[Feature] Optim PaddleOCR-VL (#4873)

* [Feature] Optim PaddleOCR-VL

* fix bug

custom_ops/gpu_ops/cpp_extensions.cc

@@ -1059,6 +1059,15 @@ std::vector<paddle::Tensor> UpdateAttnMaskOffsets(
     const paddle::Tensor& decode_states,
     const paddle::Tensor& mask_rollback);
 
+std::vector<paddle::Tensor> FusedNeoxRopeEmbedding(
+    const paddle::Tensor& qkv,
+    const paddle::Tensor& cos_emb,
+    const paddle::Tensor& sin_emb,
+    const int num_heads,
+    const int head_dim);
+
+std::vector<paddle::Tensor> GeluTanh(paddle::Tensor& input);
+
 PYBIND11_MODULE(fastdeploy_ops, m) {
   m.def("get_expert_token_num",
         &GetExpertTokenNum,
@@ -1648,4 +1657,10 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
   m.def("update_attn_mask_offsets",
         &UpdateAttnMaskOffsets,
         "update attention mask");
+
+  m.def("fused_neox_rope_embedding",
+        &FusedNeoxRopeEmbedding,
+        "fused_neox_rope_embedding function");
+
+  m.def("gelu_tanh", &GeluTanh, "gelu_tanh function");
 }

