[Model][MiniCPM] support MiniCPM

Signed-off-by: MengqingCao <[email protected]>

Author: MengqingCao

vllm_ascend/patch/__init__.py

@@ -125,6 +125,20 @@
 #    Future Plan:
 #       Revert it when the related pr is merged in vllm.
 #
+# ** File: worker/patch_common/patch_minicpm.py **
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+#   1. `vllm.model_executor.models.minicpm.MiniCPMAttention.forward`
+#    Why:
+#       The forward func of MiniCPMAttention in vllm do a datatype convert
+#       (original datatype --> float32) to ensure the precision on cuda.
+#       However float32 is not supported in cann rope op, thus we keep this patch
+#    How：
+#       Removed the dtype convert operations in forward
+#    Related PR (if no, explain why): 1. refused by vllm. 2. vllm doesn't support 3. prepare to submit....
+#       NO, only for npu due to rope op.
+#    Future Plan:
+#       Keep this patch in vllm-ascend.
+#
 # ** File: worker/patch_common/patch_multi_step_worker.py **
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 #   1. `vllm.spec_decode.multi_step_worker.MultiStepWorker.sampler_output`

vllm_ascend/patch/worker/patch_common/__init__.py

@@ -16,5 +16,6 @@
 #
 
 import vllm_ascend.patch.worker.patch_common.patch_metrics  # noqa
+import vllm_ascend.patch.worker.patch_common.patch_minicpm  # noqa
 import vllm_ascend.patch.worker.patch_common.patch_multi_step_worker  # noqa
 import vllm_ascend.patch.worker.patch_common.patch_spec_decode_worker  # noqa

vllm_ascend/patch/worker/patch_common/patch_minicpm.py

@@ -0,0 +1,36 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# This file is a part of the vllm-ascend project.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import torch
+from vllm.model_executor.models.minicpm import MiniCPMAttention
+
+
+def forward(
+    self,
+    positions: torch.Tensor,
+    hidden_states: torch.Tensor,
+) -> torch.Tensor:
+    qkv, _ = self.qkv_proj(hidden_states)
+    q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+    q, k = self.rotary_emb(positions, q, k)
+    attn_output = self.attn(q, k, v)
+    output, _ = self.o_proj(attn_output)
+    return output
+
+
+# The type conversion in the forward function is deleted to support the rope operator.
+MiniCPMAttention.forward = forward