[PyTorch][Inference][EC2] Upgrade torchserve to 0.12.0 in PT 2.4 (#4401)

* Upgrade TS to 0.12.0 and disable autopatch

* Update ['dlc_developer_config.toml']

dlc_developer_config.toml:
{   'build': {   'build_frameworks': ['pytorch'],
                 'build_inference': True,
                 'build_training': False},
    'buildspec_override': {   'dlc-pr-pytorch-inference': 'pytorch/inference/buildspec-2-4-ec2.yml'},
    'dev': {   'deep_canary_mode': False,
               'graviton_mode': False,
               'neuronx_mode': False},
    'test': {   'ec2_tests': True,
                'ecs_tests': True,
                'eks_tests': True,
                'sagemaker_local_tests': True,
                'sagemaker_remote_tests': True,
                'sanity_tests': True}}

* update core_packages for gpu images

* Update ['dlc_developer_config.toml']

dlc_developer_config.toml:
{   'build': {   'build_frameworks': ['pytorch'],
                 'build_inference': True,
                 'build_training': False},
    'buildspec_override': {   'dlc-pr-pytorch-graviton-inference': 'pytorch/inference/buildspec-graviton-2-4-ec2.yml'},
    'dev': {   'deep_canary_mode': False,
               'graviton_mode': True,
               'neuronx_mode': False},
    'test': {   'ec2_tests': True,
                'ecs_tests': True,
                'eks_tests': True,
                'sagemaker_local_tests': True,
                'sagemaker_remote_tests': True,
                'sanity_tests': True}}

* Restore ['dlc_developer_config.toml']

dlc_developer_config.toml:
('Restore to '
 'https://raw.githubusercontent.com/aws/deep-learning-containers/master/dlc_developer_config.toml')

pytorch/inference/buildspec-2-4-ec2.yml

@@ -5,7 +5,7 @@ framework: &FRAMEWORK pytorch
 version: &VERSION 2.4.0
 short_version: &SHORT_VERSION "2.4"
 arch_type: x86
-autopatch_build: "True"
+# autopatch_build: "True"
 
 repository_info:
   inference_repository: &INFERENCE_REPOSITORY
@@ -40,7 +40,7 @@ images:
     python_version: &DOCKER_PYTHON_VERSION py3
     tag_python_version: &TAG_PYTHON_VERSION py311
     os_version: &OS_VERSION ubuntu22.04
-    torch_serve_version: &TORCHSERVE_VERSION 0.11.1
+    torch_serve_version: &TORCHSERVE_VERSION 0.12.0
     tag: !join [ *VERSION, "-", *DEVICE_TYPE, "-", *TAG_PYTHON_VERSION, "-", *OS_VERSION, "-ec2" ]
     latest_release_tag: !join [ *VERSION, "-", *DEVICE_TYPE, "-", *TAG_PYTHON_VERSION, "-", *OS_VERSION, "-ec2" ]
     # build_tag_override: "beta:2.4.0-cpu-py311-ubuntu22.04-ec2"
@@ -57,7 +57,7 @@ images:
     tag_python_version: &TAG_PYTHON_VERSION py311
     cuda_version: &CUDA_VERSION cu124
     os_version: &OS_VERSION ubuntu22.04
-    torch_serve_version: &TORCHSERVE_VERSION 0.11.1
+    torch_serve_version: &TORCHSERVE_VERSION 0.12.0
     tag: !join [ *VERSION, "-", *DEVICE_TYPE, "-", *TAG_PYTHON_VERSION, "-", *CUDA_VERSION, "-", *OS_VERSION, "-ec2" ]
     latest_release_tag: !join [ *VERSION, "-", *DEVICE_TYPE, "-", *TAG_PYTHON_VERSION, "-", *CUDA_VERSION, "-", *OS_VERSION, "-ec2" ]
     # build_tag_override: "beta:2.4.0-gpu-py311-cu124-ubuntu22.04-ec2"

