From 4e06f742d8478cf63711591a7289e3b6d8d38327 Mon Sep 17 00:00:00 2001 From: mariecwhite Date: Wed, 11 Oct 2023 00:20:24 +0000 Subject: [PATCH] Re-enable PyTorch Benchmarks --- .../e2e_test_framework/models/model_groups.py | 55 +--- .../e2e_test_framework/models/torch_models.py | 63 +--- .../generated_e2e_test_fetch_models.cmake | 63 ++++ .../generated_e2e_test_iree_artifacts.cmake | 288 ++++++++++++++++++ 4 files changed, 359 insertions(+), 110 deletions(-) diff --git a/build_tools/python/e2e_test_framework/models/model_groups.py b/build_tools/python/e2e_test_framework/models/model_groups.py index 6c600d9fa192..478f8f4c8fb0 100644 --- a/build_tools/python/e2e_test_framework/models/model_groups.py +++ b/build_tools/python/e2e_test_framework/models/model_groups.py @@ -9,7 +9,7 @@ from e2e_test_framework.models import ( matmul, tflite_models, - # torch_models, + torch_models, tf_models, jax_models, ) @@ -65,9 +65,6 @@ common_definitions.CpuBenchmarkConfig( model=tf_models.MINILM_L12_H384_UNCASED_INT32_SEQLEN128, threads=[1, 8] ), - # common_definitions.CpuBenchmarkConfig( - # model=torch_models.EFFICIENTNET_V2_S_FP32_TORCH, threads=[1, 8] - # ), # Large models. # TODO: These models should be running at 8, 13, 28 threads but we use 8 for now until new hardware becomes available. common_definitions.CpuBenchmarkConfig( @@ -76,9 +73,6 @@ common_definitions.CpuBenchmarkConfig( model=tf_models.BERT_LARGE_TF_FP32_SEQLEN384, threads=[8] ), - # common_definitions.CpuBenchmarkConfig( - # model=torch_models.EFFICIENTNET_B7_FP32_TORCH, threads=[8] - # ), ] # A subset of `x86_64_MODELS_AND_THREADS`. @@ -122,16 +116,10 @@ common_definitions.CpuBenchmarkConfig( model=tf_models.MINILM_L12_H384_UNCASED_INT32_SEQLEN128, threads=[8] ), - # common_definitions.CpuBenchmarkConfig( - # model=torch_models.EFFICIENTNET_V2_S_FP32_TORCH, threads=[8] - # ), # Large models. common_definitions.CpuBenchmarkConfig( model=tf_models.BERT_LARGE_TF_FP32_SEQLEN384, threads=[8] ), - # common_definitions.CpuBenchmarkConfig( - # model=torch_models.EFFICIENTNET_B7_FP32_TORCH, threads=[8] - # ), ] X86_64_BENCHMARK_CONFIG_LONG = [ @@ -184,34 +172,10 @@ # Batched Torch models. -# BERT_LARGE_TORCH_BATCHES = [ -# model -# for batch_size, model in torch_models.BERT_LARGE_384_FP32_TORCH_BATCHES.items() -# # Higher batch sizes disabled due to OOM https://github.com/openxla/iree/issues/14668. -# if batch_size < 64 -# ] -# -# BERT_LARGE_FP16_TORCH_BATCHES = [ -# model -# for batch_size, model in torch_models.BERT_LARGE_384_FP16_TORCH_BATCHES.items() -# # Batchsize 1 is included separately in CUDA_MODELS -# # Higher batch sizes disabled due to OOM https://github.com/openxla/iree/issues/14668. -# if batch_size != 1 and batch_size < 64 -# ] -# -# RESNET50_TORCH_BATCHES = [ -# model -# for batch_size, model in torch_models.RESNET50_3X224X224_FP32_TORCH_BATCHES.items() -# # Higher batch sizes disabled due to OOM https://github.com/openxla/iree/issues/14668. -# if batch_size < 2048 -# ] -# -# RESNET50_FP16_TORCH_BATCHES = list( -# model -# for batch_size, model in torch_models.RESNET50_3X224X224_FP16_TORCH_BATCHES.items() -# # Higher batch sizes disabled due to OOM https://github.com/openxla/iree/issues/14668. -# if batch_size < 2048 -# ) +BERT_LARGE_TORCH_BATCHES = [ + model + for batch_size, model in torch_models.BERT_LARGE_384_FP32_TORCH_BATCHES.items() +] # Batched Tensorflow models. BERT_LARGE_TF_BATCHES = [ @@ -267,25 +231,20 @@ # PyTorch model are disabled due to https://github.com/openxla/iree/issues/14993. # torch_models.MODEL_CLIP_TEXT_SEQLEN64_FP32_TORCH, # torch_models.MODEL_UNET_2D_FP32_TORCH, - # torch_models.EFFICIENTNET_B7_FP32_TORCH, - # torch_models.BERT_LARGE_384_FP16_TORCH_BATCHES[1], - # torch_models.EFFICIENTNET_V2_S_FP16_TORCH, ] CUDA_MODELS_LONG = ( RESNET50_TF_BATCHES + BERT_LARGE_TF_BATCHES + T5_LARGE_TF_BATCHES - # + BERT_LARGE_TORCH_BATCHES - # + RESNET50_TORCH_BATCHES - # + RESNET50_FP16_TORCH_BATCHES - # + BERT_LARGE_FP16_TORCH_BATCHES + + BERT_LARGE_TORCH_BATCHES + BERT_LARGE_JAX_BATCHES + RESNET50_JAX_BATCHES + T5_LARGE_JAX_BATCHES ) VULKAN_MODELS = [ + # PyTorch model are disabled due to https://github.com/openxla/iree/issues/14993. # torch_models.MODEL_CLIP_TEXT_SEQLEN64_FP32_TORCH, # torch_models.MODEL_UNET_2D_FP32_TORCH, ] diff --git a/build_tools/python/e2e_test_framework/models/torch_models.py b/build_tools/python/e2e_test_framework/models/torch_models.py index 0d5f41622750..558e238a7112 100644 --- a/build_tools/python/e2e_test_framework/models/torch_models.py +++ b/build_tools/python/e2e_test_framework/models/torch_models.py @@ -92,17 +92,11 @@ ID_FORMAT = string.Template("${model_id}-batch-${batch_size}") NAME_FORMAT = string.Template("${name}Batch${batch_size}") BERT_LARGE_FP32_URL = string.Template( - "https://storage.googleapis.com/iree-model-artifacts/pytorch/pt_models_20230813.929_1692010793/BERT_LARGE_FP32_PT_384XI32_BATCH${batch_size}/linalg.mlirbc" + "https://storage.googleapis.com/iree-model-artifacts/pytorch/pt_models_20231010.987_1696982151/BERT_LARGE_FP32_PT_384XI32_BATCH${batch_size}/linalg.mlirbc" ) BERT_LARGE_FP16_URL = string.Template( "https://storage.googleapis.com/iree-model-artifacts/pytorch/pt_models_20230816.932_1692245822/BERT_LARGE_FP16_PT_384XI32_BATCH${batch_size}/linalg.mlirbc" ) -RESNET50_FP32_URL = string.Template( - "https://storage.googleapis.com/iree-model-artifacts/pytorch/pt_models_20230813.929_1692010793/RESNET50_FP32_PT_3X224X224XF32_BATCH${batch_size}/linalg.mlirbc" -) -RESNET50_FP16_URL = string.Template( - "https://storage.googleapis.com/iree-model-artifacts/pytorch/pt_models_20230816.932_1692245822/RESNET50_FP16_PT_3X224X224XF16_BATCH${batch_size}/linalg.mlirbc" -) # Converted from https://huggingface.co/docs/transformers/v4.27.2/en/model_doc/bert#transformers.BertModel BERT_LARGE_384_FP32_TORCH_BATCHES = model_utils.generate_batch_models( @@ -122,58 +116,3 @@ ], batch_sizes=[1, 16, 24, 32, 48, 64, 512, 1024, 1280], ) - -# FP16 Versions -BERT_LARGE_384_FP16_TORCH_BATCHES = model_utils.generate_batch_models( - id_template=model_utils.partial_template_substitute( - ID_FORMAT, model_id=unique_ids.MODEL_BERT_LARGE_384_FP16_TORCH - ), - name_template=model_utils.partial_template_substitute( - NAME_FORMAT, name="BertLargefp16PT" - ), - tags=["fp16", "transformer", "seqlen384"], - source_type=common_definitions.ModelSourceType.EXPORTED_LINALG_MLIR, - source_url_template=BERT_LARGE_FP16_URL, - entry_function="forward", - input_type_templates=[ - string.Template("${batch_size}x384xi64"), - string.Template("${batch_size}x384xi64"), - ], - # Batches >=512 are disabled due to OOM: https://github.com/openxla/openxla-benchmark/issues/122. - # batch_sizes=[1, 16, 24, 32, 48, 64, 512, 1024, 1280], - batch_sizes=[1, 16, 24, 32, 48, 64], -) - -# Converted from https://pytorch.org/vision/main/models/generated/torchvision.models.resnet50.html -RESNET50_3X224X224_FP32_TORCH_BATCHES = model_utils.generate_batch_models( - id_template=model_utils.partial_template_substitute( - ID_FORMAT, model_id=unique_ids.MODEL_RESNET50_3X224X224_FP32_TORCH - ), - name_template=model_utils.partial_template_substitute( - NAME_FORMAT, name="Resnet50PT" - ), - tags=["fp32", "cnn"], - source_type=common_definitions.ModelSourceType.EXPORTED_LINALG_MLIR, - source_url_template=RESNET50_FP32_URL, - entry_function="forward", - input_type_templates=[string.Template("${batch_size}x3x224x224xf32")], - batch_sizes=[1, 8, 64, 128, 256, 2048], -) - -# FP16 Versions -RESNET50_3X224X224_FP16_TORCH_BATCHES = model_utils.generate_batch_models( - id_template=model_utils.partial_template_substitute( - ID_FORMAT, model_id=unique_ids.MODEL_RESNET50_3X224X224_FP16_TORCH - ), - name_template=model_utils.partial_template_substitute( - NAME_FORMAT, name="Resnet50fp16PT" - ), - tags=["fp32", "cnn"], - source_type=common_definitions.ModelSourceType.EXPORTED_LINALG_MLIR, - source_url_template=RESNET50_FP16_URL, - entry_function="forward", - input_type_templates=[string.Template("${batch_size}x3x224x224xf16")], - # Batch 2048 is disabled due to OOM: https://github.com/openxla/openxla-benchmark/issues/122 - # batch_sizes=[1, 8, 64, 128, 256, 2048], - batch_sizes=[1, 8, 64, 128, 256], -) diff --git a/tests/e2e/test_artifacts/generated_e2e_test_fetch_models.cmake b/tests/e2e/test_artifacts/generated_e2e_test_fetch_models.cmake index c88a205fc755..cdcc4e46cd7c 100644 --- a/tests/e2e/test_artifacts/generated_e2e_test_fetch_models.cmake +++ b/tests/e2e/test_artifacts/generated_e2e_test_fetch_models.cmake @@ -285,6 +285,69 @@ iree_fetch_artifact( UNPACK ) +iree_fetch_artifact( + NAME "model-BertLargePTBatch1" + SOURCE_URL "https://storage.googleapis.com/iree-model-artifacts/pytorch/pt_models_20231010.987_1696982151/BERT_LARGE_FP32_PT_384XI32_BATCH1/linalg.mlirbc" + OUTPUT "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch1.mlirbc" + UNPACK +) + +iree_fetch_artifact( + NAME "model-BertLargePTBatch16" + SOURCE_URL "https://storage.googleapis.com/iree-model-artifacts/pytorch/pt_models_20231010.987_1696982151/BERT_LARGE_FP32_PT_384XI32_BATCH16/linalg.mlirbc" + OUTPUT "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch16.mlirbc" + UNPACK +) + +iree_fetch_artifact( + NAME "model-BertLargePTBatch24" + SOURCE_URL "https://storage.googleapis.com/iree-model-artifacts/pytorch/pt_models_20231010.987_1696982151/BERT_LARGE_FP32_PT_384XI32_BATCH24/linalg.mlirbc" + OUTPUT "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch24.mlirbc" + UNPACK +) + +iree_fetch_artifact( + NAME "model-BertLargePTBatch32" + SOURCE_URL "https://storage.googleapis.com/iree-model-artifacts/pytorch/pt_models_20231010.987_1696982151/BERT_LARGE_FP32_PT_384XI32_BATCH32/linalg.mlirbc" + OUTPUT "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch32.mlirbc" + UNPACK +) + +iree_fetch_artifact( + NAME "model-BertLargePTBatch48" + SOURCE_URL "https://storage.googleapis.com/iree-model-artifacts/pytorch/pt_models_20231010.987_1696982151/BERT_LARGE_FP32_PT_384XI32_BATCH48/linalg.mlirbc" + OUTPUT "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch48.mlirbc" + UNPACK +) + +iree_fetch_artifact( + NAME "model-BertLargePTBatch64" + SOURCE_URL "https://storage.googleapis.com/iree-model-artifacts/pytorch/pt_models_20231010.987_1696982151/BERT_LARGE_FP32_PT_384XI32_BATCH64/linalg.mlirbc" + OUTPUT "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch64.mlirbc" + UNPACK +) + +iree_fetch_artifact( + NAME "model-BertLargePTBatch512" + SOURCE_URL "https://storage.googleapis.com/iree-model-artifacts/pytorch/pt_models_20231010.987_1696982151/BERT_LARGE_FP32_PT_384XI32_BATCH512/linalg.mlirbc" + OUTPUT "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch512.mlirbc" + UNPACK +) + +iree_fetch_artifact( + NAME "model-BertLargePTBatch1024" + SOURCE_URL "https://storage.googleapis.com/iree-model-artifacts/pytorch/pt_models_20231010.987_1696982151/BERT_LARGE_FP32_PT_384XI32_BATCH1024/linalg.mlirbc" + OUTPUT "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch1024.mlirbc" + UNPACK +) + +iree_fetch_artifact( + NAME "model-BertLargePTBatch1280" + SOURCE_URL "https://storage.googleapis.com/iree-model-artifacts/pytorch/pt_models_20231010.987_1696982151/BERT_LARGE_FP32_PT_384XI32_BATCH1280/linalg.mlirbc" + OUTPUT "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch1280.mlirbc" + UNPACK +) + iree_fetch_artifact( NAME "model-BERT_LARGE_JAX_384XI32_BATCH1" SOURCE_URL "https://storage.googleapis.com/iree-model-artifacts/jax/jax_models_0.4.14_1691969180/BERT_LARGE_FP32_JAX_384XI32_BATCH1/stablehlo.mlirbc" diff --git a/tests/e2e/test_artifacts/generated_e2e_test_iree_artifacts.cmake b/tests/e2e/test_artifacts/generated_e2e_test_iree_artifacts.cmake index 14242806251c..7401e1a27479 100644 --- a/tests/e2e/test_artifacts/generated_e2e_test_iree_artifacts.cmake +++ b/tests/e2e/test_artifacts/generated_e2e_test_iree_artifacts.cmake @@ -940,6 +940,114 @@ iree_bytecode_module( PUBLIC ) +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch1_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch1.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch1_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + FRIENDLY_NAME "BertLargePTBatch1(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch16_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch16.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch16_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + FRIENDLY_NAME "BertLargePTBatch16(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch24_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch24.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch24_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + FRIENDLY_NAME "BertLargePTBatch24(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch32_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch32.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch32_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + FRIENDLY_NAME "BertLargePTBatch32(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch48_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch48.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch48_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + FRIENDLY_NAME "BertLargePTBatch48(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch64_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch64.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch64_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + FRIENDLY_NAME "BertLargePTBatch64(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch512_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch512.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch512_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + FRIENDLY_NAME "BertLargePTBatch512(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch1024_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch1024.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch1024_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + FRIENDLY_NAME "BertLargePTBatch1024(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch1280_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch1280.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch1280_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + FRIENDLY_NAME "BertLargePTBatch1280(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags]" + PUBLIC +) + iree_bytecode_module( NAME "iree-module-BERT_LARGE_JAX_384XI32_BATCH1_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_" SRC "${ROOT_ARTIFACTS_DIR}/model_BERT_LARGE_JAX_384XI32_BATCH1.mlirbc" @@ -2686,6 +2794,150 @@ iree_bytecode_module( PUBLIC ) +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch1_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch1.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch1_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + "--iree-vm-emit-polyglot-zip=true" + "--iree-llvmcpu-debug-symbols=false" + "--iree-scheduling-dump-statistics-format=json" + "--iree-scheduling-dump-statistics-file=${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch1_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/scheduling_stats.json" + FRIENDLY_NAME "BertLargePTBatch1(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags,compile-stats]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch16_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch16.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch16_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + "--iree-vm-emit-polyglot-zip=true" + "--iree-llvmcpu-debug-symbols=false" + "--iree-scheduling-dump-statistics-format=json" + "--iree-scheduling-dump-statistics-file=${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch16_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/scheduling_stats.json" + FRIENDLY_NAME "BertLargePTBatch16(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags,compile-stats]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch24_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch24.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch24_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + "--iree-vm-emit-polyglot-zip=true" + "--iree-llvmcpu-debug-symbols=false" + "--iree-scheduling-dump-statistics-format=json" + "--iree-scheduling-dump-statistics-file=${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch24_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/scheduling_stats.json" + FRIENDLY_NAME "BertLargePTBatch24(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags,compile-stats]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch32_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch32.