init

transformers/benchmark/.gitignore

@@ -0,0 +1 @@
+benchmark_results/

transformers/benchmark/README.md

@@ -0,0 +1,49 @@
+# Benchmarks
+
+You might want to add new benchmarks.
+
+You will need to define a python function named `run_benchmark` in your python file and the file must be located in this `benchmark/` directory.
+
+The expected function signature is the following:
+
+```py
+def run_benchmark(logger: Logger, branch: str, commit_id: str, commit_msg: str, num_tokens_to_generate=100):
+```
+
+## Writing metrics to the database
+
+`MetricsRecorder` is thread-safe, in the sense of the python [`Thread`](https://docs.python.org/3/library/threading.html#threading.Thread). This means you can start a background thread to do the readings on the device measurements while not blocking the main thread to execute the model measurements.
+
+cf [`llama.py`](./llama.py) to see an example of this in practice.
+
+```py
+from benchmarks_entrypoint import MetricsRecorder
+import psycopg2
+
+def run_benchmark(logger: Logger, branch: str, commit_id: str, commit_msg: str, num_tokens_to_generate=100):
+  metrics_recorder = MetricsRecorder(psycopg2.connect("dbname=metrics"), logger, branch, commit_id, commit_msg)
+  benchmark_id = metrics_recorder.initialise_benchmark({"gpu_name": gpu_name, "model_id": model_id})
+    # To collect device measurements
+    metrics_recorder.collect_device_measurements(
+        benchmark_id, cpu_util, mem_megabytes, gpu_util, gpu_mem_megabytes
+    )
+    # To collect your model measurements
+    metrics_recorder.collect_model_measurements(
+        benchmark_id,
+        {
+            "model_load_time": model_load_time,
+            "first_eager_forward_pass_time_secs": first_eager_fwd_pass_time,
+            "second_eager_forward_pass_time_secs": second_eager_fwd_pass_time,
+            "first_eager_generate_time_secs": first_eager_generate_time,
+            "second_eager_generate_time_secs": second_eager_generate_time,
+            "time_to_first_token_secs": time_to_first_token,
+            "time_to_second_token_secs": time_to_second_token,
+            "time_to_third_token_secs": time_to_third_token,
+            "time_to_next_token_mean_secs": mean_time_to_next_token,
+            "first_compile_generate_time_secs": first_compile_generate_time,
+            "second_compile_generate_time_secs": second_compile_generate_time,
+            "third_compile_generate_time_secs": third_compile_generate_time,
+            "fourth_compile_generate_time_secs": fourth_compile_generate_time,
+        },
+    )
+```

