AOT mechanism.

cms-ml · Apr 10, 2024 · 33b2320 · 33b2320
1 parent 3988304
commit 33b2320
Show file tree

Hide file tree

Showing 57 changed files with 811 additions and 422 deletions.
diff --git a/.flake8 b/.flake8
@@ -1,6 +1,6 @@
 [flake8]
 
-# line length of 100 is recommended, but set it to a forgiving value
+# line length
 max-line-length = 120
 
 # codes of errors to ignore

diff --git a/.gitattributes b/.gitattributes
@@ -6,3 +6,5 @@
 *.pb filter=lfs diff=lfs merge=lfs -text
 *.pb.txt filter=lfs diff=lfs merge=lfs -text
 *.pbtxt filter=lfs diff=lfs merge=lfs -text
+*.index filter=lfs diff=lfs merge=lfs -text
+*.data-* filter=lfs diff=lfs merge=lfs -text
diff --git a/.gitignore b/.gitignore
@@ -12,8 +12,6 @@
 *.npz
 *.h5
 *.hdf5
-*.json
-*.yaml
 *.out
 *.parquet
 .coverage

diff --git a/cmssw/MLProf/RuntimeMeasurement/plugins/aot/TFAOTInference.cc b/cmssw/MLProf/RuntimeMeasurement/plugins/aot/TFAOTInference.cc
@@ -59,7 +59,7 @@ class TFAOTInference : public edm::stream::EDAnalyzer<> {
   std::normal_distribution<double> normalPDFDouble_;
 
   // aot model
-  tfaot::Model<tfaot_model::test_simple> model_;
+  // INSERT=model
 };
 
 void TFAOTInference::fillDescriptions(edm::ConfigurationDescriptions& descriptions) {
@@ -108,37 +108,10 @@ TFAOTInference::TFAOTInference(const edm::ParameterSet& config)
     throw cms::Exception("InvalidInputType") << "input type unknown: " << inputTypeStr_;
   }
 
-  // parse and register batch rules
-  for (size_t i = 0; i < batchRules_.size(); i++) {
-    std::string rule = batchRules_[i];
-
-    // extract the target batch size from the front
-    auto pos = rule.find(":");
-    if (pos == std::string::npos) {
-      throw cms::Exception("InvalidBatchRule") << "invalid batch rule format: " << batchRules_[i];
-    }
-    size_t batchSize = std::stoi(rule.substr(0, pos));
-    rule = rule.substr(pos + 1);
-
-    // loop through remaining comma-separated composite batch size
-    std::vector<size_t> sizes;
-    size_t sumSizes = 0;
-    while (!rule.empty()) {
-      pos = rule.find(",");
-      sizes.push_back(std::stoi(rule.substr(0, pos)));
-      sumSizes += sizes.back();
-      rule = rule.substr(pos + 1);
-    }
-
-    // the sum of composite batch sizes should never be smaller than the target batch size
-    if (sumSizes < batchSize) {
-      throw cms::Exception("InvalidBatchRule")
-          << "sum of composite batch sizes is smaller than target batch size: " << batchRules_[i];
-    }
-
-    // register the batch rule
-    model_.setBatchRule(batchSize, sizes, sumSizes - batchSize);
-    std::cout << "registered batch rule for batch size " << batchSize << std::endl;
+  // register batch rules
+  for (const auto& rule : batchRules_) {
+    model_.setBatchRule(rule);
+    std::cout << "registered batch rule " << rule << std::endl;
   }
 }
 

diff --git a/cmssw/MLProf/RuntimeMeasurement/test/onnx_runtime_cfg.py b/cmssw/MLProf/RuntimeMeasurement/test/onnx_runtime_cfg.py
@@ -5,19 +5,69 @@
 