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch32_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + "--iree-vm-emit-polyglot-zip=true" + "--iree-llvmcpu-debug-symbols=false" + "--iree-scheduling-dump-statistics-format=json" + "--iree-scheduling-dump-statistics-file=${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch32_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/scheduling_stats.json" + FRIENDLY_NAME "BertLargePTBatch32(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags,compile-stats]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch48_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch48.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch48_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + "--iree-vm-emit-polyglot-zip=true" + "--iree-llvmcpu-debug-symbols=false" + "--iree-scheduling-dump-statistics-format=json" + "--iree-scheduling-dump-statistics-file=${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch48_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/scheduling_stats.json" + FRIENDLY_NAME "BertLargePTBatch48(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags,compile-stats]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch64_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch64.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch64_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + "--iree-vm-emit-polyglot-zip=true" + "--iree-llvmcpu-debug-symbols=false" + "--iree-scheduling-dump-statistics-format=json" + "--iree-scheduling-dump-statistics-file=${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch64_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/scheduling_stats.json" + FRIENDLY_NAME "BertLargePTBatch64(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags,compile-stats]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch512_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch512.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch512_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + "--iree-vm-emit-polyglot-zip=true" + "--iree-llvmcpu-debug-symbols=false" + "--iree-scheduling-dump-statistics-format=json" + "--iree-scheduling-dump-statistics-file=${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch512_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/scheduling_stats.json" + FRIENDLY_NAME "BertLargePTBatch512(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags,compile-stats]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch1024_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch1024.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch1024_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + "--iree-vm-emit-polyglot-zip=true" + "--iree-llvmcpu-debug-symbols=false" + "--iree-scheduling-dump-statistics-format=json" + "--iree-scheduling-dump-statistics-file=${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch1024_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/scheduling_stats.json" + FRIENDLY_NAME "BertLargePTBatch1024(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags,compile-stats]" + PUBLIC +) + +iree_bytecode_module( + NAME "iree-module-BertLargePTBatch1280_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_" + SRC "${ROOT_ARTIFACTS_DIR}/model_BertLargePTBatch1280.mlirbc" + MODULE_FILE_NAME "${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch1280_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/module.vmfb" + FLAGS + "--iree-hal-target-backends=cuda" + "--iree-input-type=none" + "--iree-hal-cuda-llvm-target-arch=sm_80" + "--iree-vm-emit-polyglot-zip=true" + "--iree-llvmcpu-debug-symbols=false" + "--iree-scheduling-dump-statistics-format=json" + "--iree-scheduling-dump-statistics-file=${ROOT_ARTIFACTS_DIR}/iree_module_BertLargePTBatch1280_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_/scheduling_stats.json" + FRIENDLY_NAME "BertLargePTBatch1280(linalg) [cuda-sm_80-linux_gnu-cuda][default-flags,compile-stats]" + PUBLIC +) + iree_bytecode_module( NAME "iree-module-BERT_LARGE_JAX_384XI32_BATCH1_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_" SRC "${ROOT_ARTIFACTS_DIR}/model_BERT_LARGE_JAX_384XI32_BATCH1.mlirbc" @@ -3513,6 +3765,15 @@ add_dependencies(iree-benchmark-import-models-large ${PACKAGE_NAME}_model-BERT_LARGE_JAX_384XI32_BATCH24 ${PACKAGE_NAME}_model-BERT_LARGE_JAX_384XI32_BATCH32 ${PACKAGE_NAME}_model-BERT_LARGE_JAX_384XI32_BATCH48 + ${PACKAGE_NAME}_model-BertLargePTBatch1 + ${PACKAGE_NAME}_model-BertLargePTBatch1024 + ${PACKAGE_NAME}_model-BertLargePTBatch1280 + ${PACKAGE_NAME}_model-BertLargePTBatch16 + ${PACKAGE_NAME}_model-BertLargePTBatch24 + ${PACKAGE_NAME}_model-BertLargePTBatch32 + ${PACKAGE_NAME}_model-BertLargePTBatch48 + ${PACKAGE_NAME}_model-BertLargePTBatch512 + ${PACKAGE_NAME}_model-BertLargePTBatch64 ${PACKAGE_NAME}_model-BertLargeTFBatch1 ${PACKAGE_NAME}_model-BertLargeTFBatch16 ${PACKAGE_NAME}_model-BertLargeTFBatch24 @@ -3647,6 +3908,15 @@ add_dependencies(iree-benchmark-suites-comp-stats-large ${PACKAGE_NAME}_iree-module-BERT_LARGE_JAX_384XI32_BATCH24_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_ ${PACKAGE_NAME}_iree-module-BERT_LARGE_JAX_384XI32_BATCH32_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_ ${PACKAGE_NAME}_iree-module-BERT_LARGE_JAX_384XI32_BATCH48_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch1024_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch1280_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch16_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch1_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch24_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch32_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch48_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch512_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch64_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_ ${PACKAGE_NAME}_iree-module-BertLargeTFBatch16_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_ ${PACKAGE_NAME}_iree-module-BertLargeTFBatch1_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_compile-stats_ ${PACKAGE_NAME}_iree-module-BertLargeTFBatch1_stablehlo___x86_64-cascadelake-linux_gnu-llvm_cpu__default-flags_compile-stats_ @@ -3704,6 +3974,15 @@ add_dependencies(iree-benchmark-suites-cuda-large ${PACKAGE_NAME}_iree-module-BERT_LARGE_JAX_384XI32_BATCH24_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_ ${PACKAGE_NAME}_iree-module-BERT_LARGE_JAX_384XI32_BATCH32_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_ ${PACKAGE_NAME}_iree-module-BERT_LARGE_JAX_384XI32_BATCH48_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch1024_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch1280_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch16_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch1_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch24_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch32_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch48_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch512_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch64_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ ${PACKAGE_NAME}_iree-module-BertLargeTFBatch16_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_ ${PACKAGE_NAME}_iree-module-BertLargeTFBatch1_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_ ${PACKAGE_NAME}_iree-module-BertLargeTFBatch24_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_ @@ -3812,6 +4091,15 @@ add_dependencies(iree-benchmark-suites-large ${PACKAGE_NAME}_iree-module-BERT_LARGE_JAX_384XI32_BATCH24_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_ ${PACKAGE_NAME}_iree-module-BERT_LARGE_JAX_384XI32_BATCH32_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_ ${PACKAGE_NAME}_iree-module-BERT_LARGE_JAX_384XI32_BATCH48_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch1024_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch1280_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch16_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch1_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch24_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch32_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch48_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch512_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ + ${PACKAGE_NAME}_iree-module-BertLargePTBatch64_linalg___cuda-sm_80-linux_gnu-cuda__default-flags_ ${PACKAGE_NAME}_iree-module-BertLargeTFBatch16_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_ ${PACKAGE_NAME}_iree-module-BertLargeTFBatch1_stablehlo___cuda-sm_80-linux_gnu-cuda__default-flags_ ${PACKAGE_NAME}_iree-module-BertLargeTFBatch1_stablehlo___x86_64-cascadelake-linux_gnu-llvm_cpu__default-flags_