[Paddle TensorRT] Installing paddle does not depend on local tensorrt (#67525)

lizexu123 · web-flow · commit 26009f7a9608 · 2024-08-22T10:52:42.000+08:00
* kk

* fix env_dict.py.in

* 增加WITH_PIP_TENSORRT开关，设置为ON才下载tensorrt

* fix build bug

* 打开with_tensorrt=on,下载python版的tensorrt

* delete ununsed code

* fix 空格

* fix combine bug

* 修复编译

* 删除windows,trt8.6不支持pip install安装

* windows下trt版本大于10支持
diff --git a/python/env_dict.py.in b/python/env_dict.py.in
@@ -90,5 +90,8 @@ env_dict={
     'WITH_CPP_DIST':'@WITH_CPP_DIST@',
     'PADDLE_INSTALL_DIR':'@PADDLE_INSTALL_DIR@',
     'PADDLE_LIB_TEST_DIR':'@PADDLE_LIB_TEST_DIR@',
-    'WITH_PIP_CUDA_LIBRARIES':'@WITH_PIP_CUDA_LIBRARIES@'
+    'WITH_PIP_CUDA_LIBRARIES':'@WITH_PIP_CUDA_LIBRARIES@',
+    'TENSORRT_FOUND':'@TENSORRT_FOUND@',
+    'TR_INFER_RT':'@TR_INFER_RT@',
+    'TENSORRT_LIBRARY_DIR':'@TENSORRT_LIBRARY_DIR@',
 }
diff --git a/python/paddle/tensorrt/converter.py b/python/paddle/tensorrt/converter.py
@@ -27,6 +27,9 @@
 from .register import converter_registry
 from .util import map_dtype
 
+version = trt.__version__
+version_list = list(map(int, version.split('.')))
+
 
 def get_cache_path():
     home_path = os.path.expanduser("~")
@@ -232,7 +235,11 @@ def convert_subgraph_to_trt(self, program, group_op):
 
         config = builder.create_builder_config()
         config.add_optimization_profile(profile)
-        config.builder_optimization_level = 5
+
+        if version_list[0] > 8 or (
+            version_list[0] == 8 and version_list[1] >= 6
+        ):
+            config.builder_optimization_level = 5
         config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, 1 << 30)
         trt_engine = builder.build_engine(network, config)
         trt_params = paddle.base.libpaddle.TRTEngineParams()
diff --git a/python/paddle/tensorrt/impls/core.py b/python/paddle/tensorrt/impls/core.py
@@ -193,7 +193,9 @@ def softmax_converter(network, paddle_op, inputs):
     return softmax_layer
 
 
-@converter_registry.register("pd_op.layer_norm", trt_version="8.x")
+@converter_registry.register(
+    "pd_op.layer_norm", trt_version="trt_version_gt=8.6"
+)
 def layernorm_converter(network, paddle_op, inputs):
     input_a, scale, bias = inputs
 
diff --git a/python/setup.py.in b/python/setup.py.in
@@ -10,6 +10,7 @@ import errno
 import platform
 import glob
 import shlex
+import ctypes
 
 from contextlib import contextmanager
 from setuptools import Command
@@ -389,7 +390,6 @@ def cinn() -> str:
             'with_mkl': '@WITH_MKL@',
             'cinn': get_cinn_version(),
             'with_pip_cuda_libraries': '@WITH_PIP_CUDA_LIBRARIES@'})
-
 write_version_py(filename='@PADDLE_BINARY_DIR@/python/paddle/version/__init__.py')
 
 def write_cuda_env_config_py(filename='paddle/cuda_env.py'):
@@ -436,8 +436,51 @@ def is_transpiler():
 write_distributed_training_mode_py(filename='@PADDLE_BINARY_DIR@/python/paddle/incubate/distributed/fleet/parameter_server/version.py')
 
