[alanhuangyoo]

What does this PR do?

GLM-4V training dies at the first attention layer on every transformers version this repo
supports.

glm4v_attn_forward replaces Glm4vTextAttention.forward whenever use_remove_padding or
ulysses_sp_size > 1 (monkey_patch.py:473), so this is the ordinary GLM-4V path, not a corner:

from transformers.models.glm4v.modeling_glm4v import apply_multimodal_rotary_pos_emb, repeat_kv

transformers 5.3.0
ImportError: cannot import name 'apply_multimodal_rotary_pos_emb'
             from 'transformers.models.glm4v.modeling_glm4v'

transformers 5.0 replaced apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section) with
apply_rotary_pos_emb(q, k, cos, sin) and moved the sectioning into
Glm4vTextRotaryEmbedding, which reads mrope_section from config.rope_parameters and folds
it into cos/sin before the attention ever sees them. Glm4vTextAttention.forward upstream is
now just:

cos, sin = position_embeddings
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

Absent in every supported version — requirements.txt asks for
transformers>=5.5.3,!=5.6.0,<5.11, and the extras pin 5.3.0 / 5.5.3:

transformers	apply_multimodal_rotary_pos_emb in modeling_glm4v
4.57.1	yes
5.0.0, 5.3.0, 5.5.3, 5.10.4, 5.16.1	no

It came in with #5724 (transformers upgrade to 5.3.0), which updated the config-side
rope_scaling → rope_parameters migration a few lines above but not the function that moved
with it.

qwen2_vl.py uses the same call and is not affected: modeling_qwen2_vl still exports
apply_multimodal_rotary_pos_emb. Only GLM-4V was migrated.

The change

Call the function upstream calls, and drop the mrope_section lookup that fed the old one.

That lookup was dead twice over — it reads self.rope_scaling or self.rope_parameter off the
attention module, and on transformers 5 the attention module has neither (the config
attribute is rope_parameters, plural, and lives on the config):

transformers 5.3.0, Glm4vTextAttention
  self.rope_scaling    -> <ABSENT>
  self.rope_parameter  -> <ABSENT>

so the assert under it would have fired had the import ever got that far.

Restoring the old function instead would be wrong, not just unavailable: the 4.x call expects
cos/sin with a leading 3-section dim, and applying it to what transformers 5 produces raises
IndexError: index 1 is out of bounds for dimension 0 with size 1. The sectioning genuinely
moved.

Checklist Before Starting

• Search for similar PRs. No PR or issue mentions apply_multimodal_rotary_pos_emb on this
  repo; the one hit, RuntimeError: Tensor size mismatch on apply_multimodal_rotary_pos_emb with ulysses_sequence_parallel_size=2 #2603, predates the 5.x upgrade and is closed.
• Format the PR title as [{modules}] {type}: {description}.

Test

tests/models/test_glm4v_rotary_on_cpu.py, following the on_cpu.py convention so it runs in
cpu_unit_tests.yml. It builds a real Glm4vTextAttention and Glm4vTextRotaryEmbedding, runs
glm4v_attn_forward with the CUDA-only flash-attention call patched out, and asserts the q/k
reaching attention equal apply_rotary_pos_emb(q, k, cos, sin) — the same rotary upstream
applies — with KV repeated to the query head count. A second case records why the argument had
to go rather than be restored.

Load-bearing, run against a clean origin/main worktree in the same environment:

$ pytest tests/models/test_glm4v_rotary_on_cpu.py          # origin/main
>       from transformers.models.glm4v.modeling_glm4v import apply_multimodal_rotary_pos_emb, repeat_kv
E       ImportError: cannot import name 'apply_multimodal_rotary_pos_emb' ...
1 failed, 1 passed

$ pytest tests/models/test_glm4v_rotary_on_cpu.py          # this branch
2 passed

$ ruff check / ruff format --check                          # ruff 0.12.2, as pinned in .pre-commit-config.yaml
All checks passed! / 2 files already formatted

$ python3 tests/special_sanity/check_license.py --directories .
exit 0

The flash-attention call itself is CUDA-only and unchanged by this PR, so it stays covered by
the existing GPU tests.

API and Usage Example

No API change.

Design & Code Changes

One import and one call site in verl/models/transformers/glm4v.py; the mrope_section
derivation is removed with them.

Checklist Before Submitting

• Read the Contribute Guide.
• Apply pre-commit checks.
• Add / update the documentation. Not applicable.
• Add unit or end-to-end test(s).

[wuxibin89]

CI not needed.

[alanhuangyoo]

Removed. The import and rotary-call fix in glm4v.py stands on its own — apply_multimodal_rotary_pos_emb does not exist in any transformers this repo supports, so the GLM-4V path raises ImportError at the first attention layer with remove_padding or ulysses_sp_size > 1.

[alanhuangyoo]

The vllm / vllm_ascend reds here are not from this PR — it touches one file,
verl/models/transformers/glm4v.py, and neither job exercises it.

tests/workers/rollout/rollout_vllm/test_vllm_abort.py::test_abort_all_requests_propagates_pause_failure
    AttributeError: 'vLLMHttpServer' object has no attribute '_resume_event'
    verl/workers/rollout/vllm_rollout/vllm_async_server.py:942

That test builds the server with vLLMHttpServer.__new__(vLLMHttpServer), so __init__ never
runs and neither _submission_paused nor _resume_event is set. abort_all_requests started
touching both in #7511 ("gate rollout submission during weight-sync drain for DP>1"), which
lands before the pause_generation call the test is actually asserting on. The test file has
not been updated since.

Happy to send that as a separate one-liner if it is not already being handled.