 # setup minimal options
 options = VarParsing("python")
+options.setDefault("maxEvents", 1)
+options.register(
+    "csvFile",
+    "results.csv",
+    VarParsing.multiplicity.singleton,
+    VarParsing.varType.string,
+    "path of the csv file to save results",
+)
 options.register(
     "batchSize",
     1,
     VarParsing.multiplicity.singleton,
     VarParsing.varType.int,
-    "Batch size to be tested",
+    "batch size to be tested",
 )
 options.register(
-    "csvFile",
-    "results.csv",
+    "inputType",
+    "",
+    VarParsing.multiplicity.singleton,
+    VarParsing.varType.string,
+    "input type; 'random', 'incremental', 'zeros', or 'ones'",
+)
+options.register(
+    "nCalls",
+    100,
+    VarParsing.multiplicity.singleton,
+    VarParsing.varType.int,
+    "number of evaluation calls for averaging",
+)
+options.register(
+    "graphPath",
+    "",
     VarParsing.multiplicity.singleton,
     VarParsing.varType.string,
-    "The path of the csv file to save results",
+    "path to the graph file",
+)
+options.register(
+    "inputTensorNames",
+    [],
+    VarParsing.multiplicity.list,
+    VarParsing.varType.string,
+    "names of the input tensors",
+)
+options.register(
+    "outputTensorNames",
+    [],
+    VarParsing.multiplicity.list,
+    VarParsing.varType.string,
+    "Names of the output tensors",
+)
+options.register(
+    "inputRanks",
+    [],
+    VarParsing.multiplicity.list,
+    VarParsing.varType.int,
+    "ranks of the input tensors",
+)
+options.register(
+    "flatInputSizes",
+    [],
+    VarParsing.multiplicity.list,
+    VarParsing.varType.int,
+    "sizes of the flattened input tensors; the amount should match the sum of ranks",
 )
 options.parseArguments()
 
@@ -29,11 +79,11 @@
 process.load("FWCore.MessageService.MessageLogger_cfi")
 process.MessageLogger.cerr.FwkReport.reportEvery = 1
 process.maxEvents = cms.untracked.PSet(
-    input=cms.untracked.int32(__N_EVENTS__),  # noqa
+    input=cms.untracked.int32(options.maxEvents),
 )
 process.source = cms.Source(
     "PoolSource",
-    fileNames=cms.untracked.vstring(*__INPUT_FILES__),  # noqa
+    fileNames=cms.untracked.vstring(options.inputFiles),
 )
 
 # process options
@@ -44,20 +94,20 @@
 
 # multi-threading options
 process.options.numberOfThreads = cms.untracked.uint32(1)
-process.options.numberOfStreams = cms.untracked.uint32(0)
+process.options.numberOfStreams = cms.untracked.uint32(1)
 process.options.numberOfConcurrentLuminosityBlocks = cms.untracked.uint32(1)
 
 # setup the plugin
 process.load("MLProf.RuntimeMeasurement.onnxInference_cfi")
-process.onnxInference.graphPath = cms.string("__GRAPH_PATH__")
-process.onnxInference.inputTensorNames = cms.vstring(__INPUT_TENSOR_NAMES__)  # noqa
-process.onnxInference.outputTensorNames = cms.vstring(__OUTPUT_TENSOR_NAMES__)  # noqa
+process.onnxInference.graphPath = cms.string(options.graphPath)
+process.onnxInference.inputTensorNames = cms.vstring(options.inputTensorNames)
+process.onnxInference.outputTensorNames = cms.vstring(options.outputTensorNames)
 process.onnxInference.outputFile = cms.string(options.csvFile)
-process.onnxInference.inputType = cms.string("__INPUT_TYPE__")
-process.onnxInference.inputRanks = cms.vint32(__INPUT_RANKS__)  # noqa
-process.onnxInference.flatInputSizes = cms.vint32(__FLAT_INPUT_SIZES__)  # noqa
+process.onnxInference.inputType = cms.string(options.inputType)
+process.onnxInference.inputRanks = cms.vint32(options.inputRanks)
+process.onnxInference.flatInputSizes = cms.vint32(options.flatInputSizes)
 process.onnxInference.batchSize = cms.int32(options.batchSize)
-process.onnxInference.nCalls = cms.int32(__N_CALLS__)  # noqa
+process.onnxInference.nCalls = cms.int32(options.nCalls)
 
 # define what to run in the path
 process.p = cms.Path(process.onnxInference)
diff --git a/cmssw/MLProf/RuntimeMeasurement/test/tf_runtime_cfg.py b/cmssw/MLProf/RuntimeMeasurement/test/tf_runtime_cfg.py
@@ -5,19 +5,69 @@
 
