Implement pass search

TODO:
Remove old implementation and update documentation

.gitignore

@@ -126,6 +126,7 @@ celerybeat.pid
 # Environments
 .env
 .venv
+.vs
 env/
 venv/
 ENV/

examples/bert/bert.py

@@ -0,0 +1,27 @@
+# -------------------------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+# --------------------------------------------------------------------------
+import argparse
+import json
+from pathlib import Path
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--optimize",
+        action="store_true",
+        help="If set, run transformers optimization pass",
+    )
+    args = parser.parse_args()
+
+    input_filename = "bert_cuda_gpu.template.json"
+    with Path(input_filename).open("r") as f:
+        config = json.load(f)
+
+    if not args.optimize:
+        del config["passes"]["transformers_optimization"]
+
+    output_filename = input_filename.replace(".template", "")
+    with Path(output_filename).open("w") as strm:
+        json.dump(config, fp=strm, indent=4)

examples/bert/bert_cuda_gpu.json → examples/bert/bert_cuda_gpu.template.json

@@ -51,11 +51,7 @@
         "transformers_optimization": { "type": "OrtTransformersOptimization", "float16": true },
         "session_params_tuning": { "type": "OrtSessionParamsTuning", "data_config": "glue_mrpc", "io_bind": true }
     },
-    "pass_flows": [
-        [ "conversion", "transformers_optimization", "session_params_tuning" ],
-        [ "conversion", "session_params_tuning" ]
-    ],
-    "search_strategy": { "execution_order": "joint", "search_algorithm": "tpe", "num_samples": 3, "seed": 0 },
+    "search_strategy": { "execution_order": "joint", "sampler": "tpe", "max_samples": 3, "seed": 0 },
     "host": "local_system",
     "target": "local_system",
     "evaluator": "common_evaluator",

examples/bert/bert_inc_dynamic_ptq_cpu.json

@@ -34,7 +34,7 @@
         "transformers_optimization": { "type": "OrtTransformersOptimization", "model_type": "bert" },
         "dynamic_quantization": { "type": "IncDynamicQuantization" }
     },
-    "search_strategy": { "execution_order": "joint", "search_algorithm": "exhaustive" },
+    "search_strategy": { "execution_order": "joint", "sampler": "sequential" },
     "evaluator": "common_evaluator",
     "cache_dir": "cache",
     "output_dir": "models/bert_inc_dynamic_ptq_cpu"

examples/bert/bert_inc_ptq_cpu.json

@@ -64,7 +64,7 @@
             }
         }
     },
-    "search_strategy": { "execution_order": "joint", "search_algorithm": "exhaustive" },
+    "search_strategy": { "execution_order": "joint", "sampler": "sequential" },
     "evaluator": "common_evaluator",
     "cache_dir": "cache",
     "output_dir": "models/bert_inc_ptq_cpu"

examples/bert/bert_ptq_cpu.json

@@ -67,7 +67,7 @@
         },
         "session_params_tuning": { "type": "OrtSessionParamsTuning", "data_config": "glue_mrpc" }
     },
-    "search_strategy": { "execution_order": "joint", "search_algorithm": "tpe", "num_samples": 3, "seed": 0 },
+    "search_strategy": { "execution_order": "joint", "sampler": "tpe", "max_samples": 3, "seed": 0 },
     "evaluator": "common_evaluator",
     "host": "local_system",
     "target": "local_system",

examples/bert/notebook/bert_auto_opt_gpu.json

@@ -43,7 +43,7 @@
             ]
         }
     },
-    "search_strategy": { "execution_order": "joint", "search_algorithm": "tpe", "num_samples": 1, "seed": 0 },
+    "search_strategy": { "execution_order": "joint", "sampler": "tpe", "max_samples": 1, "seed": 0 },
     "evaluator": "common_evaluator",
     "host": "local_system",
     "target": "local_system",

examples/bert/notebook/multi_ep_search.ipynb

@@ -141,17 +141,15 @@
     "#### Engine and search strategy\n",
     "\n",
     "Engine is used to manage the optimization process where we run optimization on host device, and run evaluation on target device.\n",
-    "Search strategy is used to search the optimal optimization among different EPs. In this notebook, we use `joint` as the `execution_order` and `tpe` as the `search_algorithm`. We set the `num_samples` to 1 and `seed` to 0.\n",
+    "Search strategy is used to search the optimal optimization among different EPs. In this notebook, we use `joint` as the `execution_order` and `tpe` as the `sampler`. We set the `max_samples` to 1 and `seed` to 0.\n",
     "\n",
     "```json\n",
     "\"engine\": {\n",
     "    \"search_strategy\": {\n",
     "        \"execution_order\": \"joint\",\n",
-    "        \"search_algorithm\": \"tpe\",\n",
-    "        \"search_algorithm_config\": {\n",
-    "            \"num_samples\": 1,\n",
-    "            \"seed\": 0\n",
-    "        }\n",
+    "        \"sampler\": \"tpe\",\n",
+    "        \"max_samples\": 1,\n",
+    "        \"seed\": 0\n",
     "    },\n",
     "    \"evaluator\": \"common_evaluator\",\n",
     "    \"host\": \"local_system\",\n",

