modelbeast/docs/FABLE_STARTUP_PACKET.md

# FABLE STARTUP PACKET — TRELLIS.2 + Hunyuan MLX build

> Read this once, top to bottom, before touching anything. It consolidates five prep-only recon
> passes (all inspection, **no GPU job launched, no installs, no production code touched**). Every
> path here is on **m3ultra** (`m3ultra@100.89.131.57`) unless noted. Where a recon was uncertain,
> it says **(UNVERIFIED)** — do not treat those as facts.
>
> Two baselines you will see quoted, do **not** conflate them:
> - **124.8 s** = our current torch-MPS *production* path on the anatomy bench (the number to beat).
> - **1004.6 s / 3.09M tris / 17.82 GB** = PR#175's reported **MPS** E2E on an M4 Max 36 GB
>   (different hardware, different mesh size, and NOT the MLX path). Not comparable to 124.8 s.

---

## 1 · STATE ON DISK — what prep already put on the box

Everything below already exists. **Do not redo it.**

### Staging dir: `~/Documents/trellis2-mlx-staging/`
| Path | What it is | Status |
|---|---|---|
| `trellis2-apple/` | Full clone of `pedronaugusto/trellis2-apple`, HEAD `6055b86` (2026-04-22), full 17-commit history deepened, MIT. 39 MB. **Nothing installed.** | read-only clone |
| `prep_local_draft.py` | 9638 B preprocessing draft (§4). `py_compile` OK under rmbg venv. Not wired in. | draft, do-not-run-yet-in-prod |
| `convert_weights.py` | Draft torch→MLX weight converter (§5). Loads ~14 GB — **do not run now.** | draft |
| `parity_dump_hook.py` | Draft monkeypatch fixture dumper (§5). Loads models+GPU — **do not run now.** | draft |
| `inspect_safetensors.py` | Stdlib safetensors header parser (used to produce the §5 layout table). | utility, safe |
| `__pycache__/` | Compile artifacts from the py_compile checks. | ignore |

### Venvs & operators already on the box (do not reinstall)
- **rmbg venv** `~/Documents/MODELBEAST/venvs/rmbg/` (uv, CPython 3.11). Has torch 2.13.0, torchvision 0.28.0, transformers 5.13.1, timm, kornia, einops, safetensors, hf_hub, Pillow, numpy, scipy, scikit-image. **No** cv2 / rembg / birefnet-pkg / upscaler. Installed by `~/Documents/MODELBEAST/scripts/install_rmbg.sh`.
- **`bg_remove_local` operator** `~/Documents/MODELBEAST/server/operators/bg_remove_local/run.py` — runs RMBG-2.0, pins `venvs/rmbg/bin/python`.

### Weights already in HF cache
- `microsoft/TRELLIS.2-4B` — `~/.cache/huggingface/hub/models--microsoft--TRELLIS.2-4B/snapshots/af44b45.../ckpts/` (7 of 8 pipeline modules).
- `sparse_structure_decoder` (8th module) comes from `microsoft/TRELLIS-image-large` snapshot `25e0d31.../ckpts/ss_dec_conv3d_16l8_fp16.safetensors`.
- `briaai/RMBG-2.0` — 844 MB, present.
- Hunyuan MLX converted weights — `~/.cache/huggingface/hub/models--dgrauet--hunyuan3d-2.1-mlx/`.
- **NOT present:** `ZhengPeng7/BiRefNet` (Apache/MIT, ~900 MB) — must be fetched once for the commercial preprocessing path.

---

## 2 · TRELLIS.2 MLX — ADOPT-vs-BUILD DECISION

### VERDICT: **ADOPT** the Jourloy PR#175 `mlx_backend/` tree as the MLX vendor base. Do **not** build from scratch, and do **not** wholesale-swap the 124.8 s production path — gate any promotion on a measured A0 run.

**Why PR#175 and not the already-cloned `trellis2-apple`:** the two share the *identical* 16-file / ~3,114-LOC `mlx_backend/` (same author — `pedronaugusto` co-authors PR#175). PR#175 (`Jourloy/TRELLIS.2`, opened 2026-07-17, +7,419/−258 across 62 files) is a **strict superset**: same MLX modules **plus** the correctness oracle we lack (`tests/test_mlx_parity.py`, 4 unit parities at rtol 1e-6–3e-5), a backend probe/fallback resolver (`trellis2/backends.py`), a full CLI (`scripts/generate_asset.py`), macOS setup/probe/weight tooling, and the `--backend mlx-experimental` flag so optimized modules drop in behind a flag while MPS keeps E2E green. So PR#175 dominates `trellis2-apple` on every axis. Keep `vendor/trellis-mac` (shivampkumar, no MLX) only as the torch-MPS E2E oracle.

**Why NOT a blind swap (both recons agree):** the stock MLX path will very likely **regress** as-is, for two structural reasons:
- **Un-fused sparse conv.** `mlx_backend/sparse_conv.py` is the naive gather→matmul→sum, 512 MB-chunked — it does **not** use the `flex_gemm` Metal fast path, which our recon measured at **1.72× faster** on this shape. This is the ~10× crux, reimplemented slow in MLX.
- **CPU host bounce.** `mlx_backend/adapters.py` does `torch→numpy→mx→numpy→torch` on **every** model call — per sampler step (default ~12 steps × 3 stages × 2 for CFG) and at every sparse-conv boundary. The upstream torch pipeline still owns sampling orchestration + mesh extraction; MLX models are injected as adapters.
- **Zero benchmarks in either repo.** No evidence the MLX path beats — or matches — 124.8 s. Do **not** benchmark the stock `--backend mlx-experimental` path as representative; it's a correct scaffold, not a fast path.
- **Parity gap in `trellis2-apple`, closed by PR#175.** `trellis2-apple` has NO MLX parity test; PR#175 adds `test_mlx_parity.py`. Adopt PR#175 specifically to get that oracle.

