---
license: apache-2.0
tags:
- executorch
- xnnpack
- pte
- on-device
- text-generation
base_model:
- Qwen/Qwen3.5-2B
---
# Qwen3.5-2B — ExecuTorch XNNPACK 8da4w + 8-bit embedding

`qwen3_5_2b_xnnpack_8da4w_e8.pte` (1483.7 MB)

- **Source**: Qwen/Qwen3.5-2B
- **License**: Apache-2.0
- **Quantization**: 8da4w linear + 8-bit embedding (`embedding_quantize: "8,0"`). The
  248320-row embedding is the reason: left in fp32 it is a gigabyte on its own, and at this
  vocabulary size that is most of what a small model weighs.
- **Export**: executorch 1.4.0 `export_llm`, static shape (seq_len=1), max_seq_length 2048,
  XNNPACK extended_ops
- **Config**: `llm_params/qwen3_5_2b_xnnpack_8da4w_e8.yaml`

## Verification (Mac arm64, 2026-08-21)

`llm_params/gen_static.py`, token-by-token prefill then greedy decode, a fresh process per
prompt so no answer is read through the previous one's cache:

| prompt | answer | decode |
|---|---|---|
| capital of France? | "The capital of France is **Paris**." | 33.25 tok/s |
| 日本の首都は? | 「日本の首都は**東京都**です。」 | 33.25 tok/s |
| 17 times 4? | opens a `<think>` block and works through it | 33.25 tok/s |

All three rates are from one run with nothing else on the machine, which matters more than
it sounds: the same file measured 16.6 tok/s while an export was running alongside.

Chat template: ChatML, bos 248045, eos [248046, 248044].

Not measured on a phone. The
[0.8B](https://huggingface.co/mlboydaisuke/Qwen3.5-0.8B-ExecuTorch) was measured at 10.5
tok/s on an iPhone 17 Pro, but on its earlier build, before
`use_sdpa_with_kv_cache` was turned on.

## The setting that doubles it

`use_sdpa_with_kv_cache: True`. Upstream's own `qwen3_5` config has it off with no reason
given, while the equally hybrid `lfm2` config has it on. Measured on the 2B in one run:
**8.20 tok/s off, 16.64 on**.

That it is only a kernel swap and not a different model was checked three ways: three of
four prompts came out token-identical; the first-step logits agree at **corr 0.995** with
the same top-1; and 338 tokens deep they agree at **corr 0.990**, same top-1 and top-2. The
fourth prompt differs where a thinking model decides how long to think, which is a near-tie
either way.

## Running it

```bash
python llm_params/gen_static.py \
  --pte qwen3_5_2b_xnnpack_8da4w_e8.pte \
  --tokenizer tokenizer.json \
  --prompt $'<|im_start|>user\nWhat is the capital of France?<|im_end|>\n<|im_start|>assistant\n' \
  --eos_ids "[248046,248044]"
```

The 8-bit embedding needs `from executorch.kernels import quantized` before the program is
loaded. Without it the method will not even load — `kernel
'quantized_decomposed::embedding_byte.dtype_out' not found` — which reads like a broken
export rather than a runtime missing its kernels.

(conversion scripts: [executorch-models](https://github.com/john-rocky/executorch-models) ·
iOS sample: [executorch-samples](https://github.com/john-rocky/executorch-samples))