Fix ort inputs filtering (#129)

.gitignore

@@ -168,6 +168,7 @@ data/
 version.txt
 
 .engine/
+actions-runner-duplicate/
 actions-runner/
 experiments/
 amdsmi/

Makefile

@@ -1,9 +1,5 @@
 # List of targets that are not associated with files
-.PHONY:	quality style install \
-		build_docker_cpu, build_docker_cuda, build_docker_rocm, \ 
-		test_cli_cpu_pytorch, test_cli_rocm_pytorch, \
-		test_cli_cpu_neural_compressor, test_cli_cpu_onnxruntime, test_cli_cpu_openvino, \
-		test_api_cpu, test_api_cuda, test_api_rocm, test_api_misc
+.PHONY:	quality style install build_docker_cpu build_docker_cuda build_docker_rocm test_cli_cpu_neural_compressor test_cli_cpu_onnxruntime test_cli_cpu_openvino test_cli_cpu_pytorch test_cli_rocm_pytorch test_cli_cuda_pytorch test_api_cpu test_api_cuda test_api_rocm test_api_misc
 
 quality:
 	ruff check .
@@ -28,6 +24,7 @@ build_docker_rocm:
 test_cli_cpu_neural_compressor:
 	docker run \
 	--rm \
+	--pid=host \
 	--entrypoint /bin/bash \
 	--volume $(PWD):/workspace \
 	--workdir /workspace \
@@ -36,6 +33,7 @@ test_cli_cpu_neural_compressor:
 test_cli_cpu_onnxruntime:
 	docker run \
 	--rm \
+	--pid=host \
 	--entrypoint /bin/bash \
 	--volume $(PWD):/workspace \
 	--workdir /workspace \
@@ -44,6 +42,7 @@ test_cli_cpu_onnxruntime:
 test_cli_cpu_openvino:
 	docker run \
 	--rm \
+	--pid=host \
 	--entrypoint /bin/bash \
 	--volume $(PWD):/workspace \
 	--workdir /workspace \
@@ -52,6 +51,7 @@ test_cli_cpu_openvino:
 test_cli_cpu_pytorch:
 	docker run \
 	--rm \
+	--pid=host \
 	--entrypoint /bin/bash \
 	--volume $(PWD):/workspace \
 	--workdir /workspace \
@@ -60,6 +60,7 @@ test_cli_cpu_pytorch:
 test_cli_rocm_pytorch:
 	docker run \
 	--rm \
+	--pid=host \
 	--device=/dev/kfd \
 	--device /dev/dri/renderD128 \
 	--device /dev/dri/renderD129 \
@@ -72,6 +73,7 @@ test_cli_rocm_pytorch:
 test_cli_cuda_pytorch:
 	docker run \
 	--rm \
+	--pid=host \
 	--gpus '"device=0,1"' \
 	--entrypoint /bin/bash \
 	--volume $(PWD):/workspace \
@@ -81,6 +83,7 @@ test_cli_cuda_pytorch:
 test_api_cpu:
 	docker run \
 	--rm \
+	--pid=host \
 	--entrypoint /bin/bash \
 	--volume $(PWD):/workspace \
 	--workdir /workspace \
@@ -89,6 +92,7 @@ test_api_cpu:
 test_api_cuda:
 	docker run \
 	--rm \
+	--pid=host \
 	--gpus '"device=0,1"' \
 	--entrypoint /bin/bash \
 	--volume $(PWD):/workspace \
@@ -98,6 +102,7 @@ test_api_cuda:
 test_api_rocm:
 	docker run \
 	--rm \
+	--pid=host \
 	--device=/dev/kfd \
 	--device /dev/dri/renderD128 \
 	--device /dev/dri/renderD129 \
@@ -110,6 +115,7 @@ test_api_rocm:
 test_api_misc:
 	docker run \
 	--rm \
+	--pid=host \
 	--entrypoint /bin/bash \
 	--volume $(PWD):/workspace \
 	--workdir /workspace \

optimum_benchmark/backends/base.py

@@ -10,7 +10,7 @@
 
 from ..task_utils import get_automodel_class_for_task
 from .config import BackendConfigT
-from .diffusers_utils import extract_diffusers_shapes_from_config, get_diffusers_pretrained_config
+from .diffusers_utils import extract_diffusers_shapes_from_model, get_diffusers_pretrained_config
 from .timm_utils import extract_timm_shapes_from_config, get_timm_pre_processor, get_timm_pretrained_config
 from .transformers_utils import (
     PretrainedProcessor,
@@ -41,7 +41,7 @@ def __init__(self, config: BackendConfigT):
 
         if self.config.library == "diffusers":
             self.pretrained_config = get_diffusers_pretrained_config(self.config.model, **self.config.hub_kwargs)
-            self.model_shapes = extract_diffusers_shapes_from_config(self.config.model, **self.config.hub_kwargs)
+            self.model_shapes = extract_diffusers_shapes_from_model(self.config.model, **self.config.hub_kwargs)
             self.model_type = self.config.task
             self.generation_config = None
             self.pre_processor = None

optimum_benchmark/backends/diffusers_utils.py

@@ -12,7 +12,7 @@ def get_diffusers_pretrained_config(model: str, **kwargs) -> Dict[str, int]:
     return diffusers.DiffusionPipeline.load_config(model, **kwargs)
 