custom_ops/gpu_ops/fused_neox_rope_embedding.cu

@@ -0,0 +1,140 @@
+// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+//
+// 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.
+
+#include "helper.h"
+#include "paddle/extension.h"
+
+template <typename T, int VecSize = 1>
+__global__ void FusedNeoxRopeEmbeddingKernel(const T *__restrict__ qkv,
+                                             const float *__restrict__ cos_emb,
+                                             const float *__restrict__ sin_emb,
+                                             T *__restrict__ q,
+                                             T *__restrict__ k,
+                                             T *__restrict__ v,
+                                             const int64_t elem_cnt,
+                                             const int num_head,
+                                             const int last_dim) {
+  using LoadT = AlignedVector<T, VecSize>;
+  using LoadEmbT = AlignedVector<float, VecSize>;
+  LoadT left_vec;
+  LoadT right_vec;
+  LoadEmbT cos_emb_vec;
+  LoadEmbT sin_emb_vec;
+  int64_t global_thread_idx = blockDim.x * blockIdx.x + threadIdx.x;
+  const int half_lastdim = last_dim / 2;
+  const int hidden_size = num_head * half_lastdim;
+  const int full_hidden_size = num_head * last_dim;
+  const int offset = 3 * hidden_size;
+  for (int64_t linear_index = global_thread_idx * VecSize,
+               step = gridDim.x * blockDim.x * VecSize;
+       linear_index < elem_cnt;
+       linear_index += step) {
+    const int token_idx = linear_index / offset;
+    const int bias = linear_index % offset;
+    const int qkv_id = bias / hidden_size;
+    const int qkv_bias = bias % hidden_size;
+    const int hi = qkv_bias / half_lastdim;
+    const int h_bias = qkv_bias % half_lastdim;
+    const int base_idx_left = token_idx * 3 * full_hidden_size +
+                              qkv_id * full_hidden_size + hi * last_dim +
+                              h_bias;
+    const int base_idx_right = base_idx_left + half_lastdim;
+    const int emb_idx = token_idx * last_dim + h_bias;
+    const int base_split_idx_left =
+        token_idx * full_hidden_size + hi * last_dim + h_bias;
+    const int base_split_idx_right = base_split_idx_left + half_lastdim;
+
+    // q,k,v output
+    T *out_p = nullptr;
+    if (qkv_id == 0) {
+      out_p = q;
+    } else if (qkv_id == 1) {
+      out_p = k;
+    } else {
+      out_p = v;
+    }
+
+    Load<T, VecSize>(&qkv[base_idx_left], &left_vec);
+    Load<T, VecSize>(&qkv[base_idx_right], &right_vec);
+    // do rope
+    if (qkv_id < 2) {
+      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
+#pragma unroll
+      for (int i = 0; i < VecSize; i++) {
+        float input_left = static_cast<float>(left_vec[i]);
+        float input_right = static_cast<float>(right_vec[i]);
+        const float cos_tmp = cos_emb_vec[i];
+        const float sin_tmp = sin_emb_vec[i];
+        left_vec[i] =
+            static_cast<T>(input_left * cos_tmp - input_right * sin_tmp);
+        right_vec[i] =
+            static_cast<T>(input_right * cos_tmp + input_left * sin_tmp);
+
+        int cur_idx_1 = base_split_idx_left + i;
+        int cur_idx_2 = base_split_idx_right + i;
+      }
+    }
+    Store<T, VecSize>(left_vec, &out_p[base_split_idx_left]);
+    Store<T, VecSize>(right_vec, &out_p[base_split_idx_right]);
+  }
+}
+
+std::vector<paddle::Tensor> FusedNeoxRopeEmbedding(
+    const paddle::Tensor &qkv,
+    const paddle::Tensor &cos_emb,
+    const paddle::Tensor &sin_emb,
+    const int num_heads,
+    const int head_dim) {
+  typedef PDTraits<paddle::DataType::BFLOAT16> traits_;
+  typedef typename traits_::DataType DataType_;
+  typedef typename traits_::data_t data_t;
+
+  const auto &qkv_dims = qkv.dims();
+  const int token_num = qkv_dims.size() == 2 ? qkv_dims[0] : qkv_dims[1];
+
+  auto stream = qkv.stream();
+  paddle::Tensor q = GetEmptyTensor(
+      {token_num, num_heads, head_dim}, qkv.dtype(), qkv.place());
+  paddle::Tensor k = GetEmptyTensor(
+      {token_num, num_heads, head_dim}, qkv.dtype(), qkv.place());
+  paddle::Tensor v = GetEmptyTensor(
+      {token_num, num_heads, head_dim}, qkv.dtype(), qkv.place());
+
+  int64_t elem_nums = token_num * num_heads * head_dim * 3 / 2;
+  constexpr int PackSize = 4;
+  const int pack_num = elem_nums / PackSize;
+  const int blocksize = 128;
+  int grid_size = 1;
+  GetNumBlocks<128>(pack_num, &grid_size);
+
+  FusedNeoxRopeEmbeddingKernel<DataType_, PackSize>
+      <<<grid_size, blocksize, 0, stream>>>(
+          reinterpret_cast<const DataType_ *>(qkv.data<data_t>()),
+          cos_emb.data<float>(),
+          sin_emb.data<float>(),
+          reinterpret_cast<DataType_ *>(q.data<data_t>()),
+          reinterpret_cast<DataType_ *>(k.data<data_t>()),
+          reinterpret_cast<DataType_ *>(v.data<data_t>()),
+          elem_nums,
+          num_heads,
+          head_dim);
+  return {q, k, v};
+}
+
+PD_BUILD_STATIC_OP(fused_neox_rope_embedding)
+    .Inputs({"qkv", "cos_emb", "sin_emb"})
+    .Outputs({"q", "k", "v"})
+    .Attrs({"num_heads: int", "head_dim: int"})
+    .SetKernelFn(PD_KERNEL(FusedNeoxRopeEmbedding));

