mlboydaisuke/Qwen3.5-0.8B-ExecuTorch

Qwen3.5-0.8B — ExecuTorch XNNPACK 8da4w + 8-bit embedding

qwen3_5_0_8b_xnnpack_8da4w_e8.pte (646.9 MB — down from 1413 MB without embedding quantization; the 248320x1024 fp32 embedding was about a gigabyte of the first export)

• Source: Qwen/Qwen3.5-0.8B
• License: Apache-2.0
• Quantization: 8da4w linear + 8-bit embedding (embedding_quantize: "8,0")
• Export: executorch 1.4.0 export_llm, static shape (seq_len=1), max_seq_length 2048, XNNPACK extended_ops
• Config: llm_params/qwen3_5_0_8b_xnnpack_8da4w_e8.yaml

Verification (Mac arm64, 2026-08-21)

llm_params/gen_static.py, token-by-token prefill then greedy decode:

prompt	output	decode
capital of France?	"The capital of France is Paris."	38.06 tok/s
日本の首都は?	「日本の首都は 東京 です。」	38.06 tok/s

Prefill tok/s is a sequential-prefill reference only, this being a static export. Chat template: ChatML, bos 248045, eos [248046, 248044].

The iPhone number below predates the use_sdpa_with_kv_cache change and was measured on the earlier build, so it is not this file's.

iPhone 17 Pro (ETBench, XNNPACK CPU, default threads), earlier build: decode 10.5 tok/s, ttft 0.58 s, load 0.7 s — the same speed as the 1413 MB fp32-embedding export (9.2-11.1 tok/s) at 46% of its size. A hot device throttles to about 7 tok/s; that was a cool run.

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.