 
-def extract_diffusers_shapes_from_config(model: str, **kwargs) -> Dict[str, int]:
+def extract_diffusers_shapes_from_model(model: str, **kwargs) -> Dict[str, int]:
     config = diffusers.DiffusionPipeline.load_config(model, **kwargs)
 
     shapes = {}

optimum_benchmark/backends/onnxruntime/backend.py

@@ -332,13 +332,14 @@ def prepare_for_inference(self, **kwargs) -> None:
 
     def prepare_inputs(self, inputs: Dict[str, Any]) -> Dict[str, Any]:
         if self.config.library == "diffusers":
-            return {"prompt": inputs["prompt"]}
+            return inputs
 
         LOGGER.info(f"\t+ Moving inputs tensors to device {self.config.device}")
         for key, value in list(inputs.items()):
             if key in self.inputs_names:
                 inputs[key] = value.to(self.config.device)
             else:
+                LOGGER.warning(f"Input {key} is not in expected inputs names. Removing it.")
                 inputs.pop(key)
 
         return inputs

optimum_benchmark/backends/transformers_utils.py

@@ -42,7 +42,7 @@ def get_transformers_pre_processor(model: str, **kwargs) -> Optional["Pretrained
     try:
         # sometimes contains information about the model's input shapes that are not available in the config
         return AutoProcessor.from_pretrained(model, **kwargs)
-    except ValueError:
+    except Exception:
         return None
 
 

optimum_benchmark/benchmarks/inference/benchmark.py

@@ -11,6 +11,7 @@
 from ..base import Benchmark
 from ..report import BenchmarkMeasurements, BenchmarkReport
 from .config import InferenceConfig
+from .inputs_utils import extract_text_generation_inputs
 
 if is_torch_distributed_available():
     import torch.distributed
@@ -72,37 +73,35 @@ def run(self, backend: Backend[BackendConfigT]) -> None:
                     "The batch size must be divisible by the number of processes in a distributed environment"
                 )
             self.config.input_shapes["batch_size"] //= torch.distributed.get_world_size()
-            if backend.config.device == "cuda" and backend.config.task in TEXT_GENERATION_TASKS:
-                TEXT_GENERATION_TASKS["synced_gpus"] = True
 
         LOGGER.info("\t+ Creating input generator")
         self.input_generator = InputGenerator(
             task=backend.config.task, model_shapes=backend.model_shapes, input_shapes=self.config.input_shapes
         )
 
         if backend.config.task in TEXT_GENERATION_TASKS:
-            LOGGER.info("\t+ Generating and preparing Text Generation input")
-            self.forward_inputs = self.input_generator(mode="forward")
-            self.generate_input = self.input_generator(mode="generate")
+            LOGGER.info("\t+ Generating and preparing Text Generation inputs")
+            self.forward_inputs = self.input_generator()
             self.forward_inputs = backend.prepare_inputs(self.forward_inputs)
-            self.generate_input = backend.prepare_inputs(self.generate_input)
+            self.generate_inputs = extract_text_generation_inputs(self.forward_inputs)
             LOGGER.info("\t+ Updating Text Generation kwargs with default values")
             self.config.generate_kwargs = {**TEXT_GENERATION_KWARGS, **self.config.generate_kwargs}
             LOGGER.info("\t+ Initializing Text Generation report")
             self.report = TextGenerationReport(prefill=BenchmarkMeasurements(), decode=BenchmarkMeasurements())
 
         elif backend.config.task in IMAGE_DIFFUSION_TASKS:
-            LOGGER.info("\t+ Generating and preparing Image Diffusion input")
-            self.diffuse_input = self.input_generator(mode="call")
-            self.diffuse_input = backend.prepare_inputs(self.diffuse_input)
+            LOGGER.info("\t+ Generating Image Diffusion inputs")
+            self.call_inputs = self.input_generator()
+            self.call_inputs = backend.prepare_inputs(self.call_inputs)
+            self.call_inputs = {"prompt": self.call_inputs["prompt"]}
             LOGGER.info("\t+ Updating Image Diffusion kwargs with default values")
-            self.config.forward_kwargs = {**IMAGE_DIFFUSION_KWARGS, **self.config.forward_kwargs}
+            self.config.call_kwargs = {**IMAGE_DIFFUSION_KWARGS, **self.config.call_kwargs}
             LOGGER.info("\t+ Initializing Image Diffusion report")
             self.report = ImageDiffusionReport(call=BenchmarkMeasurements())
 
