diffract-core 0.2.1__tar.gz → 0.3.0__tar.gz

{diffract_core-0.2.1 → diffract_core-0.3.0}/.gitignore

@@ -75,6 +75,7 @@ instance/
 # Sphinx documentation
 docs/_build/
 docs/reference/_generated/
+docs/reference/metrics/_generated_table.md
 
 # PyBuilder
 target/

{diffract_core-0.2.1 → diffract_core-0.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffract-core
-Version: 0.2.1
+Version: 0.3.0
 Summary: Spectral analysis of neural network weights and their evolution over training
 Project-URL: Homepage, https://github.com/Risk-AI-Research/diffract
 Project-URL: Documentation, https://risk-ai-research.github.io/diffract/
@@ -17,10 +17,11 @@ Keywords: deep-learning,llm,neural-networks,random-matrix-theory,spectral-analys
 Classifier: Development Status :: 3 - Alpha
 Classifier: Intended Audience :: Science/Research
 Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
 Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Classifier: Topic :: Software Development :: Libraries :: Python Modules
 Classifier: Typing :: Typed
-Requires-Python: <3.13,>=3.12
+Requires-Python: >=3.12
 Requires-Dist: dependency-injector>=4.44.0
 Requires-Dist: h5py>=3.9.0
 Requires-Dist: numpy<3,>=1.26
@@ -52,6 +53,7 @@ Provides-Extra: dev
 Requires-Dist: bandit[toml]>=1.7.5; extra == 'dev'
 Requires-Dist: fsspec>=2024.2.0; extra == 'dev'
 Requires-Dist: hydra-core>=1.3.0; extra == 'dev'
+Requires-Dist: hypothesis>=6.0.0; extra == 'dev'
 Requires-Dist: interrogate>=1.7.0; extra == 'dev'
 Requires-Dist: kaleido>=0.2.1; extra == 'dev'
 Requires-Dist: mypy>=1.10.1; extra == 'dev'
@@ -80,6 +82,7 @@ Requires-Dist: sphinx-autodoc-typehints>=1.25.0; extra == 'docs'
 Requires-Dist: sphinx-copybutton>=0.5.0; extra == 'docs'
 Requires-Dist: sphinx-design>=0.6.0; extra == 'docs'
 Requires-Dist: sphinx>=7.0.0; extra == 'docs'
+Requires-Dist: sphinxcontrib-bibtex>=2.6.0; extra == 'docs'
 Requires-Dist: sphinxext-opengraph>=0.9.0; extra == 'docs'
 Provides-Extra: flax
 Requires-Dist: flax>=0.10.0; extra == 'flax'
@@ -127,7 +130,8 @@ Description-Content-Type: text/markdown
 [![Docs](https://img.shields.io/badge/docs-github.io-blue)](https://risk-ai-research.github.io/diffract/)
 [![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
 
-Official library of ["Diffract: Spectral View of LLM Domain Adaptation"](https://openreview.net/forum?id=XBUHoiAGDE)
+Official library of
+["Diffract: Spectral View of LLM Domain Adaptation"](https://openreview.net/forum?id=XBUHoiAGDE)
 (**ICML 2026 Oral**).
 
 Diffract is a Python package for analyzing deep neural network weights and tracking
@@ -144,7 +148,7 @@ dictionaries of NumPy weight matrices.
 
 ## 🚀 Quick Start
 
-Diffract requires Python 3.12. The core package installs without any deep
+Diffract requires Python 3.12 or newer. The core package installs without any deep
 learning framework; heavy dependencies are opt-in extras:
 
 ```bash
@@ -155,10 +159,9 @@ pip install "diffract-core[taichi]"      # + accelerated heavy-tailed fits and p
 pip install "diffract-core[all]"         # torch + viz + taichi + pandas/polars exports
 ```
 
-Further extras: `frameworks` (TensorFlow, Flax, ONNX), `pandas` / `polars`
-(DataFrame exports), `zarr` (cloud storage), `redis` (shared cache),
-`notebooks` (tooling for the example notebooks). The quotes around
-`"diffract-core[...]"` matter in zsh.
+Further extras: `frameworks` (TensorFlow, Flax, ONNX), `pandas` / `polars` (DataFrame
+exports), `zarr` (cloud storage), `redis` (shared cache), `notebooks` (tooling for the
+example notebooks). The quotes around `"diffract-core[...]"` matter in zsh.
 
