fix: fix dapo-deepseek-v3-64n8g.v2 and dsv3 tokenizer

[yuki-97]

1. Fix dapo-deepseek-v3-64n8g.v2

Update num_prompts_per_step * num_generations_per_prompt from 64 * 8 to 32 * 16.

The dynamic-sampling fix in #2478 filters more samples than before. With num_generations_per_prompt=8, samples are easily filtered and most runs approach the max allowed number of batches (10), triggering the 4h timeout. Some runs also fail with:

ValueError: Dynamic sampling has reached the maximum allowed number of batches (10). Consider evaluating the complexity of your data or adjusting the num_prompts_per_step or num_generations_per_prompt parameters to enhance the diversity of the samples.

Raising num_generations_per_prompt to 16 gives enough within-group diversity.

2. Fix DSV3 tokenizer

The Moonlight tokenizer fix in #3116 breaks DSV3. Narrow the scope of the Moonlight patch so it only kicks in when the default AutoTokenizer.from_pretrained path fails.

Tests

Validated below tests passed.

• dapo-deepseek-v3-64n8g.v2
• grpo-moonlight-16b-automodel-1n8g-ep8

Note

dapo-deepseek-v3-64n8g.v2 is fixed in r0.7.0 branch, but will easily meet the below error in main branch. tracked in #3155.

(MegatronPolicyWorker pid=1391870) [2026-07-10 06:19:36] pool0-00008:1391870:1393239 [7] init.cc:904 NCCL WARN Duplicate GPU detected : rank 15 and rank 7 both on CUDA device e4000

[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.

[yuki-97]

/ok to test da580b9

[ZhiyuLi-Nvidia]

Thank you LGTM.