mlx-bun 0.4.0 → 1.0.0

package/docs/archive/investigations/audio-input-plan.md

@@ -1,356 +0,0 @@
-# Audio input — design & implementation plan (gemma-4 audio tower)
-
-Status: **A0–A4 DONE** (2026-07-07: served end to end on e4b, oracle-EXACT
-greedy offline AND over HTTP; A5 bench + 12B coverage open).
-Scope: audio **understanding** — audio-in, text-out through the chat API.
-TTS / STT-as-an-endpoint / STS stay out of scope (see
-docs/design/omlx-adoption-map.md); §5 lists non-goals explicitly.
-
-## 0. TL;DR
-
-e4b ships a 12-block Conformer audio tower that we already have on disk —
-the `optiq_vision.safetensors` sidecar in the local e4b OptiQ-4bit snapshot
-contains 752 `audio_tower.*`/`embed_audio.*` tensors, and its `config.json`
-carries the full `audio_config` plus the audio token ids. mlx-lm strips all
-of this and rejects audio requests, so (as with vision) **the oracle is
-optiq's internal machinery**, which is complete end-to-end but not exposed
-by optiq's own serve frontend. The implementation is a fourth tower next to
-`src/vision/{siglip,embedder,diffusion-vision}.ts`, reusing every existing
-multimodal seam: sidecar lazy-load, `<|placeholder|>` splice, merged
-embeddings into `forwardEmbeddings`, serial-lane routing. New primitives
-needed: a `mlx_conv2d` binding (SSCP subsampler) and a small TS-side
-mel-spectrogram extractor (USM pipeline). Everything else is pattern-match
-on the vision port.
-
-## 1. What the ancestors do (surveyed 2026-07-06, pinned oracle venv)
-
-### mlx-lm 0.31.3 — nothing, and actively strips
-
-- `models/gemma3n.py:602` and `models/gemma4.py:55` `sanitize()` pop
-  `audio_tower`, `embed_audio` (and the vision keys) from the weights at
-  load. The upstream capability is deliberately discarded.
-- `server.py:118–150` `process_message_content()` raises
-  `ValueError("Only 'text' content type is supported.")` for any non-text
-  content part — `input_audio` requests hard-fail.
-- No mel/spectrogram/feature code anywhere in the package; no mlx-vlm /
-  mlx-audio / mlx-whisper installed in the oracle venv.
-
-**Parity-tree consequence:** there is no mlx-lm arm for audio. Like the
-vision port, correctness gates come from optiq's internals; anything no
-ancestor does (e.g. ever batching audio requests) is novel territory →
-KL + quality gates, not bit-parity (see three-tier doctrine in PLAN.md).
-
-### mlx-optiq — complete machinery, unexposed at the serve layer
-
-All under `optiq/vlm/_mlxvlm/models/gemma4/` in the oracle venv:
-
-- **Feature extraction** (`audio_feature_extractor.py`,
-  `Gemma4AudioFeatureExtractor`) — USM pipeline: 16 kHz mono input,
-  frame 20 ms / hop 10 ms, magnitude STFT, 128 HTK-scale mel bins over
-  0–8 kHz, `log(mel + 1e-3)` floor, optional per-bin mean/stddev
-  normalization; `preemphasis=0.0`, `fft_overdrive=False`, `dither=0.0`
-  by default. Runs CPU-side (numpy). Output
-  `input_features [B,T,128]` + validity mask.
-- **Tower** (`audio.py`, `AudioEncoder`) —
-  `SubSampleConvProjection`: two Conv2d stages (channels `(128, 32)`,
-  each 2× time stride) then linear to `hidden_size=1024`; then
-  **12 Conformer blocks**: FF (residual_weight 0.5) → chunked local
-  attention (chunk 12, left context 13 chunks, right 0, logit softcap
-  50.0, invalid logits −1e9) → depthwise light Conv1d (kernel 5) → FF →
-  clamp + RMSNorm; then output projection to `output_proj_dims=1536`.
-  `use_clipped_linears=True`: linears clamp activations to recorded
-  per-tensor `input_min/max`/`output_min/max` ranges (those stats ship in
-  the sidecar — confirmed present in our local e4b sidecar).
