Skip to content

Commit

Permalink
Results from GH action on NVIDIA_RTX4090x2
Browse files Browse the repository at this point in the history
  • Loading branch information
@arjunsuresh
arjunsuresh committed Dec 23, 2024
1 parent 3912fed commit 13a16ab
Show file tree
Hide file tree
Showing 81 changed files with 43,556 additions and 0 deletions.
135 changes: 135 additions & 0 deletions ...a_original-gpu-tensorrt-vdefault-default_config/retinanet/multistream/README.md
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,135 @@
This experiment is generated using the [MLCommons Collective Mind automation framework (CM)](https://github.com/mlcommons/cm4mlops).

*Check [CM MLPerf docs](https://docs.mlcommons.org/inference) for more details.*

## Host platform

* OS version: Linux-6.8.0-49-generic-x86_64-with-glibc2.29
* CPU version: x86_64
* Python version: 3.8.10 (default, Nov 7 2024, 13:10:47)
[GCC 9.4.0]
* MLCommons CM version: 3.5.2

## CM Run Command

See [CM installation guide](https://docs.mlcommons.org/inference/install/).

```bash
pip install -U cmind

cm rm cache -f

cm pull repo mlcommons@mlperf-automations --checkout=225220c7d9bb7e66e5b9a1e1ebfc3e0180fbd094

cm run script \
--tags=app,mlperf,inference,generic,_nvidia,_retinanet,_tensorrt,_cuda,_valid,_r4.1-dev_default,_multistream \
--quiet=true \
--env.CM_QUIET=yes \
--env.CM_MLPERF_IMPLEMENTATION=nvidia \
--env.CM_MLPERF_MODEL=retinanet \
--env.CM_MLPERF_RUN_STYLE=valid \
--env.CM_MLPERF_SKIP_SUBMISSION_GENERATION=False \
--env.CM_DOCKER_PRIVILEGED_MODE=True \
--env.CM_MLPERF_BACKEND=tensorrt \
--env.CM_MLPERF_SUBMISSION_SYSTEM_TYPE=datacenter,edge \
--env.CM_MLPERF_CLEAN_ALL=True \
--env.CM_MLPERF_DEVICE=cuda \
--env.CM_MLPERF_SUBMISSION_DIVISION=closed \
--env.CM_MLPERF_USE_DOCKER=True \
--env.CM_NVIDIA_GPU_NAME=rtx_4090 \
--env.CM_HW_NAME=RTX4090x2 \
--env.CM_RUN_MLPERF_SUBMISSION_PREPROCESSOR=yes \
--env.CM_MLPERF_INFERENCE_PULL_CODE_CHANGES=yes \
--env.CM_MLPERF_INFERENCE_PULL_SRC_CHANGES=yes \
--env.OUTPUT_BASE_DIR=/home/arjun/gh_action_results \
--env.CM_MLPERF_INFERENCE_SUBMISSION_DIR=/home/arjun/gh_action_submissions \
--env.CM_MLPERF_SUBMITTER=MLCommons \
--env.CM_USE_DATASET_FROM_HOST=yes \
--env.CM_USE_MODEL_FROM_HOST=yes \
--env.CM_MLPERF_LOADGEN_ALL_SCENARIOS=yes \
--env.CM_MLPERF_LOADGEN_COMPLIANCE=yes \
--env.CM_MLPERF_SUBMISSION_RUN=yes \
--env.CM_RUN_MLPERF_ACCURACY=on \
--env.CM_RUN_SUBMISSION_CHECKER=yes \
--env.CM_TAR_SUBMISSION_DIR=yes \
--env.CM_MLPERF_SUBMISSION_GENERATION_STYLE=full \
--env.CM_MLPERF_INFERENCE_VERSION=5.0-dev \
--env.CM_RUN_MLPERF_INFERENCE_APP_DEFAULTS=r4.1-dev_default \
--env.CM_MLPERF_LOADGEN_ALL_MODES=yes \
--env.CM_MLPERF_INFERENCE_SOURCE_VERSION=5.0.4 \
--env.CM_MLPERF_LAST_RELEASE=v5.0 \
--env.CM_TMP_PIP_VERSION_STRING= \
--env.CM_MODEL=retinanet \
--env.CM_MLPERF_CLEAN_SUBMISSION_DIR=yes \
--env.CM_RERUN=yes \
--env.CM_MLPERF_LOADGEN_EXTRA_OPTIONS= \
