mlx-bun 0.2.0 → 0.4.0

package/docs/reference/cli.md

@@ -1,119 +1,222 @@
 # CLI reference
 
-Every mlx-bun verb. Commands are shown as `mlx-bun <verb>`; from a clone the
-identical command is `bun src/cli.ts <verb>`. `mlx-bun help <verb>` (or
-`mlx-bun <verb> --help`) prints the full flag list for any command.
-
-Model arguments are **substring queries** against the registry (`e4b`, `26B`,
-`12B-it`); a query matching more than one model errors out and lists the
-candidates — just make it more specific.
-
-Bare `mlx-bun` with no verb runs `serve` — the appliance path: first run
-downloads a starter model, serves it, and opens the chat UI.
-
-> All verbs below ship in current releases; the release notes in
-> `docs/planning/` record when each verb first appeared.
-
-## Serving & chat
+Every mlx-bun verb, from `src/cli.ts`. Commands are shown as
+`mlx-bun <verb>`; from a clone the identical command is
+`bun src/cli.ts <verb>`. `mlx-bun help <verb>` (or `mlx-bun <verb> --help`)
+prints the full flag list for any command; `mlx-bun --version` (also `-v` or
+`mlx-bun version`) prints the version.
+
+**Model arguments are substring queries** against the registry (`e4b`,
+`12B`, `qwen`, a repo substring); a query that matches more than one model
+errors and lists the candidates. A directory containing `config.json` is
+loaded directly wherever a model is accepted (`serve`, `pi`, `benchmark`,
+`embed`, `fuse`, `convert`, `perplexity`).
+
+**Bare `mlx-bun`** (no verb, or only options such as `mlx-bun --port 9000`)
+runs `serve`. Explicit verbs and `--help`/`--version` still win.
+
+**Unknown flags are ignored**, not rejected — `mlx-bun serve --kv-bits 4`
+starts a server as if the flag were absent. The exceptions are `fuse` and
+`convert`, which refuse the mlx-lm flags they don't implement (listed under
+each verb).
+
+## Verb index
+
+| Verb | Does | mlx-lm counterpart |
+|---|---|---|
+| `serve` | OpenAI/Anthropic-compatible server + web app | `mlx_lm.server` |
+| `pi` | built-in coding agent on a local model | — (`mlx_lm.chat` is the nearest) |
+| `harness pi` | connect your own pi install to the local server | — |
+| `generate` (`gen`) | one-shot generation, no server | `mlx_lm.generate` |
+| `embed` | text embeddings, no server | — |
+| `get` | download a model (resumable, verified) | `mlx_lm.manage` (partial) |
+| `scan` | re-index the Hugging Face cache | `mlx_lm.manage --scan` |
+| `ls` | list downloaded models | `mlx_lm.manage --scan` |
+| `gc` | reclaim superseded snapshots + dead blobs | `mlx_lm.manage --delete` (different unit) |
+| `fit` | will it fit? memory + speed assessment | — (`optiq latency`) |
+| `train` | LoRA fine-tune (SFT / DPO / ORPO) | `mlx_lm.lora --train` |
+| `train-watch` | live dashboard for a training run | — |
+| `fuse` | merge an adapter into the base weights | `mlx_lm.fuse` |
+| `convert` | quantize an HF model into a local MLX snapshot | `mlx_lm.convert` (+ `mlx_lm.dynamic_quant`, `optiq convert`) |
+| `upload` | push a model directory to the Hub | `mlx_lm.upload` |
+| `perplexity` | perplexity over a local dataset | `mlx_lm.perplexity` |
+| `memory` (`setup`) | the assistant's personal wiki | — |
+| `benchmark` (`bench`) | decode/prefill speed of our stack | `mlx_lm.benchmark` |
+| `evals` | recorded benchmark runs | — |
+| `help` | overview or per-verb help | — |
+
+## Serving & agents
 
 ### `serve` — run the server
 
-Start the OpenAI/Anthropic-compatible server plus the browser web app (chat,
-status, and model tools).
+Start the OpenAI/Anthropic-compatible server plus the browser web app.
 
 ```sh
-mlx-bun serve                              # default model, http://localhost:8080
-mlx-bun serve e4b                          # pick a model by substring
-mlx-bun serve gemma --memory-budget 18     # ...with admission control (GB)
-mlx-bun serve e4b --no-open                # don't open the browser chat UI
-mlx-bun serve GLM-5.2 --context-length 4096 # bounded Colibri runtime, native MTP
+mlx-bun serve                       # default model
+mlx-bun serve e4b                   # pick a model by substring
+mlx-bun serve --model ./fused_model # explicit path (mlx_lm.server spelling)
+mlx-bun serve e4b --no-open         # don't open the chat UI
+mlx-bun serve GLM-5.2 --context-length 4096
 ```
 
-Common flags (full list in [server-config.md](server-config.md)):
-
-| Flag | Effect |
+`serve` takes the model as a positional, `--query <q>`, or `--model <path|query>`
+(`--model` wins). The flag set below is **shared with `mlx-bun pi`**. Defaults,
+semantics, and which combinations compose live in
+**[server-config.md](server-config.md)** — this page only names them.
+
+Library requests support delayed uniform or per-layer affine conversion with
+qualified shared drafting providers, including rotating caches. CLI affine
+quantization starts at zero.
+
+The unreleased shared executor supports qualified Qwen MTP and start-zero
+TurboQuant at one or several active rows; `--batch 1` still selects the
+legacy serial control. No new CLI flag is required for the shared interfaces.
+Full-attention and rotating targets, including Llama, MiniCPM and Gemma, use shared bf16, uniform KV4/KV8, per-layer affine KV or TurboQuant
+prompt lookup and standalone drafting, including generated RAM/SSD state.
+Gemma assistant drafting also shares these target layouts and caches its last true target
+hidden alongside target KV through the same RAM/SSD interfaces.
+DeepSpec and DSpark use the same executor and persist their projected context through those
+interfaces. DeepSpec accepts bf16 or affine-quantized weights; DSpark retains
+its Markov/RNN head and checkpoint confidence policy.
+
+Shared Qwen MTP retains completed decode state through the existing
+`--prompt-cache` and `--ssd-cache` settings.
+
+| Group | Flags |
 |---|---|
-| `--host <addr>` | Interface to bind (default `127.0.0.1`, loopback only; `0.0.0.0` to expose) |
-| `--port <n>` | Listen port (default 8080) |
-| `--memory-budget <GB>` | Admission ceiling for requests (bills the active `--kv-quant` scheme; clamps `max_tokens`, 400s only prompts that leave no room — never refuses to start) |
-| `--kv-budget <GB>` | Aggregate KV budget across concurrent batch rows: over-budget joiners queue until rows finish; a request over the budget alone is rejected (off unless set) |
-| `--prompt-cache <GB>` | RAM prefix-KV cache, byte-capped LRU (default 8; `0` disables). Non-consuming prefix sharing: a hit clones instead of consuming the donor entry |
-| `--isolate` | Run the inference engine as a crash-isolated child process on a unix socket; this process becomes a pure reverse-proxy (instant under GPU load, auto-respawns on engine crash) — default off ([runtime-isolation.md](../design/runtime-isolation.md)) |
-| `--model-pool <n>` | With `--isolate`: cap on resident per-model engine children (default 1); over the cap the LRU engine drains its cache to the SSD tier and exits, respawning on demand |
-| `--ssd-cache <dir>` | SSD cold tier under the prompt cache: idle/evicted prefix-KV spills to disk and restores on reattach, surviving restarts (default off — [ssd-kv-cold-tier.md](../design/ssd-kv-cold-tier.md)) |
-| `--ssd-cache-max <GB>` | SSD tier byte cap, oldest-mtime evicted first (default 32; `--ssd-cache` only) |
-| `--ssd-cache-verify` | Verify every tensor hash on restore, not just the header (integrity paranoia; roughly doubles restore reads; `--ssd-cache` only) |
-| `--ssd-demote-idle <sec>` | Idle prompt-cache entries spill to the SSD tier and free GPU memory after this long (default 300 with `--ssd-cache`; `0` disables) |
-| `--kv-quant config\|off\|4\|8\|turbo[:k<bits>v<bits>]` | KV cache quantization: per-layer `kv_config.json`, bf16, uniform bits, or TurboQuant (rotation-based, default `k8v3` — [turboquant-kv.md](../design/turboquant-kv.md); full-attention layers only, solo-only/serial) (default `off` — quantized KV trades 5–20% decode speed for memory headroom, so it's opt-in) |
-| `--paged-kv` | OPTIONAL vLLM-style paged KV cache (block pool + gather before the stock SDPA; env `MLX_BUN_PAGED_KV=1`) — bit-exact with the plain path, default off. v1: serial only (pins `--batch 1` unless `--batch` is given; refuses `--batch N>1`, `--kv-quant`, `--draft-model`), Gemma4-family, bf16; bypasses the prompt cache and runs uncompiled decode. Expect a small decode cost at batch=1 — v1 ships the abstraction for batched/CoW follow-ups ([paged-kv-cache.md](../design/paged-kv-cache.md)) |
-| `--paged-kv-block-size <n>` | Tokens per KV block (default 256 = the plain cache's growth step; `--paged-kv` only) |
-| `--adapter <dir>` | Mount a LoRA adapter at startup (`--adapter-path` accepted as the mlx_lm.server alias) |
-| `--draft-model <path\|query>` | Speculative decoding: a drafter proposes, the main model verifies — exact results, faster decode when drafts land. The artifact's kind is auto-detected: a full same-tokenizer model (mlx_lm.server parity), a Gemma `-assistant` KV-borrowing drafter, a DSpark checkpoint trained here, or one of DeepSeek's released DSpark drafters (DeepSpec `Gemma4DSparkModel`, e.g. `dspark_gemma4_12b_block7`). Serial lane only: with `--batch N` a mounted draft routes every request serial |
-| `--draft-kind <kind>` | Override draft-artifact detection: `two-model` \| `assistant` \| `dspark` \| `deepspec` \| `mtp` \| `ngram` (default: auto). `mtp` is a NATIVE multi-token-prediction head split from the target's own release (e.g. `mlx-community/Qwen3.8-27B-MTP-bf16`; auto-detected from a `*_mtp` `model_type`) — it binds the target's embeddings/lm-head and taps its pre-final-norm hiddens, and the serve loop rolls the target's recurrent DeltaNet caches back across rejected drafts via bit-exact snapshot/replay. `--draft-kind mtp` WITHOUT `--draft-model` resolves the companion from the artifact's own `mtp/` subfolder (single-repo bundles, e.g. the TQ releases) and refuses with a clear error when the model bundles none. `ngram` is MODEL-FREE prompt lookup — no `--draft-model` (mounting one alongside it is refused): drafts are copied from the request's own prompt+generation whenever its trailing tokens have occurred earlier in that stream, verified by the target like any other draft, so results stay exact at any temperature. Zero weights, zero memory; best on agentic/RAG/code-edit traffic that re-emits spans already in context (prior art: prompt-lookup decoding / vLLM's `ngram` proposer) |
-| `--num-draft-tokens <n>` | Drafts per verify round (default 3; `ngram` defaults to 10 — drafting is free; a DSpark draft pins to its trained block width) |
-| `--ngram-max <k>` / `--ngram-min <k>` | Prompt-lookup match-window bounds, `--draft-kind ngram` only: the longest/shortest trailing k-gram searched (defaults 3 / 1, the reference values) |
-| `--mtp on\|off` | GLM-5.2 checkpoint-native MTP (default on). Native MTP uses the exact serial verify lane; `off` enables ordinary continuous batching (batched per-row MTP is post-release) |
-| `--context-length <n>` | GLM-5.2 context reserved by the header-only resource equation (default 4096) |
-| `--batch <n>` | Continuous-scheduler concurrency cap, mlx-lm parity at the same B (default 8; a lone admitted request uses the gated B=1 fast path, while `--batch 1` pins the preserved strict serial executor) |
-| `--temperature` / `--top-p` / `--top-k` / `--max-tokens` | Server-wide sampling defaults (per-request fields still win; GLM-5.2 defaults to the planned 128 generated tokens) |
-| `--l1` / `--l2` | Parity tier alias: bit-exact to mlx-lm / bit-exact to mlx-optiq. **No tier = `--l1`** (the default since 2026-07-05 — output-changing levers are opt-in until they beat the L1 baseline in a paired A/B). Each expands to per-fork flags (`--compiled-decode`, `--compiled-activations`, `--fused-sdpa`, `--kv-quant`); a fork flag overrides one. `--l3` was removed 2026-07-05 and now errors (the Lab replaces it — [unified-engine-frontier-plan.md](../design/unified-engine-frontier-plan.md)). See [server-config.md](server-config.md#fidelity-tiers-and-the-decode-route---l1----l2). |
-| `--allow-private-media` | Let `image_url`/`audio_url` parts fetch from private/loopback/link-local hosts (blocked by default — SSRF guard; the 10 s timeout and 64 MB cap on remote media apply either way) |
-| `--no-open` | Don't auto-open the chat UI |
+| Network | `--host`, `--port`, `--no-open`, `--allow-private-media` |
+| Memory & caches | `--memory-budget`, `--kv-budget`, `--prompt-cache`, `--ssd-cache`, `--ssd-cache-max`, `--ssd-cache-verify`, `--ssd-demote-idle`, `--generation-checkpoint` |
+| Process model | `--isolate`, `--model-pool`, `--unix` (internal — the engine half of `--isolate`) |
+| Scheduling | `--batch` (`--decode-concurrency` accepted as the mlx_lm.server alias) |
+| KV cache | `--kv-quant`, `--paged-kv`, `--paged-kv-block-size` |
+| Adapters (compatible sets can batch) | `--adapter` (`--adapter-path` alias) |
+| Speculative decoding (Qwen methods and full-attention/rotating lookup/standalone and Gemma assistant/DeepSpec/DSpark drafting share execution) | `--draft-model`, `--draft-kind`, `--num-draft-tokens`, `--ngram-max`, `--ngram-min` |
+| GLM-5.2 | `--mtp`, `--context-length` |
+| Sampling defaults | `--temperature` (`--temp` alias), `--top-p`, `--top-k`, `--max-tokens`, `--thinking`, `--hlg-sampling`, `--hlg-width`, `--hlg-shoulder`, `--hlg-toe`, `--hlg-pivot-offset` |
+| Parity tier | `--l1`, `--l2` (`--l3` errors — removed) |
+| Kill switches | `--compiled-decode`, `--compiled-activations`, `--fused-sdpa`, `--force-wire`, `--expert-offload` |
+
+Serving treats memory estimates as advisory by default and attempts the request.
+`--memory-budget` opts into estimated admission limits; `--kv-budget` opts
+into aggregate KV limits. `--max-tokens` supplies an explicit output cap.
 