**What is safe to reuse directly (low risk, no sparse-conv tax):** `dinov3.py`, `norm.py` (two-pass LayerNorm32 PT-parity), `rope.py` (interleaved complex-mul parity), `transformer_block.py`, `structure_decoder.py` (dense `mx.conv_general`). **Rewrite/replace first:** `sparse_conv.py` (fused Metal gather-GEMM-scatter), the `adapters.py` boundary (move the Euler/CFG loop *inside* MLX to kill the bounce), `sparse_ops.py` (numpy `np.where` host sync in upsample).

**Known smells to verify before trusting a module (UNVERIFIED in recon):** (1) `structure_decoder.py` `_remap_structure_decoder_weights` is a **no-op** feeding dict-stored conv weights — untested whether `nn.Module.load_weights` populates plain-dict children; (2) hand-rolled softplus/quad_lerp in `vae_decoders.py:1204` FlexiDualGrid decode; (3) full **non-windowed** attention (`attention.py`) may diverge from upstream windowed sparse attn; (4) docstrings claim "custom Metal kernel" but code is plain MLX built-in SDPA — **misleading comments, no hand-written kernel** anywhere in `mlx_backend/`. `mx.compile` *is* already applied to the DiT block loop and structure-decoder graph (with try/except fallback).

**Upstreamability: LOW — plan to fork, not to depend on merge.** PR#175 is unmerged, CLA-gated, 0 reviews, only the CLA bot has commented. Issue #74 (non-CUDA backends) has no maintainer response. Vendor it and **pin the commit**; basing our kernels on #175's layout keeps our work a clean, eventually-upstreamable diff.

### First 3 concrete commands
```bash
# 1. Get the superset fork (has test_mlx_parity.py; the cloned trellis2-apple does not)
cd ~/Documents/trellis2-mlx-staging && \
  git clone https://github.com/Jourloy/TRELLIS.2 trellis2-jourloy && cd trellis2-jourloy

# 2. Run the correctness ORACLE first — cheap, no heavy GPU E2E, validates
#    LayerNorm32 / apply_rope / SDPA / MlxSparseConv3d vs torch before any gen.
#    (Needs an isolated venv with mlx + the deps; test no-ops off-Mac via importorskip.)
pytest tests/test_mlx_parity.py -v

# 3. The single GPU job — the measured A0 gate. Isolated venv, the 3 pedronaugusto
#    Metal pkgs --no-build-isolation, reuse the existing TRELLIS.2-4B weights.
#    ONE gen, anatomy bench, seed 42, 1024-cascade; compare wall-clock to 124.8 s
#    and quality to the fal renders in ~/Documents/trellis2-bench/.
pip install -r requirements_macos.txt --no-build-isolation   # torch>=2.11, mlx>=0.31, mtlgemm/mtldiffrast/cumesh
# then: scripts/generate_asset.py --backend mlx-experimental --seed 42 --pipeline-type 1024_cascade <anatomy_img>
```
**Decision rule after A0:** within ~1.5× of 124.8 s AND parity holds → patch-and-tune, adopt as production path behind the flag. If 3–5× slower (the likely outcome given un-fused conv + host bounce) → keep torch-MPS in `nodes.json` for production, lift `mlx_backend/` as the module-by-module scaffold, and land the kernel rewrites (fused Metal sparse conv → in-MLX sampler loop → parity fixtures) in that order. Also: fork to Gitea `monster/trellis2-mrp-mlx`, keep `origin` upstream for rebases (house pattern).

---

## 3 · HUNYUAN TUNE — edit-target list (ready to apply)

All paths under `~/Documents/MODELBEAST/vendor/hunyuan3d-mlx/`. Static inspection only (GPU was busy). Ship each edit with a PT-parity test (≤1e-4 rel) + a BENCHMARKS.md row.

**Headline: `head_dim` pad-to-64 does NOT apply** — every attention site is already a 64-multiple (64 or 128). The corridorkey padding win is **out of scope**; do not build pad wrappers. `mx.compile` count in the whole MLX port = **0** — that is the entire opportunity surface (corridorkey precedent: fixed-shape compile = 1.47× M3 / 1.11× M1, bit-identical, an 8-line change `vendor/corridorkey-mlx@4c660df`).

| # | Target | file:line | Change | Expected | Risk |
|---|---|---|---|---|---|
| 1 | **Paint UNet forward** (called 3×/step, ~15 steps) | `hy3dpaint/hunyuanpaintpbr_mlx/unet/unet_mlx.py:158` (`__call__`); loop `inference.py:254`, calls `:314,328,339`; steps default `inference.py:72` | Wrap forward in `mx.compile` (add `compile=True` ctor flag, `self._fwd = mx.compile(self._forward_impl)`). No MoE / no `.item()` → compile-safe. | ~1.1–1.3× on ~112 s paint | low; last chunk may have fewer views → 1 retrace (pad final chunk to fixed `n_chunk`) |
| 2 | **DiT dense blocks 0–14** (100 forwards = 50 steps × CFG 2) | `hy3dshape/.../denoisers/hunyuandit_mlx.py:379` (`HunYuanDiTBlock.__call__`); driver `pipeline_mlx.py:160`, loop `:143`, steps `:95` | `mx.compile` the dense-block forward (fixed shape `(B∈{1,2},4097,2048)`). Leave 6 MoE blocks (15–20) eager until #3. | ~1.1–1.3× on ~149 s shape (dense = 15/21 depth) | low |
| 3 | **MoE `.item()` sync tax** (PREREQUISITE for full-DiT compile) | `mlx_arsenal/moe/moe.py:110` — `if not mx.any(w>0).item(): continue` | Remove the `.item()` early-skip; always run all 8 experts, let `w*out` (w=0) zero them. Deletes ~4800 syncs/gen AND makes the block static → unlocks compiling whole `HunYuanDiTPlain.__call__` (`hunyuandit_mlx.py:514`). | −4800 syncs/gen + unblocks #2 to 21/21 blocks | **`mlx_arsenal` is a SHARED dep** — do NOT edit in place. Vendor a local `MoELayer` subclass override, or upstream to arsenal. `MoEGate` routing is already static/compile-clean. |
| 4 | **Paint DINOv2-giant attn** (manual→fused, once/gen ×40 layers) | `hy3dpaint/hunyuanpaintpbr_mlx/dino_mlx.py:58–60` (manual `q@kᵀ` softmax, fp32) | Replace with `mx.fast.scaled_dot_product_attention(q,k,v,scale=self.scale)` + existing transpose/reshape. `head_dim=64` hits fast path; fp32 supported. | one-shot; kills N² materialization (fused d64 = 1.51× on M3) | low |
| 5 | **Paint VAE decode** (once/gen) | `hy3dpaint/hunyuanpaintpbr_mlx/vae_mlx.py` decode; manual softmax `:81–84` (`head_dim=512`) | `mx.compile` the decode (fuses the softmax). Prefer compile over `mx.fast` swap — `head_dim=512` likely **misses** MLX fused sdpa (kernel ≤256), would fall back anyway. | one-shot, minor | low |
| 6 | **(CONDITIONAL) Shape-VAE geo cross-attn** | `hy3dshape/.../autoencoders/model_mlx.py:206` (`CrossAttentionBlock`); driver `_query_sdf_volume` `:506`, chunk `:510` | `mx.compile` per-chunk cross-attn (fixed 10000-query × 4096-latent). Only final ragged chunk retraces. | small–moderate | **Verify** decode isn't dominated by the Metal SDF/marching-cubes step before investing |

**Rasterizer: NO opportunity.** `hy3dpaint/DifferentiableRenderer/mesh_render_mlx.py` is a thin adapter over the `mlx_arsenal.rasterize` Metal kernel (already fused); Python side is glue + numpy round-trips. Skip for both levers.

**DTYPE: keep fp16 as the global default.** Port is fp16 everywhere (one deliberate exception: paint DINOv2-giant upcast to **fp32** in `load_model.py` — accumulation-depth fix, not a format preference). bf16 buys **zero** speed on M3/M4/M5 (equal TFLOPS) and is **~20% slower on M1** — and M1 Ultra (`johnking@100.91.239.7`) is the second production lane and hunyuan's headline "runs on M1" selling point. Do NOT globally switch to bf16. (This is the opposite of the TRELLIS.2 port, where bf16 range matters.)

**Execution order:** 1 → 3 → 2(dense, then whole-forward once 3 lands) → 4/5/6 (one-shot cleanups last).

---

## 4 · PREPROCESSING — draft, plug-in point, license, TODOs

**North star** (`meshgod/config.py:65-70`, `STYLE_3D`): a single isolated object, centered, fully visible (no crop); plain flat neutral light-gray bg; even soft shadowless studio light; neutral symmetrical A-pose.

**Draft: `~/Documents/trellis2-mlx-staging/prep_local_draft.py`** (9638 B, `py_compile` OK under rmbg venv, uses only PIL/numpy/torch/torchvision/transformers, not wired in). Signature `prep_local(in_path, out_path, commercial=False)`:
1. `_maybe_upscale` — if `min(w,h) < 1024`, Lanczos up to short-edge 1024, on RGB, **BEFORE** matte.
2. `_matte` — BiRefNet if `commercial` else RMBG-2.0; real inference mirroring `bg_remove_local/run.py` exactly (`Resize((1024,1024))→ToTensor→Normalize(imagenet)`; `model(inp)[-1].sigmoid()`; → `L` mask); model cached per-process.
3–7. `_compose_center_square` — composite onto `NEUTRAL_GRAY=(200,200,200)` RGB (not alpha) via mask, tight bbox crop from thresholded mask (`MASK_THRESH=8`), center, square-pad to `SUBJECT_FRAC=0.88`, resize 1024 Lanczos.

Run (GPU-free-time session): `venvs/rmbg/bin/python prep_local_draft.py --in X --out Y [--commercial]`.

**Order lesson (verified live 2026-07-13):** upscale FIRST, matte LAST — SeedVR strips alpha (RGBA→RGB), so matting first gets flattened. Draft honors this.