 ### Development Install
 
@@ -166,7 +169,7 @@ Further extras: `frameworks` (TensorFlow, Flax, ONNX), `pandas` / `polars`
 # Install uv if you haven't already
 curl -LsSf https://astral.sh/uv/install.sh | sh
 
-# Clone and install (uv provisions Python 3.12 automatically)
+# Clone and install (uv provisions the right Python automatically)
 git clone https://github.com/Risk-AI-Research/diffract.git
 cd diffract
 uv sync --extra dev --extra torch
@@ -219,7 +222,7 @@ internal structure:
 
 <br>
 
-## 🔑 Key Features
+## Key Features
 
 - **Session-based API**: Simple `models.add`, `compute.apply`, and
   `results.export_metrics` workflow.
@@ -232,6 +235,7 @@ internal structure:
   HDF5, SQLite, Zarr, and hybrid backends.
 
 - **Kernel Apply Levels**: Kernels can work at multiple levels:
+
   - **PARAMETER** - Operate on individual weight matrices.
   - **IN_MODEL** - Aggregate within a single model.
   - **CROSS_MODEL** - Compare or aggregate across models.
@@ -257,7 +261,9 @@ session = Session()
 
 with session:
     session.models.add(torch_model)  # torch.nn.Module
-    session.models.add(torch_state_dict, model_id="checkpoint")  # Dict[str, torch.Tensor]
+    session.models.add(
+        torch_state_dict, model_id="checkpoint"
+    )  # Dict[str, torch.Tensor]
     session.models.add(numpy_weights, model_id="raw-weights")  # Dict[str, np.ndarray]
     session.models.add(onnx_model, model_id="onnx-model")  # onnx.ModelProto
     session.models.add(flax_model, model_id="flax-model")  # flax.linen.Module
@@ -273,6 +279,10 @@ session.compute.apply("frob_norm", "stable_rank")
 session.compute.apply("pl_ks")  # has many dependencies—all resolved automatically
 ```
 
+Every built-in metric—its formula, apply level, required inputs, and configuration—is
+catalogued in the
+[metrics reference](https://risk-ai-research.github.io/diffract/reference/metrics/).
+
 ### Filtering Parameters
 
 Filter computations by model, parameter type, or name:
@@ -309,9 +319,7 @@ Export results in various formats (`pandas`, `polars`, `dict`, `json`, or `list`
 
 ```python
 scalars_df = session.results.export_metrics("stable_rank", export_format="pandas")
-aggregates_df = session.results.export_aggregates(
-    "stable_rank", export_format="pandas"
-)
+aggregates_df = session.results.export_aggregates("stable_rank", export_format="pandas")
 
 # Other formats work the same way
 results = session.results.export_metrics("stable_rank", export_format="polars")
@@ -351,7 +359,7 @@ List and configure kernels at runtime:
 ```python
 session.compute.list_available_kernels(verbose=True)
 session.compute.list_available_metrics(verbose=True)
-session.compute.configure_kernel("hard_rank", threshold=1e-6)
+session.compute.configure_kernel("hard_rank", rtol=1e-6)
 ```
 
 ### Session Management
@@ -402,8 +410,11 @@ You can also override the registered name and output fields:
 
 ```python
 with session:
+
     @session.compute.kernel(name="scaled_metric", produce_fields=["scaled_result"])
-    def custom_analysis(frob_norm: float, stable_rank: float, *, weight: float = 0.5) -> float:
+    def custom_analysis(
+        frob_norm: float, stable_rank: float, *, weight: float = 0.5
+    ) -> float:
         """Custom analysis combining multiple metrics."""
         return weight * frob_norm + (1 - weight) * stable_rank
 ```
@@ -448,9 +459,9 @@ Diffract offers built-in **profiles** for common setups:
 from diffract import Session
 
 # Use a profile (recommended for most users)
-session = Session(profile="ram")      # fast, temporary
-session = Session(profile="local")    # persistent, simple
-session = Session(profile="hybrid")   # persistent, optimized for large arrays
+session = Session(profile="ram")  # fast, temporary
+session = Session(profile="local")  # persistent, simple
+session = Session(profile="hybrid")  # persistent, optimized for large arrays
 
 # Or use a custom config file for full control
 session = Session(config_path="my_config.ini")
@@ -496,12 +507,14 @@ max_workers = 4
 
 ## 📚 Documentation
 
-The documentation site is sourced from `docs/` and built with Sphinx + MyST. Install the
-tooling via `uv sync --extra docs` and run `make docs` to render the HTML locally.
+The documentation is hosted at
+[risk-ai-research.github.io/diffract](https://risk-ai-research.github.io/diffract/). It
+is sourced from `docs/` and built with Sphinx + MyST; `uv sync --extra docs` and
+`make docs` render the HTML locally.
 
 <br>
 
-## 📝 Citation
+## Citation
 
 If you use Diffract or build on the paper, please cite:
 
@@ -517,14 +530,14 @@ If you use Diffract or build on the paper, please cite:
 
 <br>
 
-## ❤️ Contributions
+## Contributions
 
-Contributions are welcome! Fork the repo, create a feature branch, and submit a PR. Use
-`make lint` and `make test` to validate your changes.
+Contributions are welcome — see [CONTRIBUTING.md](CONTRIBUTING.md) for the development
+setup, the checks CI runs, and the design principles.
 
 <br>
 
-## 📄 License
+## License
 
 Licensed under the Apache License 2.0 — see [LICENSE](LICENSE). Copyright 2026 Risk AI
 Research.

{diffract_core-0.2.1 → diffract_core-0.3.0}/README.md

@@ -6,7 +6,8 @@
 [![Docs](https://img.shields.io/badge/docs-github.io-blue)](https://risk-ai-research.github.io/diffract/)
 [![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
 
-Official library of ["Diffract: Spectral View of LLM Domain Adaptation"](https://openreview.net/forum?id=XBUHoiAGDE)
+Official library of
+["Diffract: Spectral View of LLM Domain Adaptation"](https://openreview.net/forum?id=XBUHoiAGDE)
 (**ICML 2026 Oral**).
 
 Diffract is a Python package for analyzing deep neural network weights and tracking
@@ -23,7 +24,7 @@ dictionaries of NumPy weight matrices.
 
 ## 🚀 Quick Start
 
-Diffract requires Python 3.12. The core package installs without any deep
+Diffract requires Python 3.12 or newer. The core package installs without any deep
 learning framework; heavy dependencies are opt-in extras:
 
 ```bash
@@ -34,10 +35,9 @@ pip install "diffract-core[taichi]"      # + accelerated heavy-tailed fits and p
 pip install "diffract-core[all]"         # torch + viz + taichi + pandas/polars exports
 ```
 
-Further extras: `frameworks` (TensorFlow, Flax, ONNX), `pandas` / `polars`
-(DataFrame exports), `zarr` (cloud storage), `redis` (shared cache),
-`notebooks` (tooling for the example notebooks). The quotes around
-`"diffract-core[...]"` matter in zsh.
+Further extras: `frameworks` (TensorFlow, Flax, ONNX), `pandas` / `polars` (DataFrame
+exports), `zarr` (cloud storage), `redis` (shared cache), `notebooks` (tooling for the
+example notebooks). The quotes around `"diffract-core[...]"` matter in zsh.
 
 ### Development Install
 
@@ -45,7 +45,7 @@ Further extras: `frameworks` (TensorFlow, Flax, ONNX), `pandas` / `polars`
 # Install uv if you haven't already
 curl -LsSf https://astral.sh/uv/install.sh | sh
 
-# Clone and install (uv provisions Python 3.12 automatically)
+# Clone and install (uv provisions the right Python automatically)
 git clone https://github.com/Risk-AI-Research/diffract.git
 cd diffract
 uv sync --extra dev --extra torch
@@ -98,7 +98,7 @@ internal structure:
 
 <br>
 
-## 🔑 Key Features
+## Key Features
 
 - **Session-based API**: Simple `models.add`, `compute.apply`, and
   `results.export_metrics` workflow.
@@ -111,6 +111,7 @@ internal structure:
   HDF5, SQLite, Zarr, and hybrid backends.
 
 - **Kernel Apply Levels**: Kernels can work at multiple levels:
+
   - **PARAMETER** - Operate on individual weight matrices.
   - **IN_MODEL** - Aggregate within a single model.
   - **CROSS_MODEL** - Compare or aggregate across models.
@@ -136,7 +137,9 @@ session = Session()
 
 with session:
     session.models.add(torch_model)  # torch.nn.Module
-    session.models.add(torch_state_dict, model_id="checkpoint")  # Dict[str, torch.Tensor]
+    session.models.add(
+        torch_state_dict, model_id="checkpoint"
+    )  # Dict[str, torch.Tensor]
     session.models.add(numpy_weights, model_id="raw-weights")  # Dict[str, np.ndarray]
     session.models.add(onnx_model, model_id="onnx-model")  # onnx.ModelProto
     session.models.add(flax_model, model_id="flax-model")  # flax.linen.Module
@@ -152,6 +155,10 @@ session.compute.apply("frob_norm", "stable_rank")
 session.compute.apply("pl_ks")  # has many dependencies—all resolved automatically
 ```
 
+Every built-in metric—its formula, apply level, required inputs, and configuration—is
+catalogued in the
+[metrics reference](https://risk-ai-research.github.io/diffract/reference/metrics/).
+
 ### Filtering Parameters
 
 Filter computations by model, parameter type, or name:
@@ -188,9 +195,7 @@ Export results in various formats (`pandas`, `polars`, `dict`, `json`, or `list`
 
 ```python
 scalars_df = session.results.export_metrics("stable_rank", export_format="pandas")
-aggregates_df = session.results.export_aggregates(
-    "stable_rank", export_format="pandas"
-)
+aggregates_df = session.results.export_aggregates("stable_rank", export_format="pandas")
 
 # Other formats work the same way
 results = session.results.export_metrics("stable_rank", export_format="polars")
@@ -230,7 +235,7 @@ List and configure kernels at runtime:
 ```python
 session.compute.list_available_kernels(verbose=True)
 session.compute.list_available_metrics(verbose=True)
-session.compute.configure_kernel("hard_rank", threshold=1e-6)
+session.compute.configure_kernel("hard_rank", rtol=1e-6)
 ```
 
 ### Session Management
@@ -281,8 +286,11 @@ You can also override the registered name and output fields:
 
 ```python
 with session:
+
     @session.compute.kernel(name="scaled_metric", produce_fields=["scaled_result"])
-    def custom_analysis(frob_norm: float, stable_rank: float, *, weight: float = 0.5) -> float:
+    def custom_analysis(
+        frob_norm: float, stable_rank: float, *, weight: float = 0.5
+    ) -> float:
         """Custom analysis combining multiple metrics."""
         return weight * frob_norm + (1 - weight) * stable_rank
 ```
@@ -327,9 +335,9 @@ Diffract offers built-in **profiles** for common setups:
 from diffract import Session
 
