gpt-oss sft megatron does not support sequence packing

[jordane95]

Describe the bug

When set sequence_packing=true and use Megatron backend for fine-tuning gpt-oss. The debug info says no attention backend available.
If sequence_packing=False, only UnfusedAttention is available, FlashAttention disabled for softmax_type = learnable, FusedAttention disabled as no backend supports the provided input.

This is different from the fine-tuning in megatron-bridge, which support FusedAttention with learnable softmax.

Update: The different backends enabled are due to different versions of cuDNN in docker image. Latest nemo-rl nano image has cuDNN version = 91002, but in nemo image cuDNN version is 91310.

Steps/Code to reproduce bug

Please list minimal steps or code snippet for us to be able to reproduce the bug.

A helpful guide on on how to craft a minimal bug report http://matthewrocklin.com/blog/work/2018/02/28/minimal-bug-reports.

Expected behavior

A clear and concise description of what you expected to happen.

Additional context

Add any other context about the problem here.

[coderabbitai]

📝 CodeRabbit Plan Mode

Generate an implementation plan and prompts that you can use with your favorite coding agent.

• Create Plan

Examples

• Example 1
• Example 2

---

🔗 Related PRs

#809 - fix: crash when sequence packing is enabled for gemma 1b. [merged]

👤 Suggested Assignees

• joyang-nv

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord or schedule a call!

[guyueh1]

@jordane95 Yes we are aware that the latest nemo-rl container can't use fused attention for gpt-oss because the cuDNN version is too old; We have a way to manually bump the cudnn version in the container so that you can use fused attention, it incurs no change to nemo-rl code but you need to install a newer cuDNN and properly expose it to LD library path. @seonjinn can share the details.

If you don't like the workaround, you can expect the newer cuDNN version will come when we bump to a newer torch version, but that might take a while.

These are all for non sequence-packing case. For sequence packing case, we don't have fused attention kernels in cudnn, so even if you bump cudnn to the latest it will still use unfused attention.