--env.CM_MLPERF_LOADGEN_MODE=performance \
--env.CM_MLPERF_LOADGEN_SCENARIO=MultiStream \
--env.CM_MLPERF_LOADGEN_SCENARIOS,=SingleStream,Offline,MultiStream,Server \
--env.CM_MLPERF_LOADGEN_MODES,=performance,accuracy \
--env.CM_OUTPUT_FOLDER_NAME=valid_results \
--env.CM_DOCKER_REUSE_EXISTING_CONTAINER=yes \
--env.CM_DOCKER_DETACHED_MODE=yes \
--env.CM_MLPERF_INFERENCE_RESULTS_DIR_=/home/arjun/gh_action_results/valid_results \
--env.CM_DOCKER_CONTAINER_ID=e891b1c381b6 \
--env.CM_MLPERF_LOADGEN_COMPLIANCE_TEST=TEST01 \
--add_deps_recursive.compiler.tags=gcc \
--add_deps_recursive.coco2014-original.tags=_full \
--add_deps_recursive.coco2014-preprocessed.tags=_full \
--add_deps_recursive.imagenet-original.tags=_full \
--add_deps_recursive.imagenet-preprocessed.tags=_full \
--add_deps_recursive.openimages-original.tags=_full \
--add_deps_recursive.openimages-preprocessed.tags=_full \
--add_deps_recursive.openorca-original.tags=_full \
--add_deps_recursive.openorca-preprocessed.tags=_full \
--add_deps_recursive.coco2014-dataset.tags=_full \
--add_deps_recursive.igbh-dataset.tags=_full \
--add_deps_recursive.get-mlperf-inference-results-dir.tags=_version.r4_1-dev \
--add_deps_recursive.get-mlperf-inference-submission-dir.tags=_version.r4_1-dev \
--add_deps_recursive.mlperf-inference-nvidia-scratch-space.tags=_version.r4_1-dev \
--adr.compiler.tags=gcc \
--adr.coco2014-original.tags=_full \
--adr.coco2014-preprocessed.tags=_full \
--adr.imagenet-original.tags=_full \
--adr.imagenet-preprocessed.tags=_full \
--adr.openimages-original.tags=_full \
--adr.openimages-preprocessed.tags=_full \
--adr.openorca-original.tags=_full \
--adr.openorca-preprocessed.tags=_full \
--adr.coco2014-dataset.tags=_full \
--adr.igbh-dataset.tags=_full \
--adr.get-mlperf-inference-results-dir.tags=_version.r4_1-dev \
--adr.get-mlperf-inference-submission-dir.tags=_version.r4_1-dev \
--adr.mlperf-inference-nvidia-scratch-space.tags=_version.r4_1-dev \
--v=False \
--print_env=False \
--print_deps=False \
--dump_version_info=True \
--env.CM_DATASET_OPENIMAGES_PATH=/home/cmuser/CM/repos/local/cache/cae305b311f24865/install/validation/data \
--env.CM_OPENIMAGES_CALIBRATION_DATASET_PATH=/home/cmuser/CM/repos/local/cache/f5639d7dd9ba459b/install/calibration/data \
--env.CM_DATASET_OPENIMAGES_ANNOTATIONS_DIR_PATH=/home/cmuser/CM/repos/local/cache/e44ef68a8a294e31 \
--env.OUTPUT_BASE_DIR=/cm-mount/home/arjun/gh_action_results \
--env.CM_MLPERF_INFERENCE_SUBMISSION_DIR=/cm-mount/home/arjun/gh_action_submissions \
--env.MLPERF_SCRATCH_PATH=/home/cmuser/CM/repos/local/cache/4db00c74da1e44c8
```
*Note that if you want to use the [latest automation recipes](https://docs.mlcommons.org/inference) for MLPerf (CM scripts),
you should simply reload mlcommons@mlperf-automations without checkout and clean CM cache as follows:*

```bash
cm rm repo mlcommons@mlperf-automations
cm pull repo mlcommons@mlperf-automations
cm rm cache -f