 # Use a profile (recommended for most users)
-session = Session(profile="ram")      # fast, temporary
-session = Session(profile="local")    # persistent, simple
-session = Session(profile="hybrid")   # persistent, optimized for large arrays
+session = Session(profile="ram")  # fast, temporary
+session = Session(profile="local")  # persistent, simple
+session = Session(profile="hybrid")  # persistent, optimized for large arrays
 
 # Or use a custom config file for full control
 session = Session(config_path="my_config.ini")
@@ -375,12 +383,14 @@ max_workers = 4
 
 ## 📚 Documentation
 
-The documentation site is sourced from `docs/` and built with Sphinx + MyST. Install the
-tooling via `uv sync --extra docs` and run `make docs` to render the HTML locally.
+The documentation is hosted at
+[risk-ai-research.github.io/diffract](https://risk-ai-research.github.io/diffract/). It
+is sourced from `docs/` and built with Sphinx + MyST; `uv sync --extra docs` and
+`make docs` render the HTML locally.
 
 <br>
 
-## 📝 Citation
+## Citation
 
 If you use Diffract or build on the paper, please cite:
 
@@ -396,14 +406,14 @@ If you use Diffract or build on the paper, please cite:
 
 <br>
 
-## ❤️ Contributions
+## Contributions
 
-Contributions are welcome! Fork the repo, create a feature branch, and submit a PR. Use
-`make lint` and `make test` to validate your changes.
+Contributions are welcome — see [CONTRIBUTING.md](CONTRIBUTING.md) for the development
+setup, the checks CI runs, and the design principles.
 
 <br>
 
-## 📄 License
+## License
 
 Licensed under the Apache License 2.0 — see [LICENSE](LICENSE). Copyright 2026 Risk AI
 Research.

{diffract_core-0.2.1 → diffract_core-0.3.0}/pyproject.toml

@@ -11,12 +11,13 @@ maintainers = [
 ]
 license = "Apache-2.0"
 license-files = ["LICENSE", "NOTICE"]
-requires-python = ">=3.12,<3.13"
+requires-python = ">=3.12"
 keywords = ["spectral-analysis", "random-matrix-theory", "deep-learning", "llm", "neural-networks"]
 classifiers = [
     "Development Status :: 3 - Alpha",
     "Intended Audience :: Science/Research",
     "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
     "Topic :: Software Development :: Libraries :: Python Modules",
     "Typing :: Typed",
@@ -104,9 +105,11 @@ docs = [
     "sphinx-design >=0.6.0",
     "furo >=2024.8.0",
     "sphinxext-opengraph >=0.9.0",
+    "sphinxcontrib-bibtex >=2.6.0",
 ]
 dev = [
     "bandit[toml] >=1.7.5",
+    "hypothesis >=6.0.0",
     "interrogate >=1.7.0",
     "mypy >=1.10.1",
     "pandas >=2.1.0",
@@ -169,8 +172,10 @@ filterwarnings = [
     "ignore::DeprecationWarning",
     "ignore::UserWarning",
     "ignore::PendingDeprecationWarning",
-    # taichi 1.7.x ships a docstring with an invalid escape sequence
-    "ignore::SyntaxWarning:taichi.tools.image",
+    # taichi 1.7.x ships a docstring with an invalid escape sequence; the
+    # compile-time warning is attributed to the importing frame rather than
+    # the module that carries it, so it is matched by message.
+    "ignore:invalid escape sequence:SyntaxWarning",
 ]
 log_cli = true
 log_cli_level = "INFO"
@@ -200,7 +205,6 @@ omit = [
     "*/.venv/*",
     "*/venv/*",
     "*/.tox/*",
-    "src/diffract/core/compute/kernels/*",
 ]
 branch = true
 parallel = true
@@ -224,7 +228,7 @@ exclude_lines = [
 ]
 show_missing = true
 precision = 2
-fail_under = 65
+fail_under = 67
 skip_covered = false
 skip_empty = false
 

