fix(glm_image): build the target-shape scaffold without HF's private formatter

Nightly 3034 (Simple · Model Executor Test) failed all six
test_generation_prompt_ids_preserve_hf_target_scaffold cases with

    AttributeError: type object 'GlmImageProcessor' has no attribute '_build_prompt_with_target_shape'

The 0.29 rewrite of GlmImageMultiModalProcessor.apply reached for
GlmImageProcessor._build_prompt_with_target_shape, a private transformers
method. The test suite stubs transformers.models.glm_image.processing_glm_image
with a bare class to import glm_image_ar without the real package, so in the
full CPU run the private attribute is not there; the same dependency would
break at runtime on any transformers build that renames or removes it.

Format the scaffold in Omni instead: _build_target_shape_scaffold mirrors the
HF formula (identical in transformers 5.13 and 5.14: target grid, plus the
preview grid for text-to-image, then the image bos token) using only the
processor's public grid_bos_token / grid_eos_token / bos_token attributes.
The test keeps its hard-coded expected suffix, which is what HF produces for
512x768, and no longer imports the HF class.

Validation: tests/model_executor/models/glm_image passes (25 tests) both alone
and inside the full `tests/model_executor -m "core_model and cpu"` run.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MpXe2LeRrUQGjZfKFaocMZ
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>

tests/model_executor/models/glm_image/test_glm_image_ar.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Unit tests for GLM-Image AR model: DataParser, processor, and M-RoPE."""
 
 import importlib.util
@@ -94,18 +94,15 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 @pytest.mark.parametrize("num_images", [0, 1, 2])
 @pytest.mark.parametrize("preformatted", [False, True])
 def test_generation_prompt_ids_preserve_hf_target_scaffold(num_images, preformatted):
-    from functools import partial
     from types import SimpleNamespace
 
-    from transformers.models.glm_image.processing_glm_image import GlmImageProcessor
     from vllm.multimodal.parse import MultiModalDataItems
     from vllm.multimodal.processing import BaseMultiModalProcessor, ProcessorInputs
 
     processor = object.__new__(GlmImageMultiModalProcessor)
-    symbols = SimpleNamespace(grid_bos_token="<sop>", grid_eos_token="<eop>", bos_token="<image_start>")
-    hf_processor = SimpleNamespace(
-        _build_prompt_with_target_shape=partial(GlmImageProcessor._build_prompt_with_target_shape, symbols)
-    )
+    # Only the public token attributes of GlmImageProcessor are needed; the
+    # expected suffix below is what HF's own formatter produces for 512x768.
+    hf_processor = SimpleNamespace(grid_bos_token="<sop>", grid_eos_token="<eop>", bos_token="<image_start>")
     tokenizer = SimpleNamespace(encode=lambda text, **kwargs: list(text.encode()))
     processor.info = SimpleNamespace(
         get_tokenizer=lambda: tokenizer,

vllm_omni/model_executor/models/glm_image/glm_image_ar.py

@@ -341,6 +341,30 @@ class GlmImageDummyInputsBuilder(BaseDummyInputsBuilder[GlmImageProcessingInfo])
         }
 
 
+_GLM_IMAGE_GRID_FACTOR = 32
+
+
+def _build_target_shape_scaffold(processor, *, height: int, width: int, is_text_to_image: bool) -> str:
+    """Return the grid scaffold HF appends to a GLM-Image generation prompt.
+
+    Mirrors ``GlmImageProcessor._build_prompt_with_target_shape`` (identical in
+    transformers 5.13 and 5.14) using only the processor's public token
+    attributes, so this does not depend on a private HF method that a stub or
+    a future release may not provide.
+    """
+    factor = _GLM_IMAGE_GRID_FACTOR
+    token_h = (height // factor * factor) // factor
+    token_w = (width // factor * factor) // factor
+    grid_bos, grid_eos, bos = processor.grid_bos_token, processor.grid_eos_token, processor.bos_token
+    scaffold = f"{grid_bos}{token_h} {token_w}{grid_eos}"
+    if is_text_to_image:
+        ratio = token_h / token_w
+        prev_token_h = int(math.sqrt(ratio) * (factor // 2))
+        prev_token_w = int(math.sqrt(1 / ratio) * (factor // 2))
+        scaffold += f"{grid_bos}{prev_token_h} {prev_token_w}{grid_eos}"
+    return scaffold + bos
+
+
 class GlmImageMultiModalProcessor(BaseMultiModalProcessor[GlmImageProcessingInfo]):
     """
     Multimodal processor for GLM-Image.
@@ -374,12 +398,13 @@ class GlmImageMultiModalProcessor(BaseMultiModalProcessor[GlmImageProcessingInfo
 
         processor = self.info.get_hf_processor()
         tokenizer = self.info.get_tokenizer()
-        # Reuse HF's shape formatter without decoding/re-tokenizing the user's
-        # tokens. Text-to-image needs both target and preview grids; i2i needs
-        # only the target. Grid metadata alone does not supply these AR tokens.
+        # Append HF's target-shape scaffold without decoding/re-tokenizing the
+        # user's tokens. Text-to-image needs both target and preview grids; i2i
+        # needs only the target. Grid metadata alone does not supply these AR
+        # tokens.
         target_grid = self._build_generation_grids(inputs.hf_processor_mm_kwargs)[0]
-        suffix, *_ = processor._build_prompt_with_target_shape(
-            "",
+        suffix = _build_target_shape_scaffold(
+            processor,
             height=int(target_grid[1]) * 32,
             width=int(target_grid[2]) * 32,
             is_text_to_image=num_images == 0,