localdeploy 0.6.0__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- localdeploy-0.6.0/.env.example +50 -0
- localdeploy-0.6.0/CHANGELOG.md +424 -0
- localdeploy-0.6.0/LICENSE +21 -0
- localdeploy-0.6.0/MANIFEST.in +8 -0
- localdeploy-0.6.0/PKG-INFO +202 -0
- localdeploy-0.6.0/README.md +154 -0
- localdeploy-0.6.0/SECURITY.md +48 -0
- localdeploy-0.6.0/api_server.py +15 -0
- localdeploy-0.6.0/benchmark.py +15 -0
- localdeploy-0.6.0/chat_cli.py +15 -0
- localdeploy-0.6.0/compare_models.py +15 -0
- localdeploy-0.6.0/config.example.json +371 -0
- localdeploy-0.6.0/localdeploy/__init__.py +3 -0
- localdeploy-0.6.0/localdeploy/backends/__init__.py +0 -0
- localdeploy-0.6.0/localdeploy/backends/llamacpp.py +143 -0
- localdeploy-0.6.0/localdeploy/backends/ollama.py +374 -0
- localdeploy-0.6.0/localdeploy/backends/openai_compatible.py +179 -0
- localdeploy-0.6.0/localdeploy/benchmark.py +1996 -0
- localdeploy-0.6.0/localdeploy/chat_cli.py +249 -0
- localdeploy-0.6.0/localdeploy/cli.py +59 -0
- localdeploy-0.6.0/localdeploy/compare_models.py +481 -0
- localdeploy-0.6.0/localdeploy/control/__init__.py +75 -0
- localdeploy-0.6.0/localdeploy/control/_config.py +180 -0
- localdeploy-0.6.0/localdeploy/control/_ollama.py +253 -0
- localdeploy-0.6.0/localdeploy/control/bakeoff.py +216 -0
- localdeploy-0.6.0/localdeploy/control/bench.py +516 -0
- localdeploy-0.6.0/localdeploy/control/calibration.py +191 -0
- localdeploy-0.6.0/localdeploy/control/community.py +161 -0
- localdeploy-0.6.0/localdeploy/control/fit.py +651 -0
- localdeploy-0.6.0/localdeploy/control/hardware.py +602 -0
- localdeploy-0.6.0/localdeploy/control/history.py +125 -0
- localdeploy-0.6.0/localdeploy/control/manifest.py +452 -0
- localdeploy-0.6.0/localdeploy/control/models.py +833 -0
- localdeploy-0.6.0/localdeploy/control/monitor.py +378 -0
- localdeploy-0.6.0/localdeploy/control/profiles.py +161 -0
- localdeploy-0.6.0/localdeploy/control/recommend.py +322 -0
- localdeploy-0.6.0/localdeploy/control/registry.py +956 -0
- localdeploy-0.6.0/localdeploy/control/report.py +407 -0
- localdeploy-0.6.0/localdeploy/control/starter.py +520 -0
- localdeploy-0.6.0/localdeploy/control/updates.py +118 -0
- localdeploy-0.6.0/localdeploy/grader_sandbox.py +268 -0
- localdeploy-0.6.0/localdeploy/server.py +1873 -0
- localdeploy-0.6.0/localdeploy/utils.py +155 -0
- localdeploy-0.6.0/localdeploy/web/favicon.png +0 -0
- localdeploy-0.6.0/localdeploy/web/index.html +891 -0
- localdeploy-0.6.0/localdeploy/web/js/app.js +63 -0
- localdeploy-0.6.0/localdeploy/web/js/benchmark-views.js +73 -0
- localdeploy-0.6.0/localdeploy/web/js/benchmark.js +1691 -0
- localdeploy-0.6.0/localdeploy/web/js/chat.js +779 -0
- localdeploy-0.6.0/localdeploy/web/js/models.js +2758 -0
- localdeploy-0.6.0/localdeploy/web/js/shared.js +454 -0
- localdeploy-0.6.0/localdeploy/web/js/system.js +626 -0
- localdeploy-0.6.0/localdeploy/web/logo.svg +22 -0
- localdeploy-0.6.0/localdeploy/web/styles.css +2032 -0
- localdeploy-0.6.0/localdeploy.egg-info/PKG-INFO +202 -0
- localdeploy-0.6.0/localdeploy.egg-info/SOURCES.txt +61 -0
- localdeploy-0.6.0/localdeploy.egg-info/dependency_links.txt +1 -0
- localdeploy-0.6.0/localdeploy.egg-info/entry_points.txt +2 -0
- localdeploy-0.6.0/localdeploy.egg-info/requires.txt +23 -0
- localdeploy-0.6.0/localdeploy.egg-info/top_level.txt +5 -0
- localdeploy-0.6.0/pyproject.toml +72 -0
- localdeploy-0.6.0/requirements.txt +9 -0
- localdeploy-0.6.0/setup.cfg +4 -0
|
@@ -0,0 +1,50 @@
|
|
|
1
|
+
OLLAMA_BASE_URL=http://127.0.0.1:11434
|
|
2
|
+
# LocalDeploy only talks to loopback backends. This also disables Ollama's
|
|
3
|
+
# separate cloud-model feature when LocalDeploy starts Ollama for you.
|
|
4
|
+
OLLAMA_NO_CLOUD=true
|
|
5
|
+
# The macOS/Linux launcher starts Ollama when it is installed but not reachable.
|
|
6
|
+
# Set this to false when another service manages Ollama. Windows also starts an
|
|
7
|
+
# installed Ollama service when needed and offers a separate stop switch.
|
|
8
|
+
START_OLLAMA=true
|
|
9
|
+
LLAMACPP_BASE_URL=http://127.0.0.1:8080
|
|
10
|
+
LMSTUDIO_BASE_URL=http://127.0.0.1:1234
|
|
11
|
+
DOCKER_MODEL_RUNNER_BASE_URL=http://127.0.0.1:12434
|
|
12
|
+
# VLLM_BASE_URL=http://127.0.0.1:8001
|
|
13
|
+
# OPENAI_COMPAT_BASE_URL=http://127.0.0.1:8001
|
|
14
|
+
LLAMACPP_SERVER_PATH=
|
|
15
|
+