-- **Embedder** (`merge.py`, `MultimodalEmbedder`) — RMSNormNoScale +
-  bias-free Linear into text hidden size; embeddings scattered into the
-  input sequence at `audio_token_id` positions via `masked_scatter`
-  (`gemma4.py get_input_embeddings`).
-- **Tokens & splicing** (`processing_gemma4.py:264–299`) —
-  `<|audio|>` expands to `boa + audio_token×n + eoa` with
-  `n = min(ceil(duration_ms / 40), 750)` (40 ms/token, ≤30 s ≙ 750 soft
-  tokens). Ids (confirmed in our local e4b config.json):
-  `boa=256000`, `audio=258881`, `eoa=258883`.
-- **Request parsing** (`processing_gemma4.py:690`) — accepts OpenAI-style
-  parts `{"audio", "input_audio", "audio_url"}`.
-- **Sidecar & quant policy** — `vlm/sidecar.py` extracts
-  `audio_tower.*`/`embed_audio.*` into the same `optiq_vision.safetensors`
-  as vision, kept **bf16**; `backends/mlx_backend.py:491` skips
-  `audio_tower.` in the quantization predicate. Calibration recipe (and a
-  worked example of driving the audio path programmatically):
-  `calibration/datasets.py load_audio_calibration` (LibriSpeech clean).
-- **The gap:** `vlm/gemma4/frontend.py` only wires **vision** into optiq
-  serve; audio works only through the internal model class. Golden
-  generation therefore scripts the internal model directly (the
-  calibration loader shows exactly how) — same move as the vision port.
-
-## 2. What we already have (verified on this machine, 2026-07-06)
-
-- **e4b OptiQ-4bit** (snapshot `fcdb12…`): sidecar contains **752 audio
-  tensors** (`audio_tower.*` + `embed_audio.*`, incl. clipped-linear
-  min/max stats); `config.json` has `audio_config`
-  (`model_type=gemma4_audio`, 12 layers, hidden 1024, heads 8, conv
-  kernel 5, subsampling channels (128,32), logit cap 50, residual 0.5,
-  `output_proj_dims=1536`, `use_clipped_linears=true`) plus
-  `audio_token_id/boa_token_id/eoa_token_id`; `tokenizer_config.json` has
-  `audio_token: "<|audio|>"`. **No download needed — audio-ready today.**
-- **12B OptiQ-4bit**: `audio_config` present but the local sidecar holds
-  only **1** audio tensor → needs a sidecar rebuild via optiq's
-  `build_vision_sidecar` (its selective-download path pulls only the
-  audio shards, not the full base model).
-- **26B-A4B, DiffusionGemma, bf16 assistants**: no `audio_config` — no
-  audio for these, ever (architecture, not a porting gap).
-- **Vision pipeline is the template** — every seam already exists:
-  content-part detection `src/server.ts:2234`, extraction + splice + merge
-  `src/vision/prompt.ts`, lazy tower `src/server.ts:326–350`, sidecar
-  detection `src/registry.ts:487` / `src/config.ts:287`, embeddings
-  prefill `src/generate.ts:468` → `forwardEmbeddings`
-  `src/model/gemma4.ts:876–896` (per-layer-input id zeroing + bidir mask).
-- **FFI gaps**: `mlx_conv1d` is bound (`src/mlx/ffi.ts:173`, Qwen3.5
-  DeltaNet, depthwise) but there is **no conv2d and no FFT** binding.
-  SSCP needs conv2d; mel extraction needs an FFT (resolved TS-side, §3.1).
-
-## 3. Design
-
-### 3.1 New files (mirror of `src/vision/`)
-
-- **`src/audio/decode.ts`** — WAV parser (PCM16/24/32/float32) → mono
-  mix-down → resample to 16 kHz (windowed-sinc or linear; gate vs oracle
-  resampler on the fixture). Non-WAV formats (mp3/m4a/flac/ogg): shell out
-  to macOS `afconvert` (built into darwin, our only platform) to produce
-  WAV — zero new dependencies. **Duration (and thus soft-token count) is
-  computed from decoded samples**, so format never affects splicing.
