Measured on device (edge-compat): Galaxy S26 · LiteRT-LM prebuilt-adac974b · GPU · decode 27.4 tok/s · prefill 579 tok/s · TTFT 1.80 s · all 1518 ops delegated (2026-10-05); Mac Studio M4 Max · LiteRT-LM 0.17.1 · GPU · decode 150.7 tok/s · prefill 2332 tok/s · TTFT 116 ms · all 1518 ops delegated (2026-10-05); Galaxy S26 · LiteRT-LM prebuilt-adac974b · CPU · decode 14.4 tok/s · prefill 237 tok/s · TTFT 4.43 s (2026-10-05). Record: https://github.com/john-rocky/edge-compat/blob/main/cards/hy-mt2-1.8b-int8/CARD.md

Hy-MT2-1.8B — LiteRT-LM

tencent/Hy-MT2-1.8B converted to the LiteRT-LM (.litertlm) format for on-device inference with Google's LiteRT-LM runtime.

Hy-MT2-1.8B is Tencent's "fast-thinking" multilingual translation model: 33 languages, tuned to follow translation instructions rather than chat. It is a dense hunyuan_v1_dense stack — 32 layers, hidden 2048, 16 query / 4 KV heads (GQA) with QK-norm, head_dim 128, vocab 120,818, tied embeddings — 2.04B parameters.

File	Recipe	Size
Hy-MT2-1.8B_int8.litertlm	export-time dynamic int8 (linears + embedding)	1.83 GB

2026-10-06: GPU graph rewrite. Only the prefill/decode graph of Hy-MT2-1.8B_int8.litertlm changed. Its two DYNAMIC_UPDATE_SLICE KV-cache writes per layer are now one STABLEHLO_COMPOSITE odml.cache_update. Its BATCH_MATMUL(adjY) attention products are now odml.runtime_bmm. A signature input param_tensor INT32[1,1,1,7] was added; the runtime fills start/end. A graph exported with litert-torch's --apply_gpu_composites has this same shape, and the previous file was exported without that flag. The attention mask is applied as an ADD on a [bk, g, T, C] view of the logits, with the FLOAT32 mask broadcast; the mask input stays FLOAT32. A 16-token prefill signature, prefill_16, was added as a copy of prefill_128 that shares every weight buffer. Every other bundle section and the graph's weight region are byte-identical to the previous file (checked per section and per buffer). The new file is 1,827,026,224 bytes, sha256 275ff52bf04c4d577e1bf48733daa77d873d78b0637841574d7a10e1560a011a. The previous file was 1,815,622,960 bytes, sha256 529e6d378df5869d89a5a08717c06604105a32d8a4dab4800175d6baabc4da50. The checks compare three files: the previous file, an intermediate file with the rewrite but without prefill_16, and the new file. On the Mac GPU, answers are byte-identical from the previous file to the intermediate (8/8 questions) and from the intermediate to the new file (9/9 prompts), with fp32 activations and with the default fp16 activations alike. On the CPU, the new file's answers match the intermediate's byte for byte (9/9 prompts), and prefill_16 gives bit-identical output to prefill_128 on the same tokens (2/2 cases). On the Mac CPU without the weight cache, the new file takes longer to load and uses more memory; see Performance. The speed rows below ran both files in the same window, with the Mac protocol under Performance. The Raspberry Pi 5 rows further down were measured on the previous file.

Mac Studio M4 Max, GPU, fp16 activations	previous file	new file
8 questions, correct final answers	6/8	6/8
Prefill, 256-token prompt	2014 tok/s	2332 tok/s
Decode, 256 tokens	103.5 tok/s	150.7 tok/s
TTFT, 16-token prompt	0.076 s	0.031 s

2026-09-21: chat template updated to accept the 0.18 content-parts form (string form unchanged); weights, tokenizer and executor metadata byte-identical.

Correctness

All numbers below were measured on the previous file (litert-lm 0.16.0, Apple M4 Max). The 2026-10-06 note above lists the checks on the new file.