LLAMACPP_STARTUP_TIMEOUT_SECONDS=180
|
|
16
|
+
REQUIRE_GPU_ONLY=false
|
|
17
|
+
DEFAULT_BACKEND=ollama
|
|
18
|
+
# Pin a specific profile as the default for /chat and the CLIs. Leave unset to
|
|
19
|
+
# use config.json's default_profile, which is set automatically on first pull.
|
|
20
|
+
# DEFAULT_MODEL_PROFILE=
|
|
21
|
+
CONFIG_PATH=config.json
|
|
22
|
+
API_HOST=127.0.0.1
|
|
23
|
+
API_PORT=8000
|
|
24
|
+
REQUEST_TIMEOUT_SECONDS=180
|
|
25
|
+
SLOW_RESPONSE_SECONDS=60
|
|
26
|
+
# Seconds to wait for a model to warm into memory on Start/Switch. Default scales
|
|
27
|
+
# with the device (longer for CPU offload, which is slow for large models). Set
|
|
28
|
+
# this to override both paths with a fixed value.
|
|
29
|
+
# OLLAMA_LOAD_TIMEOUT=600
|
|
30
|
+
# Upper bound (GB) on a direct-URL GGUF import (Get a model > Pull / Import).
|
|
31
|
+
# MAX_GGUF_IMPORT_GB=128
|
|
32
|
+
GLOBAL_MAX_PROMPT_CHARS=20000
|
|
33
|
+
GLOBAL_MAX_OUTPUT_TOKENS=2048
|
|
34
|
+
GLOBAL_MAX_IMAGES=8
|
|
35
|
+
GLOBAL_MAX_IMAGE_MB=10
|
|
36
|
+
ENABLE_LLAMA_CPP=false
|
|
37
|
+
ENABLE_CORS=false
|
|
38
|
+
ENABLE_WEB_UI=true
|
|
39
|
+
# OFFLINE=true blocks LocalDeploy's outbound internet calls (model search and
|
|
40
|
+
# update checks). Local backend calls to loopback are unaffected. Verify with:
|
|
41
|
+
# python scripts/egress_selftest.py
|
|
42
|
+
OFFLINE=false
|
|
43
|
+
# Optional local shared token. This is not internet-facing authentication: there
|
|
44
|
+
# is no TLS, per-user identity, tenant isolation, or rate limiting. When set, the
|
|
45
|
+
# HTTP API requires the token via the 'X-API-Token' header, 'Authorization:
|
|
46
|
+
# Bearer <token>', or '?token=<token>'. Open the UI once at /ui?token=<token>.
|
|
47
|
+
API_TOKEN=
|
|
48
|
+
# A non-loopback API_HOST always prints a startup warning, even with a token.
|
|
49
|
+
# This legacy switch hard-fails only when a non-loopback bind also has no token.
|
|
50
|
+
REQUIRE_TOKEN_ON_LAN=false
|
|
@@ -0,0 +1,424 @@
|
|
|
1
|
+
# Changelog
|
|
2
|
+
|
|
3
|
+
All notable changes to this project should be documented here.
|
|
4
|
+
|
|
5
|
+
## 0.6.0 - 2026-07-19
|
|
6
|
+
|
|
7
|
+
### ModelScope search and direct GGUF import
|
|
8
|
+
|
|
9
|
+
- Added ModelScope as a third model source alongside Ollama and Hugging Face in the unified
|
|
10
|
+
catalog search, pulled through Ollama's native `modelscope.cn/<repo>:<file>.gguf` support
|
|
11
|
+
(Ollama 0.30+) - no new pull mechanism needed. Curated each repo's exposed quant variants down
|
|
12
|
+
to a standard ladder (Q3/Q4/Q5/Q6/Q8/F16) instead of listing every file a repo ships (some
|
|
13
|
+
unsloth-style repos carry 20+ near-duplicate quant/IQ variants), rounded ModelScope's
|
|
14
|
+
full-precision parameter counts for display (`9.197093888B` to `9.2B`), and fixed a quant-label
|
|
15
|
+
regex gap that left period-separated filenames (`Model.Q4_K_M.gguf`) with no visible quant badge.
|
|
16
|
+
- Added GGUF import for models that aren't in any registry: a local file already on disk (becomes
|
|
17
|
+
a llama.cpp run profile, no download) or a direct URL such as a GitHub Release (downloaded,
|
|
18
|
+
registered with Ollama via `/api/create`, and given a run profile - so Deploy/Unload/Delete work
|
|
19
|
+
exactly like a pulled model).
|
|
20
|
+
|
|
21
|
+
### Fixed
|
|
22
|
+
|
|
23
|
+
- `chat.js`'s "Open model catalog" button (shown when no models are installed) called an
|
|
24
|
+
undefined function and did nothing; it now navigates correctly.
|
|
25
|
+
- Every tab switch re-ran hardware detection and an internet update-check on top of the model/
|
|
26
|
+
profile refresh already needed for correctness, and switching tabs quickly could stack up
|
|
27
|
+
several overlapping bursts of requests since navigation didn't wait for the previous switch's
|
|
28
|
+
refresh to finish. Tab navigation now only refreshes Ollama reachability; hardware/update checks
|
|
29
|
+
stay on their original triggers (startup, the explicit button, and actions that actually change
|
|
30
|
+
VRAM usage).
|
|
31
|
+
- `start.ps1` could report "LocalDeploy did not start" on a successful launch if no default browser
|
|
32
|
+
was configured, because opening the browser ran under the same fail-fast error handling as the
|
|
33
|
+
rest of the script.