 Endpoints: `/v1/chat/completions`, `/v1/completions`, `/v1/messages`,
-`/v1/responses`, `/v1/models`, `/v1/adapters`, `/health`, `/stats`, `/fit`,
-`/library`, `/downloads` — web app at `/` (lands on chat); `/chat`, `/status`,
-`/quantize`, `/finetune`, `/dataset` deep-link its views. Full
-schemas in [server-api.md](server-api.md).
+`/v1/responses`, `/v1/embeddings`, `/v1/models`, `/v1/adapters`, `/health`,
+`/stats`, `/fit`, `/library`, `/downloads`; the web app at `/` (`/chat`,
+`/status`, `/quantize`, `/finetune`, `/dataset` deep-link its views). Schemas
+in [server-api.md](server-api.md).
 
 ### `pi` — built-in coding agent
 
 Run the pi coding agent's terminal UI in-process against a local model —
-nothing to install. Reuses a healthy running server, otherwise starts one for
-the session.
+pi is bundled, nothing to install. Reuses a healthy server on the chosen
+port; otherwise starts one that lives exactly as long as the session.
 
 ```sh
-mlx-bun pi                          # interactive agent session
-mlx-bun pi -p "explain this repo"   # one-shot: print the reply and exit
+mlx-bun pi                               # interactive session
+mlx-bun pi fix the failing test          # pre-fill the first turn
+mlx-bun pi -p "explain this repo"        # one-shot: print the reply and exit
+mlx-bun pi -p --json "summarize README"  # one-shot with the full event stream
+mlx-bun pi --mode rpc                    # JSONL RPC over stdio
+mlx-bun pi -q qwen --port 8090           # model + server flags for the session server
+git diff | mlx-bun pi -p "review this"   # piped stdin makes it one-shot too
 ```
 
+| Flag | Meaning |
+|---|---|
+| `-q`, `--query <q>` | model to serve when starting a session server |
+| `-p`, `--print [msg]` | one-shot print mode (also reads piped stdin) |
+| `--json` | with `-p`: emit the full event stream as JSON |
+| `--mode json\|rpc\|text` | `json` = print + JSON, `rpc` = JSONL RPC over stdio, `text` = print |
+| `--verbose` | verbose agent logging |
+| every `serve` flag | applies to the session server; ignored (with a note) when reusing a running one |
+
+Other pi flags are dropped with a note — the full pi surface (`--continue`,
+`--resume`, extensions, themes) lives in your own pi via `harness pi`.
+`mlx-bun help pi` is our help; `mlx-bun pi --help` prints it only when it is
+the sole argument.
+
 ### `harness pi` — connect your own pi
 
-For people who already use
-[pi](https://www.npmjs.com/package/@earendil-works/pi-coding-agent): installs a small,
-reversible discovery extension that registers the local server as a pi
-provider.
+For people who already run [pi](https://www.npmjs.com/package/@earendil-works/pi-coding-agent):
+writes one reversible extension, `~/.pi/agent/extensions/mlx-bun-provider.ts`,
+that registers `mlx-bun` as a pi provider. Models are discovered live from
+`/v1/models` at pi startup (the list baked at install time is the fallback
+when the server is down), exposed under the stable id `mlx-bun/local`. It
+never touches your existing pi config.
 
 ```sh
-mlx-bun harness pi              # then: pi --provider mlx-bun
-mlx-bun harness pi --remove     # disconnect
+mlx-bun harness pi                                  # install; requires pi on PATH
+mlx-bun harness pi --base-url http://127.0.0.1:8090/v1
+mlx-bun harness pi --remove                         # delete the extension
+pi --provider mlx-bun                               # then run pi as usual
+pi --models "mlx-bun/*"                             # or scope cycling
 ```
 
+#### LoRA adapters from pi: the adapter extension
+
+`scripts/packaging/pi-extensions/mlx-bun-adapter.ts` ships with the repo and
+adds an `/adapter` command to pi. Install by copying (or symlinking) it into
+`~/.pi/agent/extensions/`, then `/reload` inside pi:
+
+```sh
+cp scripts/packaging/pi-extensions/mlx-bun-adapter.ts ~/.pi/agent/extensions/
+MLX_BUN_URL=http://127.0.0.1:8080 pi --provider mlx-bun
+```
+
+| In pi | Effect |
+|---|---|
+| `/adapter` or `/adapter list` | list adapters on disk (`/v1/adapters/available`), which are loaded, and the active one |
+| `/adapter <id>` | mount it if needed (`POST /v1/adapters`) and inject `adapter: <id>` into every request |
+| `/adapter off` | back to the base model (no field sent) |
+
+The extension reads the server URL from `MLX_BUN_URL` (its built-in fallback
+is port 8090, so set it to match your server). Default is no adapter; the
+mechanism is pi's `before_provider_request` hook, the same one the web chat's
+adapter selector uses. Serving-side detail: [server-api.md](server-api.md#adapters-lora-hot-swap)
+and [docs/design/web-chat-redesign.md](../design/web-chat-redesign.md).
+
 ### `generate` (alias `gen`) — one-shot generation
 
-Load the model, generate, print, exit — no server. Renders through the chat
-template by default; `--raw` feeds the prompt verbatim. Sampling and
-decode-path levers mirror `serve` (`--temperature`, `--seed`, `--l1/--l2`,
-`--kv-quant`, …).
+Load the model, generate, print, exit. The prompt is rendered through the
+chat template unless `--raw`.
 
 ```sh
-mlx-bun generate gemma "a haiku about metal shaders"
+mlx-bun generate e4b "a haiku about metal shaders"
 mlx-bun gen e4b --prompt "…" --max-tokens 512 --seed 42
-mlx-bun generate GLM-5.2 --prompt "Explain DSA" --max-tokens 128
+mlx-bun generate qwen --raw --prompt "def fib(n):" --temperature 0
 ```
 
+| Flag | Meaning |
+|---|---|
+| `[query]` / `--query <q>` | model (auto-picks when omitted) |
+| `--prompt "…"` | the prompt (or pass it as the positional after the model) |
+| `--raw` | skip the chat template |
+| `--max-tokens <n>` | completion cap |
+| `--temperature <n>` (`--temp` alias), `--top-p`, `--top-k`, `--seed` | sampling |
+| `--l1` / `--l2`, `--kv-quant`, `--fused-sdpa`, `--compiled-decode`, `--compiled-activations` | decode route, same meaning as for `serve` |
+
 ### `embed` — text embeddings
 
-One-shot embeddings from a local Qwen3-Embedding model (last-token pooled,
-L2-normalized — bit-exact vs mlx-lm). One JSON vector per input line.
+Embeddings from a local Qwen3-Embedding model, no server. One JSON vector
+per input line.
 
 ```sh
 mlx-bun embed --text "hello world"
-echo -e "line one\nline two" | mlx-bun embed --json
+mlx-bun embed Qwen3-Embedding "hello world"
+printf 'line one\nline two\n' | mlx-bun embed --json
+mlx-bun embed --instruct "Retrieve passages that answer the question" --text "what is a KV cache?"
 ```
 
-For a server endpoint, `mlx-bun serve <embedding-model>` then
-`POST /v1/embeddings`.
+| Flag | Meaning |
+|---|---|
+| `[query]` / `--query <q>` | embedding model; omitted = the first downloaded embedding model (errors with a `get` suggestion when none) |
+| `--text "…"` | text to embed (omit to read one text per stdin line) |
+| `--instruct "…"` | Qwen3-Embedding query instruction; omit for plain document embeddings |
+| `--json` | one OpenAI-style `{object:"list",data:[…]}` object instead of bare arrays |
+
+Server equivalent: `mlx-bun serve <embedding-model>` then
+[`POST /v1/embeddings`](server-api.md#post-v1embeddings-openai-embeddings-api).
 
 ## Models & disk
 
@@ -121,26 +224,26 @@ Full guide: [models.md](models.md).
 
 ### `get` — download a model
 
-Resumable, checksum-verified download into the standard Hugging Face cache
-(plain HTTPS, no Xet — no 0% stalls). Before payload transfer, the downloader
-credits complete blobs and partial prefixes, then requires enough free space
-for the exact remainder plus a fixed 1 GiB reserve.
+Resumable (Range requests against partial blobs), sha256-verified download
+into the standard Hugging Face cache over plain HTTPS (no Xet). Complete
+blobs and partial prefixes are credited before transfer, and the volume must
+fit the remainder plus a 1 GiB reserve. Uses `HF_TOKEN` / `hf auth login`
+credentials when present.
 
 ```sh
 mlx-bun get mlx-community/gemma-4-12B-it-OptiQ-4bit
-mlx-bun get 12B                  # no "/" = registry query; re-gets/refreshes the match
-mlx-bun get mateogrgic/GLM-5.2-colibri-int4-with-int8-mtp \
-  --revision 3cc8db99b1b13fc79325d987ba3c1c430766b3b8
+mlx-bun get 12B                          # no "/" = registry query; re-gets/refreshes the match
+mlx-bun get mlx-community/MiniCPM5-1B-OptiQ-4bit --revision main
 ```
 
-The GLM artifact is roughly 357 GiB; select `HF_HUB_CACHE` on a sufficiently
-large fast volume before starting. Rerunning the same command resumes verified
-`.incomplete` prefixes. See the [32 GB GLM quickstart](models.md#glm-52-on-a-32-gb-mac).
+`--revision <rev>` selects a git revision. When upstream has pushed a new
+revision the previous snapshot stays on disk — `get` says so and `gc`
+reclaims it. GLM-5.2 specifics: [models.md](models.md#glm-52-on-a-32-gb-mac).
 
 ### `scan` — index your cache
 
-Re-index the HF cache into the registry (config + safetensors headers only,
-never tensor bytes) so `ls`, `serve`, and `fit` can find models by substring.
+Re-index the HF cache into the registry (config + safetensors headers, never
+tensor bytes) so `ls`, `serve`, and `fit` resolve models by substring.
 
 ```sh
 mlx-bun scan
@@ -148,105 +251,139 @@ mlx-bun scan
 
 ### `ls` — list models
 
+One row per repo (canonical revision, `refs/main`); superseded snapshots are
+hidden.
+
 ```sh
-mlx-bun ls                          # size, params, quant, capabilities
-mlx-bun ls --vision --max-size 10GB # filter
-mlx-bun ls --all-revisions          # one row per cached snapshot
+mlx-bun ls                          # size, params, quant, license, capabilities
+mlx-bun ls gemma                    # filter by substring
+mlx-bun ls --vision --max-size 10GB # vision-capable, under 10 GB
+mlx-bun ls --all-revisions          # one row per cached snapshot (canonical marked *)
 ```
 
 ### `gc` — reclaim disk
 
-Delete superseded snapshots and the blobs only they reference. Prints the plan
-by default; `--yes` actually deletes. Same planner (`src/registry.ts`
-`planGc`/`executeGc`) also backs `GET /api/gc/plan` / `POST /api/gc/execute`
-on a running `serve` — see [server-api.md](server-api.md#get-apigcplan--post-apigcexecute).
+Delete superseded snapshots and the blobs only they reference. Prints the
+plan by default.
 
 ```sh
-mlx-bun gc          # dry-run: what would be reclaimed
-mlx-bun gc --yes    # delete
+mlx-bun gc              # dry run: what would be reclaimed
+mlx-bun gc --yes        # delete
+mlx-bun gc --yes --force  # also prune snapshots holding files the canonical revision lacks
 ```
 
+`--dry-run` never deletes even with `--yes`. The same planner backs
+[`GET /api/gc/plan` / `POST /api/gc/execute`](server-api.md#get-apigcplan--post-apigcexecute).
+
 ### `fit` — memory contract
 
-Deterministic memory assessment: does it fit, what's the max context, predicted
-tok/s. Weights bytes from safetensors headers, KV bytes/token from the config,
-calibrated prefill transient, wired-memory ceiling. `--kv-quant` mirrors serve's
-flag: it bills the quantized cache at its true bytes/element, so the reported
-window matches what serving with the same flag admits (a 4-bit cache holds
-~3.5× the context of bf16 in the same budget).
+Deterministic memory assessment: does it fit, max safe context, predicted
+tok/s — from safetensors headers, the config's KV bytes/token, a calibrated
+prefill transient, and the wired-memory ceiling.
 
 ```sh
-mlx-bun fit gemma --ctx 32768          # for this machine
+mlx-bun fit gemma --ctx 32768          # this machine
 mlx-bun fit gemma --ctx 8192 --skus    # across the Apple Silicon lineup
-mlx-bun fit qwen --kv-quant 4          # window with a 4-bit-quantized KV cache
+mlx-bun fit qwen --kv-quant 4          # bill a 4-bit-quantized KV cache
+mlx-bun fit GLM-5.2                    # streamed-artifact plan (resident weights, expert slabs, KV, reserves)
 ```
 
-The direct Colibri GLM-5.2 artifact is handled by its streamed-runtime
-contract, not by the generic resident-weights estimator. `fit GLM-5.2`
-defaults to the supported 4,096-token / 128-generation-token plan and reports
-the artifact's disk footprint separately from resident weights, the main and
-MTP expert slabs, KV, reserves, and the 25 GiB process ceiling. Its throughput
-line is the measured warm quality-preserving result; the direct Colibri oracle
-and the 2 tok/s aspiration are labeled separately. `--skus` explains that the
-generic resident-model matrix does not apply to this streamed artifact.
-
-```sh
-mlx-bun fit GLM-5.2
-mlx-bun fit GLM-5.2 --ctx 4096
-```
+| Flag | Meaning |
+|---|---|
+| `--ctx <tokens>` | context size to assess |
+| `--kv-quant 4\|8\|config\|off` | bill the KV cache at the quantized bytes/element (`config` needs a model that ships `kv_config.json`; `turbo` is not accepted here) |
+| `--skus` | also print the Apple silicon SKU matrix (not applicable to the streamed GLM-5.2 artifact) |
 
 ## Training & model creation
 
 Full reference: [training.md](training.md); guided walkthrough:
-[orpo-quickstart.md](orpo-quickstart.md).
+[docs/reference/training.md](./training.md).
 
 ### `train` — LoRA fine-tune
 
-SFT / DPO / ORPO on your data, with the full memory stack (flash-CCE head,
-prefix-sharing, segmented backward) on by default. Saves a mountable adapter.
+SFT / DPO / ORPO on `<dir>/train.jsonl` (+ optional `valid.jsonl`). Rows are
+`{prompt, chosen, rejected}` for dpo/orpo or `{messages|text}` for sft. Runs
+in the foreground and saves a mountable adapter.
 
 ```sh
-mlx-bun train e4b --data ./prefs                     # ORPO (default) on {prompt,chosen,rejected}
-mlx-bun train e4b --data ./prefs --method sft        # SFT on {messages|text}
-mlx-bun train e4b --data ./prefs --dry-run           # inspect the plan, don't train
+mlx-bun train e4b --data ./prefs                     # ORPO (default)
+mlx-bun train e4b --data ./prefs --method sft
+mlx-bun train e4b --data ./prefs --dry-run           # inspect data + resolved plan, don't train
+mlx-bun train cpm5 --data ./prefs --iters 400 --save-every 50 --adapter ~/adapters/cpm5-orpo
+nohup mlx-bun train e4b --data ./prefs &             # long runs: detach from your shell
 ```
 
+| Flag | Meaning |
+|---|---|
+| `<model>` / `--query <q>` | model to fine-tune (auto-picks when omitted) |
+| `--data <dir>` | dataset dir (required) |
+| `--method sft\|dpo\|orpo` | objective |
+| `--adapter <dir>` | output adapter dir (default under `~/.cache/mlx-bun/mlx-bun-finetunes/`) |
+| `--iters`, `--lr`, `--rank`, `--scale`, `--seq`, `--batch`, `--grad-accum`, `--grad-clip`, `--seed`, `--val-size` | loop hyperparameters (method-dependent defaults: `mlx-bun help train`) |
+| `--lambda <f>`, `--sft-scope full\|response` | ORPO odds-ratio weight and chosen-NLL scope |
+| `--seg <n>`, `--no-segment` | segmented backward (layers per segment / off) |
+| `--no-flash`, `--no-prefix` | disable the flash-CCE head / prefix-sharing |
+| `--save-every <n>`, `--resume <dir>` | checkpoint cadence / warm-start LoRA weights |
+| `--dry-run` | print the plan, don't train |
+
 ### `train-watch` — live training dashboard
 
-Tails an adapter dir's `metrics.jsonl` and renders live loss / speed / memory.
+Tails `<adapter-dir>/metrics.jsonl` (written by `train` and the web
+fine-tune tile) and renders live loss / speed / memory.
 
 ```sh
 mlx-bun train-watch ~/.cache/mlx-bun/mlx-bun-finetunes/orpo-e4b
+mlx-bun train-watch --adapter ~/adapters/cpm5-orpo
 ```
 
 ### `fuse` — merge an adapter into the base
 
-Folds LoRA deltas into the base weights and writes a standalone snapshot; a
-quantized base keeps its exact quantization layout. Flag names match
-`mlx_lm.fuse`.
+Folds LoRA deltas into the base weights and writes a standalone snapshot
+(weights + config + tokenizer); a quantized base keeps its per-module
+quantization layout. Flag names follow `mlx_lm.fuse`.
 