custom_ops/gpu_ops/gelu_tanh.cu

@@ -0,0 +1,106 @@
+// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+//
+// 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.
+
+#include "helper.h"
+#include "paddle/extension.h"
+
+__forceinline__ __device__ float tanh_ptx(float x) {
+  float y;
+  asm volatile("tanh.approx.f32 %0, %1;" : "=f"(y) : "f"(x));
+  return y;
+}
+
+__device__ __forceinline__ float gelu_tanh_func(const float& val) {
+  const float cdf =
+      0.5f * (1.0f + tanh_ptx((0.7978845608028654f *
+                               (val + 0.044715f * val * val * val))));
+  return val * cdf;
+}
+
+template <typename T>
+__global__ void gelu_tanh_kernel(T* __restrict__ out,
+                                 const T* __restrict__ input,
+                                 const int d) {
+  constexpr uint32_t kVecSize = 16 / sizeof(T);
+  const int64_t token_idx = blockIdx.x;
+  const int64_t thread_idx = threadIdx.x;
+  const int64_t stride = blockDim.x;
+  const int64_t offset = token_idx * d;
+  using vec_t = AlignedVector<T, kVecSize>;
+#if (__CUDACC_VER_MAJOR__ >= 12 && defined(__CUDA_ARCH__) && \
+     (__CUDA_ARCH__ >= 900))
+  asm volatile("griddepcontrol.wait;");
+#endif
+
+#pragma unroll 1
+  for (uint32_t idx = thread_idx; idx < d / kVecSize; idx += stride) {
+    vec_t x_vec;
+    Load(input + offset + idx * kVecSize, &x_vec);
+#pragma unroll
+    for (uint32_t i = 0; i < kVecSize; ++i) {
+      x_vec[i] = static_cast<T>(gelu_tanh_func(static_cast<float>(x_vec[i])));
+    }
+    Store(x_vec, out + token_idx * d + idx * kVecSize);
+  }
+
+  const int64_t remaining_offset = d - d % (stride * kVecSize);
+  // process the remaining elements
+#pragma unroll 1
+  for (int64_t idx = thread_idx; idx < d % (stride * kVecSize); idx += stride) {
+    float x = static_cast<float>(input[offset + remaining_offset + idx]);
+    out[token_idx * d + remaining_offset + idx] =
+        static_cast<T>(gelu_tanh_func(x));
+  }
+
+#if (__CUDACC_VER_MAJOR__ >= 12 && defined(__CUDA_ARCH__) && \
+     (__CUDA_ARCH__ >= 900))
+  asm volatile("griddepcontrol.launch_dependents;");
+#endif
+}
+
+std::vector<paddle::Tensor> GeluTanh(paddle::Tensor& input) {
+  int d = input.dims()[1];
+  int64_t num_tokens = input.dims()[0];
+  cudaStream_t stream = input.stream();
+
+  paddle::Tensor output =
+      GetEmptyTensor(input.dims(), input.dtype(), input.place());
+
+  DISPATCH_FLOAT_FP6_DTYPE(input.dtype(), scalar_t, {
+    uint32_t vec_size = 16 / sizeof(scalar_t);
+    cudaLaunchConfig_t config;
+    config.gridDim = num_tokens;
+    config.blockDim = std::min(d / vec_size, 1024U);
+    config.dynamicSmemBytes = 0;
+    config.stream = stream;
+    cudaLaunchAttribute attrs[1];
+    attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
+    attrs[0].val.programmaticStreamSerializationAllowed = false;
+    config.numAttrs = 1;
+    config.attrs = attrs;
+
+    cudaLaunchKernelEx(&config,
+                       gelu_tanh_kernel<scalar_t>,
+                       output.data<scalar_t>(),
+                       input.data<scalar_t>(),
+                       d);
+  });
+
+  return {output};
+}
+
+PD_BUILD_STATIC_OP(gelu_tanh)
+    .Inputs({"input"})
+    .Outputs({"output"})
+    .SetKernelFn(PD_KERNEL(GeluTanh));

custom_ops/setup_ops.py

@@ -306,6 +306,8 @@ elif paddle.is_compiled_with_cuda():
         "gpu_ops/limit_thinking_content_length_v1.cu",
         "gpu_ops/limit_thinking_content_length_v2.cu",
         "gpu_ops/update_attn_mask_offsets.cu",
+        "gpu_ops/fused_neox_rope_embedding.cu",
+        "gpu_ops/gelu_tanh.cu",
     ]
 
     # pd_disaggregation

docs/best_practices/PaddleOCR-VL-0.9B.md

@@ -5,8 +5,8 @@
 ## 1. Environment Preparation
 ### 1.1 Support Status
 Recommended Hardware Configuration:
-- GPU Memory: 12GB or more
-- Shared Memory: 2GB or more
+- GPU Memory: 8GB or more
+- Shared Memory: 4GB or more
 
