Read rope_theta from rope_parameters across Inference V2

deepspeed/inference/v2/model_implementations/exaone4/model.py

@@ -94,8 +94,7 @@ def positional_embedding_type(self) -> PositionalEmbeddingType:
 
     @property
     def positional_embedding_config(self) -> Optional[RotateHalfConfig]:
-        rope_theta = getattr(self._config, "rope_theta", 1000000.0)
-        return RotateHalfConfig(theta_base=rope_theta)
+        return RotateHalfConfig(theta_base=self.rope_theta)
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)

deepspeed/inference/v2/model_implementations/inference_transformer_base.py

@@ -165,6 +165,26 @@ def positional_embedding_config(self) -> Optional[RotateHalfConfig]:
     Derived helpers
     """
 
+    @property
+    def rope_theta(self) -> float:
+        """The rotary base, read from wherever the installed transformers keeps it.
+
+        transformers 5.0 folded the rotary settings into ``config.rope_parameters`` and
+        dropped the ``rope_theta`` attribute, so reading the attribute alone raises
+        against a stock config on 5.x. ``exaone4_5`` already reads both spellings; this
+        is the same lookup for every model that reaches it through this base.
+        """
+        theta = getattr(self._config, "rope_theta", None)
+        if theta is not None:
+            return theta
+
+        rope_parameters = getattr(self._config, "rope_parameters", None) or getattr(self._config, "rope_scaling", None)
+        if isinstance(rope_parameters, dict) and rope_parameters.get("rope_theta") is not None:
+            return rope_parameters["rope_theta"]
+
+        raise ValueError(f"{type(self._config).__name__} carries no rope_theta, either as an "
+                         "attribute or in rope_parameters/rope_scaling.")
+
     @cached_property
     def n_heads_q_local(self) -> int:
         """

deepspeed/inference/v2/model_implementations/llama_v2/model.py

@@ -107,7 +107,7 @@ def positional_embedding_type(self) -> PositionalEmbeddingType:
 
     @property
     def positional_embedding_config(self) -> Optional[RotateHalfConfig]:
-        return RotateHalfConfig(theta_base=self._config.rope_theta)
+        return RotateHalfConfig(theta_base=self.rope_theta)
 
     """
     Forward implementations

deepspeed/inference/v2/model_implementations/mistral/model.py

@@ -106,7 +106,7 @@ def positional_embedding_type(self) -> PositionalEmbeddingType:
 
     @property
     def positional_embedding_config(self) -> Optional[RotateHalfConfig]:
-        return RotateHalfConfig(theta_base=self._config.rope_theta)
+        return RotateHalfConfig(theta_base=self.rope_theta)
 
     """
     Forward implementations

deepspeed/inference/v2/model_implementations/mixtral/model.py

@@ -114,7 +114,7 @@ def positional_embedding_config(self) -> Optional[RotateHalfConfig]:
         """
         The positional embedding configuration for the model.
         """
-        return RotateHalfConfig(theta_base=self._config.rope_theta)
+        return RotateHalfConfig(theta_base=self.rope_theta)
 
     """
     Inherited from `DSMoETransformerModelBase`

deepspeed/inference/v2/model_implementations/phi/model.py

@@ -97,7 +97,7 @@ def positional_embedding_type(self) -> PositionalEmbeddingType:
     @property
     def positional_embedding_config(self) -> Optional[RotateHalfConfig]:
         rotary_dim = int(self._config.partial_rotary_factor * self.head_size)
-        return RotateHalfConfig(rotate_dim=rotary_dim, theta_base=self._config.rope_theta)
+        return RotateHalfConfig(rotate_dim=rotary_dim, theta_base=self.rope_theta)
 
     """
     Forward implementations

deepspeed/inference/v2/model_implementations/phi3/model.py

@@ -106,7 +106,7 @@ def positional_embedding_type(self) -> PositionalEmbeddingType:
 
     @property
     def positional_embedding_config(self) -> Optional[RotateHalfConfig]:
-        return RotateHalfConfig(theta_base=self._config.rope_theta)
+        return RotateHalfConfig(theta_base=self.rope_theta)
 
