[Fork][Draft] Integration stack 20260824: rebuilt FlashInfer mm-prefix/NVFP4/workspace + fixes

tests/config/test_mrope_cache_bound_config.py

@@ -0,0 +1,103 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""``ModelConfig.mrope_cache_max_position`` resolution.
+
+The bound is only handed to the RoPE layer when the concrete model class
+certifies its unpruned positions and no config-dependent transform can push a
+position past the sequence length. It also participates in the model
+compilation hash, so a compiled artifact built with one bound is never reused
+for another.
+"""
+
+import types
+
+import pytest
+
+from vllm.config.model import ModelConfig
+
+
+def _model_config(certified: bool, pruning_rate: float | None) -> ModelConfig:
+    """Build the smallest object that exercises the property's branches."""
+    config = ModelConfig.__new__(ModelConfig)
+    object.__setattr__(
+        config,
+        "_model_info",
+        types.SimpleNamespace(mrope_positions_are_sequence_bounded=certified),
+    )
+    if pruning_rate is None:
+        multimodal_config = None
+    else:
+        multimodal_config = types.SimpleNamespace(
+            is_multimodal_pruning_enabled=lambda: pruning_rate > 0
+        )
+    object.__setattr__(config, "multimodal_config", multimodal_config)
+    object.__setattr__(config, "max_model_len", 262144)
+    return config
+
+
+def test_certified_model_without_pruning_gets_the_sequence_bound():
+    config = _model_config(certified=True, pruning_rate=None)
+
+    assert config.mrope_cache_max_position == 262144
+
+
+def test_uncertified_model_keeps_the_legacy_cache():
+    config = _model_config(certified=False, pruning_rate=None)
+
+    assert config.mrope_cache_max_position is None
+
+
+@pytest.mark.parametrize("pruning_rate", [0.0, 0.5])
+def test_pruning_falls_back_to_the_legacy_cache(pruning_rate):
+    """Pruned positions can exceed the reduced sequence length."""
+    config = _model_config(certified=True, pruning_rate=pruning_rate)
+
+    expected = 262144 if pruning_rate == 0.0 else None
+    assert config.mrope_cache_max_position == expected
+
+
+def test_uncertified_model_ignores_pruning_state():
+    config = _model_config(certified=False, pruning_rate=0.5)
+
+    assert config.mrope_cache_max_position is None
+
+
+def test_bound_participates_in_the_model_hash(monkeypatch):
+    """A compiled artifact built for one bound must not be reused for another."""
+    import vllm.config.utils as config_utils
+
+    monkeypatch.setattr(
+        config_utils,
+        "get_hash_factors",
+        lambda _config, _ignored: {"base": "unchanged"},
+    )
+    monkeypatch.setattr(config_utils, "hash_factors", lambda factors: repr(factors))
+
+    unbounded = _model_config(certified=False, pruning_rate=None)
+    bounded = _model_config(certified=True, pruning_rate=None)
+
+    unbounded_hash = unbounded.compute_hash()
+    bounded_hash = bounded.compute_hash()
+
+    assert "mrope_cache_max_position" in unbounded_hash
+    assert unbounded_hash != bounded_hash
+
+
+def test_get_rope_accepts_the_bound_from_model_config():
+    """The value handed to the RoPE factory is the property, unmodified.
+
+    ``Qwen3NextAttention`` builds its projections before calling ``get_rope``,
+    so constructing it needs a distributed environment; the hand-off is pinned
+    here at the boundary instead, and `test_mrope_cache_bounds.py` covers the
+    factory's own behaviour for both values.
+    """
+    import inspect
+
+    from vllm.model_executor.layers.rotary_embedding import get_rope
+
+    parameter = inspect.signature(get_rope).parameters["mrope_cache_max_position"]
+
+    assert parameter.default is None
+    assert _model_config(
+        certified=True, pruning_rate=None
+    ).mrope_cache_max_position == (262144)