**Plug-in point:** `~/Documents/meshgod/meshgod/server.py:350` (local lane, `server.py:346-350`):
```python
paths[i] = stage3_reconstruct.mb_bg_remove(p, pp)   # <- replace with prep_local(p, pp, commercial=...)
```
`prep_local` supersedes `mb_bg_remove` (does the full north-star transform in-process on rmbg venv, not just a farm-round-trip cutout to RMBG-2.0). Secondary optional precede-point: fal lane `server.py:359` (run `prep_local` locally, then skip fal `clean_bg`). Companion anchors to update: `config.py:143` (`BG_REMOVE_MODEL`), `web/index.html:229` label.

**LICENSE RULE (hard):**
| Model | HF id | License | Commercial ship? | Local? |
|---|---|---|---|---|
| RMBG-2.0 | `briaai/RMBG-2.0` | CC-BY-NC 4.0 | **NO** — non-commercial / personal one-offs only | Yes (844 MB) |
| BiRefNet | `ZhengPeng7/BiRefNet` | MIT | **YES** — safe for game/product assets | **No — fetch once** |
Default for anything game-bound = **BiRefNet**. RMBG-2.0 stays default only for local/personal one-offs. (Note: RMBG-2.0 *is itself* a BiRefNet-architecture checkpoint — same code, differs only in weights + license.)

**Open TODOs (in draft):**
1. **BiRefNet weights fetch** — first `--commercial` run must `huggingface-cli download ZhengPeng7/BiRefNet` (~900 MB, MIT, ungated). Until then only RMBG-2.0 works.
2. **Better upscaler** — PIL Lanczos is the zero-dep default (never invents detail). Swap for local faithful SR: MODELBEAST `seedvr2_upscale` operator (weights present: `models--numz--SeedVR2_comfyUI`) or Real-ESRGAN. Must stay BEFORE `_matte`.
3. **Lighting/exposure normalization (D2)** — optional auto-levels + gray-world WB on RGB before matte (fights TRELLIS dark-patch mode). Stubbed so v0 silhouette stays identical to the fal path for A/B.
4. **Tunables to A/B once GPU free:** `NEUTRAL_GRAY=(200,200,200)`, `SUBJECT_FRAC=0.88`, `MASK_THRESH=8` (all chosen from north-star text).
5. **Wire-in / flag plumbing** — add `req.commercial`, replace `server.py:350`, expose in `web/index.html`, decide local-lane upscale policy (`server.py:511` currently forces `upscale=False`).

---

## 5 · WEIGHT CONVERSION + PARITY

Drafts in staging: `convert_weights.py`, `parity_dump_hook.py`, `inspect_safetensors.py`. Both loaders touch ~14 GB / GPU — **run in a later session, not now.**

### Conversion pattern (from Hunyuan's `convert_realesrgan.py:34-58`)
`torch.load(weights_only=True)` → **keys pass through unchanged** (module-tree names already match MLX list indexing, no remap dict) → **the only tensor mutation is conv transpose** (PyTorch `(O,I,H,W)`→MLX channels-last `(O,H,W,I)`, detected by `ndim==4`; generalizes to 5D Conv3d as `(0,2,3,4,1)`) → `mx.save_safetensors`.

### TRELLIS.2-4B layout — the transpose decision (resolved from source, not guessed)
There are **two Conv3d conventions on disk; the correct action is opposite for each:**
| pipeline key | file (ckpts/) | dtype | conv action |
|---|---|---|---|
| sparse_structure_flow_model | ss_flow_img_dit_1_3B_64_bf16 (640 t, 2.58 GB) | BF16 | none (pure Linear DiT) |
| **sparse_structure_decoder** (from TRELLIS-image-large) | ss_dec_conv3d_16l8_fp16 (74 t) | F32/F16 | **20 dense Conv3d → TRANSPOSE `(0,2,3,4,1)`** |
| shape_slat_flow_model_512 / _1024 | slat_flow_img2shape_dit_1_3B_{512,1024}_bf16 | BF16 | none |
| shape_slat_decoder | shape_dec_next_dc_f16c32_fp16 (292 t, 948 MB) | F16 | **40 sparse Conv3d → NO transpose (already channels-last)** |
| tex_slat_flow_model_512 / _1024 | slat_flow_imgshape2tex_dit_1_3B_{512,1024}_bf16 | BF16 | none |
| tex_slat_decoder | tex_dec_next_dc_f16c32_fp16 (284 t, 948 MB) | F16 | **40 sparse Conv3d → NO transpose** |

- **`ss_dec_conv3d` = dense `nn.Conv3d`**, on-disk `(Co,Ci,Kd,Kh,Kw)` → needs the 5D transpose. All 20 5-D tensors.
- **`*_next_dc` = o-voxel `SparseConv3d`**, on-disk **already** `(Co,Kd,Kh,Kw,Ci)` → **load verbatim, NO transpose.** Proven by source: `conv_flex_gemm.py:33-34` and `conv_none.py:40-41` both `permute(0,2,3,4,1)` at construction, and the checkpoint was saved from that path. MLX forward must reshape `(Co,Kd,Kh,Kw,Ci)→(Kvol,Ci,Co)` for the per-offset gather-GEMM (contract at `conv_none.py:106`).
- **All 5 DiTs: pure `nn.Linear` (ndim 2) + attention + RMSNorm** — zero conv, zero transpose; only a BF16→target-dtype cast (**fp16 on M1** per fleet matrix). Block arch (30 blocks, identical across all 5): hidden 1536, 12 heads × **head_dim 128**, MLP 8192, cross-ctx 1024, adaLN-zero `modulation [9216]`. **head_dim 128 ≥ 64 → fused-SDPA fast path applies, NO padding.** Keep fused `to_qkv [4608,1536]` / `to_kv [3072,1024]` packed and slice in-kernel.