transformers/benchmark/benches/llama.py

@@ -0,0 +1,354 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+import sys
+from logging import Logger
+from threading import Event, Thread
+from time import perf_counter, sleep
+from typing import Optional
+
+
+# Add the parent directory to Python path to import benchmarks_entrypoint
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+import gpustat
+import psutil
+import psycopg2
+from benchmarks_entrypoint import MetricsRecorder
+
+
+# Optional heavy ML dependencies - only required when actually running the benchmark
+try:
+    import torch
+
+    from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig, StaticCache
+
+    TRANSFORMERS_AVAILABLE = True
+except ImportError:
+    TRANSFORMERS_AVAILABLE = False
+    torch = None
+    AutoModelForCausalLM = None
+    AutoTokenizer = None
+    GenerationConfig = None
+    StaticCache = None
+
+os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
+os.environ["TOKENIZERS_PARALLELISM"] = "1"
+
+# Only set torch precision if torch is available
+if TRANSFORMERS_AVAILABLE:
+    torch.set_float32_matmul_precision("high")
+
+
+def collect_metrics(benchmark_id, continue_metric_collection, metrics_recorder):
+    p = psutil.Process(os.getpid())
+    while not continue_metric_collection.is_set():
+        with p.oneshot():
+            cpu_util = p.cpu_percent()
+            mem_megabytes = p.memory_info().rss / (1024 * 1024)
+        gpu_stats = gpustat.GPUStatCollection.new_query()
+        gpu_util = gpu_stats[0]["utilization.gpu"]
+        gpu_mem_megabytes = gpu_stats[0]["memory.used"]
+        metrics_recorder.collect_device_measurements(
+            benchmark_id, cpu_util, mem_megabytes, gpu_util, gpu_mem_megabytes
+        )
+        sleep(0.01)
+
+
+def run_benchmark(
+    logger: Logger,
+    repository: str,
+    branch: str,
+    commit_id: str,
+    commit_msg: str,
+    metrics_recorder=None,
+    num_tokens_to_generate=100,
+):
+    # Check if required ML dependencies are available
+    if not TRANSFORMERS_AVAILABLE:
+        logger.error("Transformers and torch are required to run the LLaMA benchmark. Please install them with:")
+        logger.error("pip install torch transformers")
+        logger.error("Skipping LLaMA benchmark due to missing dependencies.")
+        return
+
+    continue_metric_collection = Event()
+    metrics_thread = None
+    model_id = "meta-llama/Llama-2-7b-hf"
+
+    # If no metrics_recorder is provided, create one for backward compatibility
+    if metrics_recorder is None:
+        try:
+            metrics_recorder = MetricsRecorder(
+                psycopg2.connect("dbname=metrics"), logger, repository, branch, commit_id, commit_msg, True
+            )
+            should_close_recorder = True
+        except Exception as e:
+            logger.error(f"Failed to create metrics recorder: {e}")
+            return
+    else:
+        should_close_recorder = False
+    try:
+        gpu_stats = gpustat.GPUStatCollection.new_query()
+        gpu_name = gpu_stats[0]["name"]
+        benchmark_id = metrics_recorder.initialise_benchmark({"gpu_name": gpu_name, "model_id": model_id})
+        logger.info(f"running benchmark #{benchmark_id} on {gpu_name} for {model_id}")
+        metrics_thread = Thread(
+            target=collect_metrics,
+            args=[benchmark_id, continue_metric_collection, metrics_recorder],
+        )
+        metrics_thread.start()
+        logger.info("started background thread to fetch device metrics")
+
+        os.environ["TOKENIZERS_PARALLELISM"] = "false"  # silence warnings when compiling
+
+        device = "cuda"
+
+        logger.info("downloading weights")
+        # This is to avoid counting download in model load time measurement
+        model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.float16)
+        gen_config = GenerationConfig(do_sample=False, top_p=1, temperature=1)
+        logger.info("loading model")
+        start = perf_counter()
+        model = AutoModelForCausalLM.from_pretrained(
+            model_id, dtype=torch.float16, generation_config=gen_config
+        ).eval()
+        model.to(device)
+        torch.cuda.synchronize()
+        end = perf_counter()
+        model_load_time = end - start
+        logger.info(f"loaded model in: {model_load_time}s")
+
+        tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+        prompt = "Why dogs are so cute?"
+        inputs = tokenizer(prompt, return_tensors="pt").to(device)
+
+        # Specify the max length (including both the prompt and the response)
+        # When calling `generate` with `cache_implementation="static" later, this is also used to create a `StaticCache` object
+        # with sequence length = `max_length`. The longer the more you will re-use it
+        seq_length = inputs["input_ids"].shape[1]
+        model.generation_config.max_length = seq_length + num_tokens_to_generate
+        batch_size = inputs["input_ids"].shape[0]
+
+        # Copied from the gpt-fast repo
+        def multinomial_sample_one_no_sync(probs_sort):  # Does multinomial sampling without a cuda synchronization
+            q = torch.empty_like(probs_sort).exponential_(1)
+            return torch.argmax(probs_sort / q, dim=-1, keepdim=True).to(dtype=torch.int)
+
+        def logits_to_probs(logits, temperature: float = 1.0, top_k: Optional[int] = None):
+            logits = logits / max(temperature, 1e-5)
+
+            if top_k is not None:
+                v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
+                pivot = v.select(-1, -1).unsqueeze(-1)
+                logits = torch.where(logits < pivot, -float("Inf"), logits)
+            probs = torch.nn.functional.softmax(logits, dim=-1)
+            return probs
+
+        def sample(logits, temperature: float = 1.0, top_k: Optional[int] = None):
+            probs = logits_to_probs(logits[0, -1], temperature, top_k)
+            idx_next = multinomial_sample_one_no_sync(probs)
+            return idx_next, probs
+
+        # First eager forward pass
+        logger.info("running first eager forward pass")
+        start = perf_counter()
+        _ = model(**inputs)
+        torch.cuda.synchronize()
+        end = perf_counter()
+        first_eager_fwd_pass_time = end - start
+        logger.info(f"completed first eager forward pass in: {first_eager_fwd_pass_time}s")
+
+        # Second eager forward pass (should be faster)
+        logger.info("running second eager forward pass")
+        start = perf_counter()
+        _ = model(**inputs)
+        torch.cuda.synchronize()
+        end = perf_counter()
+        second_eager_fwd_pass_time = end - start
+        logger.info(f"completed second eager forward pass in: {second_eager_fwd_pass_time}s")
+
+        # First eager generation
+        logger.info("running first eager generation")
+        start = perf_counter()
+        output = model.generate(**inputs)
+        torch.cuda.synchronize()
+        end = perf_counter()
+        first_eager_generate_time = end - start
+        logger.info(f"completed first eager generation in: {first_eager_generate_time}s")
+        logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
