[None][fix] Pass max_seq_len to the Qwen-VL attention helper in the MiniCPM-V 4.6 vision encoder

tensorrt_llm/_torch/models/modeling_minicpmv4_6.py

@@ -549,7 +549,9 @@ def _make_attn_metadata(self, seq_lens: List[int]) -> AttentionMetadata:
             max_num_tokens=sum(seq_lens) + 1,
             kv_cache_manager=None,
         )
-        return _prepare_qwen_vl_vision_attn_metadata(seq_lens, attn_metadata)
+        return _prepare_qwen_vl_vision_attn_metadata(
+            seq_lens, attn_metadata, max_seq_len=max(seq_lens)
+        )
 
     @staticmethod
     def _grid_seq_lens(target_sizes: torch.Tensor) -> List[int]:

tests/unittest/_torch/modeling/test_modeling_minicpmv4_6.py

@@ -26,13 +26,15 @@
   with ``None.itemsize``).
 * The self-contained ``MiniCPMV4_6VisionConfig`` window helpers.
 * The runtime ``transformers>=5.7.0`` guard used by the input processor.
+* The vision encoder's call into the shared Qwen-VL attention-metadata helper.
 
 A single ``transformers>=5.7.0``-gated test asserts the native config is present
 once the pin is bumped (at which point the local shim can be removed).
 """
 
 import copy
 import json
+from types import SimpleNamespace
 
 import pytest
 import torch
@@ -254,6 +256,26 @@ def test_passes_on_supported_transformers(self, monkeypatch, version):
         mod._ensure_transformers_supports_minicpmv4_6()
 
 
+# ---------------------------------------------------------------------------
+# Vision encoder attention metadata
+# ---------------------------------------------------------------------------
+def test_vision_attn_metadata_passes_max_seq_len(monkeypatch):
+    from tensorrt_llm._torch.models import modeling_minicpmv4_6 as mod
+
+    calls = []
+
+    def capture_prepare_attn_metadata(seq_lens, attn_metadata, *, max_seq_len):
+        calls.append((seq_lens, max_seq_len))
+        return attn_metadata
+
+    monkeypatch.setattr(mod, "_prepare_qwen_vl_vision_attn_metadata", capture_prepare_attn_metadata)
+    vision_model = SimpleNamespace(metadata_cls=lambda **kwargs: SimpleNamespace(**kwargs))
+
+    mod.MiniCPMV4_6VisionModel._make_attn_metadata(vision_model, [64, 16])
+
+    assert calls == [([64, 16], 64)]
+
+
 # ---------------------------------------------------------------------------
 # Native transformers config (only once the pin is bumped to >=5.7.0)
 # ---------------------------------------------------------------------------