[cuichenx]

Summary

• add explicit GLM-4.5V assistant and next-role boundaries for loss-mask construction
• trim only the template's empty thinking prefix while preserving real reasoning and tool-call content
• delimit the final assistant span with a temporary EOS sentinel that is removed before the mask is returned

NVBug 6468521

Root cause

GLM-4.5V's chat template has no Jinja generation block and ends its final assistant turn without a terminator. The boundary-based assistant-mask path therefore had neither a Hugging Face assistant mask nor a known marker configuration and failed on the first collated batch.

Validation

• reproduced the failure with the stock 26.08 RC2 production collator on CORD-v2
• verified the fixed collator on the same input: shape (1, 2227) and 786 assistant-loss tokens, matching the historical 26.06 control
• uv run python -m pytest tests/unit_tests/data/collators/test_model_collators.py tests/unit_tests/data/test_conversation_processing.py — 131 passed
• uv run pre-commit run --all-files
• independent review completed on the exact final signed head

This PR intentionally targets main only. The 26.08 release branch has not been cut, so no mapped release label or cherry-pick PR is available.

AI assistance

This PR was authored by an AI coding agent under user direction and independently reviewed by a separate AI agent.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[claude]

LGTM. GLM-4.5V-specific assistant loss masking is a focused, correct change.

Correctness

• The virtual-EOS pattern in _build_glm4v_assistant_loss_mask is sound: EOS is appended to delimit the final assistant span, then the trailing element is trimmed so it is never returned or trained. The ValueError guard when eos_token_id is None is the right fail-fast behavior.
• The old infer_assistant_mask_boundary_config import is fully removed with no dangling references in src or tests.
• The trim_leading_token_sequences handling for the empty think prefix correctly guards against an empty tokenization result before wrapping in a tuple.

Tests: the new parametrized test exercises multi-turn masking, a sequence already ending in EOS, and an observation-role boundary, and the monkeypatch targets in the existing packing test were updated to the new private helper names.

Suggested test cases

• test_glm4v_assistant_mask_uses_role_boundaries_and_virtual_final_terminator
• test_glm4v_collate_packs_mm_token_type_ids_and_restores_padding

No perf tests impacted.