fix(qwen3.5): freeze vision modules in text-only recipes

[seonjinn]

Problem

Qwen3.5 text-only Megatron recipes instantiate the vision modules but did not freeze them in the shared base recipe. Text batches never execute those modules, so DDP saw only 307 of 640 parameters in one gradient bucket and failed when the next training step reset the bucket.

Change

Freeze the vision model and projection in the shared Qwen3.5 text recipe before DDP and optimizer setup. BF16, FP8, and MXFP8 child recipes now inherit the same trainable parameter set. Remove the duplicate FP8-only settings.

Test

• Regression test before the fix: 1 failed, 2 passed (freeze_config missing)
• GB200 regression test after the fix: 4 passed
• Qwen3.5 Async NCCL Reshard, MXFP8 training (fp8_param=true) + MXFP8 rollout, FlashInfer TRTLLM: 2/2 steps completed; generation KL error was 0.0012 at both steps
• 20-step follow-up is queued

W&B: https://wandb.ai/nvidia/nemo-rl-mxfp8-training/runs/oz68ga7p

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.