 # setup minimal options
 options = VarParsing("python")
+options.setDefault("maxEvents", 1)
+options.register(
+    "csvFile",
+    "results.csv",
+    VarParsing.multiplicity.singleton,
+    VarParsing.varType.string,
+    "path of the csv file to save results",
+)
 options.register(
     "batchSize",
     1,
     VarParsing.multiplicity.singleton,
     VarParsing.varType.int,
-    "Batch size to be tested",
+    "batch size to be tested",
 )
 options.register(
-    "csvFile",
-    "results.csv",
+    "inputType",
+    "",
+    VarParsing.multiplicity.singleton,
+    VarParsing.varType.string,
+    "input type; 'random', 'incremental', 'zeros', or 'ones'",
+)
+options.register(
+    "nCalls",
+    100,
+    VarParsing.multiplicity.singleton,
+    VarParsing.varType.int,
+    "number of evaluation calls for averaging",
+)
+options.register(
+    "graphPath",
+    "",
     VarParsing.multiplicity.singleton,
     VarParsing.varType.string,
-    "The path of the csv file to save results",
+    "path to the graph file",
+)
+options.register(
+    "inputTensorNames",
+    [],
+    VarParsing.multiplicity.list,
+    VarParsing.varType.string,
+    "names of the input tensors",
+)
+options.register(
+    "outputTensorNames",
+    [],
+    VarParsing.multiplicity.list,
+    VarParsing.varType.string,
+    "Names of the output tensors",
+)
+options.register(
+    "inputRanks",
+    [],
+    VarParsing.multiplicity.list,
+    VarParsing.varType.int,
+    "ranks of the input tensors",
+)
+options.register(
+    "flatInputSizes",
+    [],
+    VarParsing.multiplicity.list,
+    VarParsing.varType.int,
+    "sizes of the flattened input tensors; the amount should match the sum of ranks",
 )
 options.parseArguments()
 
@@ -29,11 +79,11 @@
 process.load("FWCore.MessageService.MessageLogger_cfi")
 process.MessageLogger.cerr.FwkReport.reportEvery = 1
 process.maxEvents = cms.untracked.PSet(
-    input=cms.untracked.int32(__N_EVENTS__),  # noqa
+    input=cms.untracked.int32(options.maxEvents),
 )
 process.source = cms.Source(
     "PoolSource",
-    fileNames=cms.untracked.vstring(*__INPUT_FILES__),  # noqa
+    fileNames=cms.untracked.vstring(options.inputFiles),
 )
 
 # process options
@@ -44,20 +94,20 @@
 
 # multi-threading options
 process.options.numberOfThreads = cms.untracked.uint32(1)
-process.options.numberOfStreams = cms.untracked.uint32(0)
+process.options.numberOfStreams = cms.untracked.uint32(1)
 process.options.numberOfConcurrentLuminosityBlocks = cms.untracked.uint32(1)
 
 # setup the plugin
 process.load("MLProf.RuntimeMeasurement.tfInference_cfi")
-process.tfInference.graphPath = cms.string("__GRAPH_PATH__")
-process.tfInference.inputTensorNames = cms.vstring(__INPUT_TENSOR_NAMES__)  # noqa
-process.tfInference.outputTensorNames = cms.vstring(__OUTPUT_TENSOR_NAMES__)  # noqa
+process.tfInference.graphPath = cms.string(options.graphPath)
+process.tfInference.inputTensorNames = cms.vstring(options.inputTensorNames)
+process.tfInference.outputTensorNames = cms.vstring(options.outputTensorNames)
 process.tfInference.outputFile = cms.string(options.csvFile)
-process.tfInference.inputType = cms.string("__INPUT_TYPE__")
-process.tfInference.inputRanks = cms.vint32(__INPUT_RANKS__)  # noqa
-process.tfInference.flatInputSizes = cms.vint32(__FLAT_INPUT_SIZES__)  # noqa
+process.tfInference.inputType = cms.string(options.inputType)
+process.tfInference.inputRanks = cms.vint32(options.inputRanks)
+process.tfInference.flatInputSizes = cms.vint32(options.flatInputSizes)
 process.tfInference.batchSize = cms.int32(options.batchSize)
-process.tfInference.nCalls = cms.int32(__N_CALLS__)  # noqa
+process.tfInference.nCalls = cms.int32(options.nCalls)
 