 ```sh
 mlx-bun fuse e4b --adapter ./adapters --save-path ./fused_model
+mlx-bun fuse --model ./base-snapshot --adapter-path ./adapters
 mlx-bun serve ./fused_model
 ```
 
+| Flag | Meaning |
+|---|---|
+| `<model>` / `--model` | base model (query or snapshot path) |
+| `--adapter <dir>` (`--adapter-path` alias) | adapter directory (`adapters.safetensors` + `adapter_config.json`); default `adapters` |
+| `--save-path <dir>` | output directory; default `fused_model` |
+
+Refused with an error (mlx_lm.fuse flags not implemented): `--de-quantize`,
+`--dequantize`, `--export-gguf`, `--gguf-path`, `--upload-repo`.
+
 ### `convert` — quantize a model
 
-Quantize an HF model (local path, downloaded model, or repo id) into a local
-MLX snapshot — uniform 4/8-bit, or mixed-precision via `--target-bpw`
-(sensitivity sweep + per-layer bit knapsack). `--upload-repo` pushes the result.
+Quantize an HF model (local path, downloaded model, or repo id — an
+un-downloaded repo is fetched first) into a local MLX snapshot: uniform
+affine 4/8-bit, or mixed precision via `--target-bpw` (sensitivity sweep +
+per-layer bit knapsack). Same engine as the web `/api/quantize`.
 
 ```sh
-mlx-bun convert Qwen/Qwen3-4B -q                      # uniform 4-bit
-mlx-bun convert Qwen/Qwen3-4B --target-bpw 4.5        # mixed precision
-mlx-bun convert ./qwen3.5-bf16 -q --rotate-weights    # fold then quantize
+mlx-bun convert Qwen/Qwen3-4B -q                              # uniform 4-bit → ./mlx_model
+mlx-bun convert --hf-path Qwen/Qwen3-4B -q --q-bits 8 --mlx-path ./qwen3-8bit
+mlx-bun convert Qwen/Qwen3-4B --target-bpw 4.5                # mixed precision (implies -q)
+mlx-bun convert ./qwen3.5-bf16 -q --rotate-weights            # fold the TurboQuant rotation, then quantize
+mlx-bun convert Qwen/Qwen3-4B -q --upload-repo you/qwen3-4b-4bit
 ```
 
-`--rotate-weights` makes the offline rotation part of the same conversion
-pipeline and records its recipe in `optiq_metadata.json`. The adapter is
-selected from the model/tensor schema (Llama, Qwen3.5 trunk, or Qwen MTP);
-`--rotation-seed` defaults to `42`. Folding requires full-precision source
-weights and an MTP companion must use the same seed as its trunk.
+| Flag | Meaning |
+|---|---|
+| `--hf-path <src>` (`--model` or the bare positional) | source model |
+| `--mlx-path <dir>` | output directory; must not already exist (mlx_lm.convert behavior) |
+| `-q`, `--quantize`, `--q-bits 4\|8`, `--q-group-size 32\|64` | uniform affine quantization |
+| `--target-bpw <f>`, `--candidate-bits <list>`, `--calibration-mix optiq\|<jsonl>`, `--n-calibration <n>` | mixed precision (writes per-module bits into `config.json`) |
+| `--rotate-weights`, `--rotation-seed <n>` | fold the offline TurboQuant rotation first (auto-detects Llama / Qwen3.5 trunk / Qwen MTP; an MTP companion must use its trunk's seed); recipe recorded in `optiq_metadata.json` |
+| `--upload-repo <id>` | push the result afterwards (token checked before any work starts) |
+
+Refused with an error: `--dtype`, `-d`/`--dequantize`, `--quant-predicate`,
+`--q-mode` other than `affine`, and plain non-quantizing conversion (pass `-q`
+or `--target-bpw`).
 
 ### `upload` — push to the Hugging Face Hub
 
@@ -255,57 +392,194 @@ git-LFS batch protocol, commits. Flag names match `mlx_lm.upload`.
 
 ```sh
 mlx-bun upload --path ./fused_model --upload-repo you/my-model
+mlx-bun upload --path ./mlx_model --upload-repo you/my-model --private
 ```
 
+Needs a WRITE token from `~/.mlx-bun/hf.json` (web UI Settings), `$HF_TOKEN`,
+or `~/.cache/huggingface/token` (`hf auth login`); errors before uploading
+anything when none is found. `--private` is an mlx-bun extension.
+
 ### `perplexity` — evaluate on your data
 
-`mlx_lm.perplexity` methodology exactly, over a **local** `.jsonl`/`.txt` file.
+`mlx_lm.perplexity` methodology exactly (seeded sample order, concatenate,
+non-overlapping rows, f32 CE at every position, delta-method standard
+error) over a **local** `.jsonl` (`{"text": …}` rows) or `.txt` file — never
+an HF dataset download.
 
 ```sh
 mlx-bun perplexity e4b --data-path ./corpus.jsonl
+mlx-bun perplexity ./fused_model --data-path ./corpus.txt --sequence-length 1024 --num-samples -1 --batch-size 4 --seed 7
 ```
 
+Flags: `<model>` (`--model` / `--query` accepted), `--data-path <file>`
+(required), `--sequence-length <n>`, `--num-samples <n>` (`-1` = all),
+`--batch-size <n>`, `--seed <n>`.
+
 ## Memory
 
-### `memory` — the personal wiki
+### `memory` (alias `setup`) — the personal wiki
 
-A local, durable memory for the assistant: a git-tracked Markdown wiki at
-`~/.mlx-bun/wiki` that loads into every `mlx-bun pi` / `mlx-bun serve` session.
-Full guide: [memory.md](memory.md).
+A local, git-tracked Markdown wiki at `~/.mlx-bun/wiki` that loads into
+every `mlx-bun pi` / `mlx-bun serve` session. Full guide:
+[memory.md](memory.md). `mlx-bun memory` with no subcommand runs `status`.
 
