hypernix 0.45.2__tar.gz → 0.45.3__tar.gz

{hypernix-0.45.2 → hypernix-0.45.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hypernix
-Version: 0.45.2
+Version: 0.45.3
 Summary: Download and quantize the HyperNix PyTorch model to GGUF (fp32/fp16/Q8_0/Q6_K/Q4_K_M).
 Home-page: https://github.com/minerofthesoal/hypernix-pip
 Author: HyperNix contributors
@@ -53,11 +53,6 @@ Dynamic: license-file
 **End-to-end toolkit for the `ray0rf1re/hyper-nix.1` family of PyTorch
 language models.**
 
-`hypernix` started as a one-shot GGUF converter and has since grown into
-a small but opinionated framework covering the whole lifecycle — from a
-blank directory to a trained, quantized, uploaded HuggingFace snapshot.
-Every subsystem is a plain Python module you can pick up and use in
-isolation.
 
 | Subsystem | What it does |
 |---|---|
@@ -325,6 +320,7 @@ Topic-focused reference guides live in the `wiki/` directory:
 - [`wiki/Training.md`](wiki/Training.md) — scratch training, expansion, and fine-tuning flows
 - [`wiki/CLI.md`](wiki/CLI.md) — full CLI cheat sheet
 - [`wiki/Quantization.md`](wiki/Quantization.md) — GGUF conversion + k-quant pipeline
+- [`wiki/Changelog.md`](wiki/Changelog.md) — full per-release version history
 
 ## How the GGUF pipeline works
 
@@ -340,9 +336,9 @@ k-quant in the plan.
 
 ## Platform notes
 
-- **Linux**: full support, every distro tested (Ubuntu, Debian, Arch, Fedora, openSUSE, Alpine, NixOS).
-- **macOS**: Metal for inference, Homebrew for `llama-quantize`.
-- **Windows**: native support; doctor accepts Windows; `llama-quantize` auto-downloads Windows binaries; use scoop / chocolatey for system deps.
+- **Linux**: full support, every distro tested on: (Ubuntu, Debian, Arch.)
+- **macOS**: Metal for inference, Homebrew for `llama-quantize`. (untested)
+- **Windows**: native support; doctor accepts Windows; `llama-quantize` auto-downloads Windows binaries; use scoop / chocolatey for system deps. (untested)
 - **Pascal (GTX 1080 / 1080 Ti / Titan Xp)**: install torch from the CUDA 11.8 index first (see above). Use `OldFreezer` or `auto_freezer()`; `pascal_safe_dtype()` picks fp16. `hypernix.freezer.pascal_mode_hints()` returns the full Pascal cheat sheet.
 
 ## Build / release

{hypernix-0.45.2 → hypernix-0.45.3}/README.md

@@ -7,11 +7,6 @@
 **End-to-end toolkit for the `ray0rf1re/hyper-nix.1` family of PyTorch
 language models.**
 
-`hypernix` started as a one-shot GGUF converter and has since grown into
-a small but opinionated framework covering the whole lifecycle — from a
-blank directory to a trained, quantized, uploaded HuggingFace snapshot.
-Every subsystem is a plain Python module you can pick up and use in
-isolation.
 
 | Subsystem | What it does |
 |---|---|
@@ -279,6 +274,7 @@ Topic-focused reference guides live in the `wiki/` directory:
 - [`wiki/Training.md`](wiki/Training.md) — scratch training, expansion, and fine-tuning flows
 - [`wiki/CLI.md`](wiki/CLI.md) — full CLI cheat sheet
 - [`wiki/Quantization.md`](wiki/Quantization.md) — GGUF conversion + k-quant pipeline
+- [`wiki/Changelog.md`](wiki/Changelog.md) — full per-release version history
 
 ## How the GGUF pipeline works
 
@@ -294,9 +290,9 @@ k-quant in the plan.
 