{diffract_core-0.2.1 → diffract_core-0.3.0}/ruff.toml

@@ -187,8 +187,16 @@ dummy-variable-rgx = "^(_+|(_+[a-zA-Z0-9_]*[a-zA-Z0-9]+?))$"
     "PLR0913", # Too many arguments
     "PLR0915", # Too many statements
     "ARG001", # Unused function argument
+    "ARG002", # Unused method argument (fixtures, mock signatures)
+    "ARG005", # Unused lambda argument (stub implementations)
     "FBT001", # Boolean positional argument
     "PT011",  # pytest.raises() is too broad
+    "PLR2004", # Literal expectations are the substance of an assertion
+    "SLF001", # Tests exercise private members deliberately
+    "INP001", # Test directories are namespace packages by design
+    "PLC0415", # Imports inside tests control DI wiring order
+    "S108",   # /tmp only appears as literal data in config-parsing asserts
+    "S110",   # try/except/pass in teardown paths
 ]
 "notebooks/*" = [
     "D",      # Docstring requirements

{diffract_core-0.2.1 → diffract_core-0.3.0}/src/diffract/__init__.py

@@ -22,7 +22,7 @@ from __future__ import annotations
 import importlib
 from typing import TYPE_CHECKING, Any
 
-__version__ = "0.2.1"
+__version__ = "0.3.0"
 
 from .containers import list_profiles
 from .core.data.nn.params.schema import ParameterType

{diffract_core-0.2.1 → diffract_core-0.3.0}/src/diffract/configs/hybrid.ini

@@ -71,4 +71,5 @@ max_workers = 8
 skip_not_implemented_types = true
 
 [export]
-default_export_format = "pandas"
+# Formats beyond "dict"/"json"/"list" need their optional extra installed.
+default_export_format = "dict"

{diffract_core-0.2.1 → diffract_core-0.3.0}/src/diffract/configs/sqlite.ini

@@ -59,4 +59,5 @@ max_workers = 8
 skip_not_implemented_types = true
 
 [export]
-default_export_format = "pandas"
+# Formats beyond "dict"/"json"/"list" need their optional extra installed.
+default_export_format = "dict"

{diffract_core-0.2.1 → diffract_core-0.3.0}/src/diffract/containers.py

@@ -25,6 +25,7 @@ from .core.compute.containers import (
     ComputeSingletonContainerWiringConfig,
 )
 from .core.data.metadata.containers import MetadataContainer
+from .core.data.metadata.sqlite_index import IN_MEMORY_DATABASE
 from .core.data.nn.containers import ModelParametersContainer
 from .core.export.containers import ExportContainer
 from .core.parallel import ParallelSingletonContainer
@@ -42,6 +43,10 @@ PROFILES: dict[str, str] = {
     "hybrid": "configs/hybrid.ini",
 }
 
+# Backend sentinels that occupy a "path" key without naming a file. They are
+# carried verbatim so a backend can recognize them by equality.
+_NON_FILESYSTEM_PATH_VALUES: frozenset[str] = frozenset({IN_MEMORY_DATABASE})
+
 
 def list_profiles() -> list[str]:
     """Return available profile names."""
@@ -371,6 +376,8 @@ def create_main_container(
                 if k != key or not isinstance(v, str):
                     continue
                 value = v.strip().strip('"')
+                if value in _NON_FILESYSTEM_PATH_VALUES:
+                    continue
                 if value.startswith(("ext://", "/", "~")):
                     continue
                 cur[k] = str(base_path / value)

{diffract_core-0.2.1 → diffract_core-0.3.0}/src/diffract/core/cache/interface.py

@@ -70,9 +70,11 @@ class ICacheManager(Protocol):
         Args:
             obj_uid: Unique identifier for the cached object.
             field_name: Name of the field to store.
-            value: Value to cache (must be pickle-serializable).
+            value: Value to cache. Must be a numpy array, a JSON-serializable
+                value, or bytes.
 
         Raises:
+            ValueError: If the value is not a supported kind.
             Exception: If storage operation fails.
         """
         ...