 ```sh
-mlx-bun memory init              # create the wiki (idempotent); import + schedule
-mlx-bun memory status            # path, article count, git + schedule state
-mlx-bun memory open [article]    # open in Obsidian (falls back to Finder)
-mlx-bun memory search <query>    # search articles from the terminal
-mlx-bun memory synthesize        # conversations → cross-linked articles, all local
-mlx-bun memory schedule --at 03:00
+mlx-bun memory init                       # create the wiki (idempotent); offers import + nightly job
+mlx-bun memory status                     # path, article count, git + schedule state
+mlx-bun memory open                       # open the vault in Obsidian (falls back to Finder)
+mlx-bun memory open Archie_Project        # open one article
+mlx-bun memory list                       # article titles + read-only Reference docs
+mlx-bun memory search kv cache            # search from the terminal
+mlx-bun memory read Archie_Project        # print an article (stem)
+mlx-bun memory toc Archie_Project         # headings + anchors
+mlx-bun memory section Archie_Project decisions   # one section by anchor
+mlx-bun memory links Archie_Project       # outbound + inbound wikilinks
+mlx-bun memory synthesize                 # conversations → articles, the full local DAG
+mlx-bun memory synthesize --dry-run       # plan only, never writes the vault
+mlx-bun memory segment --limit 20         # one stage worker (also: extract, route, synthesize-stage)
+mlx-bun memory extract --convs a1b2,c3d4  # restrict a stage to specific conversations
+mlx-bun memory link --limit 50            # deterministic cross-linking pass (no model)
+mlx-bun memory schedule --at 03:00        # install the nightly launchd job
+mlx-bun memory unschedule                 # remove it
 ```
 
-`setup` is an alias for `memory`.
+| Subcommand | Aliases | Meaning |
+|---|---|---|
+| `init` | `setup` | create the wiki; offers to import an existing vault and install the nightly job (TTY prompts; non-interactive shells take the defaults) |
+| `status` | | path, counts, git + schedule state |
+| `open [article]` | `browse` | open the vault or one article |
+| `list` | | article titles + `Reference/` docs |
+| `search <query>` | | ranked article hits with sample lines |
+| `read <article>` | | print an article |
+| `toc <article>` | | headings + anchors |
+| `section <article> <anchor>` | | print one section |
+| `links <article>` | | resolved outbound + inbound wikilinks |
+| `synthesize` | `pipeline`, `all` | run the full synthesis DAG; `--dry-run` plans without writing. `--since` and `--model` are parsed but the pipeline does not consume them yet (`src/memory/pipeline.ts` marks `model` reserved) |
+| `segment` · `extract` · `route` · `synthesize-stage` | `stage-synthesize` | run one decomposed, resumable stage worker (`--limit <n>` except `route`; `--convs a,b`) |
+| `link` | | inline-link first mentions + rebuild `## See also` (`--limit <n>`; no model) |
+| `schedule` | | install the nightly launchd job (`--at HH:MM`) |
+| `unschedule` | | remove it |
+
+An unknown subcommand prints the help and exits non-zero, so launchd jobs
+and scripts can detect typos.
 
 ## Measurement
 
 ### `benchmark` (alias `bench`) — measure our stack
 
-Runs mlx-bun by itself on this machine and records to the eval DB.
+Runs mlx-bun by itself (greedy decode, in-process) and records the median
+to the eval DB.
 
 ```sh
 mlx-bun bench e4b --tokens 256 --runs 3
+mlx-bun bench --model ./fused_model --prompt-tokens 4096 --kv-quant turbo:k8v3
+```
+
+| Flag | Meaning |
+|---|---|
+| `[query]` / `--query` / `--model <path\|query>` | model (auto-picks when omitted) |
+| `--tokens <n>` | tokens decoded per run |
+| `--runs <n>` | runs (median reported) |
+| `--prompt-tokens <n>` | pad the prompt to about n tokens (long-context decode) |
+| `--kv-quant config\|off\|4\|8\|turbo[:k<bits>v<bits>]` | KV scheme for the run |
+| `--compiled-decode`, `--compiled-activations`, `--fused-sdpa` | kill switches; A/B by running twice |
+
+For **quotable, cross-stack** numbers use the preflight-gated harness from
+the repo — it spawns the real CLI at its real defaults next to the mlx-lm
+and optiq reference servers and writes a dated report:
+
+```sh
+bun scripts/bench-serve.ts all
+bun scripts/bench-serve.ts all --models cpm5,e4b --arms mlx-bun,mlx-lm --out report.md
+bun scripts/bench-serve.ts all --models qwen27b --arms mlx-bun-serial,mlx-bun --logprobs --top-logprobs 3
+# Same KV settings on both executors
+bun scripts/bench-serve.ts all --models qwen27b --arms mlx-bun-serial,mlx-bun --kv-quant 4
+# Same adapter settings on both execution arms:
+bun scripts/bench-serve.ts all --models e4b --arms mlx-bun-serial,mlx-bun --adapter fixtures/adapters/upper
 ```
 
-For quotable **cross-stack** numbers use `./benchmark.sh` from the repo — it
-preflight-gates on an idle machine and runs the mlx-lm/optiq comparison legs,
-writing `benchmarks-h2h-<date>.md`.
+Curated results: [benchmarks.md](benchmarks.md).
 
 ### `evals` — recorded runs
 