```

## Results

Platform: RTX4090x2-nvidia_original-gpu-tensorrt-vdefault-default_config

Model Precision: int8

### Accuracy Results
`mAP`: `37.356`, Required accuracy for closed division `>= 37.1745`

### Performance Results
`Samples per query`: `5654438.0`
7 changes: 7 additions & 0 deletions ...retinanet/multistream/RTX4090x2-nvidia_original-gpu-tensorrt-vdefault-default_config.json
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
{
"starting_weights_filename": "https://zenodo.org/record/6617981/files/resnext50_32x4d_fpn.pth",
"retraining": "no",
"input_data_types": "int8",
"weight_data_types": "int8",
"weight_transformations": "quantization, affine fusion"
}
120 changes: 120 additions & 0 deletions ..._original-gpu-tensorrt-vdefault-default_config/retinanet/multistream/accuracy_console.out
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,120 @@
[2024-12-23 02:36:38,287 main.py:229 INFO] Detected system ID: KnownSystem.RTX4090x2
[2024-12-23 02:36:38,369 harness.py:249 INFO] The harness will load 2 plugins: ['build/plugins/NMSOptPlugin/libnmsoptplugin.so', 'build/plugins/retinanetConcatPlugin/libretinanetconcatplugin.so']
[2024-12-23 02:36:38,369 generate_conf_files.py:107 INFO] Generated measurements/ entries for RTX4090x2_TRT/retinanet/MultiStream
[2024-12-23 02:36:38,369 __init__.py:46 INFO] Running command: ./build/bin/harness_default --plugins="build/plugins/NMSOptPlugin/libnmsoptplugin.so,build/plugins/retinanetConcatPlugin/libretinanetconcatplugin.so" --logfile_outdir="/cm-mount/home/arjun/gh_action_results/valid_results/RTX4090x2-nvidia_original-gpu-tensorrt-vdefault-default_config/retinanet/multistream/accuracy" --logfile_prefix="mlperf_log_" --performance_sample_count=64 --test_mode="AccuracyOnly" --gpu_copy_streams=1 --gpu_inference_streams=1 --use_deque_limit=true --gpu_batch_size=2 --map_path="data_maps/open-images-v6-mlperf/val_map.txt" --mlperf_conf_path="/home/cmuser/CM/repos/local/cache/5860c00d55d14786/inference/mlperf.conf" --tensor_path="build/preprocessed_data/open-images-v6-mlperf/validation/Retinanet/int8_linear" --use_graphs=true --user_conf_path="/home/cmuser/CM/repos/mlcommons@mlperf-automations/script/generate-mlperf-inference-user-conf/tmp/c44de86dadc4420199ad25d6fa563275.conf" --gpu_engines="./build/engines/RTX4090x2/retinanet/MultiStream/retinanet-MultiStream-gpu-b2-int8.lwis_k_99_MaxP.plan" --max_dlas=0 --scenario MultiStream --model retinanet --response_postprocess openimageeffnms
[2024-12-23 02:36:38,369 __init__.py:53 INFO] Overriding Environment
benchmark : Benchmark.Retinanet
buffer_manager_thread_count : 0
data_dir : /home/cmuser/CM/repos/local/cache/4db00c74da1e44c8/data
disable_beta1_smallk : True
gpu_batch_size : 2
gpu_copy_streams : 1
gpu_inference_streams : 1
input_dtype : int8
input_format : linear
log_dir : /home/cmuser/CM/repos/local/cache/94a57f78972843c6/repo/closed/NVIDIA/build/logs/2024.12.23-02.36.37
map_path : data_maps/open-images-v6-mlperf/val_map.txt
mlperf_conf_path : /home/cmuser/CM/repos/local/cache/5860c00d55d14786/inference/mlperf.conf
multi_stream_expected_latency_ns : 0
multi_stream_samples_per_query : 8
multi_stream_target_latency_percentile : 99
precision : int8
preprocessed_data_dir : /home/cmuser/CM/repos/local/cache/4db00c74da1e44c8/preprocessed_data
scenario : Scenario.MultiStream
system : SystemConfiguration(host_cpu_conf=CPUConfiguration(layout={CPU(name='Intel(R) Xeon(R) w7-2495X', architecture=<CPUArchitecture.x86_64: AliasedName(name='x86_64', aliases=(), patterns=())>, core_count=24, threads_per_core=2): 1}), host_mem_conf=MemoryConfiguration(host_memory_capacity=Memory(quantity=197.334532, byte_suffix=<ByteSuffix.GB: (1000, 3)>, _num_bytes=197334532000), comparison_tolerance=0.05), accelerator_conf=AcceleratorConfiguration(layout=defaultdict(<class 'int'>, {GPU(name='NVIDIA GeForce RTX 4090', accelerator_type=<AcceleratorType.Discrete: AliasedName(name='Discrete', aliases=(), patterns=())>, vram=Memory(quantity=23.98828125, byte_suffix=<ByteSuffix.GiB: (1024, 3)>, _num_bytes=25757220864), max_power_limit=450.0, pci_id='0x268410DE', compute_sm=89): 1, GPU(name='NVIDIA GeForce RTX 4090', accelerator_type=<AcceleratorType.Discrete: AliasedName(name='Discrete', aliases=(), patterns=())>, vram=Memory(quantity=23.98828125, byte_suffix=<ByteSuffix.GiB: (1024, 3)>, _num_bytes=25757220864), max_power_limit=500.0, pci_id='0x268410DE', compute_sm=89): 1})), numa_conf=NUMAConfiguration(numa_nodes={}, num_numa_nodes=1), system_id='RTX4090x2')