 ### 1.2 Install Fastdeploy
 
@@ -18,38 +18,38 @@ Installation process reference documentation [FastDeploy GPU Install](../get_sta
 ```shell
 python -m fastdeploy.entrypoints.openai.api_server \
     --model PaddlePaddle/PaddleOCR-VL \
-    --port 8180 \
-    --metrics-port 8181 \
-    --engine-worker-queue-port 8182 \
+    --port 8185 \
+    --metrics-port 8186 \
+    --engine-worker-queue-port 8187 \
     --max-model-len 16384 \
     --max-num-batched-tokens 16384 \
-    --gpu-memory-utilization 0.9 \
-    --max-num-seqs 128
+    --gpu-memory-utilization 0.8 \
+    --max-num-seqs 256
 ```
 
 **Example 2:** Deploying a 16K Context Service on a Single RTX 4090 GPU
 ```shell
 python -m fastdeploy.entrypoints.openai.api_server \
     --model PaddlePaddle/PaddleOCR-VL \
-    --port 8180 \
-    --metrics-port 8181 \
-    --engine-worker-queue-port 8182 \
+    --port 8185 \
+    --metrics-port 8186 \
+    --engine-worker-queue-port 8187 \
     --max-model-len 16384 \
     --max-num-batched-tokens 16384 \
-    --gpu-memory-utilization 0.8 \
-    --max-num-seqs 196
+    --gpu-memory-utilization 0.7 \
+    --max-num-seqs 256
 ```
 
 **Example 3:** Deploying a 16K Context Service on a Single A100 GPU
 ```shell
 python -m fastdeploy.entrypoints.openai.api_server \
     --model PaddlePaddle/PaddleOCR-VL \
-    --port 8180 \
-    --metrics-port 8181 \
-    --engine-worker-queue-port 8182 \
+    --port 8185 \
+    --metrics-port 8186 \
+    --engine-worker-queue-port 8187 \
     --max-model-len 16384 \
     --max-num-batched-tokens 16384 \
-    --gpu-memory-utilization 0.8 \
+    --gpu-memory-utilization 0.7 \
     --max-num-seqs 256
 ```
 
@@ -71,7 +71,7 @@ An example is a set of configurations that can run stably while also delivering
 > **Available GPU memory ratio during initialization**
 - **Parameters：** `--gpu-memory-utilization`
 - **Description：** Controls the available GPU memory for FastDeploy service initialization. The default value is 0.9, meaning 10% of the memory is reserved for backup.
-- **Recommendation：** It is recommended to use 0.8. If an "out of memory" error occurs during stress testing, you may attempt to reduce this value.
+- **Recommendation：** It is recommended to use 0.7. If an "out of memory" error occurs during stress testing, you may attempt to reduce this value.
 
 #### 2.2.2 Chunked Prefill
 - **Parameters：** `--max-num-batched-tokens`

docs/zh/best_practices/PaddleOCR-VL-0.9B.md

@@ -5,8 +5,8 @@
 ## 一、环境准备
 ### 1.1 支持情况
 推荐硬件配置：
-- 显存：12GB显存及以上
-- 共享内存：2G及以上
+- 显存：8GB显存及以上
+- 共享内存：4G及以上
 
 ### 1.2 安装fastdeploy
 
@@ -18,12 +18,12 @@
 ```shell
 python -m fastdeploy.entrypoints.openai.api_server \
     --model PaddlePaddle/PaddleOCR-VL \
-    --port 8180 \
-    --metrics-port 8181 \
-    --engine-worker-queue-port 8182 \
+    --port 8185 \
+    --metrics-port 8186 \
+    --engine-worker-queue-port 8187 \
     --max-model-len 16384 \
     --max-num-batched-tokens 16384 \
-    --gpu-memory-utilization 0.9 \
+    --gpu-memory-utilization 0.8 \
     --max-num-seqs 128
 ```
 
@@ -31,25 +31,25 @@ python -m fastdeploy.entrypoints.openai.api_server \
 ```shell
 python -m fastdeploy.entrypoints.openai.api_server \
     --model PaddlePaddle/PaddleOCR-VL \
-    --port 8180 \
-    --metrics-port 8181 \
-    --engine-worker-queue-port 8182 \
+    --port 8185 \
+    --metrics-port 8186 \
+    --engine-worker-queue-port 8187 \
     --max-model-len 16384 \
     --max-num-batched-tokens 16384 \
-    --gpu-memory-utilization 0.8 \
-    --max-num-seqs 196
+    --gpu-memory-utilization 0.7 \
+    --max-num-seqs 256
 ```
 
