mlx-bun 0.0.13 → 0.3.0

package/docs/design/memory-inference-path.md

@@ -1,422 +0,0 @@
-# Memory inference path rework — The Dreaming
-
-Status: BUILT + VERIFIED (2026-06-28). Templating fix + single-model reuse
-land and are ready; batching is implemented but is a measured *loss* for
-the real memory workload (see "Verification results" below). Owner: memory
-pipeline.
-
-> ## Verification results (2026-06-28, M-series GPU, e4b OptiQ-4bit)
->
-> Ran on `~/.cache/.../chunk-data-le4000/valid.jsonl` (HELD-OUT, not
-> training rows) + the live vault (`memory.sqlite`, 2096 convs).
->
-> 1. **tsc = 0** (whole repo). Non-GPU suite green:
->    `tests/memory-templating.test.ts` → 7/7 pass (4 CPU message-shape, 2
->    CPU training-parity, 1 GPU batched-order).
-> 2. **Templating matches training — CONFIRMED.** `memoryPromptIds("chunk",
->    {system: CHUNK_SYSTEM, user}, …)` is **byte/id-identical** to
->    `encodeSftRow`'s prompt region for a real train row (the CPU parity
->    test asserts `ids.length` + `ids` equal and the rendered text starts
->    `<bos><|turn>system`). `CHUNK_SYSTEM` equals `messages[0].content`
->    verbatim. The chunk stage now decodes ON-distribution.
-> 3. **Chunk adapter on-distribution — CONFIRMED.** On the 4 held-out
->    multi-chunk valid rows (7/10/11/35) the corrected `system+user` path
->    produced valid JSON with **all anchors in range** (4/4 each, except
->    row 11 which collapsed 2→1 — same as the old path) and faithful
->    labels (e.g. "Wonder Valley homestead land size and payment").
->    Caveat: the OLD user-only path on the SAME adapter *also* produced
->    valid in-range JSON on these rows — the adapter generalizes, so the
->    behavioral delta is minor (label phrasing). The load-bearing win is
->    the byte-exact training-contract match, not a dramatic quality jump on
->    these easy/short held-out rows (the valid set is ~90% single-chunk).
-> 4. **Batching — IMPLEMENTED, but NOT a speedup for memory stages.**
->    - 8 extract-style prompts (~400-token prefills, maxTokens=96):
->      **batched 9345ms vs serial 5001ms = 0.54x (≈1.9x SLOWER)**, 8/8
->      exact match.
->    - 8 short replies (maxTokens=24): batched 2085ms vs serial 1067ms =
->      0.51x SLOWER, **7/8** match (row 6 "H2O" vs "Water" — the documented
->      near-tie decode-path divergence, `runner.ts:104-108`, not an
->      ordering/contamination bug).
->    - 16 short uniform one-liners (maxTokens=64): serial(width=1) 6168ms
->      vs batched 5139ms (width=8) = **~1.2x faster** — batching only wins
->      for many short *uniform* prompts where prefill-pad waste is small.
->
->    Root cause: the bit-exact serial path prefills each sequence with one
->    efficient single-sequence forward; the `BatchScheduler` pads to the
->    batch's max prompt length, so heterogeneous long prefills (exactly the
->    EXTRACT/chunk shape) waste compute and lose to serial.
->
-> **Recommendation:** ship the templating fix + single-model reuse (the
-> real wins); **default `MLX_BUN_MEMORY_BATCH=1`** (serial bit-exact) until
-> the scheduler grows length-bucketed / ragged prefill — current batching
-> regresses the EXTRACT stage AND can diverge on near-ties. Re-segmenting
-> the corpus on the corrected on-distribution path is *principled but
-> low-urgency* (old output was already valid in-range; delta is modest).
-
-*(Original design statement below — superseded; the rework is BUILT, see the
-banner at the top of this file.)*
-
-Scope: the local-model call seam used by every Dreaming stage
-(`src/memory/*` → `callLocal` → `src/eval/runner.ts generateText`). Two
-defects, one rework.
-
-## Problem (verified)
-
-1. **Off-distribution chunk decode (correctness).** The `memory-chunk`
-   adapter (`e4b-chunk-300`) was SFT'd on rows shaped
-   `{messages:[{role:"system",…},{role:"user",…},{role:"assistant",…}]}`
-   and training renders the *full* message list, system turn included
-   (`src/train/dataset.ts:86-88`, `:98-101`). The e4b chat template emits
-   a **distinct** `<|turn>system\n…<turn|>\n` block when
-   `messages[0].role in ['system','developer']`
-   (`chat_template.jinja:179-204`) — this is the Gemma-4 `<|turn>…<turn|>`
-   format that supports a native system role. But live inference renders
-   the prompt as a **single user turn**:
-   `tm.template!.render([{ role: "user", content: body }], …)`
-   (`src/eval/runner.ts:139-141`), so the entire system segment is
-   dropped. Verified rendered shapes on the real e4b snapshot:
-   - training: `"<bos><|turn>system\n{SYS}<turn|>\n<|turn>user\n{USER}<turn|>\n<|turn>model\n"`
-   - live:     `"<bos><|turn>user\n{USER}<turn|>\n<|turn>model\n"`
-
-   The chunk adapter is decoded with its trained system prompt missing — a
-   real distributional gap, not cosmetic.
-
-2. **Single-loaded + serial (throughput).** `callLocal`
-   (`src/memory/model.ts:62-75`) loads its **own** `TaskModel` per
-   `(stage, adapter-present)` key (`:46-57`) and decodes serially through
-   `greedyDecodeBitExact` (`runner.ts:109-131`), one forward loop per
-   call. The hot stages are embarrassingly parallel (EXTRACT: one 128-tok
-   call per chunk over thousands; ROUTE/SECTION: 4–16-tok binary gates;
-   SEGMENT: one 2048-tok call per conversation) but run one at a time.
-
-The server already solves (1): `server.ts:977` renders the **whole**
-`normalizeMessages(req.messages)` array (system + user + …) through the
-same `ChatTemplate`, so system lands as its distinct block. The fix is to
-make the memory path do the same, and to route batchable stages through
-the in-process continuous-batching engine.
-
-## (a) Routing decision — in-process gateway, NOT the HTTP server
-
-**Decision: drive an in-process `BatchScheduler` (via a thin
-`MemoryExecutor`) over ONE shared `RuntimeModel`.** Not the HTTP server.
-
-Rationale:
-- **No-start-servers rule** (CLAUDE.md): an agent/bootstrap job cannot
-  `mlx-bun serve`; the HTTP route presupposes an externally-managed
-  server. Coupling the offline bootstrap to a running server is fragile.
-- **The gateway/scheduler are pure objects.** `GenerationGateway`'s
-  constructor is `new GenerationGateway(model, batch, serialRun)`
-  (`generation-gateway.ts:77-83`) — no HTTP. `BatchScheduler`'s is
-  `new BatchScheduler(model, { maxBatch, lock })`
-  (`batch-scheduler.ts:101`). Both are constructible in-process over a
-  `loadTaskModel(...).model`.
-- **Per-request adapter does not batch on the server anyway.** Any request
-  with an adapter is forced to the serial lane on the server too
-  (`willBatch` excludes `hasAdapters`, `generation-gateway.ts:103`;
-  `server.ts:1570-1577`). So HTTP buys us nothing for the one adapter we
-  have (chunk) and adds a process boundary.
-- **The workload is exactly what continuous batching targets** — a long
-  offline sweep over all eligible conversations/chunks
-  (`chunk.ts:220-286`; the extract/route/section hot loops).
-
-**Fallback:** when only one row is in flight (N==1) or batching is
-disabled, `MemoryExecutor` decodes through the **existing bit-exact greedy
-path** (`greedyDecodeBitExact`, `runner.ts:109-131`) so single calls keep
-token-for-token parity. The in-process scheduler is the throughput lane,
-not a hard dependency.
-
-We construct the executor directly over `BatchScheduler.submit` rather
-than `GenerationGateway.run`, because `gateway.willBatch` hard-excludes
-adapters (`:103`) — routing the chunk stage through `run()` would silently
-fall back to serial. Driving `scheduler.submit()` directly lets us manage
-`loraState.active` ourselves for the chunk batch (see (c)).
-
-## (b) Templating — system vs user per stage
-
-The new seam renders a **message array** through the same template path
-the server uses, instead of wrapping `body` as a lone user turn.
-
-Render contract (mirrors `server.ts:977-980`):
-```
-const rendered = template.render(messages, { addGenerationPrompt: true });
-const ids = tokenizer.encode(rendered, /* addSpecialTokens */ false); // template emits BOS
-// BOS-dedupe identical to promptIdsFor (server.ts:979-980)
-```
-`messages` is `[{role:"system",content:SYS}, {role:"user",content:USER}]`
-(system omitted when a stage has none → user-only, today's shape). This is
-byte-identical to how `encodeSftRow` builds the prompt region
-(`dataset.ts:98-101`, `render(promptMessages, {addGenerationPrompt:true})`).
-
-**Per-stage system/user split:**
-
-| stage string | adapter | system (instruction + policy) | user (content to operate on) | maxTokens |
-|---|---|---|---|---|
-| `chunk` | `memory-chunk`→e4b-chunk-300 | **exact trained system** (below) | `CHUNK_PROMPT` body + inlined Meta pages + `Here is the conversation:` + `formatConversation(...)` | 2048 |
-| `entity` | base | extraction instruction + Entities policy | chunk/text to extract from | 128 |
-| `route` | base | `You answer only 'yes' or 'no'.` + Bucketing policy | the disambiguate/engagesAsSubject question | 4 |
-| `section` | base | binary / new-section instruction | title/section/gist content | 4 / 16 |
-| `synthesis` | base | editorial/synthesis instruction + editorial policy | entity/chunks/article-body | 8–512 |
-| `editor` | base | section-editorial instruction + Editorial_Guidelines | section body to edit | 160–512 |
-
-**Chunk stage system prompt is load-bearing and MUST match training
-byte-for-byte.** Add a constant to `src/memory/chunk.ts` reproducing the
-trained `messages[0].content` verbatim (from
-`~/.cache/mlx-bun/mlx-bun-finetunes/chunk-data-le4000/train.jsonl`):
-
-```
-export const CHUNK_SYSTEM = `You are a precise conversation-segmentation engine, not a coding assistant and not a chat assistant. You receive ONE conversation transcript and return its topic segmentation.
-
-Absolute output contract (overrides any conversational instinct):
-- Your entire response MUST be a single JSON object, parseable by JSON.parse with no edits.
-- No prose, no preamble, no explanation, no follow-up question, no markdown code fences, no tool calls. The first character you emit is "{" and the last is "}".
-- Never ask for clarification or more data — the transcript in the message IS the data. If it has no meaningful content, return {"chunks": []}.
-- Use only message UUIDs present in the provided transcript; never invent or alter a UUID.
-
-Follow the segmentation rules and JSON schema in the message exactly.`;
-```
-
-The chunk SEGMENT call becomes `callLocal("chunk", { system: CHUNK_SYSTEM,
-user: promptHead() + formatConversation(...) }, {maxTokens})`.
-
-**Secondary: chunk USER-turn content drift.** The live `CHUNK_PROMPT`
-(`chunk.ts:30-69`) also diverges from the trained user turn on four axes
-and these matter for true on-distribution behavior, but are a SEPARATE
-fix from the role bug:
-- (a) `chunk.ts:32` adds `Use ONLY content from the conversation provided
-  below; never copy any name…` — absent in training.
-- (b) `META_POLICY_PAGES = ["Chunking","Topics_to_Ignore"]` (`chunk.ts:72`)
-  inlines **2** pages; training inlined **8** (Article_Conventions,
-  Bucketing, Buckets, Category_Definitions, Chunking,
-  Editorial_Guidelines, Synthesis_Pipeline, Topics_to_Ignore).
-- (c) label guidance rewritten to abstract GOOD/BAD shape vs trained
-  inline examples (`chunk.ts:55-57`).
-- (d) training POLICY line names the vault path `/Users/joshrossi/Dreaming/Meta/`;
-  live drops it.
-
-Recommendation: land the **role fix first** (biggest, cheapest win), then
-reconcile the user turn to the trained shape *or* retrain `e4b-chunk-300`
-against the current `CHUNK_PROMPT`. Pick one; do not leave it
-half-aligned. The role fix alone moves the chunk stage substantially
-closer to its training distribution.