+
+        # Second eager generation (should be faster)
+        logger.info("running second eager generation")
+        start = perf_counter()
+        output = model.generate(**inputs)
+        torch.cuda.synchronize()
+        end = perf_counter()
+        second_eager_generate_time = end - start
+        logger.info(f"completed second eager generation in: {second_eager_generate_time}s")
+        logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
+
+        logger.info("running generation timing loop")
+
+        input_pos = torch.arange(0, seq_length, device=device)
+        inputs = inputs["input_ids"]
+
+        start = perf_counter()
+        with torch.nn.attention.sdpa_kernel(torch.nn.attention.SDPBackend.MATH):
+            logits = model(inputs, position_ids=input_pos).logits
+        next_token, probs = sample(logits, temperature=0.6, top_k=5)
+        torch.cuda.synchronize()
+        end = perf_counter()
+        time_to_first_token = end - start
+
+        input_pos = torch.tensor([seq_length], device=device, dtype=torch.int)
+        next_token = next_token.clone()
+        start = perf_counter()
+        with torch.nn.attention.sdpa_kernel(torch.nn.attention.SDPBackend.MATH):
+            logits = model(next_token, position_ids=input_pos).logits
+        next_token, probs = sample(logits, temperature=0.6, top_k=5)
+        torch.cuda.synchronize()
+        end = perf_counter()
+        time_to_second_token = end - start
+
+        input_pos = torch.tensor([seq_length + 1], device=device, dtype=torch.int)
+        next_token = next_token.clone()
+        start = perf_counter()
+        with torch.nn.attention.sdpa_kernel(torch.nn.attention.SDPBackend.MATH):
+            logits = model(next_token, position_ids=input_pos).logits
+        next_token, probs = sample(logits, temperature=0.6, top_k=5)
+        torch.cuda.synchronize()
+        end = perf_counter()
+        time_to_third_token = end - start
+
+        logger.info("running longer generation timing loop")
+
+        total_time = 0
+        for i in range(20):
+            input_pos = torch.tensor([seq_length + 2 + i], device=device, dtype=torch.int)
+            next_token = next_token.clone()
+            start = perf_counter()
+            with torch.nn.attention.sdpa_kernel(torch.nn.attention.SDPBackend.MATH):
+                logits = model(next_token, position_ids=input_pos).logits
+            next_token, probs = sample(logits, temperature=0.6, top_k=5)
+            torch.cuda.synchronize()
+            end = perf_counter()
+            total_time += end - start
+
+        mean_time_to_next_token = total_time / 20
+
+        logger.info("running compilation benchmarks")
+
+        # Now compile the model
+        model = torch.compile(model, mode="max-autotune", fullgraph=True)
+
+        # StaticCache for generation
+        with torch.device(device):
+            model.setup_caches(max_batch_size=batch_size, max_seq_len=seq_length + num_tokens_to_generate)
+
+        input_pos = torch.arange(0, seq_length, device=device)
+        inputs = tokenizer(prompt, return_tensors="pt").to(device)["input_ids"]
+
+        logger.info("compiling model")
+
+        model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.float16, generation_config=gen_config)
+        model.to(device)
+        model = torch.compile(model, mode="max-autotune", fullgraph=True)
+
+        past_key_values = StaticCache(
+            model.config,
+            max_batch_size=batch_size,
+            device=device,
+            dtype=torch.float16,
+            max_cache_len=seq_length + 128,
+        )
+        # 1st call
+        start = perf_counter()
+        output = model.generate(**inputs, past_key_values=past_key_values)
+        end = perf_counter()
+        first_compile_generate_time = end - start
+        logger.info(f"completed first compile generation in: {first_compile_generate_time}s")
+        logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
+
+        past_key_values = StaticCache(
+            model.config,
+            max_batch_size=batch_size,
+            device=device,
+            dtype=torch.float16,
+            max_cache_len=seq_length + 128,
+        )
+        # 2nd call
+        start = perf_counter()
+        output = model.generate(**inputs, past_key_values=past_key_values)
+        end = perf_counter()
+        second_compile_generate_time = end - start
+        logger.info(f"completed second compile generation in: {second_compile_generate_time}s")
+        logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
+
+        past_key_values = StaticCache(
+            model.config,
+            max_batch_size=batch_size,
+            device=device,
+            dtype=torch.float16,
+            max_cache_len=seq_length + 128,
+        )
+        # 3rd call
+        start = perf_counter()
+        output = model.generate(**inputs, past_key_values=past_key_values)
+        end = perf_counter()
+        third_compile_generate_time = end - start
+        logger.info(f"completed third compile generation in: {third_compile_generate_time}s")
+        logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
+
+        past_key_values = StaticCache(
+            model.config,
+            max_batch_size=batch_size,
+            device=device,
+            dtype=torch.float16,
+            max_cache_len=seq_length + 128,
+        )
+        # 4th call
+        start = perf_counter()
+        output = model.generate(**inputs, past_key_values=past_key_values)
+        end = perf_counter()
+        fourth_compile_generate_time = end - start
+        logger.info(f"completed fourth compile generation in: {fourth_compile_generate_time}s")
+        logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
+
+        metrics_recorder.collect_model_measurements(
+            benchmark_id,
+            {
+                "model_load_time": model_load_time,
+                "first_eager_forward_pass_time_secs": first_eager_fwd_pass_time,
+                "second_eager_forward_pass_time_secs": second_eager_fwd_pass_time,
+                "first_eager_generate_time_secs": first_eager_generate_time,
+                "second_eager_generate_time_secs": second_eager_generate_time,
+                "time_to_first_token_secs": time_to_first_token,
+                "time_to_second_token_secs": time_to_second_token,
+                "time_to_third_token_secs": time_to_third_token,
+                "time_to_next_token_mean_secs": mean_time_to_next_token,
+                "first_compile_generate_time_secs": first_compile_generate_time,
+                "second_compile_generate_time_secs": second_compile_generate_time,
+                "third_compile_generate_time_secs": third_compile_generate_time,
+                "fourth_compile_generate_time_secs": fourth_compile_generate_time,
+            },
+        )
+    except Exception as e:
+        logger.error(f"Caught exception: {e}")
+    continue_metric_collection.set()
+    if metrics_thread is not None:
+        metrics_thread.join()
+
+    # Only close the recorder if we created it locally
+    if should_close_recorder:
+        metrics_recorder.close()

