hypernix 0.72.5.dev2__tar.gz → 0.72.5.post1__tar.gz

{hypernix-0.72.5.dev2 → hypernix-0.72.5.post1}/.github/workflows/ios.yml

@@ -49,7 +49,12 @@ on:
       local_engine:
         description: "Build llama.cpp into the app so it can run models on-device."
         type: boolean
-        default: false
+        # True for a *called* build, unlike the dispatch default below.
+        # The only reason to call this workflow is to ship the result,
+        # and an IPA that cannot run a model is not worth shipping. The
+        # expensive default is the safe one here; the cheap default
+        # belongs on the PR path, where it is.
+        default: true
       version:
         description: "Marketing version to stamp into the build"
         type: string
@@ -131,8 +136,50 @@ jobs:
       # macOS runner and is not worth paying on a PR that touched a
       # view. Without it the app builds and ships with EchoRunner,
       # which tells the user this build has no local engine.
+      # Every trigger gets a stated answer, because the obvious
+      # expression does not have one.
+      #
+      # `if: ${{ inputs.local_engine }}` was the whole engine-skipping
+      # bug, and it looks correct. The `inputs` context exists *only*
+      # for `workflow_dispatch` and `workflow_call`; on a `push` or a
+      # `pull_request` it is not populated at all, so
+      # `inputs.local_engine` is null, null is falsy, and the step was
+      # skipped on every single commit to main -- silently, with a grey
+      # "skipped" in the log that looks exactly like a deliberate one.
+      # `--require-engine` was gated on the same expression, so the
+      # guard that exists to catch this was switched off by it too.
+      #
+      # Every commit to main therefore shipped an app whose on-device
+      # inference was EchoRunner. That is the "it always skips the
+      # inference engine build" report, and setting the workflow_call
+      # default to true (which fixed the release path) could not have
+      # fixed it: the release path was never the one skipping.
+      - name: Decide whether to build the engine
+        id: engine
+        env:
+          REQUESTED: ${{ inputs.local_engine }}
+          EVENT: ${{ github.event_name }}
+        run: |
+          set -euo pipefail
+          # REQUESTED is "true", "false", or empty when there is no
+          # `inputs` context to read. Empty is the case that has to be
+          # decided here rather than defaulted to false by accident.
+          case "$REQUESTED" in
+            true)  want=true ;;
+            false) want=false ;;
+            *)
+              # No input: the trigger decides. main is what people
+              # install from and it gets a real engine; a pull request
+              # does not, because 15-25 minutes on every push is what
+              # makes somebody turn this off permanently.
+              if [ "$EVENT" = "push" ]; then want=true; else want=false; fi
+              ;;
+          esac
+          echo "Engine requested: '${REQUESTED:-<no inputs context>}' on $EVENT -> $want"
+          echo "build=$want" >> "$GITHUB_OUTPUT"
+
       - name: Build the inference engine
-        if: ${{ inputs.local_engine }}
+        if: steps.engine.outputs.build == 'true'
         working-directory: ios
         run: |
           set -euo pipefail
@@ -151,7 +198,11 @@ jobs:
           # XcodeGen cannot make that call: its `optional: true` means
           # weak linking, not "skip when absent", so a spec that always
           # names the framework fails to build on a checkout without it.
-          python3 scripts/prepare_project.py
+          # --require-engine when the build asked for one, so a failed
+          # or skipped engine step fails here instead of shipping an IPA
+          # in which every on-device feature quietly does nothing.
+          python3 scripts/prepare_project.py \
+            ${{ steps.engine.outputs.build == 'true' && '--require-engine' || '' }}
           xcodegen generate --spec project.generated.yml
           ls -la HyperLink.xcodeproj
 
@@ -186,6 +237,21 @@ jobs:
           fi
           echo "Testing on simulator $device"
           set -o pipefail
+          # The whole log to a file, and only a summary to the console.
+          #
+          # This used to be `| tail -60`, which discards exactly what is
+          # needed when the build fails: xcodebuild prints a short
+          # "Testing failed:" summary of *deduplicated* messages and
+          # then the per-file errors with their line numbers, and the
+          # tail keeps the wrong end. A compile failure would report six
+          # unique messages and one file:line, so fixing it took a round
+          # trip per error.
+          #
+          # Now every `error:` line is echoed first, in full, and the
+          # complete log is uploaded as an artifact.
+          # The redirect below needs the directory to exist first.
+          mkdir -p build
+          set +e
           xcodebuild test \
             -project HyperLink.xcodeproj \
             -scheme HyperLink \
@@ -193,7 +259,28 @@ jobs:
             -destination "id=$device" \
             -resultBundlePath build/TestResults.xcresult \
             CODE_SIGNING_ALLOWED=NO \
-            | tail -60
+            > build/xcodebuild.log 2>&1
+          status=$?
+          set -e
+
+          if [ "$status" -ne 0 ]; then
+            echo "::group::Every compiler error"
+            # Sorted and deduplicated: a batch build repeats the same
+            # error once per file that triggered it, and forty copies of
+            # one message buries the other five.
+            grep -E "(error|warning): " build/xcodebuild.log \
+              | sed 's|/Users/runner/work/HyperNix-pip/HyperNix-pip/||' \
+              | sort -u \
+              | head -200 || true
+            echo "::endgroup::"
+            echo "::group::The last 200 lines"
+            tail -200 build/xcodebuild.log
+            echo "::endgroup::"
+            echo "::error::xcodebuild failed — see the groups above, or the ios-build-log artifact"
+          else
+            tail -20 build/xcodebuild.log
+          fi
+          exit "$status"
 
       - name: Archive
         working-directory: ios
@@ -315,6 +402,15 @@ jobs:
           if-no-files-found: error
           retention-days: 90
 
+      - name: Upload the build log
+        if: always() && inputs.run_tests != false
+        uses: actions/upload-artifact@v4
+        with:
+          name: ios-build-log-${{ github.run_number }}
+          path: ios/build/xcodebuild.log
+          if-no-files-found: ignore
+          retention-days: 7
+
       - name: Upload the test results
         if: always() && inputs.run_tests != false
         uses: actions/upload-artifact@v4

{hypernix-0.72.5.dev2 → hypernix-0.72.5.post1}/.github/workflows/public-release.yml

@@ -184,6 +184,19 @@ jobs:
 
           init = Path("src/hypernix/__init__.py")
           init.write_text(re.sub(r'__version__ = "[^"]+"', f'__version__ = "{v}"', init.read_text(), count=1))
+
+          # The fourth place a version string lives. install-t1.sh bakes
+          # one as the `curl | bash` fallback, for the case where there
+          # is no checkout to read and no hypernix installed yet to ask
+          # -- and a release that bumped the other three without this one
+          # shipped an installer whose banner announced the previous
+          # version over a correct install. tests/test_install_script.py
+          # fails when they drift, which is how this came to be here.
+          installer = Path("install-t1.sh")
+          installer.write_text(
+              re.sub(r'^VERSION="[^"]+"$', f'VERSION="{v}"',
+                     installer.read_text(), count=1, flags=re.M)
+          )
           PY
           echo "--- pyproject.toml version line"
           grep '^version' pyproject.toml
@@ -191,6 +204,8 @@ jobs:
           grep '^version' setup.cfg
           echo "--- __init__.py version line"
           grep __version__ src/hypernix/__init__.py
+          echo "--- install-t1.sh version line"
+          grep '^VERSION=' install-t1.sh
 
       - name: Lint
         run: ruff check src/hypernix

{hypernix-0.72.5.dev2 → hypernix-0.72.5.post1}/.github/workflows/release.yml

@@ -50,6 +50,14 @@ jobs:
     uses: ./.github/workflows/ios.yml
     with:
       run_tests: true
+      # A release IPA has to be able to run a model. Without this the
+      # build links no llama.cpp and ships EchoRunner, so every
+      # on-device feature is inert in the artifact people install --
+      # and nothing said so, because falling back is the correct
+      # behaviour for a PR build and the release used the same default.
+      # It costs 15-25 minutes and it is the difference between an app
+      # and a demo.
+      local_engine: true
       version: ""
     secrets: inherit
 

{hypernix-0.72.5.dev2 → hypernix-0.72.5.post1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hypernix
-Version: 0.72.5.dev2
+Version: 0.72.5.post1
 Summary: Download and quantize the HyperNix PyTorch model to GGUF (fp32/fp16/Q8_0/Q6_K/Q4_K_M). and train models with a massive ai toolkit 
 Home-page: https://github.com/trail-b1az3r/hypernix-pip
 Author: HyperNix contributors
@@ -78,7 +78,7 @@ Dynamic: license-file
 </p>
 
 <p align="center">
-  <img alt="PyPI" src="https://img.shields.io/badge/PyPI-v0.72.4.post20-ff2d55?style=for-the-badge&logo=pypi&logoColor=white" />
+  <img alt="PyPI" src="https://img.shields.io/badge/PyPI-v0.72.5-ff2d55?style=for-the-badge&logo=pypi&logoColor=white" />
   <img alt="Python" src="https://img.shields.io/badge/Python-3.10--3.14-00c9ff?style=for-the-badge&logo=python&logoColor=white" />
   <img alt="License" src="https://img.shields.io/badge/License-HOS%20/%20LLU-00c853?style=for-the-badge" />
 </p>

{hypernix-0.72.5.dev2 → hypernix-0.72.5.post1}/README.md

@@ -4,7 +4,7 @@
 </p>
 
 <p align="center">
-  <img alt="PyPI" src="https://img.shields.io/badge/PyPI-v0.72.4.post20-ff2d55?style=for-the-badge&logo=pypi&logoColor=white" />
+  <img alt="PyPI" src="https://img.shields.io/badge/PyPI-v0.72.5-ff2d55?style=for-the-badge&logo=pypi&logoColor=white" />
   <img alt="Python" src="https://img.shields.io/badge/Python-3.10--3.14-00c9ff?style=for-the-badge&logo=python&logoColor=white" />
   <img alt="License" src="https://img.shields.io/badge/License-HOS%20/%20LLU-00c853?style=for-the-badge" />
 </p>

{hypernix-0.72.5.dev2 → hypernix-0.72.5.post1}/bin/hypernix-t1

@@ -19,8 +19,42 @@
 
 set -euo pipefail
 
+# The version this script reports. A fallback, not the answer.
+#
+# It was a hand-maintained literal and it was stale, which is one third
+# of "the installed T1 thinks it is running an older version" — the
+# other two thirds were install-t1.sh's two constants. `hypernix-t1
+# version` now asks the installed package, because by the time this
+# script does anything useful the package is installed by definition,
+# and a number derived from the thing it describes cannot drift from it.
 VERSION="0.72.5"
 
+# The real versions, from the package this script drives.
+#
+# Deliberately not a `sed` of the source like install-t1.sh does: that
+# script runs before anything is installed and has a checkout to read,
+# and this one runs after and may have no checkout at all. Asking the
+# interpreter is the accurate answer *here* and the impossible one
+# there.
+report_versions() {
+  local py
+  py="$(python_bin 2>/dev/null)" || { printf 'hypernix-t1 %s\n' "$VERSION"; return 0; }
+  "$py" - <<'PYEOF' 2>/dev/null || printf 'hypernix-t1 %s\n' "$VERSION"
+import sys
+
+try:
+    import hypernix
+    from hypernix.t1api.version import T1_VERSION
+except Exception:
+    sys.exit(1)
+
+print(f"hypernix-t1 (hypernix {hypernix.__version__})")
+print(f"t1 api     v{T1_VERSION.short}")
+print(f"python     {'.'.join(str(p) for p in sys.version_info[:3])}")
+print(f"runs as    {sys.executable}")
+PYEOF
+}
+
 CONFIG_DIR="${T1_CONFIG_DIR:-$HOME/.hypernix/t1api}"
 ENV_FILE="$CONFIG_DIR/.env"
 PID_FILE="$CONFIG_DIR/server.pid"
@@ -695,6 +729,19 @@ cmd_training() {
     exec "$py" -m hypernix.t1api.training_cli "$@"
 }
 
+cmd_runner() {
+  # Talks to the running server over HTTP rather than loading a model
+  # here. The runner is a process the *server* owns — starting a second
+  # llama.cpp from this script would take the VRAM the server's own
+  # copy is using and the failure would land on the one that was
+  # working.
+  require_installed
+  load_env
+  local py; py="$(python_bin)"
+  T1_CONFIG_DIR="${T1_CONFIG_DIR:-$CONFIG_DIR}" \
+    exec "$py" -m hypernix.t1api.runner_cli "$@"
+}
+
 cmd_index() {
   # No require_installed: indexing reads GGUF files and writes JSON, and
   # needs neither a configured server nor the [t1api] extra. Someone who
@@ -871,11 +918,12 @@ main() {
     test)              cmd_test "$@" ;;
     key|keys)          cmd_key "$@" ;;
     index)             cmd_index "$@" ;;
+    runner)            cmd_runner "$@" ;;
     launch-script)     cmd_launch_script "$@" ;;
     training)          cmd_training "$@" ;;
     autostart)         cmd_autostart "$@" ;;
     remove|uninstall)  cmd_remove "$@" ;;
-    version|--version) printf 'hypernix-t1 %s\n' "$VERSION" ;;
+    version|--version) report_versions ;;
     help|-h|--help)    usage ;;
     *) err "Unknown command: $cmd"; usage >&2; exit 2 ;;
   esac

{hypernix-0.72.5.dev2 → hypernix-0.72.5.post1}/install-t1.sh

@@ -28,8 +28,59 @@
 
 set -euo pipefail
 
-VERSION="0.72.2.post5"
-T1_API_VERSION="1.0.26.8.1.1"
+# The version this installer announces, in the banner and in the config
+# it writes.
+#
+# These two were maintained by hand and went stale, which is the whole
+# of the "the installed T1 thinks it is running an older version"
+# report: the banner said 0.72.2.post5 / t1 v1.0.26.8.1.1 while the
+# package being installed was several releases past both. Nothing was
+# wrong with the install — the only thing that was ever wrong was the
+# number printed over it, and there is no way to tell those apart from
+# the outside.
+#
+# They are still literals because this script has to work as
+# `curl ... | bash`, where there is no checkout to read and no hypernix
+# installed yet to ask. So they are a *fallback*, `derive_versions`
+# below replaces them whenever the script is run from a clone, and
+# tests/test_install_script.py fails if the fallback drifts from the
+# package again.
+VERSION="0.72.5.post1"
+T1_API_VERSION="1.0.26.9.2.3"
+
+# Replace the baked versions with the real ones, when they can be read.
+#
+# Reads the sources rather than importing them: this runs before
+# anything is installed, and `python3 -c 'import hypernix'` at this
+# point either fails or — worse — succeeds against some *other*
+# hypernix that happens to be on the path and reports its version
+# instead of the one about to be installed.
+derive_versions() {
+  local here src package t1 year
+  here="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" 2>/dev/null && pwd)" || return 0
+  src="$here/src/hypernix"
+  [ -f "$src/__init__.py" ] || return 0
+
+  package="$(sed -n 's/^__version__ = "\(.*\)"$/\1/p' "$src/__init__.py" | head -1)"
+  [ -n "$package" ] && VERSION="$package"
+
+  # T1Version(api=1, major=0, year=2026, month=9, feature=2, fix=3)
+  # becomes 1.0.26.9.2.3 — the short spelling, two-digit year, which is
+  # what /status and every wire response carry.
+  t1="$(sed -n \
+    's/^T1_VERSION = T1Version(api=\([0-9][0-9]*\), *major=\([0-9][0-9]*\), *year=\([0-9][0-9]*\), *month=\([0-9][0-9]*\), *feature=\([0-9][0-9]*\), *fix=\([0-9][0-9]*\)).*$/\1 \2 \3 \4 \5 \6/p' \
+    "$src/t1api/version.py" | head -1)"
+  if [ -n "$t1" ]; then
+    # shellcheck disable=SC2086 # deliberate word splitting into $1..$6
+    set -- $t1
+    year="$3"
+    # 2026 -> 26. Left alone if it is already two digits.
+    [ ${#year} -gt 2 ] && year="${year#"${year%??}"}"
+    T1_API_VERSION="$1.$2.$year.$4.$5.$6"
+  fi
+}
+
+derive_versions
 