 ## Platform notes
 
-- **Linux**: full support, every distro tested (Ubuntu, Debian, Arch, Fedora, openSUSE, Alpine, NixOS).
-- **macOS**: Metal for inference, Homebrew for `llama-quantize`.
-- **Windows**: native support; doctor accepts Windows; `llama-quantize` auto-downloads Windows binaries; use scoop / chocolatey for system deps.
+- **Linux**: full support, every distro tested on: (Ubuntu, Debian, Arch.)
+- **macOS**: Metal for inference, Homebrew for `llama-quantize`. (untested)
+- **Windows**: native support; doctor accepts Windows; `llama-quantize` auto-downloads Windows binaries; use scoop / chocolatey for system deps. (untested)
 - **Pascal (GTX 1080 / 1080 Ti / Titan Xp)**: install torch from the CUDA 11.8 index first (see above). Use `OldFreezer` or `auto_freezer()`; `pascal_safe_dtype()` picks fp16. `hypernix.freezer.pascal_mode_hints()` returns the full Pascal cheat sheet.
 
 ## Build / release

{hypernix-0.45.2 → hypernix-0.45.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "hypernix"
-version = "0.45.2"
+version = "0.45.3"
 description = "Download and quantize the HyperNix PyTorch model to GGUF (fp32/fp16/Q8_0/Q6_K/Q4_K_M)."
 readme = { file = "README.md", content-type = "text/markdown" }
 license = "Apache-2.0"

{hypernix-0.45.2 → hypernix-0.45.3}/setup.cfg

@@ -1,6 +1,6 @@
 [metadata]
 name = hypernix
-version = 0.45.2
+version = 0.45.3
 description = Download and quantize the HyperNix PyTorch model to GGUF (fp32/fp16/Q8_0/Q6_K/Q4_K_M).
 long_description = file: README.md
 long_description_content_type = text/markdown

{hypernix-0.45.2 → hypernix-0.45.3}/src/hypernix/__init__.py

@@ -157,5 +157,5 @@ __all__ = [
     "verify_snapshot",
 ]
 
-__version__ = "0.45.2"
+__version__ = "0.45.3"
 DEFAULT_REPO_ID = "ray0rf1re/hyper-nix.1"

{hypernix-0.45.2 → hypernix-0.45.3}/src/hypernix/smoke_alarm.py

@@ -32,7 +32,7 @@ from __future__ import annotations
 
 import time
 from collections.abc import Callable
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from typing import Any
 
 import torch
@@ -198,6 +198,39 @@ class RadsAlarm(Alarm):
         object.__setattr__(self, "name", "RadsAlarm")
 
 
+# ---------------------------------------------------------------------------
+# Preset name resolution — shared by GasAlarm, AutoAlarm, and the
+# public factory helpers.  The ``preset=`` kwarg is the intuitive
+# single-string entry point ("give me the alarm for an i7-7700HQ" or
+# "give me the alarm for an H100"); it's resolved against GPU_PRESETS
+# first (GPU bandwidth is the dominant signal for training throughput)
+# and CPU_PRESETS as a fallback.
+# ---------------------------------------------------------------------------
+
+
+def _resolve_preset(
+    preset: str | None,
+) -> tuple[CPUPreset | None, GPUPreset | None]:
+    """Resolve a single ``preset=`` string to ``(cpu, gpu)``.
+
+    GPU matches win over CPU matches.  Raises :class:`ValueError` with
+    a useful list of valid names when the string doesn't match any
+    preset.
+    """
+    if preset is None:
+        return None, None
+    g = gpu_preset(preset)
+    if g is not None:
+        return None, g
+    c = cpu_preset(preset)
+    if c is not None:
+        return c, None
+    raise ValueError(
+        f"unknown preset {preset!r}; valid GPU presets: "
+        f"{sorted(GPU_PRESETS)}; valid CPU presets: {sorted(CPU_PRESETS)}"
+    )
+
+
 # ---------------------------------------------------------------------------
 # GasAlarm — mid tier
 # ---------------------------------------------------------------------------
@@ -205,13 +238,29 @@ class RadsAlarm(Alarm):
 @dataclass
 class GasAlarm(Alarm):
     """Mid tier.  Looks up CPU and GPU presets and scales the per-step
-    estimate by their throughput vs. the generic baseline."""
+    estimate by their throughput vs. the generic baseline.
+
+    Accepts three equivalent ways to name the hardware:
+
+    * ``cpu=CPUPreset(...)``, ``gpu=GPUPreset(...)`` — explicit objects.
+    * ``cpu_name=``, ``gpu_name=`` via :meth:`from_names` or the
+      :func:`gas_alarm` factory.
+    * ``preset="i7-7700hq"`` / ``preset="h100"`` — single-string
+      shortcut, resolved against GPU first then CPU.
+    """
 
     cpu: CPUPreset | None = None
     gpu: GPUPreset | None = None
+    preset: str | None = field(default=None, kw_only=True)
 
     def __post_init__(self) -> None:
         object.__setattr__(self, "name", "GasAlarm")
+        if self.preset is not None:
+            c, g = _resolve_preset(self.preset)
+            if c is not None and self.cpu is None:
+                self.cpu = c
+            if g is not None and self.gpu is None:
+                self.gpu = g
 
     @classmethod
     def from_names(
@@ -220,12 +269,14 @@ class GasAlarm(Alarm):
         *,
         cpu_name: str | None = None,
         gpu_name: str | None = None,
+        preset: str | None = None,
         **kwargs: Any,
     ) -> GasAlarm:
         return cls(
             time_budget_seconds=time_budget_seconds,
             cpu=cpu_preset(cpu_name) if cpu_name else None,
             gpu=gpu_preset(gpu_name) if gpu_name else None,
+            preset=preset,
             **kwargs,
         )
 
@@ -396,10 +447,20 @@ def gas_alarm(
     *,
     cpu_name: str | None = None,
     gpu_name: str | None = None,
+    preset: str | None = None,
     **kw: Any,
 ) -> GasAlarm:
+    """Construct a :class:`GasAlarm`.
+
+    ``cpu_name`` / ``gpu_name`` resolve against ``CPU_PRESETS`` /
+    ``GPU_PRESETS`` respectively.  ``preset`` is a one-string alias
+    that resolves against GPU first, then CPU — use it when you just
+    want "the alarm for my card / chip" without caring which registry
+    the name comes from.
+    """
     return GasAlarm.from_names(
-        time_budget_seconds, cpu_name=cpu_name, gpu_name=gpu_name, **kw,
+        time_budget_seconds,
+        cpu_name=cpu_name, gpu_name=gpu_name, preset=preset, **kw,
     )
 
 
@@ -423,6 +484,7 @@ def auto_alarm(
     *,
     cpu_name: str | None = None,
     gpu_name: str | None = None,
+    preset: str | None = None,
     warmup_step_fn: Callable[[], Any] | None = None,
     detect_hardware: bool = True,
     **kw: Any,
@@ -430,9 +492,24 @@ def auto_alarm(
     """Convenience entry point that returns a concrete alarm.
 
     When ``detect_hardware=True`` (the default) and no ``cpu_name`` /
-    ``gpu_name`` was supplied, :func:`detect_cpu_preset` and
-    :func:`detect_gpu_preset` are consulted first.
+    ``gpu_name`` / ``preset`` was supplied, :func:`detect_cpu_preset`
+    and :func:`detect_gpu_preset` are consulted first.
+
+    ``preset`` is the one-string shortcut — resolved against GPU
+    presets first and CPU presets second.
     """
+    if preset is not None:
+        c, g = _resolve_preset(preset)
+        if c is not None and cpu_name is None:
+            for k, v in CPU_PRESETS.items():
+                if v is c:
+                    cpu_name = k
+                    break
+        if g is not None and gpu_name is None:
+            for k, v in GPU_PRESETS.items():
+                if v is g:
+                    gpu_name = k
+                    break
     if detect_hardware and gpu_name is None:
         det_gpu = detect_gpu_preset()
         if det_gpu is not None:

{hypernix-0.45.2 → hypernix-0.45.3}/src/hypernix.egg-info/SOURCES.txt

@@ -71,6 +71,7 @@ tests/test_v031_chat_and_archs.py
 wiki/Alarms.md
 wiki/Architectures.md
 wiki/CLI.md
+wiki/Changelog.md
 wiki/Freezer.md
 wiki/Fridges.md
 wiki/Home.md

{hypernix-0.45.2 → hypernix-0.45.3}/tests/test_alarms_and_presets.py

@@ -245,6 +245,70 @@ def test_auto_alarm_with_warmup_picks_modern() -> None:
     assert a.measured_step_seconds is not None
 
 
+# Regression: GasAlarm(preset="i7-7700hq") used to raise
+# "unexpected keyword argument 'preset'".  Every constructor / factory
+# now accepts ``preset=`` as a one-string shortcut that resolves
+# against GPU_PRESETS first, then CPU_PRESETS.
+
+
+def test_gas_alarm_accepts_preset_cpu() -> None:
+    from hypernix import smoke_alarm
+
+    a = smoke_alarm.GasAlarm(time_budget_seconds=3600.0, preset="i7-7700hq")
+    assert a.cpu is not None
+    assert a.cpu.name.startswith("Intel Core i7-7700HQ")
+    assert a.gpu is None
+
+
+def test_gas_alarm_accepts_preset_gpu() -> None:
+    from hypernix import smoke_alarm
+
+    a = smoke_alarm.GasAlarm(time_budget_seconds=3600.0, preset="rtx-3080-ti")
+    assert a.gpu is not None
+    assert a.gpu.name.startswith("NVIDIA GeForce RTX 3080 Ti")
+    assert a.cpu is None
+
+
+def test_gas_alarm_factory_accepts_preset() -> None:
+    from hypernix import smoke_alarm
+
+    a = smoke_alarm.gas_alarm(time_budget_seconds=3600.0, preset="h100")
+    assert a.gpu is not None
+    assert "H100" in a.gpu.name
+
+
+def test_auto_alarm_accepts_preset() -> None:
+    from hypernix import smoke_alarm
+
+    a = smoke_alarm.auto_alarm(
+        time_budget_seconds=3600.0, preset="i7-7700hq",
+        detect_hardware=False,
+    )
+    assert isinstance(a, smoke_alarm.GasAlarm)
+    assert a.cpu is not None
+
+
+def test_gas_alarm_unknown_preset_lists_valid() -> None:
+    import pytest
+
+    from hypernix import smoke_alarm
+
+    with pytest.raises(ValueError, match="unknown preset"):
+        smoke_alarm.GasAlarm(time_budget_seconds=3600.0, preset="not-real-cpu")
+
+
+def test_preset_does_not_override_explicit_cpu() -> None:
+    """Explicit ``cpu=`` wins over a conflicting ``preset=`` hint."""
+    from hypernix import smoke_alarm
+    from hypernix.freezer import cpu_preset
+
+    explicit = cpu_preset("i7-14700k")
+    a = smoke_alarm.GasAlarm(
+        time_budget_seconds=3600.0, cpu=explicit, preset="i7-7660u",
+    )
+    assert a.cpu is explicit
+
+
 def test_auto_alarm_with_gpu_name_picks_gas() -> None:
     from hypernix import smoke_alarm
 

hypernix-0.45.3/wiki/Changelog.md

@@ -0,0 +1,248 @@
+# Changelog
+
+Full per-release notes for `hypernix`. The top-level `wiki/Home.md`
+keeps a running "recent highlights" list; this page is the canonical
+history. Semver-ish: minor bumps add features, patch bumps are bug
+fixes and UX papercuts. Dates are `YYYY-MM-DD` for PyPI-published
+releases; in-branch commits between releases are grouped under the
+next release header.
+
+## Legend
+
+- ✨ new feature
+- 🐛 bug fix
+- 🛡️ UX / error-message polish
+- 📚 documentation
+- 🔧 internal / plumbing
+
+---
+
+## 0.45.3
+
+🛡️ **`smoke_alarm.GasAlarm` accepts `preset=`.** One-string shortcut
+that resolves against `GPU_PRESETS` first, then `CPU_PRESETS`. Works
+on the class (`GasAlarm(..., preset="i7-7700hq")`), on the factory
+(`gas_alarm(..., preset="h100")`), and on the selector
+(`auto_alarm(..., preset="rtx-3080-ti")`). Unknown names raise
+`ValueError` with the full list of valid presets.
+
+🛡️ Explicit `cpu=` / `gpu=` instances still win over a conflicting
+`preset=` hint — no silent overwrite.
+
+🔧 Shared `_resolve_preset` helper in `smoke_alarm.py`.
+
+## 0.45.2
+
+🐛 **Every pan accepts `context_length=` and `max_chars=`.** Reported:
+`FryingPan(context_length=CONTEXT_LEN)` raised a bare `TypeError`.
+Both are now keyword-only fields on the `Pan` base class; when set,
+lines are truncated to fit. `context_length` is treated as
+`max_chars = context_length * 4` (English-BPE heuristic); the direct
+`max_chars=` wins when both are set. For precise chunking by tokens
+use `hypernix.food_processor` instead.
+
+## 0.45.1
+
+🐛 **Pan positional-argument fix.** `Pan` inherited `name: str` as a
+dataclass field, so `Skillet(src, "instruct")` silently set
+`name="instruct"` and left `mode="chat"`. Fix: `name` is now a
+`typing.ClassVar` on every pan — still the pan's label, no longer
+part of `__init__`. `GrillPan._seen` (internal dedupe state) marked
+`init=False`.
+
+🛡️ `pick_pan` error messages now list valid tiers / valid kwargs
+instead of raising `KeyError` or cryptic `TypeError`.
+
+## 0.45.0
+
+✨ **Espresso, blender, toaster, food_processor, smoker** — five new
+appliances, each 4 tiers. Shared interface per module.
+
+✨ **+3 microwave tiers** — now `defrost` (preheat-only) / `low_zap`
+(deterministic one-liner) / `zap` (existing) / `high_zap`
+(long-temp draft) / `chat_zap` (existing). Plus `reheat(oven,
+prior_output)` for continuation without rebuild.
+
+✨ **+2 coffee_maker tiers and one new type.**
+`FrenchPressMaker` (batch), `PercolatorMaker` (cyclic with optional
+convergence), and a new `ColdBrewMaker` (long single brew with
+mandatory JSON checkpoints, resumes cleanly after a crash).
+
+✨ **CLI `hypernix brew recipe.json`** — runs `instant_pot.brew`
+from a JSON recipe. Supports `--set KEY=VALUE` overrides with JSON
+literals.
+
+📚 `wiki/Kitchen.md` gets full sections for every new appliance.
+
+## 0.44.0
+
+✨ **Kitchen modules + pressure_cooker optimizer.** Seven new
+top-level modules (pans, microwave, table, sink, instant_pot,
+coffee_maker, pressure_cooker) covering preprocessing, throwaway
+inference, log inspection, file output, end-to-end pipelines,
+scheduled repetition, and a custom optimizer.
+
+✨ `pressure_cooker` — `torch.optim.Optimizer` subclass: AdamW +
+three-phase LR schedule (linear warmup → plateau → cosine cooldown)
++ Zhang-et-al-2019 Lookahead "pressure seal". No separate scheduler
+object; the LR lives inside the optimizer.
+
+📚 README gains a **"Who this is actually for"** section framing the
+package around the solo-GPU / consumer-card / QLoRA-to-Hub workflow,
+with an explicit disclaimer that `train()` is a smoke-tester, not a
+production trainer. New `wiki/Kitchen.md`.
+
+## 0.43.0
+
+✨ **`smoke_alarm`** — four-tier training-step planner + mid-run
+monitor. `RadsAlarm` (constants, lightest), `GasAlarm` (CPU/GPU
+presets), `ModernAlarm` (warmup-measured), `AutoAlarm` (selector).
+
+✨ **16 CPU presets** (`hypernix.freezer.CPU_PRESETS`): i7 7th gen
+(7660U / 7700HQ / 7700K), 11th–14th gen K/H/HX, Core Ultra Series 1
+(Meteor / Lunar Lake), Series 2 (Arrow Lake, AVX10).
+
+✨ **20 GPU presets** (`hypernix.freezer.GPU_PRESETS`): Hopper
+(H100/H200), Ampere workstation (A4500–A6000), RTX PRO Ada +
+Blackwell, RTX 4070 Ti Super / 4080 Super, RTX 3080 Ti, Turing
+consumer (1660 Ti, 2080, 2080 Super, 2080 Ti), Pascal (1080, 1080 Ti).
+
+📚 New `wiki/Alarms.md` with both preset tables.
+
+## 0.42.0
+
+✨ **`new_range` / `old_range` / `industrial_range`** — three
+sophistication tiers of labeling rubrics that drop into
+`mediocre_fridge.collect_responses_from(label_rule=...)`.
+
+- `new_range` — zero-dep first-fail rubric (is_empty, is_refusal,
+  math_lacks_digit, is_repetition).
+- `old_range` — weighted-mean scored rubric with `None` = "no
+  opinion", any-rule-at-0 short-circuits to BAD, references / keywords
+  / stopword-filtered overlap built in.
+- `industrial_range` — LLM-as-judge wrapper around any CodeOven;
+  pointwise + pairwise with caching.
+
+📚 New `wiki/Ranges.md`.
+
+## 0.41.0
+
+✨ **CUDA 6.1 / Pascal support.** `compute_capability`, `is_pascal`,
+`pascal_safe_dtype` (fp32 on CPU, fp16 on Pascal / Volta / Turing,
+bf16 on Ampere+), `pascal_mode_hints` (one-stop dict of recommended
+settings for sm_61).
+
+✨ **`examples/train_hypernix_1_5_gtx1080.py`** — HyperNix 1.5,
+verified 92,130,048 params, trains on an 8 GB Pascal card via
+`auto_freezer` + `flash_freezer(slow=True)`.
+
+📚 New `wiki/Pascal.md` with a full sm_61 playbook.
+
+## 0.40.0
+
+✨ **`freezer` module** — VRAM manager. `OldFreezer` (8 – 10 GB,
+batch=1, fp16, empty_cache each step), `NewFreezer` (11 GB+, batch=8,
+fp32/bf16), `FlashFreezer` (OOM-safe retry wrapper with exponential
+backoff, wait-for-free-GB, and optional slow-mode that halves
+`current_batch_size` on each retry).
+
+📚 New `wiki/Freezer.md`.
+
+## 0.36.0
+
+✨ **`old_fridge` / `mediocre_fridge` / `new_fridge`** — memory
+housekeeping (freeze/unfreeze/parameter_stats), judge-training dataset
+synthesis, and training-curve plotting.
+
+✨ `examples/train_hypernix_0_1_5_evaluator.py` — end-to-end example
+wiring ovens + all three fridges.
+
+📚 New `wiki/Fridges.md`.
+
+## 0.35.0
+
+✨ **Gemma 4, Qwen 3.5 & 3.6, GLM 5.x, Nix collection presets.** New
+entries in both `ARCH_PRESETS` (for `new_oven`) and `KNOWN_MODELS`
+(for short-name resolution). Config knobs verified against the actual
+HuggingFace repos.
+
+## 0.34.0
+
+✨ **AutoModel fallback.** `load_snapshot` routes any non-HyperNix
+`model_type` (Gemma, Phi, DeepSeek, GLM, GPT-OSS, …) through a thin
+`transformers.AutoModelForCausalLM` wrapper. New ARCH_PRESETS covering
+those families.
+
+## 0.33.0
+
+✨ **Windows + macOS support.** Cross-platform `doctor`, path
+handling, `llama-quantize` resolution.
+
+✨ **Python 3.13** support (sentencepiece 0.2.1 floor).
+
+✨ **Runtime auto-install.** `HYPERNIX_AUTO_INSTALL` env var (default
+on) lets missing runtime deps be installed lazily; `hypernix doctor
+--fix` makes it explicit.
+
+## 0.32.1
+
+🐛 Fall back to the slow tokenizer when the `tokenizers` crate is too
+old to decode a newer tokenizer.json.
+
+## 0.32.0
+
+✨ **torch 2.7+** (incl. CUDA 11.8 builds).
+
+✨ One-shot PyPI publish via GitHub Actions Trusted Publishing.
+
+## 0.31.0
+
+✨ **Chat REPL.** `hypernix chat --repo-id <short-name>` plus
+`CodeOven.chat(turns, ...)`.
+
+✨ **Nano-nano / Nano-mini / nano-nano-927** family — new entries in
+`KNOWN_MODELS`.
+
+## 0.30.0
+
+✨ **`old_oven` code-generation wrapper.** `preheat`, `CodeOven`,
+`bake_code`, `fill_middle`, `save_pt` / `load_pt`. `--auto-oven`
+top-level CLI shortcut.
+
+## 0.21.0
+
+✨ Download every file the model needs — not just weights — so the
+output directory is a self-contained snapshot.
+
+## 0.2.0
+
+✨ First subcommand-based CLI. `train` module scaffold. Fixed
+`tokenizer.ggml.merges` in GGUF output.
+
+---
+
+## Upgrading
+
+`hypernix` follows no breaking-change policy yet. Patch releases
+(`0.45.x`) are always safe to upgrade — they only fix bugs, UX
+papercuts, or improve error messages.
+
+Minor releases (`0.N.0`) add features. The usual gotcha is renamed
+kwargs from the UX-polish patches above; when in doubt, check the
+signature:
+
+```python
+import inspect
+from hypernix import smoke_alarm, pans
+
+print(inspect.signature(smoke_alarm.GasAlarm))
+print(inspect.signature(pans.FryingPan))
+```
+
+## Contributing changelog entries
+
+New features should land with a one-paragraph entry at the top of
+this file, grouped by emoji legend. Patch releases get a couple of
+bullet points; minor releases get a section per subsystem touched.
+Keep the tone utilitarian — what changed, how the caller notices,
+what to do instead if an old call stopped working.

{hypernix-0.45.2 → hypernix-0.45.3}/wiki/Home.md

@@ -18,6 +18,7 @@ see the [top-level README](../README.md).
 | [Training](Training.md) | `init_from_scratch`, `expand_checkpoint`, `train`, AutoModel fallback. |
 | [Quantization](Quantization.md) | GGUF pipeline, k-quants, `llama-quantize` integration. |
 | [CLI](CLI.md) | Every subcommand, every flag, typical invocations. |
+| [Changelog](Changelog.md) | Full per-release notes — features, fixes, UX papercuts. |
 
 ## The subsystem map
 
@@ -70,9 +71,12 @@ see the [top-level README](../README.md).
 
 ## Version history
 
-The project is semver-ish; minor bumps add features, patch bumps are
-bug fixes. Recent notable releases:
+Recent releases (see [Changelog](Changelog.md) for the full per-release
+notes going all the way back to 0.2.0):
 
+- **0.45.3** — `smoke_alarm` accepts `preset=` one-string kwarg
+- **0.45.2** — pans accept `context_length` / `max_chars` (kw-only)
+- **0.45.1** — pans init fix: positional args no longer bind to `name`
 - **0.45.0** — espresso_maker, blender, toaster, food_processor, smoker; +3 microwave tiers; +2 coffee_maker tiers + cold_brew; CLI `brew`
 - **0.44.0** — pans / microwave / table / sink / instant_pot / coffee_maker / pressure_cooker
 - **0.43.0** — `smoke_alarm` (Rads / Gas / Modern / Auto) + 16 CPU + 20 GPU presets
@@ -86,3 +90,6 @@ bug fixes. Recent notable releases:
 - **0.32.x** — CUDA 11.8 torch, slow-tokenizer fallback
 - **0.31.x** — Chat REPL, Nano-nano / Nano-mini model family
 - **0.30.x** — Code-generation oven (`old_oven.preheat`)
+
+See [Changelog.md](Changelog.md) for per-release details including
+patch versions, UX fixes, and bug reports.