acprof 0.3.0__py3-none-any.whl

acprof/config.py

@@ -0,0 +1,177 @@
+"""AC-Prof - Configuration & Constants."""
+
+from dataclasses import dataclass
+from typing import Any
+
+from acprof.extensions import CATALOG
+
+
+# ─────────────────────────────────────────────
+# 各任务族的输入缩放维度
+# ─────────────────────────────────────────────
+@dataclass
+class ScalingConfig:
+    param_name: str
+    values: list
+    csv_field: str = "input_scale"
+    description: str = ""
+
+SCALING_DIMENSIONS: dict[str, ScalingConfig] = {
+    family: ScalingConfig(**{**values, "values": list(values["values"])})
+    for family, values in CATALOG.family_scaling().items()
+}
+DEFAULT_TASK_PARAMS: dict[str, dict[str, Any]] = CATALOG.family_task_params()
+
+# ─────────────────────────────────────────────
+# 默认资源矩阵
+# ─────────────────────────────────────────────
+DEFAULT_CPU_LIST = [1, 2, 4, 8]
+DEFAULT_MEM_LIST = [2, 4, 8, 16]
+DEFAULT_GPU_LIST = ["off", "on"]
+
+# ─────────────────────────────────────────────
+# 实验参数默认值
+# ─────────────────────────────────────────────
+DEFAULT_BATCH_SIZE = 1
+DEFAULT_WARMUP = 2
+DEFAULT_REPEAT = 5
+DEFAULT_REPEAT_IN_WINDOW = 0
+DEFAULT_REPEAT_WINDOW_SECONDS = 10.0
+DEFAULT_REQUEST_TIMEOUT_SECONDS = 300.0
+CLIENT_REQUEST_TIMEOUT_EXIT_CODE = 9
+DEFAULT_SAMPLE_HZ = 20.0
+DEFAULT_IDLE_SECONDS = 20.0
+DEFAULT_IDLE_COOLDOWN_SECONDS = 5.0
+DEFAULT_COOLDOWN_SECONDS = DEFAULT_IDLE_COOLDOWN_SECONDS
+DEFAULT_COMPUTE_PROFILE_TOOL = "none"
+IDLE_DIAG_DIRNAME = "debug_idle_diag"
+SERVER_PORT = 8002
+READY_TIMEOUT_S = 180
+READY_POLL_INTERVAL_S = 0.1
+
+# ─────────────────────────────────────────────
+# CSV 输出字段
+# ─────────────────────────────────────────────
+# 保留公共字段导出及现有初始化顺序。
+from acprof.metric_registry import (  # noqa: E402
+    CSV_FIELDS as CSV_FIELDS,
+    GPU_RUNTIME_STATE_FIELDS as GPU_RUNTIME_STATE_FIELDS,
+)
+
+STATIC_META_FIELDS = [
+    "profiling_mode",
+    "capability_report",
+    "platform",
+    "collection_tier",
+    "comparability_class",
+    "environment_class",
+    "platform_runtime",
+    "schema_version",
+    "model_name",
+    "model_revision",
+    "model_source",
+    "parameter_count",
+    "parameter_bytes",
+    "precision_dtype",
+    "parameter_dtype_counts",
+    "inference_precision_by_device",
+    "static_flops",
+    "static_macs",
+    "input_format",
+    "output_format",
+    "quantized",
+    "quantization_method",
+    "quantization_config",
+    "model_license",
+    "model_metadata_source",
+    "model_resolution",
+    "task_family",
+    "pipeline_tag",
+    "runtime_backend",
+    "image_tag",
+    "image_id",
+    "image_name",
+    "runtime_environment",
+    "runtime_validation",
+    "batch_size",
+    "latency_slo",
+    "input_scale_type",
+    "workload",
+    "input_scale_plan_sha256",
+    "run_command",
+    "model_download_url",
+    "gpu",
+    "gpu_device",
+    "gpu_mem_total_bytes",
+    "host_mem_total_bytes",
+    "host_swap_total_bytes",
+    "host_swap_used_bytes_at_start",
+    "host_swap_type",
+    "host_vm_swappiness",
+    "model_cache_bytes",
+    "model_storage_mode",
+    "model_artifact_bytes",
+    "runtime_image_bytes",
+    "total_deployment_bytes",
+    "docker_image_bytes",
+    "docker_storage_total_bytes",
+    "docker_storage_available_bytes_at_start",
+    "docker_storage_filesystem",
+    "docker_storage_device",
+    "docker_storage_type",
+    "environment",
+    "cgroup_version",
+    "cgroup_collection_mode",
+    "cpu_power_source",
+    "vcpu_power_method",
+    "rapl_topology",
+    "cpu_governor",
+    "cpu_boost",
+    "compute_profile_tools",
+    "torch_profiler_eager_flop_semantics",
+    "torch_profiler_eager_attention_implementation",
+    "torch_profiler_eager_repeat_cpu",
+    "torch_profiler_eager_repeat_gpu",
+    "ncu_flop_semantics",
+    "ncu_repeat",
+    "ncu_fma_flop_weight",
+    "ncu_metrics",
+    "torch_version",
+    "transformers_version",
+    "ncu_version",
+    "gpu_compute_capability",
+    "gpu_sm_count",
+    "compute_profiles_retained",
+    "compute_profile_provenance",
+    "execution_profile_schema_version",
+    "execution_profile_tools",
+    "massif_peak_semantics",
+    "massif_repeat",
+    "massif_version",
+    "massif_sampling_strategy",
+    "massif_reference_cpu_cores",
+    "massif_reference_mem_cap_gb",
+    "massif_reused_across_resource_cases",
+    "nsys_timeline_semantics",
+    "nsys_repeat",
+    "nsys_version",
+    "nsys_sampling_strategy",
+    "nsys_reference_cpu_cores",
+    "nsys_reference_mem_cap_gb",
+    "nsys_reused_across_resource_cases",
+    "execution_profiles_retained",
+    "execution_profile_provenance",
+]
+STATIC_META_SCHEMA_VERSION = 7
+
+# ─────────────────────────────────────────────
+# Docker 镜像命名
+# ─────────────────────────────────────────────
+DOCKER_IMAGE_PREFIX = "acprof"
+
+# ─────────────────────────────────────────────
+# HF 环境变量（中国网络可选镜像）
+# ─────────────────────────────────────────────
+PYPI_MIRROR_INDEX = "https://pypi.tuna.tsinghua.edu.cn/simple"
+CONTAINER_HF_HOME = "/models/hf"
+CONTAINER_MODEL_LOCAL_PATH = "/models/model-snapshot"