pytorch/inference/buildspec-graviton-2-4-ec2.yml

@@ -5,7 +5,7 @@ framework: &FRAMEWORK pytorch
 version: &VERSION 2.4.0
 short_version: &SHORT_VERSION "2.4"
 arch_type: graviton
-autopatch_build: "True"
+# autopatch_build: "True"
 
 repository_info:
   inference_repository: &INFERENCE_REPOSITORY
@@ -40,7 +40,7 @@ images:
     python_version: &DOCKER_PYTHON_VERSION py3
     tag_python_version: &TAG_PYTHON_VERSION py311
     os_version: &OS_VERSION ubuntu22.04
-    torch_serve_version: &TORCHSERVE_VERSION 0.11.1
+    torch_serve_version: &TORCHSERVE_VERSION 0.12.0
     tag: !join [ *VERSION, "-", *DEVICE_TYPE, "-", *TAG_PYTHON_VERSION, "-", *OS_VERSION , "-ec2"]
     latest_release_tag: !join [ *VERSION, "-", *DEVICE_TYPE, "-", *TAG_PYTHON_VERSION, "-", *OS_VERSION, "-ec2" ]
     # build_tag_override: "beta:2.4.0-cpu-py311-ubuntu22.04-ec2"
@@ -57,7 +57,7 @@ images:
     tag_python_version: &TAG_PYTHON_VERSION py311
     cuda_version: &CUDA_VERSION cu124
     os_version: &OS_VERSION ubuntu22.04
-    torch_serve_version: &TORCHSERVE_VERSION 0.11.1
+    torch_serve_version: &TORCHSERVE_VERSION 0.12.0
     tag: !join [ *VERSION, "-", *DEVICE_TYPE, "-", *TAG_PYTHON_VERSION, "-", *CUDA_VERSION, "-", *OS_VERSION , "-ec2"]
     latest_release_tag: !join [ *VERSION, "-", *DEVICE_TYPE, "-", *TAG_PYTHON_VERSION, "-", *CUDA_VERSION, "-", *OS_VERSION, "-ec2" ]
     # build_tag_override: "beta:2.4.0-gpu-py311-cu124-ubuntu22.04-ec2"

pytorch/inference/docker/2.4/py3/Dockerfile.ec2.cpu.core_packages.json

@@ -34,9 +34,9 @@
     "version_specifier": "<2"
   },
   "torchserve": {
-    "version_specifier": "==0.11.1"
+    "version_specifier": "==0.12.0"
   },
   "torch-model-archiver": {
-    "version_specifier": "==0.11.1"
+    "version_specifier": "==0.12.0"
   }
 }

pytorch/inference/docker/2.4/py3/Dockerfile.ec2.graviton.cpu.core_packages.json

@@ -43,9 +43,9 @@
     "version_specifier": ">=2.32.3"
   },
   "torchserve": {
-    "version_specifier": "==0.11.1"
+    "version_specifier": "==0.12.0"
   },
   "torch-model-archiver": {
-    "version_specifier": "==0.11.1"
+    "version_specifier": "==0.12.0"
   }
 }

pytorch/inference/docker/2.4/py3/cu124/Dockerfile.ec2.gpu.core_packages.json

@@ -34,9 +34,9 @@
     "version_specifier": "<2"
   },
   "torchserve": {
-    "version_specifier": "==0.11.1"
+    "version_specifier": "==0.12.0"
   },
   "torch-model-archiver": {
-    "version_specifier": "==0.11.1"
+    "version_specifier": "==0.12.0"
   }
 }

pytorch/inference/docker/2.4/py3/cu124/Dockerfile.ec2.graviton.gpu.core_packages.json

@@ -43,9 +43,9 @@
     "version_specifier": ">=2.32.3"
   },
   "torchserve": {
-    "version_specifier": "==0.11.1"
+    "version_specifier": "==0.12.0"
   },
   "torch-model-archiver": {
-    "version_specifier": "==0.11.1"
+    "version_specifier": "==0.12.0"
   }
 }