|
|
34
|
+
- `start.ps1`/`stop.ps1` could leave a stale server running after `-Restart`: on a venv created from
|
|
35
|
+
a Conda base interpreter, this machine transparently relaunches a script under a different
|
|
36
|
+
interpreter as a child process, and the launcher only ever tracked the parent PID. Both scripts
|
|
37
|
+
now identify whoever is actually listening on the configured port and walk up its process-family
|
|
38
|
+
ancestry, so the whole chain is stopped regardless of which hop originally got recorded.
|
|
39
|
+
- `install.ps1` claimed it would `git pull` an existing install to update it but never actually did;
|
|
40
|
+
it now pulls (fast-forward only) when the checkout is clean, and skips with a clear warning
|
|
41
|
+
instead of overwriting anything if it isn't.
|
|
42
|
+
- The Docker LAN-exposure warning's closing line ("Bind API_HOST=127.0.0.1") was shown even inside
|
|
43
|
+
a container, where binding to loopback internally would break the port mapping; that line is now
|
|
44
|
+
Docker-aware, matching the reassurance already given earlier in the same message.
|
|
45
|
+
- Added `start.command` so macOS users can double-click to launch, the same way `start.bat` already
|
|
46
|
+
works on Windows (a plain `.sh` opens in a text editor on double-click instead of running).
|
|
47
|
+
- `start.sh` now offers to install Python and Ollama itself when either is missing (Homebrew on
|
|
48
|
+
macOS; apt, dnf, pacman, or zypper for Python and the official installer script for Ollama on
|
|
49
|
+
Linux), asking before running anything and never prompting on a non-interactive run - mirroring
|
|
50
|
+
the guided install `start.ps1` already offered via winget on Windows.
|
|
51
|
+
- The repository is now public; the previously-private repo made every documented install path
|
|
52
|
+
(the README one-liner, `install.ps1`'s own ZIP fallback, and anonymous `git clone`) 404 for
|
|
53
|
+
anyone without repository access.
|
|
54
|
+
|
|
55
|
+
### Desktop installers (experimental)
|
|
56
|
+
|
|
57
|
+
- Windows: `.\packaging\build.ps1 -Msi` produces a per-user `.msi` (no admin/UAC prompt), fetching
|
|
58
|
+
the WiX v3 toolset itself on first run. Verified end-to-end on a real machine: install, Start
|
|
59
|
+
Menu + Desktop shortcuts, a working Settings > Apps entry, and a fully clean uninstall.
|
|
60
|
+
- macOS: `packaging/macos.spec` builds a menu-bar-only `.app` (no Dock icon), and
|
|
61
|
+
`.github/workflows/build-macos.yml` packages it into a `.dmg` with an automated smoke test.
|
|
62
|
+
Authored without access to a Mac; the workflow is manual-trigger only until it - or the
|
|
63
|
+
resulting `.app` - has actually been run on real hardware.
|
|
64
|
+
- Neither installer is code-signed yet: Windows SmartScreen and macOS Gatekeeper will both warn on
|
|
65
|
+
first run. See `docs/PACKAGING.md` for what each warning means and the signing/notarization cost
|
|
66
|
+
and account requirements to remove them.
|
|
67
|
+
|
|
68
|
+
## 0.5.1 - 2026-07-18
|
|
69
|
+
|
|
70
|
+
### Public release audit
|
|
71
|
+
|
|
72
|
+
- Split the browser frontend into seven native ES modules and added import-graph, browser, package,
|
|
73
|
+
and static-route checks for the new layout.
|
|
74
|
+
- Fixed several audit findings in benchmark and model workflows: streamed backend failures now end
|
|
75
|
+
cleanly, benchmark queue failures no longer stay stuck as running, forced device comparisons load
|
|
76
|
+
one model at a time, imported report cards keep stable ids, category summaries include every
|
|
77
|
+
category, zero-valued limits reach validation, and single-profile reports retain backend data.
|
|
78
|
+
- Reworked the README and supporting guides in a shorter, factual voice. Added checks for local
|
|
79
|
+
documentation links and the repository's punctuation rule. Corrected two nonexistent model names
|
|
80
|
+
in the old reference table and removed recommendations that could not be kept current.
|
|
81
|
+
- Documented the first PyPI release setup. The v0.5.0 package build succeeded, but upload did not
|
|
82
|
+
because the pending trusted publisher had not been registered on PyPI.
|
|
83
|
+
- Added Ruff to local and CI checks, and pinned GitHub Actions and the Docker base image by digest.
|
|
84
|
+
- Disabled Ollama's separate cloud-model feature in the bundled Docker runtime and in Ollama
|
|
85
|
+
processes started by the Windows launcher. Clarified the boundary between LocalDeploy's offline
|
|
86
|
+
setting and the network settings of a separately managed runtime.
|
|
87
|
+
- Fixed the macOS/Linux launcher so API and Ollama addresses set only in `.env` are used for its
|
|
88
|
+
health checks, browser link, and server command.
|
|
89
|
+
- Added managed Ollama startup to the macOS/Linux launcher and a conservative `stop.sh` that only
|
|
90
|
+
stops processes launched by this checkout.
|
|
91
|
+
- Consolidated runtime dependency metadata around `requirements.txt`; development tools now live
|
|
92
|
+
only in the `.[dev]` package extra.
|
|
93
|
+
- Expanded the starter catalog with verified Qwen 3.5 local tags, added five data-only benchmark
|
|
94
|
+
questions, and added a safe `json_keys_present` grader for custom question sets.
|
|
95
|
+
- Updated client examples to use the configured default profile and optional API token. Marked shell
|
|
96
|
+
entry points executable for clones on macOS and Linux.
|
|
97
|
+
|
|
98
|
+
### Later on 0.5.1: guided recommendations, calibration, Monitor tab, reproducibility
|
|
99
|
+
|
|
100
|
+
- Recommended models is now a guided flow. Pick a use case, a priority (quality/speed/memory/
|
|
101
|
+
context), and an expected context size; `POST /registry/recommend` returns up to three labeled
|
|
102
|
+
picks (Recommended, Faster, Higher quality), each with a "why this model?" breakdown that tags
|
|
103
|
+
every reason as estimated, published, or measured on this machine without blending the three.
