[perf] feat: Add Qwen3.5 MFU flops estimation

[liangxuZhang]

What does this PR do?

• Add MFU/FLOPs estimation for qwen3_5 and qwen3_5_moe.
• Account for Qwen3.5 hybrid attention:
  • full attention projection FLOPs, including 2x q_proj output for the output gate
  • GatedDeltaNet projection and recurrence FLOPs
• Account for dense/MoE MLPs, LM head, and Qwen3.5 ViT FLOPs.
• Add regression cases in tests/utils/test_flops_counter.py.

Add concise overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ...
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, veomni, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward, fully_async, one_step_off
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)
• If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

[CLAassistant]

All committers have signed the CLA.

[gemini-code-assist]

Code Review

This pull request introduces FLOPs estimation support for Qwen3.5 and Qwen3.5-MoE models, including specialized logic for hybrid attention layers and GatedDeltaNet recurrence. The implementation covers both text and vision components, with corresponding test cases added. Feedback from the review highlights the need to include embedding parameters in the dense FLOPs calculation for consistency with other models and to adjust the causal attention factor from 12 to 6 to prevent overestimating the quadratic term in full-attention layers.

[gemini-code-assist]

The current FLOPs estimation for the dense part only accounts for the LM head parameters, but it should also include the embedding parameters. In standard LLM architectures (like Qwen2/Llama), both the embedding and the LM head are counted. For consistency with other models in this file (e.g., _estimate_qwen2_flops), you should use 2 * vocab_size * hidden_size to account for both.

Suggested change

	lm_head_N = vocab_size * hidden_size
	dense_N_flops = 6 * (mlp_N + attn_linear_N + lm_head_N) * tokens_sum
	emd_and_lm_head_N = 2 * vocab_size * hidden_size
	dense_N_flops = 6 * (mlp_N + attn_linear_N + emd_and_lm_head_N) * tokens_sum

[liangxuZhang]

fixed

[gemini-code-assist]

The factor 12 used for attn_qkv_flops is typically used for non-causal attention (like in ViT). For causal attention in decoder-only LLMs, the quadratic term is halved due to the causal mask, resulting in a factor of 6 for the forward and backward passes combined (3 * 2 * 0.5 * 4). Using 12 here will overestimate the attention FLOPs for the full-attention layers by 2x.

Suggested change

	attn_qkv_flops = 12 * seqlen_square_sum * head_dim * num_attention_heads * num_full_attn_layers
	attn_qkv_flops = 6 * seqlen_square_sum * head_dim * num_attention_heads * num_full_attn_layers

[liangxuZhang]

fixed

[liangxuZhang]

@ISEEKYAN @ArronHZG PTAL at this PR, thanks!