Fixes for 4090 and Orin (#876)

mlcommons · Jul 24, 2023 · c6dde10 · c6dde10
2 parents 4c1a536 + 4c04430
commit c6dde10
Show file tree

Hide file tree

Showing 6 changed files with 197 additions and 123 deletions.
diff --git a/cm-mlops/script/get-ml-model-retinanet/_cm.json b/cm-mlops/script/get-ml-model-retinanet/_cm.json
@@ -36,7 +36,10 @@
       },
       "update_tags_from_env_with_prefix": {
         "_url.": [ "CM_PACKAGE_URL" ]
-      }
+      },
+      "names": [
+        "dae"
+      ]
     }
   ],
   "variations": {
@@ -87,6 +90,11 @@
         "CM_UNZIP": "yes",
         "CM_ML_MODEL_FILE": "retinanet_model_10.pth",
         "CM_ML_MODEL_ACCURACY": "0.3755"
+      },
+      "add_deps_recursive": {
+        "dae": {
+          "tags": "_extract"
+        }
       }
     }
   }

diff --git a/cm-mlops/script/get-tensorrt/customize.py b/cm-mlops/script/get-tensorrt/customize.py
@@ -11,7 +11,7 @@ def preprocess(i):
     env = i['env']
 
 
-    if env.get('CM_TENSORRT_TAR_FILE_PATH','')=='' and env.get('CM_TENSORRT_REQUIRE_DEV', '') != 'yes':
+    if env.get('CM_TENSORRT_TAR_FILE_PATH','')=='' and env.get('CM_TENSORRT_REQUIRE_DEV', '') != 'yes' and env.get('CM_HOST_PLATFORM_FLAVOR', '') != 'aarch64':
 
        if os_info['platform'] == 'windows':
            extra_pre=''

diff --git a/cm-mlops/script/reproduce-mlperf-inference-nvidia/_cm.yaml b/cm-mlops/script/reproduce-mlperf-inference-nvidia/_cm.yaml
@@ -406,6 +406,30 @@ variations:
     env:
       MLPERF_NVIDIA_RUN_COMMAND: build
       CM_MLPERF_NVIDIA_HARNESS_RUN_MODE: build
+    deps:
+       - tags: get,cmake
+         version_min: "3.18"
+
+       # Detect Google Logger
+       - tags: get,generic,sys-util,_glog-dev
+
+       # Detect GFlags
+       - tags: get,generic,sys-util,_gflags-dev
+
+       # Detect libgmock-dev
+       - tags: get,generic,sys-util,_libgmock-dev
+
+       # Detect libre2-dev
+       - tags: get,generic,sys-util,_libre2-dev
+
+       # Detect libnuma-dev
+       - tags: get,generic,sys-util,_libnuma-dev
+
+       # Detect libboost-all-dev
+       - tags: get,generic,sys-util,_libboost-all-dev
+
+       # Detect rapidjson-dev
+       - tags: get,generic,sys-util,_rapidjson-dev
 
   preprocess_data:
     group: run-mode
@@ -468,6 +492,7 @@ variations:
           - batch-size
           - triton
           - build-engine-options
+
       - tags: reproduce,mlperf,inference,nvidia,harness,_download_model
         inherit_variation_tags: true
         force_cache: true
@@ -682,44 +707,28 @@ variations:
       CM_MODEL_BATCH_SIZE: "2048"
       CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
 
-  resnet50,multistream,run_harness:
-    add_deps_recursive:
-      build-engine:
-        tags: _batch_size.8
-    env:
-      CM_MODEL_BATCH_SIZE: "8"
-      CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
-
-  retinanet,multistream,run_harness:
+  gpu_memory.80,resnet50,server,run_harness:
     add_deps_recursive:
       build-engine:
-        tags: _batch_size.1
-    env:
-      CM_MODEL_BATCH_SIZE: "1"
-      CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
-
-  rnnt,multistream,run_harness:
-    add_deps_recursive:
-      build-engine:
-        tags: _batch_size.8
+        tags: _batch_size.64
     env:
-      CM_MODEL_BATCH_SIZE: "8"
+      CM_MODEL_BATCH_SIZE: "64"
       CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
 
-  3d-unet,multistream,run_harness:
+  resnet50,multistream,run_harness:
     add_deps_recursive:
       build-engine:
         tags: _batch_size.8
     env:
       CM_MODEL_BATCH_SIZE: "8"
       CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
 
-  dlrm,multistream,run_harness:
+  retinanet,multistream,run_harness:
     add_deps_recursive:
       build-engine:
-        tags: _batch_size.8
+        tags: _batch_size.2
     env:
-      CM_MODEL_BATCH_SIZE: "8"
+      CM_MODEL_BATCH_SIZE: "2"
       CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
 
   gpu_memory.16,retinanet,offline,run_harness:
@@ -754,6 +763,14 @@ variations:
       CM_MODEL_BATCH_SIZE: "32"
       CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
 
+  gpu_memory.80,retinanet,server,run_harness:
+    add_deps_recursive:
+      build-engine:
+        tags: _batch_size.8
+    env:
+      CM_MODEL_BATCH_SIZE: "8"
+      CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
+
   gpu_memory.16,rnnt,offline,run_harness:
     add_deps_recursive:
       build-engine:
@@ -863,6 +880,14 @@ variations:
       CM_MODEL_BATCH_SIZE: "64"
       CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
 
+  rtx_4090,resnet50,server,run_harness:
+    add_deps_recursive:
+      build-engine:
+        tags: _batch_size.64
+    env:
+      CM_MODEL_BATCH_SIZE: "64"
+      CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
+
   rtx_4090,retinanet,offline,run_harness:
     add_deps_recursive:
       build-engine:
@@ -871,6 +896,14 @@ variations:
       CM_MODEL_BATCH_SIZE: "2"
       CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
 
+  rtx_4090,retinanet,server,run_harness:
+    add_deps_recursive:
+      build-engine:
+        tags: _batch_size.2
+    env:
+      CM_MODEL_BATCH_SIZE: "2"
+      CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
+
   rtx_4090,bert_,offline,run_harness:
     add_deps_recursive:
       build-engine:
@@ -879,6 +912,14 @@ variations:
       CM_MODEL_BATCH_SIZE: "256"
       CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
 
+  rtx_4090,bert_,server,run_harness:
+    add_deps_recursive:
+      build-engine:
+        tags: _batch_size.256
+    env:
+      CM_MODEL_BATCH_SIZE: "256"
+      CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
+
   rtx_4090,3d-unet_,offline,run_harness:
     add_deps_recursive:
       build-engine:
@@ -887,6 +928,14 @@ variations:
       CM_MODEL_BATCH_SIZE: "8"
       CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
 
+  rtx_4090,3d-unet_,server,run_harness:
+    add_deps_recursive:
+      build-engine:
+        tags: _batch_size.8
+    env:
+      CM_MODEL_BATCH_SIZE: "8"
+      CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
+
   rtx_4090,rnnt,offline,run_harness:
     add_deps_recursive:
       build-engine:
@@ -895,6 +944,14 @@ variations:
       CM_MODEL_BATCH_SIZE: "2048"
       CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
 
+  rtx_4090,rnnt,server,run_harness:
+    add_deps_recursive:
+      build-engine:
+        tags: _batch_size.512
+    env:
+      CM_MODEL_BATCH_SIZE: "512"
+      CM_MLPERF_NVIDIA_HARNESS_GPU_BATCH_SIZE: "<<<CM_MODEL_BATCH_SIZE>>>"
+
   rtx_4090,dlrm_,offline,run_harness:
     add_deps_recursive:
       build-engine: