tracebloc 4.0.0__tar.gz → 4.0.2__tar.gz

{tracebloc-4.0.0 → tracebloc-4.0.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tracebloc
-Version: 4.0.0
+Version: 4.0.2
 Summary: Build AI with anyone. On data that can't move. SDK for the tracebloc collaborative AI workspace.
 Author-email: tracebloc <lukas@tracebloc.io>
 License-Expression: MIT
@@ -55,7 +55,7 @@ Requires-Dist: numpy==2.1.3; python_version < "3.14" and extra == "training"
 Requires-Dist: pandas==2.3.0; python_version < "3.14" and extra == "training"
 Requires-Dist: psutil==7.0.0; extra == "training"
 Requires-Dist: pillow==13.0.0; extra == "training"
-Requires-Dist: safetensors>=0.4.3; extra == "training"
+Requires-Dist: safetensors>=0.8.0; extra == "training"
 Requires-Dist: torch==2.14.1; extra == "training"
 Requires-Dist: torchvision==0.29.1; extra == "training"
 Requires-Dist: torchmetrics==1.9.0; extra == "training"
@@ -78,6 +78,7 @@ Requires-Dist: pytest==9.2.0; extra == "dev"
 Requires-Dist: pytest-cov==7.1.0; extra == "dev"
 Requires-Dist: coverage==7.17.0; extra == "dev"
 Requires-Dist: jsonschema>=4.18; extra == "dev"
+Requires-Dist: pyyaml>=6.0; extra == "dev"
 Requires-Dist: setuptools>=77.0.0; extra == "dev"
 Requires-Dist: mypy>=1.5; extra == "dev"
 Requires-Dist: ruff==0.16.0; extra == "dev"

{tracebloc-4.0.0 → tracebloc-4.0.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "tracebloc"
-version = "4.0.0"
+version = "4.0.2"
 description = "Build AI with anyone. On data that can't move. SDK for the tracebloc collaborative AI workspace."
 license = "MIT"
 license-files = ["license.txt"]
@@ -123,7 +123,7 @@ training = [
     "pandas==2.3.0; python_version < '3.14'",
     "psutil==7.0.0",
     "pillow==13.0.0",
-    "safetensors>=0.4.3",
+    "safetensors>=0.8.0",
     "torch==2.14.1",
     "torchvision==0.29.1",
     "torchmetrics==1.9.0",
@@ -152,6 +152,11 @@ dev = [
     # tests/test_check_result_contract.py validates tracebloc.checks' payload
     # against the backend schema.
     "jsonschema>=4.18",
+    # Test files read the workflow and mutation-row YAML with `import
+    # yaml` at module level, so they fail collection without it. Only [all]
+    # pulled it in (through transformers), which left `pip install -e .[dev]`
+    # short of a suite that collects.
+    "pyyaml>=6.0",
     # tests/test_wheel_packages.py asks setuptools which packages the wheel
     # ships. A 3.12+ environment has no setuptools unless something asks
     # for it (build isolation keeps the build's copy out), so ask here, at
@@ -159,7 +164,7 @@ dev = [
     "setuptools>=77.0.0",
     "mypy>=1.5",
     "ruff==0.16.0",
-    # tests/test_convert_tokenizers.py converts a SentencePiece .model and
+    # tests/convert/test_convert_tokenizers.py converts a SentencePiece .model and
     # checks its ids against sentencepiece itself; transformers reads the
     # .model through protobuf. Those tests fail, not skip, when CI lacks them.
     "sentencepiece>=0.2.0",

{tracebloc-4.0.0 → tracebloc-4.0.2}/tracebloc/client.py

@@ -23,6 +23,13 @@ are **not** automatically retried, because a server-side 500 on an
 upload might still have persisted the model and a retry would create
 a duplicate. POST failures bubble up as :class:`APIError`; callers can
 implement idempotent retry at their layer if they want.
+
+**Overload (HTTP 429):** the backend sheds load with ``429`` plus
+``Retry-After``. A rejected request was not processed, so reads
+(``GET``/``HEAD``/``OPTIONS``) and login wait the stated time and retry
+(:func:`_send_honouring_429`: each wait capped at :data:`MAX_RETRY_WAIT_SECONDS`,
+all waits at :data:`MAX_TOTAL_WAIT_SECONDS`). Past the budget, and at once for
+every other call, the failure says the service is busy (:data:`BUSY_MESSAGE`).
 """
 
 from __future__ import annotations
@@ -31,7 +38,11 @@ from typing import Any, Dict, Mapping, Optional
 
 import json as _json
 import platform
+import time
+from datetime import datetime, timezone
+from email.utils import parsedate_to_datetime
 from importlib import metadata as _metadata
+from typing import Callable
 
 import requests
 from requests.adapters import HTTPAdapter
@@ -76,13 +87,83 @@ SDK_VERSION = _sdk_version()
 USER_AGENT = f"tracebloc-sdk/{SDK_VERSION} (python/{platform.python_version()})"
 