acprof/container/__init__.py

@@ -0,0 +1 @@
+"""Container-side serving and profiling utilities."""

acprof/container/compat/__init__.py

@@ -0,0 +1 @@
+"""Bounded workarounds for explicitly reviewed runtime capabilities."""

acprof/container/compat/transformers_dynamic.py

@@ -0,0 +1,54 @@
+"""Recursive-cache shim for runtimes lacking reviewed native loader capabilities.
+
+Transformers 4.57.6 copies only direct relative imports for local snapshots.
+Reuse its resolver and cache layout while preparing the recursive closure, as
+newer upstream loaders do (Transformers dynamic_module_utils, Apache-2.0).
+Retirement criteria: docs/models/runtime.md, dynamic-module lifecycle.
+"""
+from __future__ import annotations
+
+import filecmp
+import importlib
+import shutil
+import tempfile
+from contextlib import ExitStack
+from pathlib import Path
+
+
+def load_pipeline_class_compat(root: Path, module_file: str, class_name: str) -> type:
+    """Prepare missing transitive cache files, then let Transformers import the class."""
+    from transformers.dynamic_module_utils import (
+        HF_MODULES_CACHE,
+        get_cached_module_file,
+        get_class_in_module,
+        get_relative_import_files,
+    )
+
+    # Keep snapshot filenames: resolving Hub symlinks would lose relative names
+    # in the content-addressed blob directory. Discover missing files before import.
+    dependencies = sorted({Path(name) for name in get_relative_import_files(root / module_file)})
+    relative_files = [path.relative_to(root) for path in dependencies]
+    if any(".." in path.parts for path in relative_files):
+        raise ValueError("custom pipeline relative imports must stay within the model snapshot")
+    with ExitStack() as stack:
+        source_root = root
+        sources = {root / module_file, *dependencies}
+        if any(source.is_symlink() for source in sources):
+            # Some upstream loaders resolve each source symlink before finding
+            # sibling imports. Materialize only the code closure, preserving its
+            # snapshot names and bytes; never copy weights or mutate Hub blobs.
+            directory = stack.enter_context(tempfile.TemporaryDirectory(prefix="acprof-pipeline-"))
+            source_root = Path(directory) / root.name
+            for source in sources:
+                destination = source_root / source.relative_to(root)
+                destination.parent.mkdir(parents=True, exist_ok=True)
+                shutil.copyfile(source, destination)
+        cached_module = get_cached_module_file(str(source_root), module_file, local_files_only=True)
+    cached_root = (Path(HF_MODULES_CACHE) / cached_module).parents[len(Path(module_file).parts) - 1]
+    for source, relative in zip(dependencies, relative_files, strict=True):
+        target = cached_root / relative
+        if not target.is_file() or not filecmp.cmp(source, target, shallow=False):
+            target.parent.mkdir(parents=True, exist_ok=True)
+            shutil.copyfile(source, target)
+    importlib.invalidate_caches()
+    return get_class_in_module(class_name, cached_module)