examples/bert/user_script.py

@@ -8,7 +8,7 @@
 import torch
 import torchmetrics
 import transformers
-from datasets import load_dataset, load_metric
+from datasets import load_dataset
 from datasets.utils import logging as datasets_logging
 from neural_compressor.data import DefaultDataLoader
 from torch.utils.data import Dataset
@@ -26,6 +26,11 @@
 from olive.data.registry import Registry
 from olive.model import OliveModelHandler
 
+try:
+    from datasets import load_metric
+except ImportError:
+    from evaluate import load as load_metric
+
 datasets_logging.disable_progress_bar()
 datasets_logging.set_verbosity_error()
 

examples/deberta/deberta.json

@@ -46,7 +46,7 @@
         "quantization": { "type": "OnnxQuantization", "data_config": "glue_mnli_matched" },
         "session_params_tuning": { "type": "OrtSessionParamsTuning", "data_config": "glue_mnli_matched" }
     },
-    "search_strategy": { "execution_order": "joint", "search_algorithm": "tpe", "num_samples": 3, "seed": 0 },
+    "search_strategy": { "execution_order": "joint", "sampler": "tpe", "max_samples": 3, "seed": 0 },
     "evaluator": "common_evaluator",
     "cache_dir": "cache",
     "output_dir": "models/microsoft-deberta"

examples/directml/llm/config_llm.json

@@ -90,7 +90,7 @@
             }
         }
     },
-    "search_strategy": { "execution_order": "joint", "search_algorithm": "exhaustive" },
+    "search_strategy": { "execution_order": "joint", "sampler": "sequential" },
     "evaluator": "common_evaluator",
     "evaluate_input_model": false,
     "host": "local_system",

examples/directml/stable_diffusion_xl/config_text_encoder.json

@@ -112,7 +112,6 @@
             "keep_io_types": true
         }
     },
-    "pass_flows": [ [ "convert", "optimize" ] ],
     "evaluator": "common_evaluator",
     "evaluate_input_model": false,
     "host": "local_system",

examples/directml/stable_diffusion_xl/config_text_encoder_2.json

@@ -152,7 +152,6 @@
             "keep_io_types": true
         }
     },
-    "pass_flows": [ [ "convert", "optimize" ] ],
     "evaluator": "common_evaluator",
     "evaluate_input_model": false,
     "host": "local_system",

examples/directml/stable_diffusion_xl/config_unet.json

@@ -96,7 +96,6 @@
             "keep_io_types": true
         }
     },
-    "pass_flows": [ [ "convert", "optimize" ] ],
     "evaluator": "common_evaluator",
     "evaluate_input_model": false,
     "host": "local_system",

examples/directml/stable_diffusion_xl/config_vae_decoder.json

@@ -108,7 +108,6 @@
             "use_gpu": true
         }
     },
-    "pass_flows": [ [ "convert", "optimize" ] ],
     "evaluator": "common_evaluator",
     "evaluate_input_model": false,
     "host": "local_system",

examples/directml/stable_diffusion_xl/config_vae_encoder.json

@@ -87,7 +87,6 @@
             "keep_io_types": true
         }
     },
-    "pass_flows": [ [ "convert", "optimize" ] ],
     "evaluator": "common_evaluator",
     "evaluate_input_model": false,
     "host": "local_system",

examples/directml/stable_diffusion_xl/stable_diffusion_xl.py

@@ -7,6 +7,7 @@
 import shutil
 import sys
 import warnings
+from collections import OrderedDict
 from pathlib import Path
 from typing import Dict
 
@@ -289,22 +290,25 @@ def run_inference(
 
 
 def update_config_with_provider(config: Dict, provider: str, is_fp16: bool) -> Dict:
+    used_passes = {}
     if provider == "dml":
         # DirectML EP is the default, so no need to update config.
-        return config
+        used_passes = {"convert", "optimize"}
     elif provider == "cuda":
         if version.parse(OrtVersion) < version.parse("1.17.0"):
             # disable skip_group_norm fusion since there is a shape inference bug which leads to invalid models
             config["passes"]["optimize_cuda"]["optimization_options"] = {"enable_skip_group_norm": False}
         # keep model fully in fp16 if use_fp16_fixed_vae is set
         if is_fp16:
             config["passes"]["optimize_cuda"].update({"float16": True, "keep_io_types": False})
-        config["pass_flows"] = [["convert", "optimize_cuda"]]
+        used_passes = {"convert", "optimize_cuda"}
         config["systems"]["local_system"]["accelerators"][0]["execution_providers"] = ["CUDAExecutionProvider"]
-        return config
     else:
         raise ValueError(f"Unsupported provider: {provider}")
 