transformers/benchmark/benchmark.py

@@ -0,0 +1,324 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Run benchmark using the `optimum-benchmark` library with some customization in `transformers`.
+
+Assume we are under `transformers` root directory: (make sure the commits are valid commits)
+```bash
+python benchmark/benchmark.py --config-dir benchmark/config --config-name generation --commit=9b9c7f03da625b13643e99205c691fe046461724 --metrics=decode.latency.mean,per_token.latency.mean,per_token.throughput.value backend.model=google/gemma-2b benchmark.input_shapes.sequence_length=5,7 benchmark.input_shapes.batch_size=1,2 --multirun
+```
+"""
+
+import argparse
+import glob
+import json
+import os.path
+import re
+import tempfile
+from contextlib import contextmanager
+from pathlib import Path
+
+from git import Repo
+from huggingface_hub import HfApi
+from optimum_benchmark import Benchmark
+from optimum_benchmark_wrapper import main
+
+
+PATH_TO_REPO = Path(__file__).parent.parent.resolve()
+
+
+@contextmanager
+def checkout_commit(repo: Repo, commit_id: str):
+    """
+    Context manager that checks out a given commit when entered, but gets back to the reference it was at on exit.
+    Args:
+        repo (`git.Repo`): A git repository (for instance the Transformers repo).
+        commit_id (`str`): The commit reference to checkout inside the context manager.
+    """
+    current_head = repo.head.commit if repo.head.is_detached else repo.head.ref
+
+    try:
+        repo.git.checkout(commit_id)
+        yield
+
+    finally:
+        repo.git.checkout(current_head)
+
+
+def summarize(run_dir, metrics, expand_metrics=False):
+    """Produce a summary for each optimum-benchmark launched job's output directory found in `run_dir`.
+
+    Each summary's format is as follows (for `expand_metrics=False`):
+    ```
+    {
+        "model": "google/gemma-2b",
+        "commit": "3cd6ed22e4d49219f300f5055e71e3929aba20d7",
+        "config": "benchmark.input_shapes.batch_size=1,benchmark.input_shapes.sequence_length=5",
+        "metrics": {
+            "decode.latency.mean": 1.624666809082031,
+            "per_token.latency.mean": 0.012843788806628804,
+            "per_token.throughput.value": 77.85864553330948
+        }
+    }
+    ```
+    """
+    reports = glob.glob(os.path.join(run_dir, "**/benchmark_report.json"), recursive=True)
+    report_dirs = [str(Path(report).parent) for report in reports]
+
+    summaries = []
+    for report_dir in report_dirs:
+        commit = re.search(r"/commit=([^/]+)", report_dir).groups()[0]
+
+        if not os.path.isfile(os.path.join(report_dir, "benchmark.json")):
+            continue
+        benchmark = Benchmark.from_json(os.path.join(report_dir, "benchmark.json"))
+        report = benchmark.report
+
+        model = benchmark.config.backend["model"]
+
+        # This looks like `benchmark.input_shapes.batch_size=1,benchmark.input_shapes.sequence_length=5`.
+        # (we rely on the usage of hydra's `${hydra.job.override_dirname}`.)
+        benchmark_name = re.sub(f"backend.model={model},*", "", report_dir)
+        benchmark_name = str(Path(benchmark_name).parts[-1])
+        if benchmark_name.startswith("commit="):
+            benchmark_name = benchmark.config.name
+
+        metrics_values = {}
+        # post-processing of report: show a few selected/important metric
+        for metric in metrics:
+            keys = metric.split(".")
+            value = report.to_dict()
+            current = metrics_values
+            for key in keys:
+                # Avoid KeyError when a user's specified metric has typo.
+                # TODO: Give warnings.
+                if key not in value:
+                    continue
+                value = value[key]
+
+                if expand_metrics:
+                    if isinstance(value, dict):
+                        if key not in current:
+                            current[key] = {}
+                            current = current[key]
+                    else:
+                        current[key] = value
+
+            if not expand_metrics:
+                metrics_values[metric] = value
+
+        # show some config information
+        print(f"model: {model}")
+        print(f"commit: {commit}")
+        print(f"config: {benchmark_name}")
+        if len(metrics_values) > 0:
+            print("metrics:")
+            if expand_metrics:
+                print(metrics_values)
+            else:
+                for metric, value in metrics_values.items():
+                    print(f"  - {metric}: {value}")
+        print("-" * 80)
+
+        summary = {
+            "model": model,
+            "commit": commit,
+            "config": benchmark_name,
+            "metrics": metrics_values,
+        }
+        summaries.append(summary)
+
+        with open(os.path.join(report_dir, "summary.json"), "w") as fp:
+            json.dump(summary, fp, indent=4)
+
+    return summaries
+
+
+def combine_summaries(summaries):
+    """Combine a list of summary obtained from the function `summarize`.
+
+    The combined summary's format is as follows:
+    ```
+    "google/gemma-2b": {
+        "benchmark.input_shapes.batch_size=1,benchmark.input_shapes.sequence_length=5": {
+            "3cd6ed22e4d49219f300f5055e71e3929aba20d7": {
+                "metrics": {"decode.latency.mean": 1.624666809082031}
+            },
+            "c97ee28b117c0abe8e08891f402065e4df6d72aa": {
+                "metrics": {"decode.latency.mean": 1.6278163452148438}
+            }
+        },
+        "benchmark.input_shapes.batch_size=2,benchmark.input_shapes.sequence_length=5": {
+            "3cd6ed22e4d49219f300f5055e71e3929aba20d7": {
+                "metrics": {"decode.latency.mean": 1.6947791748046876}
+            },
+            "c97ee28b117c0abe8e08891f402065e4df6d72aa": {
+                "metrics": {
+                    "decode.latency.mean": 1.6980519409179688}
+            }
+        }
+    }
+    ```
+    """
+    combined = {}
+    for summary in summaries:
+        model = summary["model"]
+        config = summary["config"]
+        commit = summary["commit"]
+
+        if model not in combined:
+            combined[model] = {}
+
+        if config not in combined[model]:
+            combined[model][config] = {}
+
+        if commit not in combined[model][config]:
+            combined[model][config][commit] = {"metrics": summary["metrics"]}
+
+    with open(os.path.join(exp_run_dir, "summary.json"), "w") as fp:
+        json.dump(combined, fp, indent=4)
+
+    print(json.dumps(combined, indent=4))
+
+    return combined
+
+
+if __name__ == "__main__":
+
+    def list_str(values):
+        return values.split(",")
+
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument("--config-dir", type=str, required=True, help="The path to the config directory.")
+    parser.add_argument("--config-name", type=str, required=True, help="The config name.")
+
+    # arguments specific to this wrapper for our own customization
+    parser.add_argument("--ensure_empty", type=bool, default=True, help="If to create a temporary directory.")
+    parser.add_argument(
+        "--commit",
+        type=list_str,
+        default="",
+        help="Comma-separated list of branch names and/or commit sha values on which the benchmark will run. If `diff` is specified, it will run on both the current head and the `main` branch.",
+    )
+    parser.add_argument("--metrics", type=str, help="The metrics to be included in the summary.")
+
+    parser.add_argument("--repo_id", type=str, default=None, help="The repository to which the file will be uploaded.")
+    parser.add_argument("--path_in_repo", type=str, default=None, help="Relative filepath in the repo.")
+    parser.add_argument("--token", type=str, default=None, help="A valid user access token (string).")
+
+    args, optimum_benchmark_args = parser.parse_known_args()
+
+    repo = Repo(PATH_TO_REPO)
+
+    metrics = [
+        "prefill.latency.mean",
+        "prefill.throughput.value",
+        "decode.latency.mean",
+        "decode.throughput.value",
+        "per_token.latency.mean",
+        "per_token.throughput.value",
+    ]
+    if args.metrics is not None:
+        metrics = args.metrics.split(",")
+
+    # Get `backend.model` in a hacky way: We want to control the experiment flow manually.
+    models = [""]
+    for idx, arg in enumerate(optimum_benchmark_args):
+        if arg.startswith("backend.model="):
+            models = arg[len("backend.model=") :]
+            models = models.split(",")
+            break
+    optimum_benchmark_args = [arg for arg in optimum_benchmark_args if not arg.startswith("backend.model=")]
+
+    # Get the commit(s)
+    current_head = str(repo.head.commit) if repo.head.is_detached else str(repo.head.ref)
+    commits = [x for x in args.commit if x != ""]
+    if len(commits) == 0:
+        commits = [current_head]
+    elif len(commits) == 1 and commits[0] == "diff":
+        # compare to `main`
+        commits = ["main", current_head]
+
+    # Get the specified run directory
+    run_dir_arg_idx, run_dir = -1, None
+    sweep_dir_arg_idx, sweep_dir = -1, None
+    for idx, arg in enumerate(optimum_benchmark_args):
+        if arg.startswith("hydra.run.dir="):
+            run_dir = arg[len("hydra.run.dir=") :]
+            run_dir_arg_idx = idx
+        elif arg.startswith("hydra.sweep.dir="):
+            sweep_dir = arg[len("hydra.sweep.dir=") :]
+            sweep_dir_arg_idx = idx
+    exp_run_dir, arg_dix, arg_name = (
+        (sweep_dir, sweep_dir_arg_idx, "hydra.sweep.dir")
+        if "--multirun" in optimum_benchmark_args
+        else (run_dir, run_dir_arg_idx, "hydra.run.dir")
+    )
+
+    # TODO: not hardcoded
+    if exp_run_dir is None and args.ensure_empty:
+        exp_run_dir = "_benchmark"
+
+    if args.ensure_empty:
+        os.makedirs(exp_run_dir, exist_ok=True)
+        exp_run_dir = tempfile.mkdtemp(dir=exp_run_dir)
+
+    run_summaries = []
+    for commit in commits:
+        with checkout_commit(repo, commit):
+            commit = str(repo.head.commit)
+
+            commit_run_dir = exp_run_dir
+            if exp_run_dir is not None:
+                commit_run_dir = os.path.join(exp_run_dir, rf"commit\={commit}")
+
+            print(f"Run benchmark on commit: {commit}")
+
+            for model in models:
+                model_arg = [f"backend.model={model}"] if model != "" else []
+                dir_args = []
+                if commit_run_dir is not None:
+                    if arg_dix > -1:
+                        optimum_benchmark_args[arg_dix] = f"{arg_name}={commit_run_dir}"
+                    else:
+                        dir_args = [
+                            f"hydra.sweep.dir={commit_run_dir}",
+                            f"hydra.run.dir={commit_run_dir}/" + "${hydra.job.override_dirname}",
+                        ]
+                main(args.config_dir, args.config_name, model_arg + dir_args + optimum_benchmark_args)
+
+            if commit_run_dir is not None:
+                # Need to remove the `\` character
+                summaries = summarize(commit_run_dir.replace("\\", ""), metrics)
+                run_summaries.extend(summaries)
+
+    # aggregate the information across the commits
+    if exp_run_dir is not None:
+        with open(os.path.join(exp_run_dir, "summaries.json"), "w") as fp:
+            json.dump(run_summaries, fp, indent=4)
+
+        combined_summary = combine_summaries(run_summaries)
+
+        if args.repo_id is not None and args.path_in_repo is not None:
+            # Upload to Hub
+            api = HfApi()
+            api.upload_folder(
+                folder_path=exp_run_dir,
+                path_in_repo=args.path_in_repo,
+                repo_id=args.repo_id,
+                repo_type="dataset",
+                token=args.token,
+            )