 ```sh
-mlx-bun evals                # table: when, model, KV mode, tok/s, TTFT, peak, commit
+mlx-bun evals                # table: when, stack, model, bench, KV, prefill, decode, TTFT, peak, commit
+mlx-bun evals --limit 50
 mlx-bun evals --raw          # JSON lines for scripts/jq
+mlx-bun evals --clear        # archive the DB to a timestamped backup and start fresh
 ```
 
-## Everything else
-
-- `help [verb]` — overview, or per-verb help (same as `<verb> --help`).
-- `--version` / `-v` — print the version.
+## mlx-lm compatibility
+
+mlx-bun borrows mlx-lm's verb and flag vocabulary so scripts drop in with a
+rename of the executable: `mlx_lm.<verb> …` becomes `mlx-bun <verb> …`.
+There is no `mlx-bun.<verb>` dotted alias (the package installs one binary,
+`mlx-bun`). Surface parity is not architecture parity: a verb run against an
+unsupported architecture errors clearly rather than guessing.
+
+### What maps to what
+
+| mlx-lm | mlx-bun | Accepted mlx-lm spellings | Not accepted |
+|---|---|---|---|
+| `mlx_lm.server` | `serve` | `--model`, `--host`, `--port`, `--adapter-path`, `--temp`, `--top-p`, `--top-k`, `--max-tokens`, `--draft-model`, `--num-draft-tokens`, `--decode-concurrency` (semantics differ — [server-config.md](server-config.md)) | `--prompt-cache-size`/`--prompt-cache-bytes` (use `--prompt-cache`), `--min-p`, `--chat-template`, `--chat-template-args`, `--use-default-chat-template`, `--trust-remote-code`, `--log-level`, `--allowed-origins`, `--pipeline`, `--prefill-step-size`, `--prompt-concurrency` — silently ignored |
+| `mlx_lm.generate` | `generate` | `--prompt`, `--max-tokens`, `--temp`, `--top-p`, `--top-k`, `--seed` | `--kv-bits`/`--kv-group-size`/`--quantized-kv-start` (use `--kv-quant`), `--min-p`, `--xtc-*`, `--prompt-cache-file`, `--max-kv-size`, `--adapter-path`, `--draft-model`, `--system-prompt`, `--ignore-chat-template` (use `--raw`), `--verbose` — silently ignored |
+| `mlx_lm.chat` | `pi` / the web chat | — | no terminal REPL verb; `pi` is the agent, `serve` opens the chat UI |
+| `mlx_lm.convert` | `convert` | `--hf-path`, `--mlx-path`, `-q`/`--quantize`, `--q-bits`, `--q-group-size`, `--q-mode affine`, `--upload-repo` | `--dtype`, `-d`/`--dequantize`, `--quant-predicate`, non-affine `--q-mode`, plain non-quantizing conversion — **refused** |
+| `mlx_lm.fuse` | `fuse` | `--model`, `--adapter-path`, `--save-path` | `--de-quantize`, `--export-gguf`, `--gguf-path`, `--upload-repo` — **refused** |
+| `mlx_lm.lora --train` | `train` | `--data`, `--iters`, `--seed` (same names); otherwise own names (`--method`, `--adapter`, `--lr`, `--rank`, `--seq`, `--batch`, `--grad-accum`, `--save-every`, `--resume`) | `--train`/`--test`, `--config` (YAML), `--fine-tune-type`, `--adapter-path`, `--learning-rate`, `--max-seq-length`, `--batch-size`, `--num-layers`, `--mask-prompt`, `--optimizer` — silently ignored |
+| `mlx_lm.upload` | `upload` | `--path`, `--upload-repo` (+ `--private`) | — |
+| `mlx_lm.perplexity` | `perplexity` | `--model`, `--sequence-length`, `--num-samples`, `--batch-size`, `--seed`; `--data-path` is a local file | HF dataset ids |
+| `mlx_lm.benchmark` | `benchmark` | `--model` | `-p`/`--prompt-tokens` short form, `-g`/`--generation-tokens`, `-b`/`--batch-size`, `-n`/`--num-trials` (ours: `--prompt-tokens`, `--tokens`, `--runs`; one stream) |
+| `mlx_lm.manage` | `get` / `scan` / `ls` / `gc` | — | `--scan`/`--delete`/`--pattern` flag forms; `gc` deletes superseded snapshots, not whole repos |
+| `mlx_lm.dynamic_quant` | `convert --target-bpw` | — | the `dynamic_quant` verb name |
+
+mlx-optiq's CLI (`optiq convert` / `serve` / `lora` / `benchmark` / `eval` /
+`latency` / `kv-cache` / `lab`) maps onto the same verbs where the concept
+exists: `optiq convert` → `convert --target-bpw`; `optiq serve --kv-config`
+→ `serve --kv-quant config` (the `--l2` preset); `optiq latency` → `fit`;
+`optiq lab` → the web app served by `serve`. `optiq eval` and `optiq
+kv-cache` have no verb (below).
+
+### mlx-lm entry points: covered / not covered
+
+Against the console scripts the pinned oracle ships
+(`mlx_lm-*.dist-info/entry_points.txt` in the oracle venv). The scoping
+behind each "not covered" row is in
+[docs/reference/cli.md](./cli.md).
+
+| Entry point | Status | Where |
+|---|---|---|
+| `mlx_lm.server` | covered | `serve` (incl. `--draft-model` speculative decoding) |
+| `mlx_lm.generate` | covered | `generate` |
+| `mlx_lm.chat` | covered differently | `pi` (agent TUI) + web chat; no plain REPL |
+| `mlx_lm.convert` | covered | `convert` (quantizing paths only) |
+| `mlx_lm.dynamic_quant` | covered | `convert --target-bpw` (sensitivity scan + per-layer bits) |
+| `mlx_lm.lora` | covered (train) | `train`; the `--test` evaluation mode is not covered |
+| `mlx_lm.fuse` | covered | `fuse` (no dequantize / GGUF / upload) |
+| `mlx_lm.upload` | covered | `upload` |
+| `mlx_lm.perplexity` | covered | `perplexity` (local files only) |
+| `mlx_lm.benchmark` | covered | `benchmark` (own flag names; `evals` views the DB) |
+| `mlx_lm.manage` | covered | `get` / `scan` / `ls` / `gc` |
+| `mlx_lm.cache_prompt` | not covered | no on-disk prompt-cache files; the server's prompt cache (+ `--ssd-cache`) is the equivalent for served traffic |
+| `mlx_lm.evaluate` | not covered | no lm-evaluation-harness backend; `benchmark`/`evals`/`perplexity` are the native measurement, and the `/v1` API accepts external harnesses |
+| `mlx_lm.awq` | not covered | activation-aware weight quantization |
+| `mlx_lm.dwq` | not covered | distilled weight quantization |
+| `mlx_lm.gptq` | not covered | Hessian-conditioned rounding |
+| `mlx_lm.share` | not covered | distributed cache share |
+| `mlx_lm` (umbrella `mlx_lm <verb>`) | covered | `mlx-bun <verb>` is the same shape |
+
+Native verbs with no mlx-lm counterpart — `pi`, `harness pi`, `embed`,
+`fit`, `train-watch`, `memory`, `evals` — are the surface that makes mlx-bun
+more than a drop-in.
+
+
+`--kv-quant turbo` composes with grouped lookup, standalone, Gemma assistant,
+DeepSpec and DSpark drafting on supported targets. Sliding layers stay bf16.
+See [batching design](../design/batching.md) for composition evidence and limits.
+
+Paged KV retains the configured `--batch` size. Gemma4 bf16 requests use the shared executor and bypass the prompt cache; quantized KV and draft combinations remain unavailable. See [server configuration](server-config.md).
+
+`--generation-checkpoint N` requires `--ssd-cache` and queues owned snapshots
+for eligible ordinary requests. Shared continuation supports eligible Qwen, Llama, MiniCPM and Gemma4 requests with or without adapters, excluding media, grammar, fill, paging, logprobs or speculation.
+Repeat the identical request to replay and resume; flush before shutdown for
+eventual queued snapshots to become durable.