  **示例3：** A100上单卡部署16K上下文的服务
 ```shell
 python -m fastdeploy.entrypoints.openai.api_server \
     --model PaddlePaddle/PaddleOCR-VL \
-    --port 8180 \
-    --metrics-port 8181 \
-    --engine-worker-queue-port 8182 \
+    --port 8185 \
+    --metrics-port 8186 \
+    --engine-worker-queue-port 8187 \
     --max-model-len 16384 \
     --max-num-batched-tokens 16384 \
-    --gpu-memory-utilization 0.8 \
+    --gpu-memory-utilization 0.7 \
     --max-num-seqs 256
 ```
 
@@ -72,7 +72,7 @@ python -m fastdeploy.entrypoints.openai.api_server \
 > **初始化时可用的显存比例**
 - **参数：** `--gpu-memory-utilization`
 - **用处：** 用于控制 FastDeploy 初始化服务的可用显存，默认0.9，即预留10%的显存备用。
-- **推荐：** 推荐使用0.8。如果服务压测时提示显存不足，可以尝试调低该值。
+- **推荐：** 推荐使用0.7。如果服务压测时提示显存不足，可以尝试调低该值。
 
 #### 2.2.2 Chunked Prefill
 - **参数：** `--max-num-batched-tokens`

docs/zh/usage/kunlunxin_xpu_deployment.md

@@ -197,7 +197,7 @@ curl -X POST "http://0.0.0.0:8188/v1/chat/completions" \
 -d '{
   "messages": [
     {"role": "user", "content": [
-              {"type": "image_url", "image_url": {"url": "https://paddle-model-ecology.bj.bcebos.com/PPOCRVL/dataset/ocr_v5_eval/handwrite_ch_rec_val/中文手写古籍_000054_crop_32.jpg"}},
+              {"type": "image_url", "image_url": {"url": "https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/ocr_demo.jpg"}},
               {"type": "text", "text": "OCR:"}
             ]}
     ],
@@ -216,7 +216,7 @@ response = client.chat.completions.create(
     model="default",
     messages=[
         {"role": "user", "content": [
-              {"type": "image_url", "image_url": {"url": "https://paddle-model-ecology.bj.bcebos.com/PPOCRVL/dataset/ocr_v5_eval/handwrite_ch_rec_val/中文手写古籍_000054_crop_32.jpg"}},
+              {"type": "image_url", "image_url": {"url": "https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/ocr_demo.jpg"}},
               {"type": "text", "text": "OCR:"}
             ]
         },

fastdeploy/model_executor/models/paddleocr_vl/paddleocr_vl.py

@@ -22,7 +22,6 @@ import paddle
 import paddle.nn as nn
 from paddleformers.transformers import PretrainedModel
 
-from fastdeploy import envs
 from fastdeploy.config import FDConfig
 from fastdeploy.model_executor.forward_meta import ForwardMeta
 from fastdeploy.model_executor.graph_optimization.decorator import (
@@ -136,12 +135,8 @@ class PaddleOCRVLForConditionalGeneration(ModelForCasualLM):
         )
 
         # Persistent buffers for CUDA graphs.
-        if envs.FD_ENABLE_MAX_PREFILL:
-            max_length = fd_config.scheduler_config.max_num_seqs * fd_config.model_config.max_model_len
-        else:
-            max_length = fd_config.model_config.max_model_len
-        self._input_embeddings = paddle.zeros(
-            [max_length, fd_config.model_config.hidden_size],
+        self._decoder_input_embeddings = paddle.zeros(
+            [fd_config.scheduler_config.max_num_seqs, fd_config.model_config.hidden_size],
             dtype=fd_config.model_config.dtype,
         )
 
@@ -247,12 +242,19 @@ class PaddleOCRVLForConditionalGeneration(ModelForCasualLM):
         input_embeddings = self.get_input_embeddings(
             ids_remove_padding=ids_remove_padding, image_features=image_features
         )
-        self._input_embeddings.copy_(input_embeddings, False)
 