 
+#: Longest single ``Retry-After`` wait, and the most an SDK call waits in total.
+MAX_RETRY_WAIT_SECONDS = 60.0
+MAX_TOTAL_WAIT_SECONDS = 180.0
+
+#: Wait when a 429 names no usable ``Retry-After``; also the shortest wait taken,
+#: so a ``Retry-After: 0`` still spends the budget and the loop ends.
+_DEFAULT_RETRY_WAIT_SECONDS = 1.0
+
+#: Methods that are re-sent on a 429 (reads); login is re-sent explicitly.
+_RETRY_ON_429_METHODS = frozenset({"GET", "HEAD", "OPTIONS"})
+
+#: What a 429 reads as once the SDK has stopped waiting, or does not wait.
+BUSY_MESSAGE = (
+    "The tracebloc service is busy right now (HTTP 429). "
+    "Please retry in a few minutes."
+)
+
+# Indirection so tests can run the waits without sleeping.
+_sleep = time.sleep
+
+
+def _retry_after_seconds(response: requests.Response) -> float:
+    """Seconds a 429 asks us to wait: ``Retry-After`` as delay-seconds or an
+    HTTP date; a default when absent or unreadable. Never negative."""
+    raw = response.headers.get("Retry-After")
+    if isinstance(raw, str) and raw.strip():
+        raw = raw.strip()
+        try:
+            return max(0.0, float(raw))
+        except ValueError:
+            pass
+        try:
+            when = parsedate_to_datetime(raw)
+            if when.tzinfo is None:
+                when = when.replace(tzinfo=timezone.utc)
+            return max(0.0, (when - datetime.now(timezone.utc)).total_seconds())
+        except (TypeError, ValueError):
+            pass
+    return _DEFAULT_RETRY_WAIT_SECONDS
+
+
+def _send_honouring_429(
+    send: Callable[[], requests.Response], *, retry: bool
+) -> requests.Response:
+    """Run ``send``; on a 429 with ``retry`` wait ``Retry-After`` and run it
+    again until the budget (:data:`MAX_TOTAL_WAIT_SECONDS`) is spent.
+
+    Returns the last response, which may still be a 429 for the caller's
+    error handling. The wait that would overrun the budget is not taken.
+    """
+    waited = 0.0
+    while True:
+        response = send()
+        if response.status_code != 429 or not retry:
+            return response
+        wait = min(
+            max(_retry_after_seconds(response), _DEFAULT_RETRY_WAIT_SECONDS),
+            MAX_RETRY_WAIT_SECONDS,
+        )
+        if waited + wait > MAX_TOTAL_WAIT_SECONDS:
+            return response
+        response.close()
+        _sleep(wait)
+        waited += wait
+
+
 def _error_message(response: requests.Response) -> str:
     """The user-facing message for a non-2xx response.
 
     A 426 is the backend refusing this SDK as older than the oldest it
     supports (``{"error": "upgrade_required", "min_version": ...}``); say
     that, with the fix, rather than the bare ``upgrade_required`` code.
+    A 429 is the backend shedding load: :data:`BUSY_MESSAGE`, never the
+    server's own wording.
     """
+    if response.status_code == 429:
+        return BUSY_MESSAGE
     if response.status_code == 426:
         try:
             body = response.json()
@@ -101,7 +182,21 @@ def _error_message(response: requests.Response) -> str:
     return _extract_error_message(response)
 