 # ---------------------------------------------------------------------------
 # Output
@@ -69,6 +120,13 @@ banner() {
 
 INTERACTIVE=1
 DRY_RUN=0
+# Index ~/.hypernix/models into the registry after installing, so models
+# already on this machine are servable — and visible in HyperLink —
+# without anybody hand-writing models.json.
+INDEX_MODELS=0
+# The same, plus a price worked out per model. Separate because pricing
+# is a policy decision and indexing is not.
+ESTIMATE_PRICES=0
 CONFIG_DIR="${T1_CONFIG_DIR:-$HOME/.hypernix/t1api}"
 INSTALL_MODE=""          # venv | user | system | skip
 ASSUME_YES=0
@@ -97,6 +155,18 @@ install-t1.sh — interactive installer and setup for the HyperNix T1 API.
                         exposes it to every network this machine is on.
   --port N              Port, answering the "Port" question up front.
   --force               Overwrite an existing .env instead of stopping.
+  --index-models        After installing, read every GGUF in
+                        ~/.hypernix/models and write it into the model
+                        registry. They become servable, switchable, and
+                        visible in HyperLink without anybody editing
+                        models.json by hand.
+  --estimate-prices     Index, and work out what each model costs to
+                        serve from its size, quantisation, parameter
+                        count, this machine's GPU, and its active-vs-total
+                        parameters if it is a mixture-of-experts. Implies
+                        --index-models. A starting point, not a market
+                        rate — and better than the zero that is otherwise
+                        written, which bills you.
   --help                This.
 
 `hypernix-t1 create` forwards its own --host/--port/--force here, so the
@@ -120,6 +190,8 @@ while [ $# -gt 0 ]; do
     --host)            shift; [ $# -gt 0 ] || die "--host needs an address"; FORCED_HOST="$1" ;;
     --port)            shift; [ $# -gt 0 ] || die "--port needs a number"; FORCED_PORT="$1" ;;
     --force)           FORCE_OVERWRITE=1 ;;
+    --index-models)    INDEX_MODELS=1 ;;
+    --estimate-prices) INDEX_MODELS=1; ESTIMATE_PRICES=1 ;;
     --trusted-network) TRUSTED_NETWORK=1; TRUSTED_NETWORK_SET=1 ;;
     --no-trusted-network) TRUSTED_NETWORK=0; TRUSTED_NETWORK_SET=1 ;;
     --trusted-network-partial-admin)
@@ -393,8 +465,12 @@ install_package() {
       ;;
     user)
       PIP_TARGET_DESC="user site-packages"
-      if [ "$DRY_RUN" = "1" ]; then dim "     would run: $PYTHON -m pip install --user $spec"; return; fi
-      "$PYTHON" -m pip install --quiet --user "$spec" || die "pip install --user $spec failed"
+      if [ "$DRY_RUN" = "1" ]; then
+        dim "     would run: $PYTHON -m pip install --user --break-system-packages $spec"
+        return
+      fi
+      "$PYTHON" -m pip install --quiet --user --break-system-packages "$spec" \
+        || die "pip install --user --break-system-packages $spec failed"
       ;;
     system)
       PIP_TARGET_DESC="system site-packages"
@@ -1390,6 +1466,53 @@ CHECKEOF
   done
 }
 