-- **`src/audio/features.ts`** — USM mel extractor, verbatim-semantics port
-  of `audio_feature_extractor.py` (megakernel copy-verbatim methodology:
-  match first, optimize never — this is ~100×128 floats, not a perf
-  path). STFT frame 320 samples / hop 160 / FFT 512
-  (`fft_overdrive=False`), 128 HTK mel bins 0–8 kHz, `log(x + 1e-3)`,
-  per-bin normalization from shipped stats (open Q3). Implemented as a
-  small radix-2 real FFT in TS, f64 accumulate → f32 out, CPU-side —
-  exactly where optiq does it (numpy), keeping the mlx graph clean.
-  *Considered and rejected for v1*: binding `mlx_fft_rfft` and doing it
-  on-device — new binding surface for no perf need.
-- **`src/audio/conformer.ts`** — the tower, bf16 mlx graph:
-  - **needs `mlx_conv2d` binding** — read the full mlx.h signature first
-    (trailing-optional-param → shifted stream arg hazard, CLAUDE.md).
-  - SSCP: Conv2d(1→128) → Conv2d(128→32), each 2× time stride → linear
-    to 1024 (exact shapes read from the sidecar weights).
-  - 12 × ConformerBlock as in §1; chunked local attention implemented as
-    an explicit dense additive mask (≤750 frames pre-subsample, ~188
-    post — dense is trivially cheap at this size) with logit softcap 50
-    and −1e9 invalid fill.
-  - Clipped linears: clamp to shipped `input_min/max`/`output_min/max` —
-    parity depends on this, port exactly (toggle test in A2).
-  - Output proj → 1536, then `embed_audio` (RMSNormNoScale + Linear →
-    text hidden), **pre-divided by embed_scale** — same convention as
-    vision towers (`forwardEmbeddings` re-multiplies).
-- **`src/audio/prompt.ts` folded into a shared multimodal prompt
-  builder** — recommendation: generalize `src/vision/prompt.ts`
-  (`buildVisionPrompt` → `buildMultimodalPrompt`) rather than add a
-  parallel audio-only path, because one message can legally carry images
-  AND audio and both must splice in document order. Vision-only and
-  audio-only fall out as special cases; the vision goldens gate the
-  refactor (spliced ids must stay byte-identical).
-
-### 3.2 Server surface
-
-- Accept OpenAI content parts:
-  `{"type":"input_audio","input_audio":{"data":"<b64>","format":"wav"|"mp3"}}`
-  (OpenAI canonical) plus optiq's `{"type":"audio"}` / `{"type":"audio_url"}`
-  aliases. `hasAudio` detection lands next to `hasImages`
-  (`src/server.ts:2234`).
-- `getAudioTower(ctx)` lazy-loads next to `getVisionTower`
-  (`src/server.ts:326–350`) — same sidecar file, separate tower object;
-  text-only sessions never pay for it. Auto-enable on
-  `audio_config` + sidecar audio tensors present, mirroring vision —
-  **no new flags** (target surface: zero CLI delta).
-- Routing: audio requests take the **serial lane**, same rule as vision —
-  embeddings-prefill doesn't batch, and `promptEmbeddings` already
-  excludes the pre-warmed prompt cache (`src/generate.ts:468`). Batching
-  is a mode, not a fallback; batched audio has no oracle → if we ever
-  want it, it goes through KL + quality gates as new territory.
-- Anthropic-protocol endpoint: the Anthropic messages API has no audio
-  input block type → clean 400 with a pointer to the OpenAI endpoint;
-  documented.
-- **Docs land in the SAME commit** (standing rule):
-  `docs/reference/server-api.md` (content-part schema, ≤30 s / 750
-  soft-token cap, 16 kHz mono internal format),
-  `docs/reference/features-matrix.md` (audio row incl. serial-lane
-  note), README capability line, `cli.md` only if a flag appears
-  (target: none).