• Chat template is byte-equal to the source: the embedded Jinja matches the repo's chat_template.jinja exactly (654 / 654 bytes).
• 8-question sanity gate: 6/8 on CPU and 6/8 on GPU, non-degenerate, with the same two misses on both backends ("Cool" for the opposite of hot, "pink" for the rhyme) — a property of this translation-tuned 1.8B, not of a backend. Arithmetic, factual and translation items are all correct.
• Translation greedy A/B vs the HF reference (PyTorch bf16, greedy, the source README's default-translation prompt): byte-identical on 1 of 3 probes; the other two are fluent alternates of the usual int8-vs-bf16 kind (e.g. spectaculaire → significative). No degeneration on any probe.
• No duplicate start token. The source chat template renders <|hy_begin_of_sentence|> itself, and the LiteRT-LM engine also prepends the metadata start_token — measured inside the runtime: [start_token]+prompt and [template BOS]+prompt generate byte-identical greedy output, so the default export was feeding BOS twice. This bundle drops the metadata start token; the on-device token stream matches the training stream. Honest note: the double-BOS variant happened to score 8/8 on the sanity gate — the stream-faithful file ships anyway, because matching the training stream is the property that generalizes.
• Stop token is <|hy_place_holder_no_2|> (id 120020), as the source declares.

Usage

Hy-MT2 expects translation instructions, not open chat. The source model card's default prompt works verbatim:

litert-lm run ./Hy-MT2-1.8B_int8.litertlm --prompt \
"Translate the following text into Japanese. Note that you should **only output the translated result without any additional explanation**:

The weather is nice today, so let's go for a walk in the park."

# GPU
litert-lm run ./Hy-MT2-1.8B_int8.litertlm --backend gpu --prompt "..."

The bundle carries the tokenizer and the stock Hy-MT2 chat template (max_num_tokens 4096).

Performance

Measured on the new file (see the 2026-10-06 note).

Apple M4 Max (macOS)

Mac Studio M4 Max, litert-lm 0.17.1 CLI, benchmark --cache no, 256 prompt and 256 decode tokens. GPU = WebGPU on Metal with fp16 activations, median of 3 processes. CPU = XNNPACK, median of 2 processes. The TTFT (16-token prompt) column comes from separate runs with a 16-token prompt and 32 decode tokens.

Backend	Prefill (256)	Decode	TTFT	TTFT (16-token prompt)	Init
GPU (WebGPU on Metal)	2332 tok/s	150.7 tok/s	0.12 s	0.031 s	1.5 s
CPU (XNNPACK)	207 tok/s	31.9 tok/s	1.27 s	0.884 s	5.3 s

Without the XNNPACK weight cache (--cache no, as in this table), the new file repacks the weights for the extra signature on the CPU. At a 16-token prompt, init goes from 3.75 s to 5.3 s and peak footprint from 4.17 GB to 5.57 GB (file without prefill_16 → new file). With the disk weight cache (the app default), init goes from 0.24 s to 0.32 s, footprint from 0.82 GB to 0.87 GB, and TTFT at 16 tokens from 0.33 s to 0.12 s.

Galaxy S26 (Android)

Samsung Galaxy S26 (SM-S942Q, Android 16), litert_lm_advanced_main from a LiteRT-LM main build of 2026-09-18. GPU = OpenCL delegate with the bundle's default fp16 activations, cold init (--disable_cache=true), 3 iterations in one process. CPU = XNNPACK with 4 threads and the XNNPACK weight cache, 2 iterations in one process. Each range runs from the lowest to the highest iteration in one process; the first GPU iteration follows the cold init. Peak is the process high-water mark (VmHWM).