 
-_DEFAULT_RETRY = Retry(
+class _Retry(Retry):
+    """urllib3 ``Retry`` that leaves a 429 alone.
+
+    A 429's ``Retry-After`` is honoured, capped, by ``_send_honouring_429``;
+    urllib3's own sleep on it is uncapped and would stack on that budget. The
+    header still governs a 503's wait.
+    """
+
+    def is_retry(self, method, status_code, has_retry_after=False):
+        if status_code == 429:
+            return False
+        return super().is_retry(method, status_code, has_retry_after)
+
+
+_DEFAULT_RETRY = _Retry(
     total=3,
     backoff_factor=0.5,
     status_forcelist=(500, 502, 503, 504),
@@ -252,17 +347,21 @@ class TraceblocClient:
         # Only passed when asked for, so every existing call is issued exactly
         # as before.
         extra = {"stream": True} if stream else {}
+        verb = method.upper()
         try:
-            response = self._session.request(
-                method=method.upper(),
-                url=url,
-                data=data,
-                json=json,
-                files=files,
-                params=params,
-                headers=self._headers(headers),
-                timeout=request_timeout,
-                **extra,
+            response = _send_honouring_429(
+                lambda: self._session.request(
+                    method=verb,
+                    url=url,
+                    data=data,
+                    json=json,
+                    files=files,
+                    params=params,
+                    headers=self._headers(headers),
+                    timeout=request_timeout,
+                    **extra,
+                ),
+                retry=verb in _RETRY_ON_429_METHODS,
             )
         except requests.RequestException as error:
             raise APIError(
@@ -312,13 +411,16 @@ class TraceblocClient:
         """
         url = self._url("api-token-auth/")
         try:
-            response = self._session.post(
-                url,
-                data={"username": username, "password": password},
-                # Not self._headers(): login must never carry a previous
-                # session's Authorization, only who is calling.
-                headers={"User-Agent": USER_AGENT},
-                timeout=self._config.timeout_seconds,
+            response = _send_honouring_429(
+                lambda: self._session.post(
+                    url,
+                    data={"username": username, "password": password},
+                    # Not self._headers(): login must never carry a previous
+                    # session's Authorization, only who is calling.
+                    headers={"User-Agent": USER_AGENT},
+                    timeout=self._config.timeout_seconds,
+                ),
+                retry=True,
             )
         except requests.RequestException as error:
             raise APIError(

{tracebloc-4.0.0 → tracebloc-4.0.2}/tracebloc/contracts/training_requirements.source.json

@@ -2,8 +2,8 @@
   "//": "Pin for tracebloc/contracts/training_requirements.txt, the training images' exact library versions. The txt is tracebloc-engine's use_cases/requirements.txt copied byte-for-byte at `sha`; `blobSha` is its git blob sha there, which tests/test_convert_training_versions.py recomputes. tracebloc.convert.training_versions reads it to refuse a conversion made under other library versions. Re-pin with scripts/sync-training-requirements.sh (put the new commit in `sha` and run it; org-config's contract-pin-bump.yml does the same), then run that test.",
   "repo": "tracebloc/tracebloc-engine",
   "ref": "develop",
-  "sha": "e5cc153e6c6ca2bd3eb3d5e42523b7d345e52cde",
+  "sha": "4c5c97741ad127eac2af32043c69a6e28b10e7f7",
   "path": "use_cases/requirements.txt",
-  "blobSha": "8e5744a92f2722c2fea81150b263cbf8aa3cf71e",
-  "fetched": "2026-10-29"
+  "blobSha": "3da50058c9374cf5fa503103d54d44d31e3c7e31",
+  "fetched": "2026-11-10"
 }

{tracebloc-4.0.0 → tracebloc-4.0.2}/tracebloc/contracts/training_requirements.txt

@@ -15,23 +15,22 @@ pandas==2.3.0
 mysql-connector-python==9.3.0
 numpy==2.1.3
 pillow==13.0.0
+# core/data/tiff_image.py reads a TIFF with every channel and its bit depth, which
+# Pillow and OpenCV do not; imagecodecs decodes its LZW / Deflate compression.
+tifffile==2026.3.3
+imagecodecs==2026.3.6
 torch==2.14.1
 torchvision==0.29.1
 torchmetrics==1.9.0
 albumentations==2.0.5
-tracebloc-telemetry==0.1.0    # RFC-BACKEND-1872 D2 — the shared contract emitter
-# Kept: a training pod on a customer cluster has no other route off it until
-# the edge collector ships (backend#1905/#1906, RFC-1872 D10).
-opencensus==0.11.4
-opencensus-context==0.1.3
-opencensus-ext-azure==1.1.15
+tracebloc-telemetry==0.2.0    # RFC-BACKEND-1872 D2 — the shared contract emitter
 timm==1.0.26
 transformers==5.16.1
 einops==0.8.2
 # SEC-03 (RFC 0004): weight I/O reads/writes SafeTensors via core/contracts/safetensors_io.py.
-# Floor (not an exact pin) kept identical across the three mirrors (SDK / engine / averaging-service);
-# left unpinned so transformers' resolver still governs the exact build.
-safetensors>=0.4.3
+# A floor, not a pin: before 0.8.0 save_file copies the whole payload to bytes, a copy the SDK's
+# memory estimate does not count. core/tests/scripts/test_safetensors_floor.py holds every family to it.
+safetensors>=0.8.0
 scikit-learn==1.8.0
 # CPU build. `xgboost` declares `nvidia-nccl-cu12` on Linux (342 MB compressed,
 # 476 MB on disk) that no CPU image can use. Same version, same `import xgboost`;
@@ -49,7 +48,7 @@ pyarrow==23.0.1
 # plotly,sensitivity,shap]`, ~300 MB on disk of dashboard/notebook extras. The
 # zoo's two EBM templates import `interpret.glassbox` only, which interpret-core
 # alone provides (hard requirements: numpy, pandas, scikit-learn, joblib).
-# `core/tests/utils/test_interpret_arm64_pin.py` holds the version floor
+# `core/tests/scripts/test_interpret_arm64_pin.py` holds the version floor
 # (backend#3456).
 interpret-core==0.6.10
 pytorch-forecasting==1.4.0

