[GRPO] Fix re-tokenization bug in tool-calling loop by concatenating token IDs

[qgallouedec]

Context

Part of the series to fix the re-tokenization bug in GRPO multi-turn tool calling (see #5224).

closes #5224
closes #5144

When the model generates a completion in a tool-calling loop, the decoded text is re-tokenized via apply_chat_template, which can produce different token IDs due to BPE merge ambiguities. To fix this, we need a token-in / token-out pipeline: tokenize once, then pass raw token IDs through every subsequent generation call — never decoding and re-tokenizing.

• Add support for raw ids in prompts in vLLM client and server #5225
• Add VLM support when passing raw token IDs to vLLM client #5227
• Move rollout_func from _generate_single_turn to _generate #5232
• [GRPO/RLOO] Tokenize before vLLM generation call #5238
• [GRPO/RLOO] Unify tokenization across all generation backends in _generate_single_turn #5239
• [GRPO/RLOO] Extract tokenize prompts from _generate_single_turn #5240
• [this one]

This is the final PR in the series. It eliminates the re-tokenization in the tool-calling loop — the actual source of the bug.

Changes

• New _get_tool_suffix_ids(tool_messages) method: Tokenizes only the tool result portion by diffing a minimal dummy conversation (2 messages vs 3 messages). This avoids re-tokenizing the full conversation history.
• _tool_call_loop: Instead of re-tokenizing prompt + completion + tool_results via apply_chat_template, builds the token sequence by concatenation: prompt_ids + completion_ids + tool_suffix_ids. The original prompt and completion token IDs are preserved exactly as they were — only the new tool result tokens are freshly tokenized.
• Removed the prefix-preserving sanity check (no longer needed since the prefix is preserved by construction).
• Removed the _tokenize_prompts call in the tool loop.

The bug and the fix

Previously, after a tool call:

1. The completion was decoded to text and appended as an assistant message
2. The full prompt + assistant + tool_results was re-tokenized via apply_chat_template
3. Due to BPE merge ambiguity, step 2 could produce different token IDs for the completion part

Now:

1. The original prompt_ids and completion_ids are kept as-is (never decoded and re-tokenized)
2. Only the tool result suffix is tokenized, using a minimal dummy conversation to extract just the template formatting
3. The full prompt is built by concatenation: prompt_ids + completion_ids + suffix_ids

Backward compatibility

No user-facing API changes. _get_tool_suffix_ids and _tool_call_loop are internal methods.

---

Note

Medium Risk
Touches GRPO/RLOO generation paths and tool-calling control flow; mistakes could change generated token sequences or break tool-loop behavior, impacting training stability.

Overview
Fixes GRPO multi-turn tool calling to be token-in/token-out. The tool loop no longer decodes and re-tokenizes prompt + completion + tool_results; it now concatenates existing prompt_ids/completion_ids with freshly tokenized tool-result suffix IDs to preserve the exact original completion tokens.

Adds internal helper _get_tool_suffix_ids() to tokenize only the tool-result formatting, updates _tool_call_loop() to carry images/multimodal_fields through regeneration without re-running _tokenize_prompts, and simplifies _generate_single_turn() return values (dropping redundant prompt_ids returns) in both grpo_trainer.py and rloo_trainer.py.

^{Written by Cursor Bugbot for commit 10708ca. This will update automatically on new commits. Configure here.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3375aeac6c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[qgallouedec]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 367a79ebc6

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".