|
|
104
|
+
- Fit estimates now calibrate themselves. Every deploy compares the pre-load VRAM estimate
|
|
105
|
+
against what Ollama actually placed, and later fit checks for the same GPU/model/quant/context
|
|
106
|
+
are corrected from that history (shown as a visible correction, never applied silently). Fit
|
|
107
|
+
checks also report a confidence level and a new `/system/fit-table` sweeps several context sizes
|
|
108
|
+
in one call.
|
|
109
|
+
- New Monitor tab: live VRAM/CPU/RAM and per-model throughput/TTFT while a model serves, a
|
|
110
|
+
numeric-only request log (prompts and responses are never recorded), rule-based alerts (sustained
|
|
111
|
+
VRAM pressure, placement mismatches, slow generation, true concurrent calls), and persisted
|
|
112
|
+
session summaries. Calibration only consumes Ollama's model-specific VRAM allocation; system-wide
|
|
113
|
+
session peaks remain diagnostic so unrelated GPU use cannot corrupt fit estimates.
|
|
114
|
+
- Benchmarking: true time-to-first-token, P90/P95/median/min/max statistics on repeated runs, a
|
|
115
|
+
context-scaling sweep, use-case benchmark packs (coding/structured/reasoning/general), and
|
|
116
|
+
regression detection in Compare (what changed between two runs' runtime/digest/quant/GPU, plus
|
|
117
|
+
the resulting tok/s/TTFT/VRAM deltas).
|
|
118
|
+
- Deployment manifests: export a running model's exact configuration as YAML/JSON, validate it
|
|
119
|
+
against a different machine (compatibility report with safe substitutions), recreate it there, and
|
|
120
|
+
generate copy-paste config for Open WebUI, Continue, Cline, curl, Python, JavaScript, and Docker
|
|
121
|
+
Compose.
|
|
122
|
+
- Update check: an opt-out (`OFFLINE=true`), no-payload GitHub release check surfaces a chip
|
|
123
|
+
when a newer version exists. A local-only anonymized benchmark-sharing preview/export was added
|
|
124
|
+
as groundwork for community sharing: there is no server to submit to yet, so nothing is sent.
|
|
125
|
+
- Fixed a UI crash that broke hardware detection, catalog search, and other buttons. Disabling a
|
|
126
|
+
just-clicked button (to show a loading state) fires a `focusout` event; a tooltip handler
|
|
127
|
+
dereferenced its trigger without checking for null and crashed before the underlying request ever
|
|
128
|
+
ran. Also fixed: a race in `benchmark.py`'s lazy import that could return a partially-initialized
|
|
129
|
+
module under concurrent first requests; the Monitor tab's "simultaneous requests" alert now uses
|
|
130
|
+
real in-flight backend calls instead of a hardcoded or recent-completion proxy; Monitor's per-model state lookup
|
|
131
|
+
used exact string matching against Ollama's reported name instead of the same fuzzy matcher used
|
|
132
|
+
elsewhere, so a bare/tagless deploy could show blank uptime; and `compare_models.py` crashed
|
|
133
|
+
instead of printing a helpful message when no profile was configured anywhere.
|
|
134
|
+
- An automated "pull top candidates, benchmark each, keep the winner" workflow was built and then
|
|
135
|
+
disabled in the UI pending a product decision (backend intact at `/system/bakeoff/run`).
|
|
136
|
+
|
|
137
|
+
- The catalog answers "will it fit?" at a glance. Since your hardware is already
|
|
138
|
+
detected, every size chip in the results is colored by fit: green fits your GPU,
|
|
139
|
+
yellow is tight/CPU-only, and red won't fit. One batched estimate runs per search
|
|
140
|
+
(`POST /system/fit-batch`); Hugging Face rows get a row-level fit badge parsed from
|
|
141
|
+
the repo name. Descriptions clamp to two lines and capability tags sit inline, so
|
|
142
|
+
the model column stops being a wall of text.
|
|
143
|
+
- Pulling a specific quant is now one click on a real tag. The quant advisor
|
|
144
|
+
fetches the family's actual published tags from ollama.com
|
|
145
|
+
(`POST /registry/library-tags`), adds a "Pull it" column with the true download size
|
|
146
|
+
(e.g. `↓ 3.0GB` for `qwen2.5:7b-instruct-q5_K_M`), marks unpublished quants honestly
|
|
147
|
+
as "not published", flags already-pulled ones, and offers an expandable list of every
|
|
148
|
+
published tag: no more guessing tag-name conventions.
|
|
149
|
+
|
|
150
|
+
- One search, every source. The Model catalog no longer asks you to pick a source:
|
|
151
|
+
one query hits the Ollama library and Hugging Face GGUF repos in parallel
|
|
152
|
+
(`POST /registry/search-models`) and renders a single table where the source is just a
|
|
153
|
+
column. Results update as you type (debounced), size chips are per-tag fit-checked
|
|
154
|
+
pulls, exact tags (`gemma3:4b`, `hf.co/org/repo`) get a verbatim "Pull exactly" row,
|
|
155
|
+
and ⚖ on any row opens the quant advisor pre-filled.
|
|
156
|
+
- Chat upgrades: attach text files (contents ride along as fenced blocks the model
|
|
157
|
+
can read and render as code blocks in your bubble), a 5-minute keep-loaded option,
|
|
158
|
+
and `keep_alive` now rides every chat request so Ollama stops silently resetting the
|
|
159
|
+
session's keep-loaded choice to its 5-minute default after the first message.
|
|
160
|
+
Assistant replies render lightweight markdown (headings, bullets, bold/italic, inline
|
|
161
|
+
code, and links). Rendering is still escape-first, with no HTML injection possible.
|
|
162
|
+
- Buttons look like buttons everywhere: one uniform shape/height/alignment for all
|
|
163
|
+
buttons (emoji and icon labels now center properly), squared-off attach control.