{tracebloc-4.0.0 → tracebloc-4.0.2}/tracebloc/convert/owncode/arguments.py

@@ -42,6 +42,13 @@ CONFIG_BYTES = 1_000_000
 
 _LITERAL = (int, float, str, bool, type(None))
 
+#: What ``ast.literal_eval`` raises on a node that is not a literal: a
+#: ``ValueError`` for a name or a call, a ``TypeError`` for a dict or set
+#: literal with an unhashable key (``{[1]: 2}``), which parses but cannot be
+#: built, and the rest for a literal too deep or too big to build. The same
+#: set ``source_tree`` and ``shake`` catch around their own ``literal_eval``.
+_NOT_LITERAL = (ValueError, TypeError, SyntaxError, MemoryError, RecursionError)
+
 
 @dataclass(frozen=True)
 class Parameter:
@@ -188,7 +195,7 @@ def _put(
 ) -> None:
     try:
         value = ast.literal_eval(node)
-    except ValueError:
+    except _NOT_LITERAL:
         return
     convert = _CONVERTERS.get(kind.id) if isinstance(kind, ast.Name) else None
     if convert is not None and isinstance(value, str):
@@ -205,7 +212,7 @@ def default_value(parameter: Parameter) -> tuple[bool, Any]:
         return False, None
     try:
         return True, ast.literal_eval(parameter.default)
-    except ValueError:
+    except _NOT_LITERAL:
         return False, None
 
 
@@ -224,7 +231,7 @@ def image_size(tree: SourceTree, sources: Sources) -> SourcedValue | None:
             if parts and parts[-1] in RESIZE_CALLS and node.args:
                 try:
                     side = _side(ast.literal_eval(node.args[0]))
-                except ValueError:
+                except _NOT_LITERAL:
                     continue
                 if side is not None:
                     detail = f"{module.relative}:{node.lineno}"

{tracebloc-4.0.0 → tracebloc-4.0.2}/tracebloc/linking/model_dataset.py

@@ -10,6 +10,7 @@ from tracebloc.training.subdataset import class_set_error, dataset_error, reques
 from tracebloc.training.image_channels import image_channel_check
 from tracebloc.utils.data_probes import get_images_count, validate_kwargs
 from tracebloc.utils.errors import format_exception
+from tracebloc.utils.staging import model_staging_dir
 from tracebloc.utils.model_params import get_model_params_count
 from tracebloc.utils.constants import (
     IMAGE_DATA_TYPES,
@@ -262,10 +263,7 @@ class LinkModelDataset(DataPlanMethods, TrainingPlanMethods, AugmentationMethods
         self._utilisation_category = kwargs["utilisation_category"]
         self.model_path = kwargs["model_path"]
         self._model_name = kwargs["model_name"]
-        self.tmp_path = os.path.join(
-            os.path.dirname(self.model_path),
-            f"tmpmodel_{self._model_name[:64]}",
-        )
+        self.tmp_path = model_staging_dir(self.model_path, self._model_name)
         if not os.path.isdir(self.tmp_path):
             os.mkdir(self.tmp_path)
         self._environment = kwargs["environment"]

tracebloc-4.0.2/tracebloc/utils/staging.py

@@ -0,0 +1,17 @@
+"""Where a model's staging directory (``tmpmodel_<name>``) lives."""
+
+import os
+
+#: Longest model name carried into the directory name.
+MAX_STAGING_NAME_LENGTH = 64
+
+
+def model_staging_dir(model_path: str, model_name: str) -> str:
+    """Path of the ``tmpmodel_<name>`` directory beside ``model_path``.
+
+    Validation and linking share it, so both must derive it here.
+    """
+    return os.path.join(
+        os.path.dirname(model_path),
+        f"tmpmodel_{model_name[:MAX_STAGING_NAME_LENGTH]}",
+    )