transformers/benchmark/benchmarks_entrypoint.py

@@ -0,0 +1,502 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import argparse
+import importlib.util
+import json
+import logging
+import os
+import sys
+import uuid
+from datetime import datetime
+
+import pandas as pd
+
+
+try:
+    from psycopg2.extensions import register_adapter
+    from psycopg2.extras import Json
+
+    register_adapter(dict, Json)
+    PSYCOPG2_AVAILABLE = True
+except ImportError:
+    PSYCOPG2_AVAILABLE = False
+
+
+class ImportModuleException(Exception):
+    pass
+
+
+class MetricsRecorder:
+    def __init__(
+        self,
+        connection,
+        logger: logging.Logger,
+        repository: str,
+        branch: str,
+        commit_id: str,
+        commit_msg: str,
+        collect_csv_data: bool = True,
+    ):
+        self.conn = connection
+        self.use_database = connection is not None
+        if self.use_database:
+            self.conn.autocommit = True
+        self.logger = logger
+        self.repository = repository
+        self.branch = branch
+        self.commit_id = commit_id
+        self.commit_msg = commit_msg
+        self.collect_csv_data = collect_csv_data
+
+        # For CSV export - store all data in pandas DataFrames (only if CSV collection is enabled)
+        if self.collect_csv_data:
+            # Initialize empty DataFrames with proper schemas
+            self.benchmarks_df = pd.DataFrame(
+                columns=[
+                    "benchmark_id",
+                    "repository",
+                    "branch",
+                    "commit_id",
+                    "commit_message",
+                    "metadata",
+                    "created_at",
+                ]
+            )
+            self.device_measurements_df = pd.DataFrame(
+                columns=["benchmark_id", "cpu_util", "mem_megabytes", "gpu_util", "gpu_mem_megabytes", "time"]
+            )
+            self.model_measurements_df = pd.DataFrame(
+                columns=[
+                    "benchmark_id",
+                    "time",
+                    "model_load_time",
+                    "first_eager_forward_pass_time_secs",
+                    "second_eager_forward_pass_time_secs",
+                    "first_eager_generate_time_secs",
+                    "second_eager_generate_time_secs",
+                    "time_to_first_token_secs",
+                    "time_to_second_token_secs",
+                    "time_to_third_token_secs",
+                    "time_to_next_token_mean_secs",
+                    "first_compile_generate_time_secs",
+                    "second_compile_generate_time_secs",
+                    "third_compile_generate_time_secs",
+                    "fourth_compile_generate_time_secs",
+                ]
+            )
+        else:
+            self.benchmarks_df = None
+            self.device_measurements_df = None
+            self.model_measurements_df = None
+
+    def initialise_benchmark(self, metadata: dict[str, str]) -> str:
+        """
+        Creates a new benchmark, returns the benchmark id (UUID)
+        """
+        # Generate a unique UUID for this benchmark
+        benchmark_id = str(uuid.uuid4())
+
+        if self.use_database:
+            with self.conn.cursor() as cur:
+                cur.execute(
+                    "INSERT INTO benchmarks (benchmark_id, repository, branch, commit_id, commit_message, metadata) VALUES (%s, %s, %s, %s, %s, %s)",
+                    (benchmark_id, self.repository, self.branch, self.commit_id, self.commit_msg, metadata),
+                )
+                self.logger.debug(f"initialised benchmark #{benchmark_id}")
+
+        # Store benchmark data for CSV export (if enabled)
+        if self.collect_csv_data:
+            # Add row to pandas DataFrame
+            new_row = pd.DataFrame(
+                [
+                    {
+                        "benchmark_id": benchmark_id,
+                        "repository": self.repository,
+                        "branch": self.branch,
+                        "commit_id": self.commit_id,
+                        "commit_message": self.commit_msg,
+                        "metadata": json.dumps(metadata),
+                        "created_at": datetime.utcnow().isoformat(),
+                    }
+                ]
+            )
+            self.benchmarks_df = pd.concat([self.benchmarks_df, new_row], ignore_index=True)
+
+        mode_info = []
+        if self.use_database:
+            mode_info.append("database")
+        if self.collect_csv_data:
+            mode_info.append("CSV")
+        mode_str = " + ".join(mode_info) if mode_info else "no storage"
+
+        self.logger.debug(f"initialised benchmark #{benchmark_id} ({mode_str} mode)")
+        return benchmark_id
+
+    def collect_device_measurements(self, benchmark_id: str, cpu_util, mem_megabytes, gpu_util, gpu_mem_megabytes):
+        """
+        Collect device metrics, such as CPU & GPU usage. These are "static", as in you cannot pass arbitrary arguments to the function.
+        """
+        # Store device measurements for CSV export (if enabled)
+        if self.collect_csv_data:
+            # Add row to pandas DataFrame
+            new_row = pd.DataFrame(
+                [
+                    {
+                        "benchmark_id": benchmark_id,
+                        "cpu_util": cpu_util,
+                        "mem_megabytes": mem_megabytes,
+                        "gpu_util": gpu_util,
+                        "gpu_mem_megabytes": gpu_mem_megabytes,
+                        "time": datetime.utcnow().isoformat(),
+                    }
+                ]
+            )
+            self.device_measurements_df = pd.concat([self.device_measurements_df, new_row], ignore_index=True)
+
+        # Store in database if available
+        if self.use_database:
+            with self.conn.cursor() as cur:
+                cur.execute(
+                    "INSERT INTO device_measurements (benchmark_id, cpu_util, mem_megabytes, gpu_util, gpu_mem_megabytes) VALUES (%s, %s, %s, %s, %s)",
+                    (benchmark_id, cpu_util, mem_megabytes, gpu_util, gpu_mem_megabytes),
+                )
+
+        self.logger.debug(
+            f"collected device measurements for benchmark #{benchmark_id} [CPU util: {cpu_util}, mem MBs: {mem_megabytes}, GPU util: {gpu_util}, GPU mem MBs: {gpu_mem_megabytes}]"
+        )
+
+    def collect_model_measurements(self, benchmark_id: str, measurements: dict[str, float]):
+        # Store model measurements for CSV export (if enabled)