-
-### 3.3 Open questions — RESOLVED against the oracle sources (2026-07-07)
-
-1. **LM attention semantics for audio tokens: strictly CAUSAL.**
-   `language.py:484–545 _make_masks`: the blockwise bidirectional overlay
-   keys off `is_vision = (mm_token_type_ids == 1) | (== 2)` only — audio
-   (type 3) never enters a block. Stronger: `use_bidirectional_vision`
-   requires `not has_audio_tokens`, so **any audio token in the prompt
-   disables the vision overlay entirely** (oracle comment: "Audio spans
-   are sequential; keep mixed image+audio prompts causal"). mlx-bun rule:
-   audio runs get NO bidir mask, and a mixed image+audio request drops
-   the image bidir mask too — pass no bidir mask at all in that case.
-2. **Per-layer-input zeroing: yes, audio included.** `gemma4.py:92–99`
-   zeroes per-layer ids for the union mask
-   `(ids==image_token) | (ids==audio_token) | (ids==video_token)`.
-   Extend the existing image-mask zeroing in
-   `src/model/gemma4.ts:876–896` to the combined multimodal mask.
-3. **Per-bin mel stats: constructor defaults, no normalization.**
-   `processing_gemma4.py:919–934`: standard HF checkpoints ship no
-   `feature_extractor` key (our e4b snapshot has no
-   processor_config.json at all) → `Gemma4AudioFeatureExtractor()` with
-   defaults; "the USM parameters are fixed for all Gemma 4 models".
-   `per_bin_mean/stddev = None` → that branch is dead for us. Also
-   pinned there: `audio_seq_length` default 750, `audio_ms_per_token`
-   default 40.
-4. **boa/eoa are splice-side.** e4b `chat_template.jinja:301,334` emits
-   bare `<|audio|>` for audio parts (both content shapes); the
-   boa + audio×n + eoa expansion is processor-side — same ownership as
-   vision's boi/eoi. Splice in our prompt builder, not the template.
-
-### 3.4 Parity strategy (tree doctrine)
-
-Oracle = optiq internal gemma4 model in the pinned venv, driven by
-`regen-audio-goldens.py (deleted; git history)` (modeled on their calibration loader;
-remember `enable_thinking` pinning for any templated prompt). Tiers follow
-the e4b vision precedent exactly:
-
-- **T0 model-free** (CI, no weights): WAV decode + mel features vs
-  oracle-dumped golden (f32 tolerance ~1e-5 — DSP is deterministic);
-  spliced ids **exact**; soft-token count **exact**.
-- **T1 golden tower**: `audio_tower`+`embed_audio` output vs oracle,
-  rel-RMSE gate (expect vision-like ~1% from bf16 composition; calibrate
-  the threshold from the first honest run, don't invent it).
-- **T2 e2e greedy** (`tests/e4b-audio.test.ts`): first-N greedy prefix
-  match vs oracle + grounded-output check (speech fixture → transcript
-  keyword present; chirp fixture → tone/beep mentioned).
-- **T3 serve**: curl e2e with base64 `input_audio` against a running
-  server (manual/gated; Josh runs servers).
-
-Fixtures (tracked in `fixtures/`, tiny): (a) 1.6 s synthesized chirp WAV
-(~50 KB, generated by a checked-in deterministic script), (b) ~2 s public
-domain speech clip (LibriSpeech clean sample, ~64 KB). Goldens:
-`goldens/e4b-audio.json` manifest tracked, `.bin` blobs regenerable and
-untracked (goldens/README.md convention).
-
-### 3.5 Perf & memory
-
-Expectation (to be *measured*, not asserted): the tower is tiny —
-≤750 mel frames → ~≤188 post-subsample positions × 12 layers × hidden
-1024 bf16 — so added TTFT should be dominated by CPU feature extraction
-plus one extra graph eval; sidecar RSS delta is the 752 bf16 tensors.
-Every number that gets quoted comes from preflight-gated `scripts/bench-serve.ts all`
-cells (tower ms, TTFT delta vs text-only, peak RSS delta) and lands in
-`benchmarks/RESULTS.md`.
-
-## 4. Work plan
-
-**Phase A0 — groundwork**
-- [ ] Bind `mlx_conv2d` (full header signature; CPU/GPU sanity test vs
-      oracle conv output)
-- [ ] `regen-audio-goldens.py (deleted; git history)` + fixtures; dump mel golden,
-      tower-output golden, spliced ids, greedy ids
-- [ ] Resolve open questions §3.3 (write the answers into this doc)
-- Exit: goldens on disk; masking/zeroing semantics documented.