+# Index the models this machine already has.
+#
+# After install_package, because it runs `hypernix-t1 index` -- which is
+# the shipped indexer rather than a second implementation here. An
+# installer that reimplemented GGUF parsing in shell would be a second
+# place for the context limit to be wrong, and the context limit is the
+# number the server then enforces.
+index_models() {
+  [ "$INDEX_MODELS" = "1" ] || return 0
+
+  local models_dir="${HOME}/.hypernix/models"
+  local registry="${MODEL_REGISTRY_PATH:-$CONFIG_DIR/models.json}"
+
+  head2 "Indexing models"
+  if [ ! -d "$models_dir" ]; then
+    dim "     $models_dir does not exist yet — nothing to index."
+    dim "     Models downloaded through HyperLink land there; run"
+    dim "     'hypernix-t1 index' once you have some."
+    return 0
+  fi
+
+  if [ "$DRY_RUN" = "1" ]; then
+    dim "     would index $models_dir into $registry"
+    return 0
+  fi
+
+  if ! py_available; then
+    warn "hypernix is not importable yet, so the index was skipped."
+    warn "Run 'hypernix-t1 index' by hand once the install settles."
+    return 0
+  fi
+
+  local args="--dir $models_dir -o $registry"
+  [ "$ESTIMATE_PRICES" = "1" ] && args="$args --estimate-prices"
+
+  # Not fatal. A registry that could not be written is worth reporting
+  # loudly and is not a reason to leave a working server uninstalled --
+  # everything else has already been set up by this point.
+  # shellcheck disable=SC2086
+  if "$PYTHON" -m hypernix.t1api.modelindex_cli $args; then
+    note_written "$registry"
+  else
+    warn "Indexing failed. The server is installed and will serve nothing"
+    warn "until the registry has an entry; run 'hypernix-t1 index' to retry."
+  fi
+}
+
 summary() {
   head2 "Done"
   say ""
@@ -1422,6 +1545,13 @@ summary() {
   fi
   [ "$WANT_TUI" = "1" ] && say "    waiter tui              ${C_DIM}# the manager dashboard${C_OFF}"
   [ "$MODEL_SOURCE" = "lmstudio" ] && say "    waiter lmstudio status  ${C_DIM}# is a model loaded?${C_OFF}"
+  # The runner is the answer to "I have models and no LM Studio", which
+  # is the state this installer leaves a machine in when --index-models
+  # found some. Worth naming here rather than leaving it to be found.
+  if [ "$INDEX_MODELS" = "1" ]; then
+    say "    hypernix-t1 runner status  ${C_DIM}# what is serving, and where its layers are${C_OFF}"
+    say "    hypernix-t1 runner load <model>  ${C_DIM}# serve one without LM Studio${C_OFF}"
+  fi
   [ "$WANT_HYPERLINK" = "1" ] && say "    waiter hyperlink pair   ${C_DIM}# connect the phone app${C_OFF}"
   [ "$WANT_SYSTEMD" = "1" ] && say "    sudo cp $CONFIG_DIR/hypernix-t1api.service /etc/systemd/system/"
   say ""
@@ -1500,6 +1630,7 @@ main() {
 
   write_env
   write_registry_template
+  index_models
   write_start_script
   write_systemd_unit
   mint_admin_key

{hypernix-0.72.5.dev2 → hypernix-0.72.5.post1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "hypernix"
-version = "0.72.5.dev2"
+version = "0.72.5.post1"
 description = "Download and quantize the HyperNix PyTorch model to GGUF (fp32/fp16/Q8_0/Q6_K/Q4_K_M). and train models with a massive ai toolkit "
 readme = { file = "README.md", content-type = "text/markdown" }
 license = { text = "HOL-1.0" }

{hypernix-0.72.5.dev2 → hypernix-0.72.5.post1}/setup.cfg

@@ -1,6 +1,6 @@
 [metadata]
 name = hypernix
-version = 0.72.5.dev2
+version = 0.72.5.post1
 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-0.72.5.dev2 → hypernix-0.72.5.post1}/src/hypernix/__init__.py

@@ -63,7 +63,7 @@ import importlib.util
 import sys
 from typing import TYPE_CHECKING, Any
 
-__version__ = "0.72.5.dev2"
+__version__ = "0.72.5.post1"
 DEFAULT_REPO_ID = "ray0rf1re/hyper-Nix.2"
 DEFAULT_MODEL = "qwen3.5-4b"  # New default model