`convert_weights.py` encodes this per-file policy (dense-transpose for `ss_dec_conv3d`, keep for `*_next_dc`, passthrough for DiTs) with a kernel-dims sanity-assert. Uses `safetensors.torch.load_file` + `mx.save_safetensors`. Publish converted safetensors to Gitea/HF so re-runs skip conversion.

### Parity fixture plan
Stage boundaries all live in `run()` of `trellis2/pipelines/trellis2_image_to_3d.py:488-596`; each fixture = the **return value of a sub-method** → clean non-invasive monkeypatch, zero vendor edits. `parity_dump_hook.py` monkeypatches the 7 sub-methods, `torch.save`s each return to `$TRELLIS2_DUMP_DIR/<name>.pt`, then `runpy`-delegates to vendor `generate.py`. Unset `TRELLIS2_DUMP_DIR` → patches are no-ops.

| # | Fixture | Return / hook (file:line in trellis2_image_to_3d.py) | Shape/notes |
|---|---|---|---|
| A/A′ | Image-cond embed 512 / 1024 | `get_cond` return `:177` (called `:539`/`:540`; disambiguate by `resolution` arg → `cond_512.pt`/`cond_1024.pt`) | `[1,N,1024]` DINOv3 ViT-L/16 tokens; neg = zeros_like |
| B | Sparse-structure coords | `sample_sparse_structure` return `:235` (call `:542`) | int `[M,4]` (b,x,y,z); subs: `z_s [1,C,64³]` `:219`, occupancy `:227` |
| C | Shape-SLat latent | `sample_shape_slat_cascade` `:364` (`:567`) / non-cascade `:275` | SparseTensor `feats [K,32]`+`coords [K,4]`, **un-normalized** |
| D | Tex-SLat latent | `sample_tex_slat` `:432` (`:574`) | SparseTensor `feats [K,32]`, un-normalized, sampled w/ `concat_cond=shape_slat` |
| E1 | Decoded shape mesh + subs | `decode_shape_slat` `:389` (`:470`) | `List[Mesh]` + `List[SparseTensor]` |
| E2 | Decoded tex voxels | `decode_tex_slat` `:453` (`:471`) | SparseTensor `feats [K,6]` PBR (base_color 0:3, metallic 3:4, roughness 4:5, alpha 5:6) |
| E3 | Final mesh | `decode_latent` `:486` (`:592`) | `MeshWithVoxel`: `.vertices`,`.faces`,`.coords`,`.attrs`,`.voxel_size`,`.origin` |

Invocation (later, GPU): `TRELLIS2_DUMP_DIR=~/Documents/trellis2-bench/parity_fixtures python parity_dump_hook.py <anatomy_img> --pipeline-type 1024_cascade --seed 42 --texture-size 2048`. Fix `seed=42` (`torch.manual_seed` `:538`). Parity target **≤1e-4 rel**. Order cheap→expensive: A/A′ → B → C (the FlexGEMM sparse-conv crux) → D → E.

**GOTCHA:** C/D fixtures are captured **after** the std/mean de-normalization inside each sampler — the MLX port must apply the same `shape_slat_normalization`/`tex_slat_normalization` (32-dim mean/std, verbatim in `pipeline.json`) at the identical point or they diverge by a constant affine. `sample_tex_slat` also **re-normalizes** `shape_slat` back down (`:407-409`) before concatenating as `concat_cond` — replicate that round-trip.

---

## 6 · RECOMMENDED FIRST HOUR

1. **Read this packet + skim the 3 staging drafts** (`prep_local_draft.py`, `convert_weights.py`, `parity_dump_hook.py`) so you know what already exists. Do NOT re-clone `trellis2-apple` or re-inventory the rmbg venv.
2. **Clone the Jourloy superset** (`git clone https://github.com/Jourloy/TRELLIS.2 ~/Documents/trellis2-mlx-staging/trellis2-jourloy`) — §2 command 1. This is the vendor base; `trellis2-apple` lacks the parity oracle.
3. **Run the parity oracle** `pytest tests/test_mlx_parity.py -v` in an isolated venv (no heavy GPU E2E). Confirms LayerNorm32/RoPE/SDPA/SparseConv parity before trusting any module. This is the cheapest signal available and needs no A0.
4. **Fork to Gitea** `monster/trellis2-mrp-mlx`, keep `origin` upstream, **pin the PR#175 head commit** (it's unmerged/CLA-gated — do not depend on merge).
5. **Queue the single A0 GPU gate** for when the box is free — the ONE gen (anatomy bench, seed 42, 1024-cascade), compare to 124.8 s and the `~/Documents/trellis2-bench/` fal renders. Do not benchmark the stock MLX path as representative before the sparse-conv + adapter rewrites; expect a regression. Apply the §2 decision rule to the result.
6. **In parallel while GPU is busy** (all CPU/inspection, no contention): start Hunyuan **Target 1** (paint UNet `mx.compile`, §3) — cleanest, no blockers — with its PT-parity test; and/or dry-run `prep_local_draft.py` on a sample image under the rmbg venv (§4). Both are safe without the A0 result.
7. **Before any commercial preprocessing run:** `huggingface-cli download ZhengPeng7/BiRefNet` (§4 TODO 1) — otherwise only the non-commercial RMBG-2.0 path works.