-
-For the **base** stages (everything except chunk), the system/user split
-is a quality choice for the base instruct model (which was trained to
-honor a system turn) — apply it, but there is no trained adapter to match,
-so byte-exactness is not required there.
-
-## (c) Batching — per-stage homogeneous-adapter batches
-
-`loraState.active` is a **single global field** on the model, set by
-`generate()`'s `adapterScoped` (`generate.ts:235-247`) and applied
-uniformly to every row; the scheduler never touches it (it calls
-`model.forwardHidden` directly — zero `loraState` refs in
-`batch-scheduler.ts`). Therefore **a batch is necessarily single-adapter
-(or all-base).** We cannot mix chunk-adapter and base rows in one batch.
-
-This fits the pipeline naturally because `callLocalBatch` is invoked
-**per stage**, and stage→adapter is fixed (`adapterDirFor`,
-`model.ts:38-41`; only `memory-chunk` exists on disk):
-
-- **Base stages (entity / route / section / synthesis / editor):** batch
-  with **zero** adapter friction — `loraState.active = []`. Submit N rows
-  to the scheduler. The easy, highest-volume win (EXTRACT first: uniform
-  128-tok prompt, one per chunk, no inter-call dependency).
-- **Chunk stage:** set `model.loraState.active = ["memory-chunk"]`
-  **once** for the whole stage batch, submit only chunk rows, reset to
-  `[]` after. Because `forwardHidden` reads the global `loraState`, all
-  rows get the adapter uniformly. The executor owns the model exclusively
-  during the batch (the scheduler holds the lock for its active span), so
-  there is no concurrent serial generation to clobber `loraState`.
-
-**Decode-path caveat.** The batched lane samples via `generate()`'s
-sampler path (greedy = argmax over `forwardHidden`/`logitsFromHidden`),
-NOT the runner's raw-`forward` bit-exact greedy. The runner warns these
-diverge on near-ties past ~32 tokens (`runner.ts:104-108`). For chunk's
-2048-tok JSON and the binary gates this is expected to be fine (chunk eval
-is well-formedness, per memory notes), but it MUST be validated against
-the greedy reference before the chunk batch is trusted (see verification).
-
-**Batching priority** (from the usage survey, batchable structure):
-1. EXTRACT (`entity`, base, 128 tok, 1/chunk over thousands) — prime win,
-   uniform prompt, no dependency.
-2. ROUTE + SECTION binaries (`route`/`section`, base, 4–16 tok) — batch
-   the per-chunk N-section grid and per-candidate gates (the trailing
-   new-section call depends on all-no — keep it after the batch).
-3. SEGMENT (`chunk`, adapter, 2048 tok, 1/conversation) — batchable across
-   conversations, but long outputs; the only adapter-bound batch.
-4. SYNTHESIZE/WIKIFY drafts (`synthesis`/`editor`, 96–512 tok) — batch
-   per-phase within an article (the N independent section drafts), NOT
-   whole-article (outline→draft and the leak-retry are data-dependent).
-5. RECONCILE — largely serial (`next` mutated in place); leave on the
-   serial fallback; only the independent binary classifiers batch.
-
-The hot loops stay where they are; we change *how a set of independent
-calls is submitted*. The cleanest refactor: where a stage today does
-`for (x of xs) await call(buildPrompt(x))`, collect `xs`, build all
-messages, and `await callLocalBatch(stage, messages[])` → results[].
-Start with EXTRACT and SEGMENT; leave dependent loops on serial `callLocal`.
-
-## (d) API surface
-
-`src/memory/model.ts` grows a messages-aware seam and a batch seam; the
-single-user-turn `generateText` overload is no longer used by the memory
-path.