+        if self.collect_csv_data:
+            # Add row to pandas DataFrame with flattened measurements
+            row_data = {"benchmark_id": benchmark_id, "time": datetime.utcnow().isoformat()}
+            # Flatten the measurements dict into the row
+            row_data.update(measurements)
+
+            new_row = pd.DataFrame([row_data])
+            self.model_measurements_df = pd.concat([self.model_measurements_df, new_row], ignore_index=True)
+
+        # Store in database if available
+        if self.use_database:
+            with self.conn.cursor() as cur:
+                cur.execute(
+                    """
+                    INSERT INTO model_measurements (
+                        benchmark_id,
+                        measurements
+                    ) VALUES (%s, %s)
+                    """,
+                    (
+                        benchmark_id,
+                        measurements,
+                    ),
+                )
+
+        self.logger.debug(f"collected model measurements for benchmark #{benchmark_id}: {measurements}")
+
+    def export_to_csv(self, output_dir: str = "benchmark_results"):
+        """
+        Export all collected data to CSV files using pandas DataFrames
+        """
+        if not self.collect_csv_data:
+            self.logger.warning("CSV data collection is disabled - no CSV files will be generated")
+            return
+
+        if not os.path.exists(output_dir):
+            os.makedirs(output_dir)
+            self.logger.info(f"Created output directory: {output_dir}")
+
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        files_created = []
+
+        # Export using pandas DataFrames
+        self._export_pandas_data(output_dir, timestamp, files_created)
+
+        self.logger.info(f"CSV export complete! Created {len(files_created)} files in {output_dir}")
+
+    def _export_pandas_data(self, output_dir: str, timestamp: str, files_created: list):
+        """
+        Export CSV files using pandas DataFrames
+        """
+        # Export benchmarks
+        benchmarks_file = os.path.join(output_dir, f"benchmarks_{timestamp}.csv")
+        self.benchmarks_df.to_csv(benchmarks_file, index=False)
+        files_created.append(benchmarks_file)
+        self.logger.info(f"Exported {len(self.benchmarks_df)} benchmark records to {benchmarks_file}")
+
+        # Export device measurements
+        device_file = os.path.join(output_dir, f"device_measurements_{timestamp}.csv")
+        self.device_measurements_df.to_csv(device_file, index=False)
+        files_created.append(device_file)
+        self.logger.info(f"Exported {len(self.device_measurements_df)} device measurement records to {device_file}")
+
+        # Export model measurements (already flattened)
+        model_file = os.path.join(output_dir, f"model_measurements_{timestamp}.csv")
+        self.model_measurements_df.to_csv(model_file, index=False)
+        files_created.append(model_file)
+        self.logger.info(f"Exported {len(self.model_measurements_df)} model measurement records to {model_file}")
+
+        # Create comprehensive summary using pandas operations
+        summary_file = os.path.join(output_dir, f"benchmark_summary_{timestamp}.csv")
+        self._create_summary(summary_file)
+        files_created.append(summary_file)
+
+    def _create_summary(self, summary_file: str):
+        """
+        Create a comprehensive summary CSV using pandas operations
+        """
+        if len(self.benchmarks_df) == 0:
+            # Create empty summary file
+            summary_df = pd.DataFrame()
+            summary_df.to_csv(summary_file, index=False)
+            self.logger.info(f"Created empty benchmark summary at {summary_file}")
+            return
+
+        # Start with benchmarks as the base
+        summary_df = self.benchmarks_df.copy()
+
+        # Add model measurements (join on benchmark_id)
+        if len(self.model_measurements_df) > 0:
+            # Drop 'time' column from model measurements to avoid conflicts
+            model_df = self.model_measurements_df.drop(columns=["time"], errors="ignore")
+            summary_df = summary_df.merge(model_df, on="benchmark_id", how="left")
+
+        # Calculate device measurement aggregates using pandas groupby
+        if len(self.device_measurements_df) > 0:
+            device_agg = (
+                self.device_measurements_df.groupby("benchmark_id")
+                .agg(
+                    {
+                        "cpu_util": ["mean", "max", "std", "count"],
+                        "mem_megabytes": ["mean", "max", "std"],
+                        "gpu_util": ["mean", "max", "std"],
+                        "gpu_mem_megabytes": ["mean", "max", "std"],
+                    }
+                )
+                .round(3)
+            )
+
+            # Flatten column names
+            device_agg.columns = [f"{col[0]}_{col[1]}" for col in device_agg.columns]
+            device_agg = device_agg.reset_index()
+
+            # Rename count column to be more descriptive
+            if "cpu_util_count" in device_agg.columns:
+                device_agg = device_agg.rename(columns={"cpu_util_count": "device_measurement_count"})
+
+            # Merge with summary
+            summary_df = summary_df.merge(device_agg, on="benchmark_id", how="left")
+
+        # Export the comprehensive summary
+        summary_df.to_csv(summary_file, index=False)
+        self.logger.info(f"Created comprehensive benchmark summary with {len(summary_df)} records at {summary_file}")
+
+    def close(self):
+        if self.use_database and self.conn:
+            self.conn.close()
+
+
+logger = logging.getLogger(__name__)
+logger.setLevel(logging.INFO)
+
+handler = logging.StreamHandler(sys.stdout)
+handler.setLevel(logging.INFO)
+formatter = logging.Formatter("[%(levelname)s - %(asctime)s] %(message)s")
+handler.setFormatter(formatter)
+logger.addHandler(handler)
+
+
+def parse_arguments() -> tuple[str, str, str, str, bool, str]:
+    """
+    Parse command line arguments for the benchmarking CLI.
+    """
+    parser = argparse.ArgumentParser(description="CLI for benchmarking the huggingface/transformers.")
+
+    parser.add_argument(
+        "repository",
+        type=str,
+        help="The repository name on which the benchmarking is performed.",
+    )
+
+    parser.add_argument(
+        "branch",
+        type=str,
+        help="The branch name on which the benchmarking is performed.",
+    )
+
+    parser.add_argument(
+        "commit_id",