 
+def find_libnvinfer():
+    """Search for libnvinfer.so file in LD_LIBRARY_PATH."""
+
+    tensorrt_library_path='@TENSORRT_LIBRARY_DIR@'
+    trt_infer_rt_path='@TR_INFER_RT@'
+
+    libnvinfer_file =os.path.join(tensorrt_library_path,trt_infer_rt_path)
+
+    if os.path.exists(libnvinfer_file):
+        return libnvinfer_file
+    else:
+        print(f"{libnvinfer_file} not found.")
+    return None
+
+
+def get_tensorrt_version():
+    try:
+        libnvinfer_path = find_libnvinfer()
+
+        if not libnvinfer_path:
+            return None
+
+        trt = ctypes.CDLL(libnvinfer_path)
+        get_version =trt.getInferLibVersion
+        get_version.restype = ctypes.c_int
+        version = get_version()
+        version_str = str(version)
+        major = version_str[:1] if len(version_str) > 1 else version_str
+        minor = version_str[1:2] if len(version_str) > 3 else version_str[1:]
+        patch = version_str[3:] if len(version_str) > 3 else ''
+
+        minor = minor if minor else '0'
+        patch = patch if patch else '0'
+        version_str = f"{major}.{minor}.{patch}"
+
+        return version_str
+
+    except Exception as e:
+        print(f"Error while getting TensorRT version: {e}")
+        return None
+
 def get_paddle_extra_install_requirements():
-    #(Note risemeup1): Paddle will install the pypi cuda package provided by Nvidia, which includes the cuda runtime, cudnn, and cublas, thereby making the operation of 'pip install paddle' no longer dependent on the installation of cuda and cudnn.
+    #(Note risemeup1): Paddle will install the pypi cuda package provided by Nvidia, which includes the cuda runtime, cudnn, and cublas. Additionally, it now supports the installation of TensorRT, further enhancing its functionality. This integration simplifies the process as the operation of 'pip install paddle' is no longer dependent on the separate installation of cuda, cudnn, or TensorRT.
+    paddle_cuda_requires = []
+    paddle_tensorrt_requires = []
     if '@WITH_PIP_CUDA_LIBRARIES@' == 'ON':
         if platform.system() == 'Linux':
             PADDLE_CUDA_INSTALL_REQUIREMENTS = {
@@ -489,6 +532,8 @@ def get_paddle_extra_install_requirements():
                     "nvidia-cusparse-cu12==12.3.1.170 "
                 ),
             }
+
+
         try:
             output = subprocess.check_output(['nvcc', '--version']).decode('utf-8')
             version_line = [line for line in output.split('\n') if 'release' in line][0]
@@ -499,9 +544,44 @@ def get_paddle_extra_install_requirements():
 
         paddle_cuda_requires = PADDLE_CUDA_INSTALL_REQUIREMENTS[cuda_major_version].split("|")
 
-        return paddle_cuda_requires
-    else:
-        return []
+    if '@TENSORRT_FOUND@' == 'ON':
+        version_str = get_tensorrt_version()
+        version_default = int(version_str.split(".")[0])
+        if platform.system() =='Linux' or (platform.system()=='Windows' and version_default>=10):
+
+            PADDLE_TENSORRT_INSTALL_REQUIREMENTS = [
+                "tensorrt==8.5.3.1",
+                "tensorrt==8.6.0",
+                "tensorrt==8.6.1.post1",
+                "tensorrt==10.3.0",
+            ]
+
+            if not version_str:
+                return paddle_cuda_requires,[]
+
+            version_main = ".".join(version_str.split(".")[:3])
+
+            matched_package = None
+            for paddle_tensorrt_requires in PADDLE_TENSORRT_INSTALL_REQUIREMENTS:
+                paddle_tensorrt_version = paddle_tensorrt_requires.split("==")[1]
+                paddle_tensorrt_main = ".".join(paddle_tensorrt_version.split(".")[:3])
+
+                if version_main == paddle_tensorrt_main:
+                    matched_package = paddle_tensorrt_requires
+                    break
+
+            if matched_package:
+                paddle_tensorrt_requires = [matched_package]
+            else:
+                print(
+                    f"No exact match found for TensorRT Version: {version_str}. We currently support TensorRT versions 8.5.3.1, 8.6.0, and 8.6.1."
+                )
+                return paddle_cuda_requires, []
+
+    return paddle_cuda_requires, paddle_tensorrt_requires
+
+
+
 
 packages=['paddle',
           'paddle.libs',
@@ -694,8 +774,9 @@ if sys.version_info >= (3,8):
         setup_requires_tmp+=[setup_requires_i]
     setup_requires = setup_requires_tmp
     if '@WITH_GPU@' == 'ON' and platform.system() in ('Linux', 'Windows') and platform.machine() in ('x86_64', 'AMD64'):
-        paddle_cuda_requires = get_paddle_extra_install_requirements()
+        paddle_cuda_requires, paddle_tensorrt_requires = get_paddle_extra_install_requirements()
         setup_requires += paddle_cuda_requires
+        setup_requires += paddle_tensorrt_requires
 