-
-**Phase A1 — decode + features**
-- [ ] `src/audio/decode.ts` (WAV + mono + resample; `afconvert` fallback)
-- [ ] `src/audio/features.ts` USM mel port
-- Exit: T0 green (mel within tolerance; token counts + spliced ids exact).
-
-**Phase A2 — tower**
-- [x] `src/audio/conformer.ts` loads `audio_tower.*`/`embed_audio.*` from
-      `optiq_vision.safetensors`; forward matches golden — rel-RMSE **2.4e-8**
-      (bit-exact: the tower runs f32 activations over bf16 weights because
-      the oracle feeds f32 mel and mlx promotes; the residual is 1 f32 ulp
-      from the test's /embed_scale·embed_scale roundtrip). Sidecar conv
-      weights are ALREADY MLX layout [C_out,kH,kW,C_in] — no transpose.
-      Golden reference point: embed_audio output BEFORE the /embed_scale
-      pre-division; features() returns pre-divided (vision convention).
-- [x] Clipped-linear toggle test (off → 90% rel-RMSE divergence, on → gate;
-      in tests/e4b-audio-tower.test.ts)
-- Exit: T1 rel-RMSE gate green. ✅ (gate 1e-6, ~40× margin)
-
-**Phase A3 — prompt + LM integration**
-- [x] Generalize prompt builder (images + audio, document order):
-      `buildMultimodalPrompt` in src/vision/prompt.ts; `<|audio|>` →
-      `boa + audio×n + eoa`; `buildVisionPrompt` kept as a byte-identical
-      wrapper (vision goldens green, e4b-vision 3/3).
-- [x] Merged embeddings + masks; per-layer id zeroing per Q2 decoupled from
-      the bidir overlay (`forwardEmbeddings` gained an optional `multimodal`
-      zeroing mask; generate() `multimodalMask` option). Merge numerics are
-      oracle-exact: the divide happens AFTER the bf16 cast with a weak
-      (bf16) embed_scale scalar — `AudioTower.features(preDivide=false)` +
-      astype + div, mirroring `features.astype(bf16) / embed_scale` (mlx
-      weak scalars take the array dtype; divide-in-f32-then-cast is up to a
-      bf16 ulp off and fails the exact gate).
-- Exit: T2 green — and stronger than planned: the FULL greedy stream +
-  decoded text match the oracle EXACTLY for both fixtures (the tower is
-  bit-exact, so no prefix allowance was needed; generate()'s
-  promptEmbeddings path prefills in one forward exactly like the golden
-  script — no tail-split gap). `tests/e4b-audio.test.ts`.
-
-**Phase A4 — serve + docs**
-- [x] `src/server.ts` audio branch, lazy tower, serial-lane routing,
-      mixed image+audio request handled. `hasAudio` detection over the
-      three part shapes; `makeAudioLoader`/`getAudioTower` mirror vision
-      (same sidecar, zero flags, auto-enable on `audio_config` +
-      sidecar) — but a request WITH audio on a towerless model is an
-      explicit 400 naming the model, never a silent text-only degrade.
-      The gateway's `Vision` payload generalized to
-      `{embeddings, imageMask?, multimodalMask?}` (bidirMask stays null
-      whenever audio is present, §3.3 Q1); `shape.hasVision = !!payload`
-      routes every media request serial and skips the prompt cache —
-      verified live: `/stats` batch.submitted_rows stays 0 for audio
-      while a concurrent text request advances it (batch default 8).
-      Non-WAV containers (mp3/m4a/flac/ogg/aiff) transcode via
-      `afconvert` (src/audio/transcode.ts, RIFF-sniffed content-based
-      passthrough, temp files under os.tmpdir(), clear 400 on failure);
-      >30 s clips truncate at 480 000 samples in decodeAudio — the
-      oracle feature extractor's own default (truncation=True,
-      max_length=480_000) verified in audio_feature_extractor.py:302.