+        type=str,
+        help="The commit hash on which the benchmarking is performed.",
+    )
+
+    parser.add_argument(
+        "commit_msg",
+        type=str,
+        help="The commit message associated with the commit, truncated to 70 characters.",
+    )
+
+    parser.add_argument("--csv", action="store_true", default=False, help="Enable CSV output files generation.")
+
+    parser.add_argument(
+        "--csv-output-dir",
+        type=str,
+        default="benchmark_results",
+        help="Directory for CSV output files (default: benchmark_results).",
+    )
+
+    args = parser.parse_args()
+
+    # CSV is disabled by default, only enabled when --csv is used
+    generate_csv = args.csv
+
+    return args.repository, args.branch, args.commit_id, args.commit_msg, generate_csv, args.csv_output_dir
+
+
+def import_from_path(module_name, file_path):
+    try:
+        spec = importlib.util.spec_from_file_location(module_name, file_path)
+        module = importlib.util.module_from_spec(spec)
+        sys.modules[module_name] = module
+        spec.loader.exec_module(module)
+        return module
+    except Exception as e:
+        raise ImportModuleException(f"failed to load python module: {e}")
+
+
+def create_database_connection():
+    """
+    Try to create a database connection. Returns None if connection fails.
+    """
+    if not PSYCOPG2_AVAILABLE:
+        logger.warning("psycopg2 not available - running in CSV-only mode")
+        return None
+
+    try:
+        import psycopg2
+
+        conn = psycopg2.connect("dbname=metrics")
+        logger.info("Successfully connected to database")
+        return conn
+    except Exception as e:
+        logger.warning(f"Failed to connect to database: {e}. Running in CSV-only mode")
+        return None
+
+
+def create_global_metrics_recorder(
+    repository: str, branch: str, commit_id: str, commit_msg: str, generate_csv: bool = False
+) -> MetricsRecorder:
+    """
+    Create a global metrics recorder that will be used across all benchmarks.
+    """
+    connection = create_database_connection()
+    recorder = MetricsRecorder(connection, logger, repository, branch, commit_id, commit_msg, generate_csv)
+
+    # Log the storage mode
+    storage_modes = []
+    if connection is not None:
+        storage_modes.append("database")
+    if generate_csv:
+        storage_modes.append("CSV")
+
+    if not storage_modes:
+        logger.warning("Running benchmarks with NO data storage (no database connection, CSV disabled)")
+        logger.warning("Use --csv flag to enable CSV output when database is unavailable")
+    else:
+        logger.info(f"Running benchmarks with: {' + '.join(storage_modes)} storage")
+
+    return recorder
+
+
+if __name__ == "__main__":
+    benchmarks_folder_path = os.path.dirname(os.path.realpath(__file__))
+    benches_folder_path = os.path.join(benchmarks_folder_path, "benches")
+
+    repository, branch, commit_id, commit_msg, generate_csv, csv_output_dir = parse_arguments()
+
+    # Create a global metrics recorder
+    global_metrics_recorder = create_global_metrics_recorder(repository, branch, commit_id, commit_msg, generate_csv)
+
+    successful_benchmarks = 0
+    failed_benchmarks = 0
+
+    # Automatically discover all benchmark modules in benches/ folder
+    benchmark_modules = []
+
+    if os.path.exists(benches_folder_path):
+        logger.debug(f"Scanning for benchmarks in: {benches_folder_path}")
+        for entry in os.scandir(benches_folder_path):
+            if not entry.name.endswith(".py"):
+                continue
+            if entry.name.startswith("__"):  # Skip __init__.py, __pycache__, etc.
+                continue
+
+            # Check if the file has a run_benchmark function
+            try:
+                logger.debug(f"checking if benches/{entry.name} has run_benchmark function")
+                module = import_from_path(entry.name.split(".")[0], entry.path)
+                if hasattr(module, "run_benchmark"):
+                    benchmark_modules.append(entry.name)
+                    logger.debug(f"discovered benchmark: {entry.name}")
+                else:
+                    logger.debug(f"skipping {entry.name} - no run_benchmark function found")
+            except Exception as e:
+                logger.debug(f"failed to check benches/{entry.name}: {e}")
+    else:
+        logger.warning(f"Benches directory not found: {benches_folder_path}")
+
+    if benchmark_modules:
+        logger.info(f"Discovered {len(benchmark_modules)} benchmark(s): {benchmark_modules}")
+    else:
+        logger.warning("No benchmark modules found in benches/ directory")
+
+    for module_name in benchmark_modules:
+        module_path = os.path.join(benches_folder_path, module_name)
+        try:
+            logger.debug(f"loading: {module_name}")
+            module = import_from_path(module_name.split(".")[0], module_path)
+            logger.info(f"running benchmarks in: {module_name}")
+
+            # Check if the module has an updated run_benchmark function that accepts metrics_recorder
+            try:
+                # Try the new signature first
+                module.run_benchmark(logger, repository, branch, commit_id, commit_msg, global_metrics_recorder)
+            except TypeError:
+                # Fall back to the old signature for backward compatibility
+                logger.warning(
+                    f"Module {module_name} using old run_benchmark signature - database connection will be created per module"
+                )
+                module.run_benchmark(logger, repository, branch, commit_id, commit_msg)
+
+            successful_benchmarks += 1
+        except ImportModuleException as e:
+            logger.error(e)
+            failed_benchmarks += 1
+        except Exception as e:
+            logger.error(f"error running benchmarks for {module_name}: {e}")
+            failed_benchmarks += 1
+
+    # Export CSV results at the end (if enabled)
+    try:
+        if generate_csv:
+            global_metrics_recorder.export_to_csv(csv_output_dir)
+            logger.info(f"CSV reports have been generated and saved to the {csv_output_dir} directory")
+        else:
+            logger.info("CSV generation disabled - no CSV files created (use --csv to enable)")
+
+        logger.info(f"Benchmark run completed. Successful: {successful_benchmarks}, Failed: {failed_benchmarks}")
+    except Exception as e:
+        logger.error(f"Failed to export CSV results: {e}")
+    finally:
+        global_metrics_recorder.close()