-
-```ts
-export interface LocalInput {
-  /** System turn (instruction + policy). Omit for a user-only prompt. */
-  system?: string;
-  /** User turn (the content to operate on). */
-  user: string;
-}
-
-/** One call. N==1 → bit-exact greedy (runner.ts:109-131). */
-export async function callLocal(
-  stage: string,
-  input: LocalInput,
-  opts?: { maxTokens?: number },
-): Promise<string>;
-
-/** N independent calls for one stage (homogeneous adapter). Submitted to
- *  the shared in-process BatchScheduler when batch>1; otherwise looped on
- *  the bit-exact greedy path. Order-preserving: out[i] ↔ inputs[i]. */
-export async function callLocalBatch(
-  stage: string,
-  inputs: LocalInput[],
-  opts?: { maxTokens?: number },
-): Promise<string[]>;
-```
-
-Backing implementation — a module-level `MemoryExecutor` (lazy singleton):
-- holds ONE `RuntimeModel` (via `loadTaskModel(MODEL_ID)`), ONE
-  `AdapterManager` with `memory-chunk` mounted once, ONE `BatchScheduler`
-  (constructed with an in-process `AsyncMutex`-backed lock), and the
-  tokenizer + template.
-- `render(input)` → messages array → `template.render(messages,
-  {addGenerationPrompt:true})` → `encode(text, false)` → BOS-dedupe →
-  `promptIds`.
-- `callLocalBatch(stage, inputs)`:
-  1. `active = adapterDirFor(stage) ? ["memory-chunk"] : []`.
-  2. `model.loraState.active = active` for the batch's lifetime.
-  3. render all inputs → promptIds; build a per-row token sink (decode via
-     `tokenizer.decode`); `Promise.all(rows.map(r =>
-     scheduler.submit({promptIds, maxTokens, eosTokenIds, sample: greedy,
-     onToken: collect})))`.
-  4. reset `model.loraState.active = []`; return decoded strings in order.
-- `callLocal` delegates to `callLocalBatch(stage, [input])[0]` when
-  batching is enabled, else the bit-exact greedy single decode.
-
-This **consolidates** the per-stage `TaskModel` instances
-(`model.ts:46-57`) into one shared model — reuse AND batching in one move.
-Stage builders change from passing a `prompt` string to passing
-`{system?, user}` (chunk supplies `CHUNK_SYSTEM`).
-
-`generateText`'s single-user render (`runner.ts:139-141`) is left intact
-for the eval suite (its own contract), but the memory path stops calling
-it. Optionally add a system-aware overload to `generateText` later to
-share the render helper; not required for this rework.
-
-## (e) Re-segment the partial import — YES
-
-The current import (720/2096 conversations segmented) was chunked with the
-**old single-user template** (no system role) AND only 2 Meta pages — i.e.
-the off-distribution shape this rework fixes. For consistency the whole
-corpus should be segmented under one (correct) prompt regime.
-
-Decision: **re-segment all 720 already-chunked conversations** after the
-role fix lands. It is cheap and safe:
-- SEGMENT is idempotent — `chunkConversations` deletes prior chunks per
-  conv and resets `chunked_at` on success (`chunk.ts:262-273`).
-- To force a re-run, clear `chunked_at` (set NULL) for the 720 so
-  `selectEligible` re-admits them; then run the batched SEGMENT.
-- Downstream stages key off chunk ids; re-segmentation changes chunk
-  boundaries/labels, so the dependent EXTRACT/ROUTE/SYNTH state for those
-  convs must be invalidated too (treat as a fresh segmentation). Sequence
-  the bootstrap so re-segmentation precedes the downstream sweep.
-
-Do NOT mix two segmentation regimes in one vault; re-segment for a clean
-baseline.
-
-## (f) Implementation steps
-
-1. **Template seam.** Add `MemoryExecutor` in `src/memory/model.ts`: one
-   shared `RuntimeModel` + tokenizer + template + `AdapterManager`
-   (mount `memory-chunk`). Implement `render({system?,user})` →
-   promptIds (render → encode(false) → BOS-dedupe, matching
-   `server.ts:979-980`).
-2. **API.** Replace `callLocal(stage, prompt, opts)` with
-   `callLocal(stage, LocalInput, opts)` + `callLocalBatch`. Single-call
-   path keeps bit-exact greedy.
-3. **Chunk system.** Add `CHUNK_SYSTEM` (verbatim trained text) to
-   `chunk.ts`; change the SEGMENT call to pass `{system: CHUNK_SYSTEM,
-   user: promptHead()+formatConversation(...)}`.