|
|
164
|
+
|
|
165
|
+
### Earlier on 0.5.1 (2026-07-17)
|
|
166
|
+
|
|
167
|
+
UX overhaul across the whole UI.
|
|
168
|
+
|
|
169
|
+
- Model catalog is now searchable. The local provider inventory gets a search box,
|
|
170
|
+
runtime and size filters, a sort control (measured tok/s, name, parameters, runtime),
|
|
171
|
+
and pagination instead of one giant unfiltered table. Provider chips show
|
|
172
|
+
reachability with per-runtime "how to enable" hints, and a Discover section links
|
|
173
|
+
Hugging Face GGUF search and the Ollama library.
|
|
174
|
+
- Disk vs VRAM are visually distinct everywhere. Model rows show `💾 X GB disk`
|
|
175
|
+
(drive space) separately from `⚡ needs ~Y GB VRAM` (memory to run), with a legend on
|
|
176
|
+
the Your models card; the serving card labels "⚡ VRAM in use" vs "💾 Size on disk".
|
|
177
|
+
- Your models polish: fit checks still run automatically on load (the button is now a
|
|
178
|
+
small ↻ re-check), icon buttons (▶ Deploy, ⚙ Tune, 🗑 Delete), and the delete dialog
|
|
179
|
+
states exactly how many GB it frees.
|
|
180
|
+
- Use via API: every served model card shows the full OpenAI-compatible endpoint URL
|
|
181
|
+
with one-click ⧉ copy buttons for the URL and a ready-to-run curl (model pre-filled).
|
|
182
|
+
- Chat redesigned as a proper chat shell: header with model picker + system-prompt
|
|
183
|
+
panel, welcome screen with clickable suggestion prompts, role avatars, fenced code
|
|
184
|
+
blocks rendered with language tag and copy button, auto-growing composer, and
|
|
185
|
+
🕒 / ⚡ metadata under each reply.
|
|
186
|
+
- One-paste Windows install: `irm …/scripts/install.ps1 | iex` downloads the repo
|
|
187
|
+
(git or ZIP: nothing preinstalled) and hands off to the guided start script.
|
|
188
|
+
- Screenshot/GIF capture scripts now pick a chat model that is actually installed, so
|
|
189
|
+
they don't record a timeout when a hardcoded model was deleted.
|
|
190
|
+
|
|
191
|
+
## 0.5.0 - 2026-07-17
|
|
192
|
+
|
|
193
|
+
- Detect NVIDIA, AMD, Intel, and Apple GPUs and estimate compatible multi-GPU placement without combining mixed runtime pools.
|
|
194
|
+
- Add a dense local-provider model catalog for Ollama, llama.cpp, LM Studio, vLLM, Docker Model Runner, and configured loopback OpenAI-compatible runtimes, including saved benchmark tok/s.
|
|
195
|
+
- Record repeated-run variance and benchmark provenance: LocalDeploy/Ollama versions, full model digest, quant, context, initial warm/cold state, native backend metrics, and hardware snapshot.
|
|
196
|
+
- Add OpenAI-compatible tool calling, `/v1/responses`, and `/v1/embeddings` with float/base64 output. LocalDeploy returns tool calls but never executes them.
|
|
197
|
+
- Clarify the single-user local-only security boundary: one shared token, no TLS, no multi-user isolation, and no claim of internet-facing readiness.
|
|
198
|
+
- Project layout: `api_server.py`, `benchmark.py`, `chat_cli.py`, and `compare_models.py`
|
|
199
|
+
moved into the `localdeploy` package (`localdeploy/server.py` etc.); the root files remain
|
|
200
|
+
as alias shims (`sys.modules` aliasing), so `python api_server.py`, `uvicorn api_server:app`,
|
|
201
|
+
and every existing import keep working unchanged. `pytest.ini` folded into `pyproject.toml`.
|
|
202
|
+
- Clean-machine startup hardening: a double-clickable `start.bat` (execution-policy safe,
|
|
203
|
+
works from a plain ZIP download), dependencies re-install automatically when
|
|
204
|
+
`requirements.txt` changes after a `git pull`, `start.ps1` heals a stale API from an older
|
|
205
|
+
checkout instead of declaring it "already running" (plus a `-Restart` switch), Ollama startup
|
|
206
|
+
is polled rather than a fixed sleep, and `start.sh` now gives actionable guidance when
|
|
207
|
+
Python 3.10+/venv/Ollama are missing.
|
|
208
|
+
|
|
209
|
+
## 0.4.0 - 2026-07-17
|
|
210
|
+
|
|
211
|
+
Public-launch release: a chat playground, quant advisor, disk usage tools, durable
|
|
212
|
+
benchmark history, and streamlining so the app only ever shows models that are
|
|
213
|
+
actually on your machine.
|
|
214
|
+
|
|
215
|
+
### New features
|
|
216
|
+
|
|
217
|
+
- Chat playground tab. Talk to any pulled model right in the UI: streaming
|
|
218
|
+
tokens over the server's own OpenAI-compatible `/v1/chat/completions`, image
|
|
219
|
+
attachments for vision-capable profiles, an optional system prompt, and a
|
|
220
|
+
Send-becomes-Stop button. The reply meta line separates model-load time
|
|
221
|
+
("first token X s") from true generation speed (tok/s).
|
|
222
|
+
- Quantization advisor (Get a model → ⚖ Quant advisor). Fit-checks every
|
|
223
|
+
common GGUF quant (Q2_K → F16) of a model size against your VRAM budget with
|
|
224
|
+
the same estimator as fit checks, and tells you when there's headroom for a
|
|
225
|
+
higher-quality tag than the usual Q4 default (`POST /system/quant-advisor`).
|
|
226
|
+
- Disk usage panel. The Your models card now shows `N models · X GB on disk`,
|
|
227
|
+
sorts by size / recency / name, and bulk-deletes selected models with a
|
|
228
|
+
freed-gigabytes preview.