transformers/benchmark/config/generation.yaml

@@ -0,0 +1,57 @@
+defaults:
+  - benchmark # inheriting benchmark schema
+  - scenario: inference
+  - launcher: process
+  - backend: pytorch
+  - _self_ # for hydra 1.1 compatibility
+
+name: pytorch_generate
+
+launcher:
+  start_method: spawn
+  device_isolation: true
+  device_isolation_action: warn
+
+backend:
+  device: cuda
+  device_ids: 0
+  no_weights: true
+  model: meta-llama/Llama-2-7b-hf
+  cache_implementation: static
+  torch_compile: true
+  dtype: float16
+  torch_compile_config:
+    backend: inductor
+    mode: reduce-overhead
+    fullgraph: true
+
+scenario:
+  input_shapes:
+    batch_size: 1
+    sequence_length: 7
+  generate_kwargs:
+    max_new_tokens: 128
+    min_new_tokens: 128
+    do_sample: false
+  memory: true
+  latency: true
+  iterations: 2
+  duration: 0
+
+
+# hydra/cli specific settings
+hydra:
+  run:
+    # where to store run results
+    dir: runs/${name}
+  job:
+    # change working directory to the run directory
+    chdir: true
+    env_set:
+      # set environment variable OVERRIDE_BENCHMARKS to 1
+      # to not skip benchmarks that have been run before
+      OVERRIDE_BENCHMARKS: 1
+      LOG_LEVEL: WARN
+  sweep:
+    dir: multirun
+    subdir: ${hydra.job.override_dirname}

transformers/benchmark/default.yml

@@ -0,0 +1,10 @@
+apiVersion: 1
+
+providers:
+  - name: 'Transformers Benchmarks'
+    orgId: 1
+    type: file
+    updateIntervalSeconds: 10
+    allowUiUpdates: true
+    options:
+      path: /etc/grafana/dashboards

transformers/benchmark/grafana_datasource.yaml

@@ -0,0 +1,17 @@
+apiVersion: 1
+datasources:
+  - name: grafana-postgresql-datasource
+    uid: be28nkzirtb0gd
+    type: postgres
+    url: $GRAFANA_POSTGRES_DATASOURCE_URL
+    user: $GRAFANA_POSTGRES_DATASOURCE_USER
+    secureJsonData:
+      password: $GRAFANA_POSTGRES_DATASOURCE_PWD
+    jsonData:
+      database: metrics
+      maxOpenConns: 100
+      maxIdleConns: 100
+      maxIdleConnsAuto: true
+      connMaxLifetime: 14400
+      postgresVersion: 1000
+      timescaledb: false

transformers/benchmark/optimum_benchmark_wrapper.py

@@ -0,0 +1,20 @@
+import argparse
+import subprocess
+
+
+def main(config_dir, config_name, args):
+    subprocess.run(
+        ["optimum-benchmark", "--config-dir", f"{config_dir}", "--config-name", f"{config_name}"]
+        + ["hydra/job_logging=disabled", "hydra/hydra_logging=disabled"]
+        + args
+    )
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument("--config-dir", type=str, required=True, help="The path to the config directory.")
+    parser.add_argument("--config-name", type=str, required=True, help="The config name.")
+    args, unknown = parser.parse_known_args()
+
+    main(args.config_dir, args.config_name, unknown)

transformers/benchmark/requirements.txt

@@ -0,0 +1,6 @@
+gpustat==1.1.1
+psutil==6.0.0
+psycopg2==2.9.9
+torch>=2.4.0
+hf_transfer
+pandas>=1.5.0