tensor_path : build/preprocessed_data/open-images-v6-mlperf/validation/Retinanet/int8_linear
test_mode : AccuracyOnly
use_deque_limit : True
use_graphs : True
user_conf_path : /home/cmuser/CM/repos/mlcommons@mlperf-automations/script/generate-mlperf-inference-user-conf/tmp/c44de86dadc4420199ad25d6fa563275.conf
system_id : RTX4090x2
config_name : RTX4090x2_retinanet_MultiStream
workload_setting : WorkloadSetting(HarnessType.LWIS, AccuracyTarget.k_99, PowerSetting.MaxP)
optimization_level : plugin-enabled
num_profiles : 1
config_ver : lwis_k_99_MaxP
accuracy_level : 99%
inference_server : lwis
skip_file_checks : False
power_limit : None
cpu_freq : None
&&&& RUNNING Default_Harness # ./build/bin/harness_default
[I] mlperf.conf path: /home/cmuser/CM/repos/local/cache/5860c00d55d14786/inference/mlperf.conf
[I] user.conf path: /home/cmuser/CM/repos/mlcommons@mlperf-automations/script/generate-mlperf-inference-user-conf/tmp/c44de86dadc4420199ad25d6fa563275.conf
Creating QSL.
Finished Creating QSL.
Setting up SUT.
[I] [TRT] Loaded engine size: 72 MiB
[I] [TRT] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +6, GPU +10, now: CPU 125, GPU 881 (MiB)
[I] [TRT] [MemUsageChange] Init cuDNN: CPU +2, GPU +10, now: CPU 127, GPU 891 (MiB)
[I] [TRT] [MemUsageChange] TensorRT-managed allocation in engine deserialization: CPU +0, GPU +68, now: CPU 0, GPU 68 (MiB)
[I] Device:0.GPU: [0] ./build/engines/RTX4090x2/retinanet/MultiStream/retinanet-MultiStream-gpu-b2-int8.lwis_k_99_MaxP.plan has been successfully loaded.
[I] [TRT] Loaded engine size: 72 MiB
[W] [TRT] Using an engine plan file across different models of devices is not recommended and is likely to affect performance or even cause errors.
[I] [TRT] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +6, GPU +10, now: CPU 159, GPU 624 (MiB)
[I] [TRT] [MemUsageChange] Init cuDNN: CPU +1, GPU +10, now: CPU 160, GPU 634 (MiB)
[I] [TRT] [MemUsageChange] TensorRT-managed allocation in engine deserialization: CPU +0, GPU +69, now: CPU 0, GPU 137 (MiB)
[I] Device:1.GPU: [0] ./build/engines/RTX4090x2/retinanet/MultiStream/retinanet-MultiStream-gpu-b2-int8.lwis_k_99_MaxP.plan has been successfully loaded.
[E] [TRT] 3: [runtime.cpp::~Runtime::401] Error Code 3: API Usage Error (Parameter check failed at: runtime/rt/runtime.cpp::~Runtime::401, condition: mEngineCounter.use_count() == 1 Destroying a runtime before destroying deserialized engines created by the runtime leads to undefined behavior.)
[I] [TRT] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +8, now: CPU 88, GPU 893 (MiB)
[I] [TRT] [MemUsageChange] Init cuDNN: CPU +0, GPU +8, now: CPU 88, GPU 901 (MiB)
[I] [TRT] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +1, GPU +1528, now: CPU 1, GPU 1665 (MiB)
[I] [TRT] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +8, now: CPU 89, GPU 636 (MiB)
[I] [TRT] [MemUsageChange] Init cuDNN: CPU +0, GPU +8, now: CPU 89, GPU 644 (MiB)
[I] [TRT] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +1528, now: CPU 1, GPU 3193 (MiB)
[I] Start creating CUDA graphs
[I] Capture 2 CUDA graphs
[I] Capture 2 CUDA graphs
[I] Finish creating CUDA graphs
[I] Creating batcher thread: 0 EnableBatcherThreadPerDevice: false
Finished setting up SUT.
Starting warmup. Running for a minimum of 5 seconds.
Finished warmup. Ran for 5.14292s.
Starting running actual test.