**Prime directives:** parity-first (every kernel/edit ships a ≤1e-4-rel PT-parity test + a BENCHMARKS.md row); do NOT edit shared `mlx_arsenal` in place (§3 Target 3); default BiRefNet for game-bound assets (§4). The GPU A0 gen is the ONLY job that must wait for a free box.

---

## 7 · PREP ROUND 2 — VERIFICATION RESULTS (2026-07-19)

> Round-2 deep-verify of §1–§6 against **live source + on-disk safetensors headers**. Still
> inspection-only: **no GPU gen, no weights loaded, no prod code touched** (`py_compile` /
> header-parse / `pytest --collect-only`). Where §7 says CONFIRMED / CORRECTION / CONTRADICTION,
> **trust §7 over the original section.**

### 7.1 · ADOPT confidence — §2 stance → **CONFIRMED (HIGH)**

- **Clone DONE & pinnable.** `~/Documents/trellis2-mlx-staging/trellis2-jourloy/`, HEAD `754d403`
  (2026-07-17, Jourloy — matches "PR#175"), branch `main` **is** the PR head, tree clean.
  **Pin `754d403`** when forking.
- **Superset CONFIRMED — stronger than §2 claimed.** All 4 named artifacts present
  (`tests/test_mlx_parity.py`, `trellis2/backends.py`, `scripts/generate_asset.py`,
  `requirements_macos.txt`) **plus 4 extra test files** (`test_backend_fallbacks`,
  `test_background_preprocess`, `test_generate_asset`, `test_model_revisions`) **and** a new
  `trellis2/model_revisions.py` (HF revision pins). `mlx_backend` = 3172 LOC / 16 files, all
  `py_compile` clean.
- **`mlx_backend` numerics = `trellis2-apple` byte-for-byte.** `diff -rq`: 14/16 files identical;
  the only 2 that differ (`dinov3.py`, `pipeline.py`) add **HF revision-pinning / offline kwargs
  only — zero algorithm change.** §2's "identical mlx_backend" holds for numerics; jourloy is a
  strict superset (adds reproducibility plumbing on top of the same math).
- **`--backend mlx-experimental` cleanly wired** (`generate_asset.py:333`→`create_mlx_pipeline`,
  `:343`) behind a real **Metal-completeness probe gate** (`:170-187` → `backends.probe_metal_backends()`)
  that refuses a half-built stack instead of silently degrading. MPS stays the E2E path.
- **§2's 4 UNVERIFIED smells — now resolved:**

  | # | smell | round-2 disposition |
  |---|---|---|
  | 1 | structure_decoder no-op remap + dict-child conv load | **BENIGN by construction** (leaves pre-alloc'd `mx.zeros`; MLX param tree recurses dict/list; `load_weights(strict=True)` replaces leaves). Residual = exact key-name match only, **not** exercised by parity test → run a 2-sec non-GPU load smoke on ss_dec. |
  | 2 | vae_decoders softplus/quad_lerp | **BENIGN** — it's the numerically-stable softplus identity `max(x,0)+log1p(exp(-|x|))` = torch softplus. Real line `vae_decoders.py:196` (§2's `:1204` misattributed). Needs one decode-boundary fixture (E1/E2). |
  | 3 | non-windowed global attention | **NEEDS-RUNTIME-CHECK / POTENTIALLY REAL — highest-value item.** `_sdpa` does full global SDPA, **no window mask**. If upstream sparse flow is windowed → numeric divergence **and** O(N²) blowup at 1024-cascade. Tiny SDPA parity test validates the primitive, not the windowing semantics. **Verify vs upstream sparse-attn config before trusting shape/tex output.** |
  | 4 | "custom Metal kernel" docstrings | **CONFIRMED BENIGN (cosmetic)** — grep `metal_kernel|@mx.custom|.metal|kernel_source` over `mlx_backend/` = zero matches; code calls MLX built-in `mx.fast.scaled_dot_product_attention`. No hand-written kernel anywhere. |

- **Bonus (confirms §2):** sparse-conv perf crux is **REAL** — `sparse_conv.py` is naive
  gather→matmul→sum, 512 MB-chunked with a per-chunk `mx.eval` host sync, O(batch·D³) dense LUT,
  **no `flex_gemm`.** Matches the ~10× claim.
- **Confidence HIGH, not lowered:** nothing contradicted the ADOPT case; several findings
  strengthened it. Two items still need a *runtime* signal (smell 3 windowed-attn, smell 1
  dict-child load) — **both cheap, non-GPU; run before the A0 gate** (see 7.6).