Backend	Prompt / decode tokens	Prefill	Decode	TTFT	Init	Peak (VmHWM)
GPU (OpenCL)	1024 / 256	576.9–580.2 tok/s	26.29–27.46 tok/s	1.80–1.81 s	4.94 s	0.83 GB
GPU (OpenCL)	16 / 32	317.7–364.2 tok/s	20.1–26.04 tok/s	0.08–0.10 s	4.91 s	0.83 GB
CPU (XNNPACK, 4 threads)	1024 / 256	213.4–260.2 tok/s	13.75–14.99 tok/s	4.00–4.87 s	3.72 s	2.96 GB

The OpenCL GPU delegate (LITERT_CL) takes every node of each transformer signature, in one partition each: prefill_128 1518/1518, prefill_16 1518/1518, decode 1390/1390. In the GPU gate on this phone, 3 of 5 prompts were answered correctly.

Conversion notes

Converted with stock litert-torch 0.9.3 through hf-to-litertlm — one command:

python scripts/convert.py tencent/Hy-MT2-1.8B

Two things route this family correctly, both measured:

• The dynamic-with-alpha rope resolves statically. transformers computes base = rope_theta * alpha^(dim/(dim-2)) once at init and never rescales below max_position_embeddings — only the leftover data-dependent cache-growth branch kills torch.export. The converter bakes the resolved base (11,158,839.925) into rope_theta and drops rope_scaling; inv_freq and teacher-forced logits are bitwise-equal to the HF reference, valid to 262,144 positions (far past this bundle's 4,096 context).
• The engine prepends the metadata start token unconditionally, so a template that renders its own BOS must not also declare one — see Correctness above.

Export took 120 s on an M4 Max. See REPRODUCE.md for the full measurement record behind every claim on this card.

Graph rewrite (2026-10-06). The current file was made from the previous one (Hub revision 18cc1a6794d7) with two tools in tools/gpu_graph/ of hf-to-litertlm. gpu_graph_retrofit.py makes the graph changes listed in the 2026-10-06 note; every original weight buffer keeps its bytes. prefill_bucket_clone.py copies an existing prefill signature (prefill_128) at a new length (16); the copy shares every weight buffer, and every existing subgraph and SignatureDef keeps its index and bytes.

python tools/gpu_graph/gpu_graph_retrofit.py previous.litertlm retrofit.litertlm --decomp exporter --mask add_bcast
python tools/gpu_graph/prefill_bucket_clone.py build retrofit.litertlm new.litertlm --lengths 16 --source prefill_128 --report bucket.json

previous.litertlm is the previous file, new.litertlm the result published here as Hy-MT2-1.8B_int8.litertlm; LITERT_LM_CLI names the litert-lm CLI the scripts use to unpack and pack the bundle.

A rebuild with these commands gives every bundle section byte-identical to this file (sha256 per section). Only the bundle header (uuid and timestamp written by litert-lm pack) differs, so the whole-file sha256 differs.

Raspberry Pi 5 (CPU) (previous file)

Measured on the previous file (sha256 529e6d378df5869d89a5a08717c06604105a32d8a4dab4800175d6baabc4da50) and not re-measured after the 2026-10-06 graph rewrite; the weights are byte-identical and the new file's Mac CPU decode and prefill are ×0.992 and ×1.010 of the previous file's, so the throughput columns are expected to hold, while peak memory may differ (see the CPU note under Performance).

Measured on a Raspberry Pi 5 Model B Rev 1.1 (8 GB, Raspberry Pi OS 64-bit) with litert-lm benchmark 0.16.1: CPU backend, 4 threads, 256 prefill + 256 decode tokens, --cache memory (the compile cache lives and dies with the process, so every invocation compiles the model from scratch; nothing is reused between runs), one warm-up plus one timed iteration per invocation, 3 invocations per file with cooldown in between. Values are the median across invocations (min–max in parentheses). No thermal throttling occurred during these runs (vcgencmd get_throttled stayed 0x0). Every file listed produced coherent text in a real generation on this backend before its numbers were recorded.

File	Prefill (tok/s)	Decode (tok/s)	TTFT	Peak RSS
Hy-MT2-1.8B_int8.litertlm	25.0 (24.9–25.8)	2.7 (2.6–2.7)	10.6 s	2.9 GB