|
|
229
|
+
- Opt-in server-side benchmark history. A toggle on the History tile mirrors
|
|
230
|
+
completed runs to `reports/benchmark-history/` as one JSON file per run
|
|
231
|
+
(`/benchmark/history` endpoints), so results survive the browser and can be
|
|
232
|
+
shared as plain files. Run ids are strictly validated (no path traversal).
|
|
233
|
+
|
|
234
|
+
### No more phantom models
|
|
235
|
+
|
|
236
|
+
- Fresh installs start truly empty. `load_config()` no longer falls back to
|
|
237
|
+
`config.example.json` when `config.json` is missing (that fallback made a fresh
|
|
238
|
+
clone list a dozen sample profiles for models that were never pulled, most
|
|
239
|
+
visibly on macOS/Linux, where `start.sh` doesn't seed a config). A missing
|
|
240
|
+
config now yields zero profiles; pulling a model auto-creates its profile.
|
|
241
|
+
`config.example.json` remains as field-reference documentation and a test fixture.
|
|
242
|
+
- `.env.example` no longer pins `DEFAULT_MODEL_PROFILE`. The pinned sample value
|
|
243
|
+
made `/chat` fail with "Unknown profile" on a fresh install until that exact model
|
|
244
|
+
was pulled, even after other models were. The config's `default_profile`
|
|
245
|
+
(auto-set on first pull) is now used.
|
|
246
|
+
- Clear first-run error. With no profiles configured, `/chat` and friends now say
|
|
247
|
+
"pull a model first" instead of "Unknown profile 'gemma3_4b_ollama_safe'".
|
|
248
|
+
- UI hides profiles whose model isn't on the machine. The benchmark model picker
|
|
249
|
+
hides them by default behind a "Show N hidden (model not pulled)" toggle, the
|
|
250
|
+
profile dropdowns annotate them, and a new Remove not-pulled profiles button
|
|
251
|
+
(Advanced → All run profiles) deletes them in one click. For llama.cpp profiles the
|
|
252
|
+
server now reports whether the GGUF file exists (`model_file_exists` on `/profiles`),
|
|
253
|
+
so dead file paths get the same treatment as un-pulled Ollama models.
|
|
254
|
+
|
|
255
|
+
### Packaging
|
|
256
|
+
|
|
257
|
+
- Python package layout. The project builds as a standard Python package
|
|
258
|
+
(`pyproject.toml`) with a `localdeploy` console command that serves the API + UI
|
|
259
|
+
and opens the browser. The web UI ships as package data (`localdeploy/web/`,
|
|
260
|
+
moved from the repo-root `web/`).
|
|
261
|
+
- App home for runtime state. Installed runs keep `.env`, `config.json`,
|
|
262
|
+
`logs/`, and `reports/` in `~/.localdeploy` (override with `LOCALDEPLOY_HOME`);
|
|
263
|
+
source checkouts keep using the repo root, so nothing changes for `git clone`
|
|
264
|
+
users. Docker now mounts that state separately from Ollama's model volume, so
|
|
265
|
+
profiles and benchmark history survive container recreation. `/favicon.ico`
|
|
266
|
+
is now served, so `/docs` page views stop logging 404s.
|
|
267
|
+
|
|
268
|
+
### Fresh coat of paint
|
|
269
|
+
|
|
270
|
+
- New logo (`localdeploy/web/logo.svg`), regenerated favicon, and a repo social banner
|
|
271
|
+
(`docs/assets/banner.png`).
|
|
272
|
+
- README rebuilt around an animated demo GIF (`docs/assets/demo.gif`), captured by the
|
|
273
|
+
new `scripts/capture_demo_gif.py`; screenshots now ship in dark (default) and light
|
|
274
|
+
themes via `scripts/capture_screenshots.py`.
|
|
275
|
+
- The web asset cache-bust test no longer pins the exact `?v=` token, so bumping asset
|
|
276
|
+
versions can't silently break CI.
|
|
277
|
+
|
|
278
|
+
### Security and reliability
|
|
279
|
+
|
|
280
|
+
- Updated `python-multipart` to `0.0.31` after dependency auditing found three
|
|
281
|
+
advisories affecting `0.0.28`; CI now audits the pinned runtime requirements.
|
|
282
|
+
- Bundled benchmark and chat clients now forward `API_TOKEN`, so enabling auth
|
|
283
|
+
no longer breaks server-initiated benchmarks or local CLI calls.
|
|
284
|
+
- The code benchmark worker now validates candidate ASTs, exposes only a small
|
|
285
|
+
builtin/import allowlist, and blocks common filesystem, process, network,
|
|
286
|
+
registry, and native-library operations in addition to its existing timeout.
|
|
287
|
+
- Removed the browser-triggered `pip install psutil` endpoint. `psutil` is a
|
|
288
|
+
required dependency, and hardware probe failures now degrade to a read-only
|
|
289
|
+
status instead of modifying the running Python environment.
|
|
290
|
+
|
|
291
|
+
## 0.3.0 - 2026-07-08
|
|
292
|
+
|
|
293
|
+
First public release.
|
|
294
|
+
|
|
295
|
+
### Repo cleanup for public release
|
|
296
|
+
|
|
297
|
+
- One start command per platform. `scripts/start.ps1` (Windows) now opens the UI by default
|
|
298
|
+
(`-NoBrowser` to skip, `-Foreground` for live logs) and `scripts/start.sh` (macOS/Linux) opens
|
|
299
|
+
the UI once the server is healthy. Removed the redundant launchers: `run.ps1` / `run.sh`
|
|
300
|
+
(curl-pipe bootstrappers that auto-installed Docker), `install.ps1` (start.ps1 already creates
|
|
301
|
+
config/venv; model pulls belong in the fit-checked UI), and `scripts/start_ui.ps1` (now the
|
|
302
|
+
default behavior).
|
|
303
|
+
- Renamed the root model-comparison CLI `test_models.py` → `compare_models.py` (it was never a
|
|
304
|
+
pytest suite; the old name needed a pytest.ini workaround).