### 7.2 · WEIGHT-CONVERT — transpose policy **VERIFIED vs disk**; converter has 2 bugs + 1 double-transpose contradiction

All 8 modules parsed **header-only** (no tensor bytes). Tensor counts match §5 **verbatim**
(640/640/640/640/640/292/284/74). Classifier found **ZERO ambiguous 5D tensors** — conventions
separate cleanly by shape.

| module(s) | dtype on disk | 5D convs | §5 policy | matches disk |
|---|---|---|---|---|
| 5 DiTs | **BF16** ×640 ea | 0 (215 Linear + 120 RMSNorm γ + 305 1D) | passthrough + dtype cast | ✅ |
| shape_dec / tex_dec | **F16** | 40 sparse ea, `(Co,K,K,K,Ci)` | **NO transpose** | ✅ |
| ss_dec | **F32×38 + F16×36** (2 F32 convs = `input_layer` + `out_layer.2`) | 20 dense, `(Co,Ci,K,K,K)` | **transpose `(0,2,3,4,1)`** | ✅ |

Transpose/convention classification in §5 holds on disk with **zero contradictions**. The
converter's `else: raise RuntimeError` on any unclassified 5D fails **loud**, so no silent
transpose-convention corruption is possible. **But three items must be fixed before any convert run:**

- 🔴 **BUG 1 (BLOCKING)** `convert_weights.py:59` `v.numpy()` on `torch.bfloat16` →
  `TypeError: unsupported ScalarType BFloat16`. **Crashes on all 5 BF16 DiTs** (the bulk of the
  pipeline); only survives on F16/F32 decoders. Loud, not silent. Fix: `v.float().numpy()`, or
  bit-reinterpret via `uint16`, or drop torch and use `mx.load`/`mx.save_safetensors`.
- 🟠 **BUG 2 (SILENT precision loss)** uniform `target_dtype` default `"bf16"` (`:46,81-83,94`)
  downcasts the decoders under default args: shape_dec/tex_dec **F16→bf16** (lose 3 mantissa bits,
  strictly worse + ~20% slower on M1); ss_dec's **2 F32 convs → bf16** (lose 16 bits on the most
  sensitive first/last SDF layers). **Contradicts §5's own dtype column (decoders = F16).** Fix:
  per-file dtype policy — DiTs→bf16 (fp16 on M1), decoders→preserve on-disk dtype.
- ⚠️ **CONTRADICTION — ss_dec DOUBLE-TRANSPOSE (silent-corruption risk).** §5/converter transpose
  ss_dec `(0,2,3,4,1)` **at load**. But the live MLX `structure_decoder.py` stores PT-format
  `(Co,Ci,kD,kH,kW)` **verbatim** and transposes **at forward-call time** (`conv3d`,
  `structure_decoder.py:14`). If the converter *also* transposes ss_dec, the decoder
  **double-transposes → garbage, silently.** **Resolve before converting:** convert ss_dec to
  PT-format on disk (NO load transpose) to match this file, OR move the transpose — pick exactly
  one, not both.
- **Port dependency (not the converter):** the MLX sparse-conv module MUST do the
  `(Co,K,K,K,Ci)→(Kvol,Ci,Co)` reshape at load (`conv_none.py:106`) or the `*_next_dc` decoders
  emit garbage. Converter correctly leaves this to the module.

### 7.3 · PARITY HOOKS — all 8 targets match live source, **ZERO corrections to the hook**

`parity_dump_hook.py` needs no edits. Source `trellis2_image_to_3d.py` is frozen at upstream
`75fbf01` (2026-06-05) + 1 benign local Mac cuda-guard (`:590-591`, downstream of every hook →
shifts no fixture line). **No post-packet drift.** Verified def/return/call (corrections to §5's
fixture table where it drifted):

| fixture | def | return | call-site(s) |
|---|---|---|---|
| get_cond (A/A′) | 164 | **181** (no-neg) / **183-186** (dict) | 539, 540 |
| sample_sparse_structure (B) | 188 | 235 | 542 |
| sample_shape_slat (C, non-cascade) | 237 | 275 | 547, 557 |
| sample_shape_slat_cascade (C′) | 277 | 364 (tuple `slat,hr_res`; `transform` takes `r[0]`) | 567, 579 |
| sample_tex_slat (D) | 391 | 432 | 551, 561, 574, 586 |
| decode_shape_slat (E1) | 366 | 389 | 470 |
| decode_tex_slat (E2) | 434 | 453 | 471 |
| decode_latent (E3) | 456 (`@torch.no_grad` :455) | 486 | 592 |

- **§5 CORRECTION:** the fixture table cites get_cond "return **:177**" — that line is the model
  call `cond = self.image_cond_model(image)`, **not** the return. Real returns are **:181 / :183-186**.
  Cosmetic (hook wraps by method name), but corrected.
- **Normalization gotcha CONFIRMED:** `shape_slat.pt` / `tex_slat.pt` captured **POST-denorm**
  (`:271-273`, `:360-362`, `:428-430`); tex sampler **re-normalizes** shape_slat back down at
  `:407-409` before `concat_cond` (:419). Port must replicate **both** the on-return de-norm and the
  denorm→renorm round-trip or fixtures diverge by a constant affine (silent ≤1e-4 failure).
- **UNCERTAIN — flag for exec session:** the hook's `__main__` `runpy`-delegates to vendor
  `generate.py`; that file's existence + CLI contract (`--pipeline-type` / `--seed` /
  `--texture-size`) was **not** checked. Verify before the dump run — a missing/renamed delegate
  means the hook installs but never fires.