     """
     Forward implementations

deepspeed/inference/v2/model_implementations/qwen_v2/model.py

@@ -100,7 +100,7 @@ def positional_embedding_type(self) -> PositionalEmbeddingType:
 
     @property
     def positional_embedding_config(self) -> Optional[RotateHalfConfig]:
-        return RotateHalfConfig(theta_base=self._config.rope_theta)
+        return RotateHalfConfig(theta_base=self.rope_theta)
 
     def make_norm_layer(self) -> None:
         """

deepspeed/inference/v2/model_implementations/qwen_v2_moe/model.py

@@ -105,7 +105,7 @@ def positional_embedding_type(self) -> PositionalEmbeddingType:
 
     @property
     def positional_embedding_config(self) -> Optional[RotateHalfConfig]:
-        return RotateHalfConfig(theta_base=self._config.rope_theta)
+        return RotateHalfConfig(theta_base=self.rope_theta)
 
     """
     Inherited from `DSMoETransformerModelBase`

tests/unit/inference/v2/model_implementations/test_rope_theta.py

@@ -0,0 +1,80 @@
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+"""Inference V2 must find rope_theta wherever the installed transformers keeps it.
+
+transformers 5.0 folded the rotary settings into ``config.rope_parameters`` and dropped
+the ``rope_theta`` attribute. Models that read the attribute directly raise against a
+stock config on 5.x; ``exaone4`` read it through a ``getattr`` default and silently used
+that default instead of the value the config actually carries.
+"""
+
+from types import SimpleNamespace
+
+import pytest
+
+from deepspeed.inference.v2.model_implementations.inference_transformer_base import DSTransformerModelBase
+
+
+class _Model:
+    """Minimal stand-in that borrows the property under test."""
+
+    rope_theta = DSTransformerModelBase.rope_theta
+
+    def __init__(self, config):
+        self._config = config
+
+
+def test_reads_the_legacy_attribute():
+    assert _Model(SimpleNamespace(rope_theta=500000.0)).rope_theta == 500000.0
+
+
+def test_reads_rope_parameters_when_the_attribute_is_gone():
+    config = SimpleNamespace(rope_parameters={"rope_theta": 10000.0, "rope_type": "default"})
+
+    assert _Model(config).rope_theta == 10000.0
+
+
+def test_reads_rope_scaling_when_that_is_the_only_dict():
+    config = SimpleNamespace(rope_scaling={"rope_theta": 1000000.0, "rope_type": "default"})
+
+    assert _Model(config).rope_theta == 1000000.0
+
+
+def test_prefers_the_attribute_when_both_are_present():
+    config = SimpleNamespace(rope_theta=500000.0, rope_parameters={"rope_theta": 10000.0})
+
+    assert _Model(config).rope_theta == 500000.0
+
+
+def test_raises_instead_of_guessing_when_nothing_carries_it():
+    # exaone4 used to fall back to 1e6 here, which is a silent 100x error against a
+    # config whose real base is 1e4.
+    with pytest.raises(ValueError, match="rope_theta"):
+        _ = _Model(SimpleNamespace()).rope_theta
+
+
+@pytest.mark.parametrize(
+    "module_name, config_name",
+    [
+        ("llama", "LlamaConfig"),
+        ("mistral", "MistralConfig"),
+        ("mixtral", "MixtralConfig"),
+        ("phi", "PhiConfig"),
+        ("phi3", "Phi3Config"),
+        ("qwen2", "Qwen2Config"),
+        ("qwen2_moe", "Qwen2MoeConfig"),
+        ("exaone4", "Exaone4Config"),
+    ],
+)
+def test_resolves_against_the_installed_transformers_configs(module_name, config_name):
+    """Every config backing a V2 model must yield a base through one spelling or the other."""
+    importlib = pytest.importorskip("importlib")
+    try:
+        module = importlib.import_module(f"transformers.models.{module_name}.configuration_{module_name}")
+        config = getattr(module, config_name)()
+    except (ImportError, AttributeError):
+        pytest.skip(f"{config_name} is not available in the installed transformers")
+
+    assert _Model(config).rope_theta > 0