{tracebloc-4.0.0 → tracebloc-4.0.2}/tracebloc/validation/checker.py

@@ -5,6 +5,7 @@ from inspect import getmembers, isclass
 from tracebloc.compat.frameworks import missing_extra_hint
 from tracebloc.utils.errors import format_exception
 from tracebloc.utils.model_loading import load_model
+from tracebloc.utils.staging import MAX_STAGING_NAME_LENGTH, model_staging_dir
 from tracebloc.validation.header_rules import missing_framework_message
 from tracebloc.validation.rewriter import ModelFileRewriter
 
@@ -27,7 +28,7 @@ class ModelChecker:
     working against a single copy of the state owned by the rewriter.
     """
 
-    MAX_MODEL_NAME_LENGTH = 64
+    MAX_MODEL_NAME_LENGTH = MAX_STAGING_NAME_LENGTH
     model = None
     notallowed = ["__MACOSX", "__pycache__"]
 
@@ -96,11 +97,8 @@ class ModelChecker:
 
     def load_model_file(self):
         # ``os.path``, not ``rsplit("/")``: a Windows path has no "/" (#898).
-        model_dir, file = os.path.split(self.model_path)
-        self.tmp_file_path = os.path.join(
-            model_dir,
-            f"tmpmodel_{self.model_name[: self.MAX_MODEL_NAME_LENGTH]}",
-        )
+        file = os.path.split(self.model_path)[1]
+        self.tmp_file_path = model_staging_dir(self.model_path, self.model_name)
         if not os.path.isdir(self.tmp_file_path):
             os.mkdir(self.tmp_file_path)
         # check if file contains the MyModel function

{tracebloc-4.0.0 → tracebloc-4.0.2}/tracebloc.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tracebloc
-Version: 4.0.0
+Version: 4.0.2
 Summary: Build AI with anyone. On data that can't move. SDK for the tracebloc collaborative AI workspace.
 Author-email: tracebloc <lukas@tracebloc.io>
 License-Expression: MIT
@@ -55,7 +55,7 @@ Requires-Dist: numpy==2.1.3; python_version < "3.14" and extra == "training"
 Requires-Dist: pandas==2.3.0; python_version < "3.14" and extra == "training"
 Requires-Dist: psutil==7.0.0; extra == "training"
 Requires-Dist: pillow==13.0.0; extra == "training"
-Requires-Dist: safetensors>=0.4.3; extra == "training"
+Requires-Dist: safetensors>=0.8.0; extra == "training"
 Requires-Dist: torch==2.14.1; extra == "training"
 Requires-Dist: torchvision==0.29.1; extra == "training"
 Requires-Dist: torchmetrics==1.9.0; extra == "training"
@@ -78,6 +78,7 @@ Requires-Dist: pytest==9.2.0; extra == "dev"
 Requires-Dist: pytest-cov==7.1.0; extra == "dev"
 Requires-Dist: coverage==7.17.0; extra == "dev"
 Requires-Dist: jsonschema>=4.18; extra == "dev"
+Requires-Dist: pyyaml>=6.0; extra == "dev"
 Requires-Dist: setuptools>=77.0.0; extra == "dev"
 Requires-Dist: mypy>=1.5; extra == "dev"
 Requires-Dist: ruff==0.16.0; extra == "dev"

{tracebloc-4.0.0 → tracebloc-4.0.2}/tracebloc.egg-info/SOURCES.txt

@@ -190,6 +190,7 @@ tracebloc/utils/model_params.py
 tracebloc/utils/registry.py
 tracebloc/utils/resources.py
 tracebloc/utils/safetensors_io.py
+tracebloc/utils/staging.py
 tracebloc/utils/tabular_classification_utils.py
 tracebloc/utils/tabular_custom_loss_utils.py
 tracebloc/utils/tabular_regression_utils.py

{tracebloc-4.0.0 → tracebloc-4.0.2}/tracebloc.egg-info/requires.txt

@@ -17,6 +17,7 @@ pytest==9.2.0
 pytest-cov==7.1.0
 coverage==7.17.0
 jsonschema>=4.18
+pyyaml>=6.0
 setuptools>=77.0.0
 mypy>=1.5
 ruff==0.16.0
@@ -64,7 +65,7 @@ xgboost-cpu>=3.2.0
 [training]
 psutil==7.0.0
 pillow==13.0.0
-safetensors>=0.4.3
+safetensors>=0.8.0
 torch==2.14.1
 torchvision==0.29.1
 torchmetrics==1.9.0