-        hidden_states = self.model(
-            input_embeddings=self._input_embeddings,
-            forward_meta=forward_meta,
-        )
+        if forward_meta.step_use_cudagraph:
+            self._decoder_input_embeddings.copy_(input_embeddings, False)
+
+            hidden_states = self.model(
+                input_embeddings=self._decoder_input_embeddings,
+                forward_meta=forward_meta,
+            )
+        else:
+            hidden_states = self.model(
+                input_embeddings=input_embeddings,
+                forward_meta=forward_meta,
+            )
 
         return hidden_states
 

fastdeploy/model_executor/models/paddleocr_vl/siglip.py

@@ -21,39 +21,13 @@ import numpy as np
 import paddle
 import paddle.nn as nn
 import paddle.nn.functional as F
-from paddleformers.transformers.activations import ACT2FN
 from paddleformers.transformers.model_utils import PretrainedModel
 
 from fastdeploy.model_executor.layers.utils import get_tensor
 from fastdeploy.model_executor.utils import slice_fn
 
 from .config import PaddleOCRVisionConfig
-
-
-def rotate_half(x):
-    Dh = x.shape[-1]
-    x1 = x[..., : Dh // 2]
-    x2 = x[..., Dh // 2 :]
-    return paddle.concat([-x2, x1], axis=-1)
-
-
-def _ensure_cos_sin_dim(cos, sin, dim_needed):
-    last = cos.shape[-1]
-    if last == dim_needed:
-        return cos, sin
-    elif last * 2 == dim_needed:
-        cos = paddle.concat([cos, cos], axis=-1)
-        sin = paddle.concat([sin, sin], axis=-1)
-        return cos, sin
-    else:
-        raise ValueError(f"Unexpected cos/sin last-dim: {last}, expected {dim_needed} or {dim_needed//2}")
-
-
-def apply_rotary_pos_emb_vision(x, cos, sin):
-    orig_dtype = x.dtype
-    x = x.astype("float32")
-    x_embed = (x * cos) + (rotate_half(x) * sin)
-    return x_embed.astype(orig_dtype)
+from .siglip_ops import get_activation_fn, neox_rope_embedding
 
 
 class SiglipAttention(nn.Layer):
@@ -147,29 +121,12 @@ class SiglipAttention(nn.Layer):
         output_attentions: Optional[bool] = False,
         cu_seqlens: Optional[List[paddle.Tensor]] = None,
         max_seqlen: Optional[paddle.Tensor] = None,
-        rope_emb: Optional[Tuple[paddle.Tensor, paddle.Tensor]] = None,  # (cos, sin)
+        cos_emb: Optional[paddle.Tensor] = None,  # (cos, sin)
+        sin_emb: Optional[paddle.Tensor] = None,  # (cos, sin)
     ):
         B, seq_length, D = hidden_states.shape
-
-        qkv = (
-            self.qkv_proj(hidden_states)
-            .reshape(
-                [
-                    seq_length,
-                    3,
-                    self.num_heads,
-                    -1,
-                ]
-            )
-            .transpose(perm=[1, 0, 2, 3])
-        )
-        q, k, v = qkv.unbind(axis=0)
-        cos, sin = rope_emb
-
-        # --------
-        q = apply_rotary_pos_emb_vision(q, cos, sin)
-        k = apply_rotary_pos_emb_vision(k, cos, sin)
-
+        qkv = self.qkv_proj(hidden_states)
+        q, k, v = neox_rope_embedding(qkv, cos_emb, sin_emb, self.num_heads, self.head_dim)
         attn_output = self.flash_attn_func(
             q,
             k,
@@ -181,11 +138,9 @@ class SiglipAttention(nn.Layer):
             causal=False,
             **self.flash_attn_kwargs,
         )[0]
-        # --------
 
         attn_output = attn_output.reshape((seq_length, -1))
         attn_output = self.out_proj(attn_output)
-
         return attn_output
 
