aye carumba

Dockerfile

@@ -34,3 +34,6 @@ COPY minimax_tool_parser.py /usr/local/lib/python3.12/dist-packages/vllm/tool_pa
 # Copy over minimax parsers with kwargs fixes
 COPY minimax_tool_parser.py /usr/local/lib/python3.12/dist-packages/vllm/tool_parsers/minimax_tool_parser.py
 COPY minimax_m2_parser.py /usr/local/lib/python3.12/dist-packages/vllm/parser/minimax_m2_parser.py
+
+# Fix DeepseekV4Config missing max_position_embeddings attr for RoPE standardization
+COPY deepseek_v4.py /usr/local/lib/python3.12/dist-packages/vllm/transformers_utils/configs/deepseek_v4.py

deepseek_v4.py

@@ -0,0 +1,17 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from transformers import PretrainedConfig
+
+
+class DeepseekV4Config(PretrainedConfig):
+    model_type = "deepseek_v4"
+
+    def __init__(
+        self,
+        max_position_embeddings: int = 163840,
+        rope_scaling: dict | None = None,
+        **kwargs,
+    ):
+        self.max_position_embeddings = max_position_embeddings
+        self.rope_scaling = rope_scaling
+        super().__init__(**kwargs)