perf: Performance script tuning

[guyueh1]

What does this PR do ?

Performance script tuning

1. use force_on_policy_ratio to skip prev_logprobs compute in dsv3 and qwen3-235b
2. use fuse loss in qwen3-235b
3. use flashinfer flags for dsv3 fp8

Issues

List issues that this PR closes (syntax):

Usage

• You can potentially add a usage example below

# Add a code snippet demonstrating how to use this

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you run the unit tests and functional tests locally? Visit our Testing Guide for how to run tests
• Did you add or update any necessary documentation? Visit our Document Development Guide for how to write, build and test the docs.

Additional Information

• ...

[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.

[youngeunkwon0405]

Hi @guyueh1, can I ask what the condition for enabling the fuse_loss is? I am curious why it is only enabled in the qwen3-235b model, but not for all models.

[guyueh1]

Hi @guyueh1, can I ask what the condition for enabling the fuse_loss is? I am curious why it is only enabled in the qwen3-235b model, but not for all models.

It is generally applicable to all cases when sequence_packing is enabled; it was already enabled in deepseek previously, I didn't add it for dense models since our release tracker stopped tracking them; But I should probably add for 30B too

[youngeunkwon0405]

Hi @guyueh1, can I ask what the condition for enabling the fuse_loss is? I am curious why it is only enabled in the qwen3-235b model, but not for all models.

It is generally applicable to all cases when sequence_packing is enabled; it was already enabled in deepseek previously, I didn't add it for dense models since our release tracker stopped tracking them; But I should probably add for 30B too

Thank you for the clarification.

[guyueh1]

/ok to test 879d913

[guyueh1]

@youngeunkwon0405 added them for the Qwen30B and a DAPO deepseek-v3;
I suggest we merge this once the fast CI is passing; it's expensive for me to run perf regression myself, the impact should be reflected in the next weekly perf CI.

[guyueh1]

/ok to test 879d913

[guyueh1]

/ok to test 8730bfe

[guyueh1]

/ok to test ca1755a

[guyueh1]

/ok to test 5e52cf9

[guyueh1]

/ok to test 27c2c8d

[guyueh1]

@youngeunkwon0405 could you approve?