No warnings encountered during test.

No errors encountered during test.
Finished running actual test.
Device Device:0.GPU processed:
6195 batches of size 2
Memcpy Calls: 0
PerSampleCudaMemcpy Calls: 0
BatchedCudaMemcpy Calls: 6195
Device Device:1.GPU processed:
6197 batches of size 2
Memcpy Calls: 0
PerSampleCudaMemcpy Calls: 0
BatchedCudaMemcpy Calls: 6197
&&&& PASSED Default_Harness # ./build/bin/harness_default
[2024-12-23 02:37:18,118 run_harness.py:166 INFO] Result: Accuracy run detected.
[2024-12-23 02:37:18,118 __init__.py:46 INFO] Running command: python3 /home/cmuser/CM/repos/local/cache/94a57f78972843c6/repo/closed/NVIDIA/build/inference/vision/classification_and_detection/tools/accuracy-openimages.py --mlperf-accuracy-file /cm-mount/home/arjun/gh_action_results/valid_results/RTX4090x2-nvidia_original-gpu-tensorrt-vdefault-default_config/retinanet/multistream/accuracy/mlperf_log_accuracy.json --openimages-dir /home/cmuser/CM/repos/local/cache/4db00c74da1e44c8/preprocessed_data/open-images-v6-mlperf --output-file build/retinanet-results.json
loading annotations into memory...
Done (t=0.52s)
creating index...
index created!
Loading and preparing results...
DONE (t=17.77s)
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=133.59s).
Accumulating evaluation results...
DONE (t=31.87s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.374
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.522
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.404
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.022
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.125
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.413
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.419
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.599
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.628
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.082
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.345
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.678
mAP=37.356%

======================== Result summaries: ========================

Loading

0 comments on commit 13a16ab

Please sign in to comment.