+    config["passes"] = OrderedDict([(k, v) for k, v in config["passes"].items() if k in used_passes])
+    return config
+
 
 def optimize(
     model_id: str,

examples/llama2/llama2.py

@@ -173,17 +173,26 @@ def get_general_config(args):
     gqa = "gqa" if args.use_gqa else "mha"
     config_name = f"llama2_{device}_{gqa}"
 
-    # add pass flows
+    # add pass names
     if not args.use_gptq:
-        template_json["pass_flows"] = [flow for flow in SUPPORTED_WORKFLOWS[device] if "gptq" not in flow[0]]
+        used_passes = [
+            pass_name
+            for passes_names in SUPPORTED_WORKFLOWS[device]
+            if "gptq" not in passes_names[0]
+            for pass_name in passes_names
+        ]
     else:
-        template_json["pass_flows"] = [flow for flow in SUPPORTED_WORKFLOWS[device] if "gptq" in flow[0]]
+        used_passes = [
+            pass_name
+            for passes_names in SUPPORTED_WORKFLOWS[device]
+            if "gptq" in passes_names[0]
+            for pass_name in passes_names
+        ]
         auto_gptq_logger = logging.getLogger("auto_gptq")
         auto_gptq_logger.addHandler(logging.StreamHandler(sys.stdout))
         auto_gptq_logger.setLevel(logging.INFO)
 
     # remove unused passes and set gqa related configs
-    used_passes = {pass_name for pass_flow in SUPPORTED_WORKFLOWS[device] for pass_name in pass_flow}
     for pass_name in list(template_json["passes"].keys()):
         if pass_name not in used_passes:
             del template_json["passes"][pass_name]

examples/llama2/llama2_lmeval.json

@@ -57,7 +57,7 @@
         }
     },
     "auto_optimizer_config": { "opt_level": 0, "disable_auto_optimizer": true, "precision": "fp16" },
-    "search_strategy": { "execution_order": "joint", "search_algorithm": "exhaustive" },
+    "search_strategy": { "execution_order": "joint", "sampler": "exhaustive" },
     "evaluator": "evaluator",
     "host": "local_system",
     "target": "local_system",

examples/llama2/llama2_model_builder.py

@@ -5,6 +5,7 @@
 
 import argparse
 import json
+from collections import OrderedDict
 
 from olive.workflows import run as olive_run
 
@@ -38,10 +39,8 @@ def main(raw_args=None):
     template_json = json.loads(template_json_str)
 
     # add pass flows
-    if args.metadata_only:
-        template_json["pass_flows"] = [["conversion", "metadata"]]
-    else:
-        template_json["pass_flows"] = [["builder", "session_params_tuning"]]
+    used_passes = {"conversion", "metadata"} if args.metadata_only else {"builder", "session_params_tuning"}
+    template_json["passes"] = OrderedDict([(k, v) for k, v in template_json["passes"].items() if k in used_passes])
     template_json["output_dir"] = f"models/{model_name}"
 
     # dump config

examples/llama2/notebook/llama2_multiep/llama2.py

@@ -0,0 +1,32 @@
+# -------------------------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+# --------------------------------------------------------------------------
+import argparse
+import json
+from pathlib import Path
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--device",
+        choices=["cpu", "gpu", "multi_ep"],
+        help="Device to use",
+    )
+    parser.add_argument(
+        "--quantize",
+        action="store_true",
+        help="If set, run transformers optimization pass",
+    )
+    args = parser.parse_args()
+
+    input_filename = f"config_{args.device}.template.json"
+    with Path(input_filename).open("r") as f:
+        config = json.load(f)
+
+    if not args.quantize:
+        del config["passes"]["blockwise_quant_int4"]
+
+    output_filename = input_filename.replace(".template", "")
+    with Path(output_filename).open("w") as strm:
+        json.dump(config, fp=strm, indent=4)

examples/mistral/mistral.py

@@ -108,7 +108,7 @@ def main(raw_args=None):
     optimized_model_dir = (
         script_dir
         / config["output_dir"]
-        / "-".join(config["pass_flows"][0])
+        / "-".join(config["passes"].keys())
         / f"{config['output_name']}_{accelerator}-{ep_header}_model"
     )
 

examples/mistral/mistral_fp16.json

@@ -45,7 +45,6 @@
             "enable_profiling": false
         }
     },
-    "pass_flows": [ [ "convert", "optimize", "session_params_tuning" ] ],
     "evaluate_input_model": false,
     "evaluator": "common_evaluator",
     "host": "local_system",

examples/mistral/mistral_int4.json

@@ -57,7 +57,6 @@
             "all_tensors_to_one_file": true
         }
     },
-    "pass_flows": [ [ "convert", "optimize", "quantization" ] ],
     "evaluate_input_model": false,
     "evaluator": "common_evaluator",
     "host": "local_system",