-      /v1/messages (Anthropic) rejects smuggled audio blocks with a 400
-      pointing to the OpenAI endpoint (protocol has no audio block).
-- [x] server-api.md (schema + curl + caps + 400s + lane note) /
-      features-matrix.md (audio row + footnote: mlx-lm strips it, optiq
-      never serves it — we expose it) / README capability line, same
-      change set; Anthropic endpoint 400 documented.
-- [x] capability discovery (post-merge CodeRabbit follow-up, 2026-07-07):
-      `audio` flag mirrors `vision` on every surface — /v1/models served
-      row, /library rows (`audioCapable` in src/registry.ts:
-      `has_audio_config AND has_audio_tower`, header-only sidecar scan),
-      and the pi ws `ready` frame. The predicate checks TENSORS, not just
-      config: the 12B stub sidecar (audio_config + embed_audio-only) reads
-      audio=false, and `makeAudioLoader` header-checks the sidecar so the
-      runtime flag can't lie either. Discovery-only in pi land — pi's
-      `Model.input` type has no "audio" modality to declare.
-- Exit: T3 green — `tests/e4b-audio-serve.test.ts`
-      (MLX_BUN_TEST_AUDIO_SERVE=1): base64 WAV over
-      /v1/chat/completions returns the EXACT golden transcription
-      ("The quick brown fox jumps over the lazy dog.", 83 prompt /
-      11 completion tokens); m4a transcode path transcribes; malformed
-      part, undecodable bytes, and Anthropic audio block all 400.
-      docs-map hygiene green. ✅ 2026-07-07
-
-**Phase A5 — bench + coverage**
-- [ ] scripts/bench-serve.ts all cells → RESULTS.md (tower ms, TTFT delta, RSS delta)
-- [ ] 12B audio cell: rebuild its sidecar via optiq `build_vision_sidecar`
-      (selective download), own goldens — per-model doctrine, every
-      (model, quant) cell gets its own validation or an explicit deferral
-- [ ] features-matrix: audio×batching = serial documented
-- Exit: numbers curated; e4b cell validated; 12B validated or deferred
-  with reason.
-
-Sizing: A0–A4 ≈ 3–4 focused sessions (the tower is the only genuinely new
-model code; everything else is seam reuse); A5 ≈ 1.
-
-## 5. Non-goals (v1)
-
-- TTS, STS, and `/v1/audio/transcriptions`-style endpoints — a
-  transcription shim over this same path is a cheap follow-up, noted, not
-  committed.
-- Streaming/real-time audio input.
-- >30 s audio: cap at 750 soft tokens exactly as the oracle does.
-- Video (`mm_token_type_ids=2`) — same seams, explicitly later.
-- Audio for 26B-A4B / DiffusionGemma / bf16 assistants (no
-  `audio_config` in those snapshots — architectural, not deferred work).
-- Batched audio prefill (novel territory; serial lane only in v1).

package/docs/archive/investigations/chunk-finetune-experiment.md

@@ -1,84 +0,0 @@
-# Fine-tuning experiment: chunking task (2026-06-16)
-
-Goal (from Josh): test mlx-bun's fine-tuning end-to-end on lucien's conversation-
-chunking task — baseline → LoRA fine-tune → re-measure the quality delta.
-
-## Headline result
-
-A LoRA fine-tune of **MiniCPM5-1B** on the chunk task, trained and served entirely
-by mlx-bun, on the 25-case frozen holdout (`lucien/benchmark/dataset/chunk.json`),
-scored with lucien's exact `scoreChunk` metric:
-
-| Metric | Baseline (stock) | Fine-tuned | Δ |
-|---|---|---|---|
-| **Mean score / 100** | **11.89** | **91.70** | **+79.8 (6.7×)** |
-| Valid JSON | 12% (3/25) | 100% (25/25) | +88 |
-| schema | 100% | 100% | — |
-| anchors | 97% | 78% | −19 |
-| labels | 100% | 96% | −4 |
-| nonempty | 100% | 100% | — |
-
-The baseline failure was pure **format**: the stock 1B model rambled past the token
-limit without emitting parseable JSON on 88% of cases. After fine-tuning, every case
-emits valid, well-formed chunk JSON and **stops** (68–456 generated tokens vs the
-baseline hitting the 1536 cap). This validates the full mlx-bun pipeline:
-**train (ops.sdpa) → save adapter (hot-swap format) → `AdapterManager.mount` → generate.**
-
-## Why not e4b (the original target)
-
-The original plan targeted `gemma-4-e4b-it-OptiQ-4bit`. The chunk data is long
-(median 4092 tokens, 0% ≤2048, 98% ≤8192), so it needs 4–8K training context. On
-this **32 GB M1 Max**, neither mlx-bun NOR the optiq reference can train a 7.4B-param
-model at that length: I ran optiq's own `train_lora` on e4b and it completed only at
-seq 512 (peak 18.24 GB, degenerate — 0 response tokens survive truncation) and
-**crashed at 2048/4096/8192** with `[METAL] Command buffer execution failed: Impacting
-Interactivity` — a memory-pressure GPU stall. It's a hardware ceiling, not a code gap.