 
 # the prefix is sys.prefix which should always be usr
diff --git a/setup.py b/setup.py
@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import ctypes
 import errno
 import fnmatch
 import glob
@@ -966,7 +967,50 @@ def get_setup_requires():
         )
 
 
+def find_libnvinfer():
+
+    trt_infer_rt_path = env_dict.get("TR_INFER_RT")
+    tensorrt_library_path = env_dict.get("TENSORRT_LIBRARY_DIR")
+
+    libnvinfer_file = os.path.join(tensorrt_library_path, trt_infer_rt_path)
+
+    if os.path.exists(libnvinfer_file):
+        return libnvinfer_file
+    else:
+        print(f"{libnvinfer_file} not found.")
+    return None
+
+
+def get_tensorrt_version():
+    try:
+
+        libnvinfer_path = find_libnvinfer()
+        if not libnvinfer_path:
+            return None
+
+        trt = ctypes.CDLL(libnvinfer_path)
+        get_version = trt.getInferLibVersion
+        get_version.restype = ctypes.c_int
+        version = get_version()
+        version_str = str(version)
+        major = version_str[:1] if len(version_str) > 1 else version_str
+        minor = version_str[1:2] if len(version_str) > 3 else version_str[1:]
+        patch = version_str[3:] if len(version_str) > 3 else ''
+
+        minor = minor if minor else '0'
+        patch = patch if patch else '0'
+        version_str = f"{major}.{minor}.{patch}"
+
+        return version_str
+
+    except Exception as e:
+        print(f"Error while getting TensorRT version: {e}")
+        return None
+
+
 def get_paddle_extra_install_requirements():
+    paddle_cuda_requires = []
+    paddle_tensorrt_requires = []
     # (Note risemeup1): Paddle will install the pypi cuda package provided by Nvidia, which includes the cuda runtime, cudnn, and cublas, thereby making the operation of 'pip install paddle' no longer dependent on the installation of cuda and cudnn.
     if env_dict.get("WITH_PIP_CUDA_LIBRARIES") == "ON":
         if platform.system() == 'Linux':
@@ -1035,9 +1079,48 @@ def get_paddle_extra_install_requirements():
             cuda_major_version
         ].split("|")
 
-        return paddle_cuda_requires
-    else:
-        return []
+    if env_dict.get("TENSORRT_FOUND") == "ON":
+        version_str = get_tensorrt_version()
+        version_default = int(version_str.split(".")[0])
+        if platform.system() == 'Linux' or (
+            platform.system() == 'Windows' and version_default >= 10
+        ):
+
+            PADDLE_TENSORRT_INSTALL_REQUIREMENTS = [
+                "tensorrt==8.5.3.1",
+                "tensorrt==8.6.0",
+                "tensorrt==8.6.1.post1",
+            ]
+
+            if not version_str:
+                return paddle_cuda_requires, []
+
+            version_main = ".".join(version_str.split(".")[:3])
+
+            matched_package = None
+            for (
+                paddle_tensorrt_requires
+            ) in PADDLE_TENSORRT_INSTALL_REQUIREMENTS:
+                paddle_tensorrt_version = paddle_tensorrt_requires.split("==")[
+                    1
+                ]
+                paddle_tensorrt_main = ".".join(
+                    paddle_tensorrt_version.split(".")[:3]
+                )
+
+                if version_main == paddle_tensorrt_main:
+                    matched_package = paddle_tensorrt_requires
+                    break
+
+            if matched_package:
+                paddle_tensorrt_requires = [matched_package]
+            else:
+                print(
+                    f"No exact match found for TensorRT Version: {version_str}. We currently support TensorRT versions 8.5.3.1, 8.6.0, and 8.6.1."
+                )
+                return paddle_cuda_requires, []
+
+    return paddle_cuda_requires, paddle_tensorrt_requires
 
 
 def get_cinn_config_jsons():
@@ -1628,8 +1711,11 @@ def get_setup_parameters():
             'AMD64',
         )
     ):
-        paddle_cuda_requires = get_paddle_extra_install_requirements()
+        paddle_cuda_requires, paddle_tensorrt_requires = (
+            get_paddle_extra_install_requirements()
+        )
         setup_requires += paddle_cuda_requires
+        setup_requires += paddle_tensorrt_requires
 
     packages = [
         'paddle',
