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.
Files changed (63) hide show
  1. localdeploy-0.6.0/.env.example +50 -0
  2. localdeploy-0.6.0/CHANGELOG.md +424 -0
  3. localdeploy-0.6.0/LICENSE +21 -0
  4. localdeploy-0.6.0/MANIFEST.in +8 -0
  5. localdeploy-0.6.0/PKG-INFO +202 -0
  6. localdeploy-0.6.0/README.md +154 -0
  7. localdeploy-0.6.0/SECURITY.md +48 -0
  8. localdeploy-0.6.0/api_server.py +15 -0
  9. localdeploy-0.6.0/benchmark.py +15 -0
  10. localdeploy-0.6.0/chat_cli.py +15 -0
  11. localdeploy-0.6.0/compare_models.py +15 -0
  12. localdeploy-0.6.0/config.example.json +371 -0
  13. localdeploy-0.6.0/localdeploy/__init__.py +3 -0
  14. localdeploy-0.6.0/localdeploy/backends/__init__.py +0 -0
  15. localdeploy-0.6.0/localdeploy/backends/llamacpp.py +143 -0
  16. localdeploy-0.6.0/localdeploy/backends/ollama.py +374 -0
  17. localdeploy-0.6.0/localdeploy/backends/openai_compatible.py +179 -0
  18. localdeploy-0.6.0/localdeploy/benchmark.py +1996 -0
  19. localdeploy-0.6.0/localdeploy/chat_cli.py +249 -0
  20. localdeploy-0.6.0/localdeploy/cli.py +59 -0
  21. localdeploy-0.6.0/localdeploy/compare_models.py +481 -0
  22. localdeploy-0.6.0/localdeploy/control/__init__.py +75 -0
  23. localdeploy-0.6.0/localdeploy/control/_config.py +180 -0
  24. localdeploy-0.6.0/localdeploy/control/_ollama.py +253 -0
  25. localdeploy-0.6.0/localdeploy/control/bakeoff.py +216 -0
  26. localdeploy-0.6.0/localdeploy/control/bench.py +516 -0
  27. localdeploy-0.6.0/localdeploy/control/calibration.py +191 -0
  28. localdeploy-0.6.0/localdeploy/control/community.py +161 -0
  29. localdeploy-0.6.0/localdeploy/control/fit.py +651 -0
  30. localdeploy-0.6.0/localdeploy/control/hardware.py +602 -0
  31. localdeploy-0.6.0/localdeploy/control/history.py +125 -0
  32. localdeploy-0.6.0/localdeploy/control/manifest.py +452 -0
  33. localdeploy-0.6.0/localdeploy/control/models.py +833 -0
  34. localdeploy-0.6.0/localdeploy/control/monitor.py +378 -0
  35. localdeploy-0.6.0/localdeploy/control/profiles.py +161 -0
  36. localdeploy-0.6.0/localdeploy/control/recommend.py +322 -0
  37. localdeploy-0.6.0/localdeploy/control/registry.py +956 -0
  38. localdeploy-0.6.0/localdeploy/control/report.py +407 -0
  39. localdeploy-0.6.0/localdeploy/control/starter.py +520 -0
  40. localdeploy-0.6.0/localdeploy/control/updates.py +118 -0
  41. localdeploy-0.6.0/localdeploy/grader_sandbox.py +268 -0
  42. localdeploy-0.6.0/localdeploy/server.py +1873 -0
  43. localdeploy-0.6.0/localdeploy/utils.py +155 -0
  44. localdeploy-0.6.0/localdeploy/web/favicon.png +0 -0
  45. localdeploy-0.6.0/localdeploy/web/index.html +891 -0
  46. localdeploy-0.6.0/localdeploy/web/js/app.js +63 -0
  47. localdeploy-0.6.0/localdeploy/web/js/benchmark-views.js +73 -0
  48. localdeploy-0.6.0/localdeploy/web/js/benchmark.js +1691 -0
  49. localdeploy-0.6.0/localdeploy/web/js/chat.js +779 -0
  50. localdeploy-0.6.0/localdeploy/web/js/models.js +2758 -0
  51. localdeploy-0.6.0/localdeploy/web/js/shared.js +454 -0
  52. localdeploy-0.6.0/localdeploy/web/js/system.js +626 -0
  53. localdeploy-0.6.0/localdeploy/web/logo.svg +22 -0
  54. localdeploy-0.6.0/localdeploy/web/styles.css +2032 -0
  55. localdeploy-0.6.0/localdeploy.egg-info/PKG-INFO +202 -0
  56. localdeploy-0.6.0/localdeploy.egg-info/SOURCES.txt +61 -0
  57. localdeploy-0.6.0/localdeploy.egg-info/dependency_links.txt +1 -0
  58. localdeploy-0.6.0/localdeploy.egg-info/entry_points.txt +2 -0
  59. localdeploy-0.6.0/localdeploy.egg-info/requires.txt +23 -0
  60. localdeploy-0.6.0/localdeploy.egg-info/top_level.txt +5 -0
  61. localdeploy-0.6.0/pyproject.toml +72 -0
  62. localdeploy-0.6.0/requirements.txt +9 -0
  63. 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.
@@ -0,0 +1,8 @@
1
+ include LICENSE README.md CHANGELOG.md SECURITY.md
2
+ include config.example.json .env.example
3
+ include requirements.txt
4
+ recursive-include localdeploy/web *
5
+ prune docs
6
+ prune tests
7
+ prune scripts
8
+ prune examples