 
@@ -327,11 +282,7 @@ class SiglipMLP(nn.Layer):
     def __init__(self, config):
         super().__init__()
         self.config = config
-        if config.hidden_act == "gelu_pytorch_tanh":
-            config.hidden_act = "gelu_new"
-
-        self.activation_fn = ACT2FN[config.hidden_act]
-
+        self.activation_fn = get_activation_fn(config.hidden_act)
         self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
         self.fc1.weight.weight_loader = self.weight_loader
         self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
@@ -353,7 +304,7 @@ class SiglipMLP(nn.Layer):
 
     def forward(self, hidden_states: paddle.Tensor) -> paddle.Tensor:
         hidden_states = self.fc1(hidden_states)
-        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.activation_fn(hidden_states[0])
         hidden_states = self.fc2(hidden_states)
         return hidden_states
 
@@ -375,7 +326,8 @@ class SiglipEncoderLayer(paddle.nn.Layer):
         output_attentions=False,
         cu_seqlens=None,
         max_seqlen=None,
-        rope_emb=None,
+        cos_emb=None,
+        sin_emb=None,
     ):
 
         residual = hidden_states
@@ -388,7 +340,8 @@ class SiglipEncoderLayer(paddle.nn.Layer):
             output_attentions=output_attentions,
             cu_seqlens=cu_seqlens,
             max_seqlen=max_seqlen,
-            rope_emb=rope_emb,
+            cos_emb=cos_emb,
+            sin_emb=sin_emb,
         )
 
         hs_post_attn = residual + x
@@ -545,13 +498,13 @@ class SiglipEncoder(nn.Layer):
 
             rope_emb = rope_emb_max_grid[pids].flatten(1)
             rope_emb = rope_emb.tile((1, 2))
-            cos = rope_emb.cos().astype("float32")
-            sin = rope_emb.sin().astype("float32")
-            cos = cos.unsqueeze(-2)
-            sin = sin.unsqueeze(-2)
-            rope_emb = (cos, sin)
+            cos_emb = rope_emb.cos().astype("float32")
+            sin_emb = rope_emb.sin().astype("float32")
+            cos_emb = cos_emb.unsqueeze(-2)
+            sin_emb = sin_emb.unsqueeze(-2)
         else:
-            rope_emb = None
+            cos_emb = None
+            sin_emb = None
 
             window_indices, cu_seqlens_within_windows = None, None
 
@@ -588,7 +541,8 @@ class SiglipEncoder(nn.Layer):
                 output_attentions=output_attentions,
                 cu_seqlens=attn_cu_seqlens,
                 max_seqlen=max_seqlen,
-                rope_emb=rope_emb,
+                cos_emb=cos_emb,
+                sin_emb=sin_emb,
             )
             hidden_states = layer_outputs[0]
 

fastdeploy/model_executor/models/paddleocr_vl/siglip_ops.py

@@ -0,0 +1,74 @@
+"""
+# Copyright (c) 2025  PaddlePaddle Authors. All Rights Reserved.
+#
+# 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.
+"""
+
+from typing import List
+
+import paddle
+from paddleformers.transformers.activations import ACT2FN
+
+from fastdeploy.platforms import current_platform
+
+if current_platform.is_cuda():
+    from fastdeploy.model_executor.ops.gpu import fused_neox_rope_embedding, gelu_tanh
+
+
+def rotate_half(x):
+    Dh = x.shape[-1]
+    x1 = x[..., : Dh // 2]
+    x2 = x[..., Dh // 2 :]
+    return paddle.concat([-x2, x1], axis=-1)
+
+
+def apply_rotary_pos_emb_vision(x, cos, sin):
+    orig_dtype = x.dtype
+    x = x.astype("float32")
+    x_embed = (x * cos) + (rotate_half(x) * sin)
+    return x_embed.astype(orig_dtype)
+
+
+def native_neox_rope_embedding(qkv, cos, sin, num_heads):
+    B, seq_length, D = qkv.shape
+    qkv = qkv.reshape(
+        [
+            seq_length,
+            3,
+            num_heads,
+            -1,
+        ]
+    ).transpose(perm=[1, 0, 2, 3])
+    q, k, v = qkv.unbind(axis=0)
+    q = apply_rotary_pos_emb_vision(q, cos, sin)
+    k = apply_rotary_pos_emb_vision(k, cos, sin)
+    return q, k, v
+
+
+def neox_rope_embedding(
+    qkv: paddle.Tensor, cos_emb: paddle.Tensor, sin_emb: paddle.Tensor, num_heads: int, head_dim: int
+) -> List[paddle.Tensor]:
+    if current_platform.is_cuda():
+        return fused_neox_rope_embedding(qkv, cos_emb, sin_emb, num_heads, head_dim)
+    else:
+        return native_neox_rope_embedding(qkv, cos_emb, sin_emb, num_heads)
+
+
+def get_activation_fn(hidden_act: str):
+    if hidden_act == "gelu_pytorch_tanh":
+        if current_platform.is_cuda():
+            return gelu_tanh
+        else:
+            return ACT2FN["gelu_new"]
+    else:
+        return ACT2FN[hidden_act]

