Qwen text encoder: sequence exceeds Metal threadgroup memory with reference video + ref images (≥150 frames)

## Error

```
h3: Qwen layer 0 causal GQA failed: causal attention sequence 10231 exceeds threadgroup memory
```

Generation aborted (exit 1) at the **very beginning** of the text encoder (layer 0, step 0/50), before denoising starts.

## Trigger condition

Combining a **reference video** (`--ref-silent-video`, ~7-8s, 1080×1920) with **multiple reference images** (`--ref-image` × 3-5) and **≥150 frames** (~6.25s at 24fps). Long prompt (~100 characters) further increases the sequence count.

The **dominant factor is the reference video** (~7500 tokens for 7.7s). Reference images contribute negligibly (~300 tokens total for 5 images). The apparent limit is **~7700 tokens** (32KB threadgroup / 4 bytes per float).

## Systematic single-variable experiments

| Combination | Frames | Result | Sequence |
|---|---|---|---|
| 5 images + video + long prompt | 22 | ✅ pass | — |
| 5 images + video + long prompt | 90 | ✅ pass | — |
| 5 images + video + long prompt | 120 | ❌ fail | 8199 |
| 5 images + video + long prompt | 160 | ❌ fail | 10231 |
| **video only** + long prompt | 160 | ❌ fail | 8184 |
| **5 images only** + long prompt | 160 | ✅ pass | — |
| 5 images + video + **short** prompt | 160 | ❌ fail | 8510 |

Downscaling all 5 images to 896px → sequence unchanged (10231). Vision token count from Qwen appears fixed-per-image, not resolution-dependent.

## MPS fallback is not safe

The existing `H3_MPS_GQA=1` fallback (`h3_gpu_gqa_mps`) fails differently: it **causes a system-wide watchdog panic** (machine freeze, 5 verified occurrences) during the `denoise` stage with these large sequences. This was tested on both M5 Max 128GB and M3 Ultra 96GB (macOS 26.4.1/26.5.2). I had to revert my MPS fallback patch because it was worse than the original error.

## Current workaround (in our webapp)

- Auto-cap frames to **90** when a `--ref-silent-video` is present (sequence stays under ~7700)
- Auto-downscale all `--ref-image` to 896px before passing to h3 (solves the jetsam memory explosion: original 2400px images × 5 caused h3 to consume 168GB unified memory on a 96GB machine)
- **Hard-block** the "reference video + reference images" combination: even with the above guards, the denoise stage triggers watchdog panics with ≥2 images + 1 video (4 verified panics, including after a fresh reboot)

90-frame cap (~3.75s at 24fps) is quite limiting — users needing longer videos must generate multiple short clips and stitch them.

## Suggested fix directions

The root cause is `h3_gpu_gqa_causal_bf16` storing the **entire causal attention score row** (sequence × 4 bytes) in threadgroup memory in a single dispatch:

1. **Tiled GQA**: split the score computation across multiple dispatches with intermediate global memory, requiring only `tile_size × sizeof(float)` of threadgroup memory per dispatch.

2. **Reference video token compression**: the Qwen encoder produces ~1000 tokens per second of video — pooling/down-sampling these before feeding them into the text encoder would directly reduce sequence length without losing semantic content.

---

*Note: PR [#44](https://github.com/antirez/h3.c/pull/44) (16-byte alignment fix) is orthogonal — it addresses a Metal validation-layer assertion, not this capacity limit.*

## Environment

- h3.c: latest `main` (commit `8b7d5b9`)
- macOS 26.5.2 / M5 Max 128GB; macOS 26.4.1 / Mac Studio M3 Ultra 96GB
- Model: `MiniMax-AI/MiniMax-H3` (FL2VA checkpoint)
- Metal: `maxThreadgroupMemoryLength` = 32768 bytes

The issue identifies h3_gpu_gqa_causal_bf16 as the failing entry point and contrasts it with h3_gpu_gqa_mps; begin by locating those implementations and the denoise path. Reproduce the reported long-sequence case, then evaluate the tiled GQA and reference-video token-compression directions. Done should avoid the threadgroup-memory failure without relying on the unsafe MPS fallback.

Scritto dal modello di indicizzazione a partire dal testo della issue.