-Pivoted to **MiniCPM5-1B** (24 layers, head_dim 128, full attention), which fits 4–8K
-training comfortably. See [[e4b-lora-training-seqlen-ceiling]].
-
-## Correctness correction: ops.sdpa vs flash
-
-Earlier in the session I concluded mlx-bun's `ops.sdpa` had a wrong dK gradient and
-built/routed a flash kernel as the "correct" training attention. **That was backwards.**
-A clean check (`scripts/sdpa-vs-manual.ts`) against **mlx-autograd of a plain
-matmul/softmax forward** (the unimpeachable ground truth) shows:
-
-```
-dQ:  sdpa-vs-manual= 0.00%   flash-vs-manual= 0.06%
-dK:  sdpa-vs-manual= 0.00%   flash-vs-manual= 106.01%   ← flash is the broken one
-dV:  sdpa-vs-manual= 0.00%   flash-vs-manual= 0.02%
-```
-
-`ops.sdpa` is correct (and fast — one fused kernel). My flash dK kernel is the buggy
-one; the earlier triangulation was fooled by a hand-derived FA-2 backward reference that
-shares the dK formula with the flash kernel. The trainer now defaults to `ops.sdpa`
-(flash opt-in via `MLX_BUN_TRAIN_ATTN=flash`). ops.sdpa's only downside is O(seq²)
-backward memory — fine at ≤4096 for a 1B model (crashes ~8192 on long examples).
-
-## Caveats (honest)
-
-- **The benchmark metric is intrinsic** (JSON validity / schema / anchor resolution /
-  label quality), not segmentation-F1 vs gold boundaries. The +79.8 reflects "produces
-  valid, well-formed output," which is exactly what was broken.
-- **anchors 97%→78%**: the fine-tuned model occasionally emits an anchor UUID that needs
-  prefix-repair; one 47K-token case failed to resolve (score 50).
-- **Under-segmentation**: the fine-tuned model emits 1 chunk on most conversations. The
-  training data was filtered to ≤4000 tokens (730/1375 examples) for speed, biasing
-  toward shorter single-topic conversations. Training on the full data at 8192 (or with
-  the flash dK fixed for O(L) memory) would likely improve multi-chunk segmentation.
-
-## Reproduce
-
-```bash
-# baseline
-bun scripts/chunk-eval.ts                                    # 11.89/100
-# filter data, fine-tune, re-measure
-MAXTOK=4000 bun scripts/chunk-filter.ts
-DATA=~/.cache/mlx-bun-finetunes/chunk-data-le4000 SEQ=4096 ITERS=300 \
-  ADAPTER=~/.cache/mlx-bun-finetunes/minicpm5-chunk-final bun scripts/chunk-finetune.ts
-ADAPTER=~/.cache/mlx-bun-finetunes/minicpm5-chunk-final bun scripts/chunk-eval.ts   # 91.70/100
-```
-
-Training: 300 iters, seq 4096, rank 16, lr 2e-4, ops.sdpa, ~26 min, peak 25.47 GB, val
-loss flat ~0.19 (no overfit). Scripts: `chunk-eval.ts`, `chunk-filter.ts`,
-`chunk-finetune.ts`, `sdpa-vs-manual.ts`.