         else:
-            LOGGER.info("\t+ Generating and preparing Inference input")
-            self.forward_inputs = self.input_generator(mode="forward")
+            LOGGER.info("\t+ Generating and preparing Inference inputs")
+            self.forward_inputs = self.input_generator()
             self.forward_inputs = backend.prepare_inputs(self.forward_inputs)
             LOGGER.info("\t+ Initializing Inference report")
             self.report = InferenceReport(forward=BenchmarkMeasurements())
@@ -111,16 +110,17 @@ def run(self, backend: Backend[BackendConfigT]) -> None:
         backend.prepare_for_inference(
             **backend.model_shapes,
             **self.config.input_shapes,
-            **self.config.forward_kwargs,
             **self.config.generate_kwargs,
+            **self.config.forward_kwargs,
+            **self.config.call_kwargs,
         )
 
         LOGGER.info("\t+ Warming up backend for Inference")
         for _ in range(self.config.warmup_runs):
             if backend.config.task in TEXT_GENERATION_TASKS:
-                _ = backend.generate(self.generate_input, {"max_new_tokens": 2, "min_new_tokens": 2})
+                _ = backend.generate(self.generate_inputs, {"max_new_tokens": 2, "min_new_tokens": 2})
             elif backend.config.task in IMAGE_DIFFUSION_TASKS:
-                _ = backend.call(self.diffuse_input, {"num_inference_steps": 2})
+                _ = backend.call(self.call_inputs, {"num_inference_steps": 2})
             else:
                 _ = backend.forward(self.forward_inputs, self.config.forward_kwargs)
 
@@ -164,8 +164,6 @@ def run(self, backend: Backend[BackendConfigT]) -> None:
             self.report.log_energy()
             self.report.log_efficiency()
 
-        self.report.log()
-
     ## Memory tracking
     def run_text_generation_memory_tracking(self, backend: Backend):
         LOGGER.info("\t+ Running memory tracking")
@@ -177,15 +175,15 @@ def run_text_generation_memory_tracking(self, backend: Backend):
 
         self.memory_tracker.reset()
         with self.memory_tracker.track():
-            _ = backend.generate(self.generate_input, self.config.generate_kwargs)
+            _ = backend.generate(self.generate_inputs, self.config.generate_kwargs)
 
         self.report.decode.memory = self.memory_tracker.get_max_memory()
 
     def run_image_diffusion_memory_tracking(self, backend: Backend):
         LOGGER.info("\t+ Running memory tracking")
         self.memory_tracker.reset()
         with self.memory_tracker.track():
-            _ = backend.call(self.diffuse_input, self.config.forward_kwargs)
+            _ = backend.call(self.call_inputs, self.config.call_kwargs)
 
         self.report.call.memory = self.memory_tracker.get_max_memory()
 
@@ -205,17 +203,21 @@ def run_text_generation_latency_tracking(self, backend: Backend):
             with self.latency_tracker.track():
                 _ = backend.forward(self.forward_inputs, self.config.forward_kwargs)
 
-        self.report.prefill.latency = self.latency_tracker.get_latency()
+        forward_latency = self.latency_tracker.get_latency()
+        forward_latency.log(prefix="forward")
+        self.report.prefill.latency = forward_latency
         self.report.prefill.throughput = self.latency_tracker.get_throughput(
             volume=self.prefill_volume, unit=PREFILL_THROUGHPUT_UNIT
         )
 
         self.latency_tracker.reset()
         while self.latency_tracker.get_elapsed_time() < self.config.duration:
             with self.latency_tracker.track():
-                _ = backend.generate(self.generate_input, self.config.generate_kwargs)
+                _ = backend.generate(self.generate_inputs, self.config.generate_kwargs)
 
-        self.report.decode.latency = self.latency_tracker.get_latency() - self.report.prefill.latency.mean
+        generate_latency = self.latency_tracker.get_latency()
+        generate_latency.log(prefix="generate")
+        self.report.decode.latency = generate_latency - self.report.prefill.latency.mean
         self.report.decode.throughput = Throughput.from_latency(
             self.report.decode.latency, self.decode_volume, unit=DECODE_THROUGHPUT_UNIT
         )
@@ -225,7 +227,7 @@ def run_image_diffusion_latency_tracking(self, backend: Backend):
         self.latency_tracker.reset()
         while self.latency_tracker.get_elapsed_time() < self.config.duration:
             with self.latency_tracker.track():
-                _ = backend.call(self.diffuse_input, self.config.forward_kwargs)
+                _ = backend.call(self.call_inputs, self.config.call_kwargs)
 
