# Sortformer CoreML — v3 (BNNS-fixed rebuild)

Rebuilt streaming Sortformer diarization models (NVIDIA `diar_streaming_sortformer_4spk-v2.1` / `-v2`),
converted to CoreML. **v3 fixes the BNNS graph-compile crash** that affected the earlier root-level models.

## Why v3

The earlier models had `chunk_pre_encoder_embs_out` as **both an input and an output** of the head
sub-model, which the macOS 26 / newer BNNS graph compiler rejects:

```
BNNS Graph Compile: Function main has tensor chunk_pre_encoder_embs_out as both an input and output.
```

That bug was a toolchain artifact (torch 2.9.x tracing folded the identity op that kept input/output
distinct). v3 is rebuilt on torch 2.7 + coremltools 9.0, producing a clean head (no alias), verified to
load and predict on `computeUnits=.all` (ANE/GPU) and numerically matched to the PyTorch reference
(speaker-argmax agreement 100%).

## Contents

```
v3/
  fp16/         full-precision (229–243 MB/variant) — default quality
  palettized/   6-bit kmeans-LUT weight palettization (93–99 MB/variant)
```

7 variants in each (Swift `ModelNames.Sortformer.Variant`):

| File | Config | chunk_len | latency | use |
|------|--------|-----------|---------|-----|
| `Sortformer_v2.1` / `_v2` | fast / default | 6 | ~0.48 s | low-latency streaming |
| `SortformerEfficient_v2.1` | efficient | 25 | ~2 s | higher-throughput streaming (~4× RTFx of default) |
| `SortformerNvidiaLow_v2.1` / `_v2` | balanced | 6 | ~0.48 s | larger FIFO for stability |
| `SortformerNvidiaHigh_v2.1` / `_v2` | high-context | 340 | ~27 s | offline / best throughput |

## fp16 vs palettized

6-bit palettization (matches Argmax's speakerkit recipe; GPU-safe LUT, **not** int8 which crashes MPSGraph):

- **Size:** ~2.5× smaller (e.g. highContext 243 → 99 MB) — fixes RAM-driven crashes on older devices.
- **Speed:** unchanged (same per-call latency on GPU).
- **Accuracy:** +0.9 pp DER avg on full AMI-SDM (16 mtg, forced-alignment GT, collar 0.25) — most meetings
  +0.0–1.0 pp, a few +2–4 pp. Use fp16 if you need the last ~1 pp; palettized to fix RAM / shrink download.

## Reference DER (AMI-SDM, forced-alignment GT, collar 0.25, M5 Pro)

| Variant | DER | RTFx |
|---------|-----|------|
| highContext offline (tuned) | ~26.5% | ~860× |
| default streaming (0.48 s) | ~29.0% | ~49× |

Absolute DER is high because AMI-SDM (single distant mic, 4 speakers) is a hard condition; same NeMo weights
as other Sortformer ports, so accuracy is model-equivalent.