### 7.4 · PREPROCESSING — BiRefNet **PRESENT**; D2 **IMPLEMENTED**

- **BiRefNet fetched (§1 "NOT present" TODO → DONE).** Cache:
  `/Users/m3ultra/.cache/huggingface/hub/models--ZhengPeng7--BiRefNet/snapshots/e2bf8e4460fc8fa32bba5ea4d94b3233d367b0e4/`.
  Size **424 MB** (`model.safetensors` = 444,473,596 B) — **§1/§4's "~900 MB" was high.**
  License **MIT** (README `license: mit`) — **§1/§4's "Apache-2.0" label was wrong.** Ungated, no login.
- **BiRefNet branch VERIFIED CORRECT (not stubbed):** RMBG-2.0 *is* a BiRefNet checkpoint (identical
  `auto_map`/`architectures`); forward `model(inp)[-1].sigmoid()` holds (`birefnet.py:2087/2090`);
  deps satisfied by the rmbg venv — **no cv2 needed** (`birefnet.py` has zero opencv refs). Only
  metadata was hardened; no code change.
- **D2 (light normalization) IMPLEMENTED** behind `normalize_light=False` (default **OFF** → v0
  silhouette stays A/B-identical to fal). New `_normalize_light` (`prep_local_draft.py:84`, PIL/numpy
  only): 1/99-pct auto-levels (flat-channel guarded) + gray-world WB; wired at step **1b** (after
  upscale, before matte — honors "matte LAST"). CLI `--normalize-light`. Functional smoke passed on a
  synthetic cast image.
- **Draft status:** `~/Documents/trellis2-mlx-staging/prep_local_draft.py` now 236 lines,
  `py_compile` OK + imports clean under rmbg venv. Signature
  `prep_local(in_path, out_path, commercial=False, normalize_light=False)`. **Staging only — still
  NOT wired into prod** (§4.5 wire-in + tunable A/B still open; need a GPU-free box for real A/B).

### 7.5 · MLX TEST VENV — oracle is **one command from running**

- **Venv READY:** `/Users/m3ultra/Documents/trellis2-mlx-staging/mlx-test-venv` (CPython **3.12.13**,
  uv). `mlx==0.32.0` (+`mlx-metal` 0.32.0), `torch==2.13.0` (prebuilt wheel, no compile),
  `numpy` 2.5.1, `pytest` 9.1.1. (safetensors/transformers/einops also installed but the oracle uses
  **none** of them — droppable for a leaner venv.)
- **Collects clean:** `--collect-only` → **4 tests collected, 0 import errors, 0 skips** (0 skips
  proves both `importorskip("mlx.core")` and `importorskip("torch")` succeeded on the box). **Not yet
  executed** — a real pass is still pending (low-risk caveat: `np.asarray(mlx_array)` under numpy
  2.5.1 surfaces only on the real run, not collection).
- **GPU/Metal?** 4 tests run microsecond ops on tiny tensors. MLX's default device is Metal (GPU), so
  they *touch* the GPU — but this is trivial library compute, **NOT a generation/inference job**, uses
  **no Metal build packages** (mtlgemm/mtldiffrast/cumesh) and no compile. Won't meaningfully contend
  for the box; force CPU with `mx.set_default_device(mx.cpu)` if ever desired.
- **EXACT run command** (use the venv python — a bare `pytest` uses the wrong interpreter, correcting
  §6/§2):
  ```
  ssh -n m3ultra@100.89.131.57 'cd ~/Documents/trellis2-mlx-staging/trellis2-jourloy && \
    ../mlx-test-venv/bin/python -m pytest tests/test_mlx_parity.py -v'
  ```
  Expected: **4 passed.** Baked tolerances: LayerNorm `2e-5`, RoPE `1e-6`, SDPA `3e-5`,
  SparseConv `2e-5` (all tighter than the ≤1e-4 target).

### 7.6 · UPDATED FIRST MOVES — deltas to §6

- **§6 step 2 (clone Jourloy): DONE** — `trellis2-jourloy/` @ `754d403`, tree clean. Skip the clone;
  just `git checkout 754d403` / pin it when forking.
- **§6 step 3 (run parity oracle): venv READY, oracle collects clean.** Now a one-command move — run
  the exact command in **7.5**. **Do NOT use the bare `pytest tests/...` from §6/§2** (wrong
  interpreter).
- **§6 step 7 / §4 TODO 1 (BiRefNet fetch): DONE** — skip. Present at the 7.4 path (424 MB, MIT).
- **NEW — before any weight-convert run (§5):** fix converter **Bug 1** (blocking bf16 crash) +
  **Bug 2** (silent decoder downcast), AND resolve the **ss_dec double-transpose** contradiction (7.2).
  Else DiT convert crashes, decoders lose precision, or ss_dec silently corrupts.
- **NEW — before the A0 GPU gate:** run the two cheap **non-GPU** runtime checks — (a) ss_dec
  dict-child weight-load smoke, and especially **(c) global-vs-windowed attention vs the upstream
  sparse-attn config** (the one smell that can be a true correctness bug, not just slow). Both fit
  CPU/inspection while the box is busy.
- **Unchanged:** §6 step 5 (single A0 GPU gate — seed 42, 1024-cascade, compare 124.8 s) is still the
  **only** job that must wait for a free box; §6 step 6 (Hunyuan Target 1) still safe in parallel.