|
|
305
|
+
- Docs: `docs/TERMINAL_TESTING.md` → `docs/CLI.md`; removed the internal `GAPS.md` and
|
|
306
|
+
`docs/VERIFICATION.md` trackers; rewrote the README around a per-platform quick start.
|
|
307
|
+
- License switched from "all rights reserved" placeholder to MIT.
|
|
308
|
+
|
|
309
|
+
### Features and fixes
|
|
310
|
+
|
|
311
|
+
- Forced CPU/GPU benchmarks now measure the device they ask for. Previously a
|
|
312
|
+
`device=cpu` (or `gpu`) run only pinned the placement at warm-up; the benchmark's own
|
|
313
|
+
`/chat` calls didn't pass `num_gpu`, so Ollama could silently re-place the model (a CPU
|
|
314
|
+
run drifting onto the GPU). `num_gpu` is now threaded through `/chat` → `execute_test` →
|
|
315
|
+
`iter_run`, so every inference call stays on the requested device end-to-end. Verified:
|
|
316
|
+
`device=cpu` reports `actual_device=cpu`, `device=gpu` reports `gpu`.
|
|
317
|
+
- Fewer spurious "status failed" benchmark rows. Forcing CPU/GPU used to hard-fail
|
|
318
|
+
whenever Ollama placed the model differently (e.g. a model too big for pure GPU lands on
|
|
319
|
+
Split). It now warns and proceeds, recording the run with its actual placement.
|
|
320
|
+
- Benchmark history/queue management: per-run delete (×) in the run library, a confirm on
|
|
321
|
+
Clear history, and the ability to dismiss finished/failed queue rows (individually or via
|
|
322
|
+
Clear finished); failed rows show their error reason inline. The running-model Kill
|
|
323
|
+
button is now Unload for consistency.
|
|
324
|
+
- Browser UI smoke tests (`tests/test_ui_playwright.py`): optional Playwright-driven checks
|
|
325
|
+
that load the real `/ui` in headless Chromium and exercise tab switching and the benchmark
|
|
326
|
+
history controls. They skip cleanly when Playwright or its browser isn't installed; added
|
|
327
|
+
`playwright` to the development dependencies.
|
|
328
|
+
- Fixed: the web UI was completely broken: `web/app.js` contained smart/curly quotes
|
|
329
|
+
(`“ ”`) used as string delimiters in the "Check New Models" function, a `SyntaxError` that
|
|
330
|
+
prevented the *entire* script from parsing (blank/dead UI). Replaced with straight quotes.
|
|
331
|
+
Added a CI step (`node --check web/app.js`) and a pytest guard so a JS parse error can never
|
|
332
|
+
ship silently again (the Python-only suite never loaded the UI before).
|
|
333
|
+
- Benchmark workspace V2: the benchmark tab is now a local experiment workspace with a
|
|
334
|
+
multi-profile Run Builder, sequential run queue, leaderboard, category heatmap, SVG
|
|
335
|
+
speed/quality scatter, per-test matrix, local browser history, selected-run comparison, and
|
|
336
|
+
report-card import/export. CPU + GPU now creates two queued benchmark batches instead of a
|
|
337
|
+
separate special-case button.
|
|
338
|
+
- Benchmark queue UX: waiting benchmark rows can now be moved up/down or removed before they
|
|
339
|
+
run, while the active run appears only in the main progress panel with elapsed time.
|
|
340
|
+
- Comparison auto-select: fresh benchmark results now join the existing selected comparison
|
|
341
|
+
set instead of replacing it, so a second run appears next to the first run automatically.
|
|
342
|
+
- Per-test matrix demoted: the per-test matrix is now collapsed as an advanced diagnostic
|
|
343
|
+
view instead of occupying the main results dashboard by default.
|
|
344
|
+
- Benchmark deployments are temporary: `/benchmark/run` now unloads each Ollama model after
|
|
345
|
+
its profile finishes, including forced CPU/GPU benchmark deployments. The benchmark tab no
|
|
346
|
+
longer leaves a model served as a side effect.
|
|
347
|
+
- Quieter device auto-detect (inline note instead of a per-run toast); the run progress bar is
|
|
348
|
+
now an ARIA `progressbar` so screen readers announce progress.
|
|
349
|
+
|
|
350
|
+
- Honest device tag: the benchmark Device tag now defaults to Auto (detect) and is
|
|
351
|
+
resolved from the model's *actual* placement (GPU/CPU/Split via `/system/status`) after the
|
|
352
|
+
run, so report cards can't be silently mislabelled. Manual GPU/CPU stays an override and warns
|
|
353
|
+
if it disagrees with what's detected.
|
|
354
|
+
- Apple Silicon (Metal) detection: `/system/hardware` now reports Apple Silicon GPUs instead
|
|
355
|
+
of claiming "CPU-only" on a Mac. Unified memory is surfaced as such (no false VRAM figure; fit
|
|
356
|
+
checks use system RAM). NVIDIA still takes precedence where present.
|
|
357
|
+
- Pull fit-gate respects the tiered warnings: a model that won't fit VRAM but runs on CPU
|
|
358
|
+
is no longer blocked behind the override: the gate now hard-blocks only the "fits nowhere"
|
|
359
|
+
(`severity: hard`) case and notes the CPU fallback for the soft case.
|
|
360
|
+
- Tune for my GPU labels skipped CPU-capable profiles as "CPU-only (skipped for GPU tuning)"
|
|
361
|
+
instead of the misleading "won't fit VRAM".
|
|
362
|
+
- Run table: detailed benchmark rows now live below the dashboard and can be filtered by
|
|
363
|
+
model, category, pass/fail, and slowest failures.
|
|
364
|
+
- Warm-up robustness: Deploy/replace actions now show a live "Loading…Ns" counter (with a
|
|
365
|
+
"large models on CPU can take a minute" hint when targeting CPU) instead of an apparently
|
|
366
|
+
frozen button. The server's load timeout scales with the device (longer for CPU offload)
|
|
367
|
+
and is overridable via `OLLAMA_LOAD_TIMEOUT`; a load timeout returns a clear "it may still be
|
|
368
|
+
loading: click Refresh status" message rather than a generic failure.
