[model]{bug-fix}: Fix wrong log prob calculations using fused kernels and USP

[joshyan1]

What does this PR do?

This PR introduces a fix to incorrect log prob calculations when using USP and fused kernels. This issue arises since we slice inputs with USP within the model when using a VLM with this patch here

def _create_ulysses_wrapped_decoder_forward(original_forward):
        def ulysses_wrapped_decoder_forward(self, *args, **kwargs):
            inputs_embeds = kwargs.get("inputs_embeds")
            position_ids = kwargs.get("position_ids")
            call_kwargs = kwargs.copy()

            current_ulysses_sp_size = get_ulysses_sequence_parallel_world_size()

            slice_now = (
                inputs_embeds is not None
                and current_ulysses_sp_size > 1
                and getattr(self, "_needs_initial_slice", True)
            )
            if slice_now:
                call_kwargs["inputs_embeds"] = slice_input_tensor(inputs_embeds, dim=1, padding=False)
                call_kwargs["position_ids"] = slice_input_tensor(position_ids, dim=-1, padding=False)
                self._needs_initial_slice = False
            try:
                return original_forward(self, *args, **call_kwargs)
            finally:
                if slice_now:
                    self._needs_initial_slice = True

        return ulysses_wrapped_decoder_forward

    original_forward = model_class.forward
    wrapped_forward = _create_ulysses_wrapped_decoder_forward(original_forward)

Later, when computing log probs, the rolled labels are created with the original input_ids, causing a mismatch for ranks using inputs that are not at the beginning of the sequence

def forward_with_torch_backend(
    self: Qwen2VLForConditionalGeneration,
    input_ids: torch.LongTensor = None,
    labels: Optional[torch.LongTensor] = None,
    temperature: float = 1.0,
    **kwargs,
) -> tuple | Qwen2VLCausalLMOutputForPPO:
    from verl.utils.experimental.torch_functional import FusedLinearForPPO

    outputs = qwen2_vl_forward(self, input_ids, **kwargs)
    hidden_states = outputs[0]

    # Loss calculations
    if labels is not None:
        rolled_labels = torch.roll(labels, shifts=-1, dims=-1)
    elif input_ids is not None:
        rolled_labels = torch.roll(input_ids, shifts=-1, dims=-1)
    else:
        raise RuntimeError("To use forward_with_torch_backend, either labels or input_ids must be provided.")

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ...
  https://github.com/volcengine/verl/pulls?q=is:pr+is:open+log+probs+
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data
  • 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).)

[gemini-code-assist]

Code Review

This pull request introduces logic to slice labels for Ulysses Sequence Parallelism (USP) within the forward_with_torch_backend and forward_with_triton_backend methods. While the functionality is correct, the implementation introduces code duplication, as the same block of code is added to both methods. My review focuses on improving maintainability by suggesting this duplicated logic be refactored into a single helper function. I've also pointed out a minor inefficiency in the new code and provided a direct suggestion to fix it.