 # define what to run in the path
 process.p = cms.Path(process.tfInference)
diff --git a/cmssw/MLProf/RuntimeMeasurement/test/tfaot_runtime_cfg.py b/cmssw/MLProf/RuntimeMeasurement/test/tfaot_runtime_cfg.py
@@ -5,26 +5,42 @@
 
 # setup minimal options
 options = VarParsing("python")
+options.setDefault("maxEvents", 1)
+options.setDefault("maxEvents", 1)
+options.register(
+    "csvFile",
+    "results.csv",
+    VarParsing.multiplicity.singleton,
+    VarParsing.varType.string,
+    "path of the csv file to save results",
+)
 options.register(
     "batchSize",
     1,
     VarParsing.multiplicity.singleton,
     VarParsing.varType.int,
-    "Batch size to be tested",
+    "batch size to be tested",
 )
 options.register(
-    "batchRules",
-    [],
-    VarParsing.multiplicity.list,
+    "inputType",
+    "",
+    VarParsing.multiplicity.singleton,
     VarParsing.varType.string,
-    "Batch rules (format 'target_size:size_1,size_2,...') to be configured",
+    "input type; 'random', 'incremental', 'zeros', or 'ones'",
 )
 options.register(
-    "csvFile",
-    "results.csv",
+    "nCalls",
+    100,
     VarParsing.multiplicity.singleton,
+    VarParsing.varType.int,
+    "number of evaluation calls for averaging",
+)
+options.register(
+    "batchRules",
+    [],
+    VarParsing.multiplicity.list,
     VarParsing.varType.string,
-    "The path of the csv file to save results",
+    "Batch rules (format 'target_size:size_1:size_2:...') to be configured",
 )
 options.parseArguments()
 
@@ -36,11 +52,11 @@
 process.load("FWCore.MessageService.MessageLogger_cfi")
 process.MessageLogger.cerr.FwkReport.reportEvery = 1
 process.maxEvents = cms.untracked.PSet(
-    input=cms.untracked.int32(__N_EVENTS__),  # noqa
+    input=cms.untracked.int32(options.maxEvents),
 )
 process.source = cms.Source(
     "PoolSource",
-    fileNames=cms.untracked.vstring(*__INPUT_FILES__),  # noqa
+    fileNames=cms.untracked.vstring(options.inputFiles),
 )
 
 # process options
@@ -51,16 +67,16 @@
 
 # multi-threading options
 process.options.numberOfThreads = cms.untracked.uint32(1)
-process.options.numberOfStreams = cms.untracked.uint32(0)
+process.options.numberOfStreams = cms.untracked.uint32(1)
 process.options.numberOfConcurrentLuminosityBlocks = cms.untracked.uint32(1)
 
 # setup the plugin
 process.load("MLProf.RuntimeMeasurement.tfaotInference_cfi")
 process.tfaotInference.outputFile = cms.string(options.csvFile)
-process.tfaotInference.inputType = cms.string("__INPUT_TYPE__")
-process.tfaotInference.batchRules = cms.vstring(options.batchRules)
+process.tfaotInference.inputType = cms.string(options.inputType)
+process.tfaotInference.batchRules = cms.vstring([r.replace(".", ",") for r in options.batchRules])
 process.tfaotInference.batchSize = cms.int32(options.batchSize)
-process.tfaotInference.nCalls = cms.int32(__N_CALLS__)  # noqa
+process.tfaotInference.nCalls = cms.int32(options.nCalls)
 
 # define what to run in the path
 process.p = cms.Path(process.tfaotInference)
-Original file line number
+Diff line change
@@ Expand Up / @@ -12,8 +12,6 @@ @@
     *.npz
     *.h5
     *.hdf5
-    *.json
-    *.yaml
     *.out
     *.parquet
     .coverage
@@ Expand Down @@