docstring updated by the picked commit.

Additional Information

• Cherry-picked with -x; applied with no conflicts. After the pick, both nemo_rl/data/multimodal_utils.py and tests/unit/data/test_multimodal_image_encoding.py are byte-identical to 55227be, so this branch carries exactly the code that passed CI on main.
• A plain git merge of 55227be was deliberately avoided: super-v3.5-posttraining diverged from main at b5916526, so a merge would have dragged in six unrelated commits (mooncake CPU RDMA data plane feat(data_plane): mooncake cpu rdma backend #2935, Qwen3.5 blockwise FP8 fix(fp8): support Qwen3.5 blockwise FP8 rollout and training #2744, MoE aux-loss PP hang fix: prevent PP hang in MoE aux-loss metric reduction #3419, tokenizer-save NFS deadlock fix: guard tokenizer save with rank 0 to avoid NFS deadlock #3504, sequence importance ratio metric fix(grpo): correct sequence importance ratio metric #3753, reference-logprob skip fix(sc): skip reference logprobs when KL is disabled #3771).
• Unit tests were not executed locally — this cluster checkout has no torch/ray environment available, and tests/unit/conftest.py imports ray at collection time. CI on this PR is the first real execution of the picked tests against this branch.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.