         self.report.call.latency = self.latency_tracker.get_latency()
         self.report.call.throughput = Throughput.from_latency(
@@ -258,7 +260,7 @@ def run_text_generation_energy_tracking(self, backend: Backend):
 
         self.energy_tracker.reset()
         with self.energy_tracker.track():
-            _ = backend.generate(self.generate_input, self.config.generate_kwargs)
+            _ = backend.generate(self.generate_inputs, self.config.generate_kwargs)
 
         self.report.decode.energy = self.energy_tracker.get_energy() - self.report.prefill.energy
         self.report.decode.efficiency = Efficiency.from_energy(
@@ -269,7 +271,7 @@ def run_image_diffusion_energy_tracking(self, backend: Backend):
         LOGGER.info("\t+ Running energy tracking")
         self.energy_tracker.reset()
         with self.energy_tracker.track():
-            _ = backend.call(self.diffuse_input, self.config.forward_kwargs)
+            _ = backend.call(self.call_inputs, self.config.call_kwargs)
 
         self.report.call.energy = self.energy_tracker.get_energy()
         self.report.call.efficiency = Efficiency.from_energy(
@@ -297,7 +299,7 @@ def prefill_volume(self) -> int:  # in tokens
 
     @property
     def call_volume(self) -> int:  # in images
-        return self.config.input_shapes["batch_size"] * self.config.forward_kwargs["num_images_per_prompt"]
+        return self.config.input_shapes["batch_size"] * self.config.call_kwargs["num_images_per_prompt"]
 
     @property
     def decode_volume(self) -> int:  # in tokens

optimum_benchmark/benchmarks/inference/inputs_utils.py

@@ -0,0 +1,17 @@
+def extract_text_generation_inputs(inputs):
+    if "pixel_values" in inputs:
+        # image input
+        text_generation_inputs = {"inputs": inputs["pixel_values"]}
+    elif "input_values" in inputs:
+        # speech input
+        text_generation_inputs = {"inputs": inputs["input_values"]}
+    elif "input_features" in inputs:
+        # waveform input
+        text_generation_inputs = {"inputs": inputs["input_features"]}
+    elif "input_ids" in inputs:
+        # text input
+        text_generation_inputs = {"inputs": inputs["input_ids"]}
+    else:
+        raise ValueError("Could not find any valid text generation inputs.")
+
+    return text_generation_inputs

optimum_benchmark/generators/input_generator.py

@@ -22,23 +22,6 @@ def __init__(self, task: str, input_shapes: Dict[str, int], model_shapes: Dict[s
                 "please submit a PR or a feature request to optimum-benchmark. "
             )
 
-    def __call__(self, mode: str) -> Dict[str, Any]:
+    def __call__(self) -> Dict[str, Any]:
         task_input = self.task_generator()
-
-        if mode == "generate":
-            if "pixel_values" in task_input:
-                # image input
-                task_input = {"inputs": task_input["pixel_values"]}
-            elif "input_values" in task_input:
-                # speech input
-                task_input = {"inputs": task_input["input_values"]}
-            elif "input_features" in task_input:
-                # waveform input
-                task_input = {"inputs": task_input["input_features"]}
-            elif "input_ids" in task_input:
-                # text input
-                task_input = {"inputs": task_input["input_ids"]}
-        elif mode == "call":
-            task_input = {"prompt": task_input["prompt"]}
-
         return task_input

optimum_benchmark/trackers/latency.py

@@ -140,14 +140,13 @@ def _cpu_latency(self):
         self.end_events.append(end)
 
     def get_elapsed_time(self) -> float:
-        # we measured in cpu to not synchronize all events
+        # we measure it in cpu to not synchronize all events
         return time.perf_counter() - self.start_time
 
     def get_latency(self) -> Latency:
         if self.backend == "pytorch" and self.device == "cuda":
-            # synchronize the last event to make sure it has been recorded
-            self.start_events[-1].synchronize()
-            self.end_events[-1].synchronize()
+            # synchronize the device to make sure all events have been recorded
+            torch.cuda.synchronize()
 
             latencies_list = [
                 self.start_events[i].elapsed_time(self.end_events[i]) / 1e3 for i in range(len(self.start_events))
@@ -210,12 +209,7 @@ def __init__(self, device: str, backend: str):
         self.reset()
 
     def reset(self):
-        if self.device == "cuda" and self.backend == "pytorch":
-            event = torch.cuda.Event(enable_timing=True)
-            event.record()
-            self.events = [event]
-        else:
-            self.events = [time.perf_counter()]
+        self.events: List[Union[float, torch.cuda.Event]] = []
 
     def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor):
         if self.device == "cuda" and self.backend == "pytorch":