-4. **Per-stage splits.** Update entity/route/section/synthesis/editor
-   builders to return `{system, user}` (move instruction+policy → system,
-   content → user). Mechanical; one stage at a time, tsc-green each.
-5. **Batch the hot stages.** Convert EXTRACT first
-   (`entity.ts`/`stages.ts:333-340`), then SEGMENT
-   (`chunk.ts:220-286`), to collect inputs and call `callLocalBatch`.
-   Leave dependent loops (reconcile, outline→draft) on serial `callLocal`.
-6. **loraState management.** In `callLocalBatch`, set/reset
-   `model.loraState.active` around the chunk batch; assert base stages
-   submit with `active=[]`.
-7. **Re-segment.** One-off: NULL `chunked_at` for the 720 imported convs,
-   invalidate downstream state, run batched SEGMENT, then the downstream
-   sweep.
-8. **tsc 0** across the repo incl. `scripts/experiments` (standing rule).
-
-## Verification plan
-
-- **Templating parity (the correctness gate).** Render the chunk
-  `{system,user}` through the new seam and assert the token ids equal the
-  training prompt-region ids for a real `train.jsonl` row: rebuild the row
-  via `encodeSftRow`'s prompt render (`dataset.ts:98-101`,
-  `render(messages,{addGenerationPrompt:true})`) and compare. Must be
-  byte/id-identical (same `<|turn>system…<turn|>` block, single BOS).
-  This is the foundation — verify it before any generation claim
-  (standing "verify the foundation" rule).
-- **System-present smoke.** Decode one conversation with the new seam and
-  confirm the rendered prompt contains the system block (not user-only);
-  diff against the old render to show the gap closed.
-- **Batching correctness.** For a base stage (EXTRACT) and the chunk
-  stage, compare `callLocalBatch([a,b,c])` against three serial
-  `callLocal` greedy decodes. Base/binary outputs should match; for chunk,
-  validate the batched (sampler-path) output against the bit-exact greedy
-  reference and accept divergence only within the documented near-tie
-  envelope (chunk eval = well-formedness: valid JSON, in-range anchors).
-- **Chunk-adapter behavior.** Confirm `loraState.active=["memory-chunk"]`
-  is set for the chunk batch and `[]` for base batches (assert in test).
-  Re-run `scripts/chunk-eval.ts` before/after the role fix; expect
-  well-formedness and boundary/label quality to improve now that decode is
-  on-distribution.
-- **No-leak / single-model.** Confirm the executor loads ONE model (not
-  per-stage instances) and peak memory does not regress vs the current
-  per-stage cache; watch for KV/loraState leaks across batches.
-- **Re-segment sanity.** After re-segmenting the 720, spot-check chunk
-  counts/labels vs the old segmentation and confirm downstream stages
-  re-run cleanly on the new boundaries.
-
-## Risks
-
-- **Decode divergence on the batched lane.** generate()'s sampler path ≠
-  the runner's raw-forward greedy past ~32 tokens on near-ties
-  (`runner.ts:104-108`). Mitigation: greedy sampler + well-formedness
-  validation; keep the bit-exact greedy path for the single-call fallback
-  and for any stage that proves sensitive.
-- **Global `loraState` clobber.** The chunk batch sets a global field; any
-  concurrent generation would corrupt it. The bootstrap runs one executor
-  with exclusive model ownership, so this is safe — but the seam must
-  never be shared with a live server process.
-- **User-turn content drift (chunk).** The role fix alone does not close
-  the 8-vs-2 Meta-pages / examples / vault-path drift. Track as a
-  follow-up: reconcile `CHUNK_PROMPT` to training or retrain the adapter.
-  Until then the chunk stage is *closer* to, but not fully on,
-  distribution.
-- **Re-segment blast radius.** Re-segmentation invalidates downstream
-  state for the 720 convs; sequencing errors could leave orphaned
-  extract/route rows. Mitigation: invalidate-then-resweep, idempotent
-  stages, verify counts.
-- **Single trained adapter assumption.** Only `memory-chunk` exists. If
-  future stages gain adapters, `callLocalBatch` already batches per stage
-  (homogeneous adapter) — but mixing adapters in one batch remains
-  impossible by construction; keep stages separate.