|
|
369
|
+
- Decision-grade run results: the benchmark workspace shows queued model/device variants,
|
|
370
|
+
live progress, Cancel, tok/s per test, inline failure reasons, response previews,
|
|
371
|
+
leaderboard ranking, per-category heatmap, and per-test matrix views.
|
|
372
|
+
- Report cards & compare carry speed: exported cards now include `avg_tokens_per_second`,
|
|
373
|
+
a per-test tok/s column, and a By category table. The A/B compare view adds a
|
|
374
|
+
tok/s (A → B) column and aggregate tok/s delta, so a "Qwen/GPU vs Qwen/CPU" diff
|
|
375
|
+
shows the speed difference directly.
|
|
376
|
+
- Better discovery (Phase 5): "New on Hugging Face" now has a search box, results-per-query
|
|
377
|
+
limit, and GGUF-only toggle: the API already accepted `queries`/`limit`/`gguf_only`, the UI
|
|
378
|
+
now surfaces them. Installed rows show quantization level, parameter size, and modified date.
|
|
379
|
+
HF candidates show download counts and likes alongside the modified date.
|
|
380
|
+
- CPU-vs-GPU benchmark comparison (Phase 6): the benchmark tab gains a *Device tag*
|
|
381
|
+
(Auto / GPU / CPU) selector. The tag is stored in the exported report card (`device` field) and
|
|
382
|
+
surfaces in the card header (`[GPU]`/`[CPU]` next to the model name in HTML and Markdown). The
|
|
383
|
+
compare view labels runs as `model/GPU` vs `model/CPU` so A/B diffs are unambiguous.
|
|
384
|
+
- Tiered fit warnings (`/system/fit-check`): adds `tier`/`severity`/`headline`/`cpu_deployable`
|
|
385
|
+
: green (comfortable), yellow (tight, or won't-fit-GPU-but-runs-on-CPU), red (too big anywhere) -
|
|
386
|
+
using system RAM to judge CPU deployability. The coarse `verdict` is unchanged (backward-compatible).
|
|
387
|
+
- Model management: `POST /models/delete` (remove a model from disk) and `POST /models/free`
|
|
388
|
+
(unload all models from memory/VRAM), with Delete buttons per installed model, a Free memory
|
|
389
|
+
button, and a Cancel button to abort an in-flight pull.
|
|
390
|
+
- Hardware panel now shows CPU + RAM (`/system/hardware`): CPU model, physical/logical cores,
|
|
391
|
+
and system RAM total/available via `psutil` (graceful fallback when absent).
|
|
392
|
+
- Choose CPU vs GPU at deploy time: the Serve panel has a *Deploy to* selector (Auto / GPU /
|
|
393
|
+
CPU). Forces Ollama `num_gpu` (0 = CPU, max = GPU); Auto is unchanged from before. `/system/status`
|
|
394
|
+
now labels each running model's placement (GPU / CPU / Split N%). Additive and backwards-compatible.
|
|
395
|
+
- Added opt-in token auth: set `API_TOKEN` to require it (`X-API-Token` /
|
|
396
|
+
`Authorization: Bearer` / `?token=`); no auth and zero overhead when unset.
|
|
397
|
+
- "Check New Models" now filters to GGUF repos and offers a one-click Pull via Ollama's
|
|
398
|
+
`hf.co/<id>` shortcut.
|
|
399
|
+
- Added an optional web UI at `/ui` (two tabs: Serve & Diagnose, Deploy & Benchmark);
|
|
400
|
+
static, no build step, gated by `ENABLE_WEB_UI` (default on). See `docs/UI.md`.
|
|
401
|
+
- Added a control-plane API: `/system/hardware`, `/system/fit-check`, `/system/status`,
|
|
402
|
+
`/system/recommend`, `/system/set-default`, `/registry/installed`, `/registry/check-updates`,
|
|
403
|
+
`/models/pull`, `/models/serve`, `/models/stop`, `/models/switch`.
|
|
404
|
+
- Added benchmark-over-HTTP: `/benchmark/example`, `/benchmark/validate`, `/benchmark/run`
|
|
405
|
+
(streamed), plus shareable report cards (`/benchmark/export`) and A/B compare
|
|
406
|
+
(`/benchmark/compare`). `benchmark.py` gained an importable `execute_test`/`iter_run` and a
|
|
407
|
+
JSON-safe grader registry shared by the CLI and the API.
|
|
408
|
+
- Added one-command Docker run (`Dockerfile`, `docker-compose.yml`) bundling Ollama + the API/UI,
|
|
409
|
+
plus a `scripts/start.sh` launcher.
|
|
410
|
+
- Added a verifiable offline mode (`OFFLINE=true`) and `scripts/egress_selftest.py`.
|
|
411
|
+
- The original CLI, OpenAI-compatible API, and loopback-only backend guard are unchanged.
|
|
412
|
+
|
|
413
|
+
## 0.2.0 - 2026-05-16
|
|
414
|
+
|
|
415
|
+
- Added OpenAI-compatible `/v1/chat/completions` and `/v1/models` endpoints for local clients.
|
|
416
|
+
- Added GitHub project hygiene files, CI validation, issue templates, and contribution/security docs.
|
|
417
|
+
- Added terminal chat helper and local benchmark documentation.
|
|
418
|
+
|
|
419
|
+
## 0.1.0 - 2026-05-16
|
|
420
|
+
|
|
421
|
+
- Initial local Ollama-first deployment server.
|
|
422
|
+
- Added optional llama.cpp profile support.
|
|
423
|
+
- Added safe model profiles for Gemma 3 4B and Gemma 3 12B testing.
|
|
424
|
+
- Added benchmark runner and Windows install helper.
|
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 iodriller and LocalDeploy contributors
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
+
furnished to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|