tests/operators/test_fused_neox_rope_embedding.py

@@ -0,0 +1,88 @@
+# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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 unittest
+
+import numpy as np
+import paddle
+
+from fastdeploy.model_executor.ops.gpu import fused_neox_rope_embedding
+
+
+def rotate_half(x):
+    Dh = x.shape[-1]
+    x1 = x[..., : Dh // 2]
+    x2 = x[..., Dh // 2 :]
+    return paddle.concat([-x2, x1], axis=-1)
+
+
+def apply_rotary_pos_emb_vision(x, cos, sin):
+    orig_dtype = x.dtype
+    x = x.astype("float32")
+    x_embed = (x * cos) + (rotate_half(x) * sin)
+    return x_embed.astype(orig_dtype)
+
+
+class TestFusedNeoxRopeEmbedding(unittest.TestCase):
+    def setUp(self):
+        paddle.set_device("gpu")
+        np.random.seed(42)
+
+    def native_neox_rope_embedding(self, qkv, cos, sin, num_heads):
+        seq_length = qkv.shape[0]
+        qkv = qkv.reshape(
+            [
+                seq_length,
+                3,
+                num_heads,
+                -1,
+            ]
+        ).transpose(perm=[1, 0, 2, 3])
+        q, k, v = qkv.unbind(axis=0)
+        q = apply_rotary_pos_emb_vision(q, cos, sin)
+        k = apply_rotary_pos_emb_vision(k, cos, sin)
+        return q, k, v
+
+    def test_fused_neox_rope_embedding(self):
+        token_num = 1024
+        hidden_size = 2048
+        head_dim = 128
+        num_heads = hidden_size // head_dim
+        qkv = paddle.randn([token_num, 3 * hidden_size]).astype("bfloat16")
+        cos_emb = paddle.rand([token_num, head_dim // 2]).tile((1, 2)).unsqueeze(1)
+        sin_emb = paddle.rand([token_num, head_dim // 2]).tile((1, 2)).unsqueeze(1)
+        q, k, v = fused_neox_rope_embedding(qkv, cos_emb, sin_emb, num_heads, head_dim)
+        q_base, k_base, v_base = self.native_neox_rope_embedding(qkv, cos_emb, sin_emb, num_heads)
+        np.testing.assert_allclose(
+            q.cast("float32").numpy(),
+            q_base.cast("float32").numpy(),
+            rtol=1e-02,
+            atol=1e-02,
+        )
+        np.testing.assert_allclose(
+            k.cast("float32").numpy(),
+            k_base.cast("float32").numpy(),
+            rtol=1e-02,
+            atol=1e-02,
+        )
+        np.testing.assert_allclose(
+            v.cast("float32").numpy(),
+            v_base.cast("float32").numpy(),
+            rtol=1e-02,
+            atol=1e-02,
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/operators/test_gelu_tanh.py

@@ -0,0 +1,42 @@
+# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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 unittest
+
+import numpy as np
+import paddle
+from paddleformers.transformers.activations import ACT2FN
+
+from fastdeploy.model_executor.ops.gpu import gelu_tanh
+
+
+class TestGeluTanh(unittest.TestCase):
+    def setUp(self):
+        paddle.set_device("gpu")
+        np.random.seed(42)
+
+    def test_gelu_tanh(self):
+        x = paddle.randn(2048, 4096)
+        y0 = ACT2FN["gelu_new"](x)
+        y1 = gelu_tanh(x)
+        np.testing.assert_allclose(
+            y0.cast("float32").numpy(),
+            y1.cast("float32").numpy(),
+            rtol=1e-04,
+            atol=1e-04,
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()