{diffract_core-0.2.1 → diffract_core-0.3.0}/src/diffract/core/cache/redis_manager.py

@@ -7,7 +7,7 @@ LRU eviction and configurable memory limits.
 Features:
     - RAM-only storage with LRU eviction
     - Configurable memory limits and TTL
-    - Pickle serialization for arbitrary Python objects
+    - Safe typed serialization via the shared storage codec
     - Connection pooling and error handling
     - Health monitoring and memory usage tracking
 
@@ -21,17 +21,41 @@ from __future__ import annotations
 
 import contextlib
 import logging
-import pickle
 from types import TracebackType
 from typing import Any, Self
 
 import diffract.core.utils.imports as import_utils
+from diffract.core.storage.serialization import decode_value, encode_value
 from diffract.core.utils.exceptions import format_exception_message
 
 from .interface import UID, ICacheManager
 
 logger = logging.getLogger(__name__)
 
+_TAG_HEADER_BYTES = 1
+_DECODE_MISS_ERRORS = (
+    ValueError,
+    EOFError,
+    OSError,
+    IndexError,
+    UnicodeDecodeError,
+)
+
+
+def _frame(payload: bytes, tag: str) -> bytes:
+    """Prefix a codec payload with its tag so the tag survives the round trip."""
+    tag_bytes = tag.encode("utf-8")
+    return bytes((len(tag_bytes),)) + tag_bytes + payload
+
+
+def _unframe(data: bytes) -> tuple[bytes, str]:
+    """Split a framed blob back into its codec payload and tag."""
+    tag_len = data[0]
+    tag = data[_TAG_HEADER_BYTES : _TAG_HEADER_BYTES + tag_len].decode("utf-8")
+    payload = data[_TAG_HEADER_BYTES + tag_len :]
+    return payload, tag
+
+
 if not import_utils.is_available("redis"):
     logger.debug("Redis not available, disabling Redis cache manager")
 
@@ -61,7 +85,7 @@ else:
 
         The manager configures Redis for RAM-only operation by disabling
         RDB snapshots and AOF logging, focusing on speed over durability.
-        All cached objects are serialized using pickle.
+        Cached values are serialized through the shared storage codec.
 
         Attributes:
             _key_prefix: Prefix for all Redis keys to avoid collisions.
@@ -213,17 +237,9 @@ else:
                 if data is None:
                     return None
                 try:
-                    return pickle.loads(data)  # noqa: S301
-                except (
-                    ModuleNotFoundError,
-                    ImportError,
-                    AttributeError,
-                    ValueError,
-                    pickle.UnpicklingError,
-                    EOFError,
-                ):
-                    # Cache may contain pickles produced by a different Python/numpy
-                    # version. Treat as a cache miss and delete the corrupted entry.
+                    payload, tag = _unframe(data)
+                    return decode_value(payload, tag)
+                except _DECODE_MISS_ERRORS:
                     with contextlib.suppress(RedisError):
                         self._redis.delete(key)
                     return None
@@ -287,19 +303,21 @@ else:
             Args:
                 obj_uid: Unique identifier for the cached object.
                 field_name: Name of the field to store.
-                value: Value to cache (must be pickle-serializable).
+                value: Value to cache. Supported kinds are NumPy arrays,
+                    ``bytes``, and JSON-serializable values.
 
             Raises:
+                ValueError: If the value is not a supported kind.
                 RedisError: If storage operation fails.
             """
+            blob = _frame(*encode_value(value))
             try:
                 key = self._make_key(obj_uid, field_name)
-                payload = pickle.dumps(value, protocol=pickle.HIGHEST_PROTOCOL)
 
                 if self._ttl_seconds is not None:
-                    self._redis.setex(key, self._ttl_seconds, payload)
+                    self._redis.setex(key, self._ttl_seconds, blob)
                 else:
-                    self._redis.set(key, payload)
+                    self._redis.set(key, blob)
 
             except RedisError:
                 logger.exception("Redis set_field error")