LFM2.5-2.6B-heretic-ExecuTorch

LFM2.5-2.6B-heretic for ExecuTorch

ExecuTorch exports of experimentalmachines/LFM2.5-2.6B-heretic (revisions b389c6aa79a6, c513fc1129cb (see each variant's source_revision)) for on-device inference with the openweights Android app or any ExecuTorch 1.4.0 runtime.

Files

Windows in this repo: XNNPACK (CPU) at 2k to 32k; MediaTek NeuroPilot MT6991 (Dimensity 9400) at 512 to 16k. The window is fixed inside the file: the runtime allocates the whole KV cache at load, so pick the largest window the device can hold (fits_phone_budget in each folder's config.json is the estimate against a 5 GB budget).

MediaTek folders also hold the token embedding table the NeuroPilot runner reads from disk, shared by every window: mtk/mt6991/LFM2.5-2.6B-heretic-neuropilot-embedding-fp32.bin.

Tokenizer: tokenizer.json, copied unchanged from the source repo. Each backend folder has a config.json listing every window as a variant with the metadata the .pte reports, and an export-report-<window>.json per file with the full export record.

Memory

• XNNPACK (CPU) at 2,048 tokens: the KV cache costs 32,768 bytes per token (fp32), 67,108,864 bytes for the whole window, allocated in full when the model loads.
• XNNPACK (CPU) at 4,096 tokens: the KV cache costs 32,768 bytes per token (fp32), 134,217,728 bytes for the whole window, allocated in full when the model loads.
• XNNPACK (CPU) at 8,192 tokens: the KV cache costs 32,768 bytes per token (fp32), 268,435,456 bytes for the whole window, allocated in full when the model loads.
• XNNPACK (CPU) at 16,384 tokens: the KV cache costs 32,768 bytes per token (fp32), 536,870,912 bytes for the whole window, allocated in full when the model loads.
• XNNPACK (CPU) at 32,768 tokens: the KV cache costs 32,768 bytes per token (fp32), 1,073,741,824 bytes for the whole window, allocated in full when the model loads.

How it was made

• XNNPACK (CPU) any arm64: ExecuTorch 1.4.0 export_llm: 8-bit dynamic activations and 4-bit weights in groups of 32, int8 per-channel embeddings, XNNPACK with extended ops, prefill chunk 2048, fp32 KV cache. Built by run 1.
• MediaTek NeuroPilot MT6991 (Dimensity 9400): ExecuTorch 1.4.0 MediaTek LLM export (examples/mediatek, lfm2.py): A16W8 (16-bit activations, 8-bit weights) calibrated on 23 samples in the model's own chat template filling the cache from 16 to 502 tokens (7 conversations from UltraChat 200k, under the app's system prompt where it fits, 7 passages from PG19 books, and MediaTek's 9 alpaca.txt prompts), 5,550 tokens observed, cut into 3 chunks, with a 128-token prompt graph and a one-token generation graph over a 512-token cache, compiled with MediaTek NeuroPilot Express SDK (mtk_converter 8.13.0+public, mtk_neuron 8.2.23) for MT6991 (Dimensity 9400). Built by run 1.
• MediaTek NeuroPilot MT6991 (Dimensity 9400): ExecuTorch 1.4.0 MediaTek LLM export (examples/mediatek, lfm2.py): A16W8 (16-bit activations, 8-bit weights) calibrated on 23 samples in the model's own chat template filling the cache from 16 to 2,038 tokens (7 conversations from UltraChat 200k, under the app's system prompt where it fits, 7 passages from PG19 books, and MediaTek's 9 alpaca.txt prompts), 17,371 tokens observed, cut into 3 chunks, with a 128-token prompt graph and a one-token generation graph over a 2048-token cache, compiled with MediaTek NeuroPilot Express SDK (mtk_converter 8.13.0+public, mtk_neuron 8.2.23) for MT6991 (Dimensity 9400). Built by run 1.
• MediaTek NeuroPilot MT6991 (Dimensity 9400): ExecuTorch 1.4.0 MediaTek LLM export (examples/mediatek, lfm2.py): A16W8 (16-bit activations, 8-bit weights) calibrated on 23 samples in the model's own chat template filling the cache from 16 to 4,086 tokens (7 conversations from UltraChat 200k, under the app's system prompt where it fits, 7 passages from PG19 books, and MediaTek's 9 alpaca.txt prompts), 32,735 tokens observed, cut into 3 chunks, with a 128-token prompt graph and a one-token generation graph over a 4096-token cache, compiled with MediaTek NeuroPilot Express SDK (mtk_converter 8.13.0+public, mtk_neuron 8.2.23) for MT6991 (Dimensity 9400). Built by run 1.
• MediaTek NeuroPilot MT6991 (Dimensity 9400): ExecuTorch 1.4.0 MediaTek LLM export (examples/mediatek, lfm2.py): A16W8 (16-bit activations, 8-bit weights) calibrated on 23 samples in the model's own chat template filling the cache from 16 to 8,182 tokens (7 conversations from UltraChat 200k, under the app's system prompt where it fits, 7 passages from PG19 books, and MediaTek's 9 alpaca.txt prompts), 63,227 tokens observed, cut into 3 chunks, with a 128-token prompt graph and a one-token generation graph over a 8192-token cache, compiled with MediaTek NeuroPilot Express SDK (mtk_converter 8.13.0+public, mtk_neuron 8.2.23) for MT6991 (Dimensity 9400). Built by run 1.
• MediaTek NeuroPilot MT6991 (Dimensity 9400): ExecuTorch 1.5.1 MediaTek LLM export (examples/mediatek, lfm2.py): A16W8 (16-bit activations, 8-bit weights) calibrated on 23 samples in the model's own chat template filling the cache from 16 to 16,374 tokens (7 conversations from UltraChat 200k, under the app's system prompt where it fits, 7 passages from PG19 books, and MediaTek's 9 alpaca.txt prompts), 124,679 tokens observed, cut into 10 chunks, with a 128-token prompt graph and a one-token generation graph over a 16384-token cache, compiled with MediaTek NeuroPilot Express SDK (mtk_converter 8.13.0+public, mtk_neuron 8.2.23) for MT6991 (Dimensity 9400). Built by run 1.

License

A quantized derivative of experimentalmachines/LFM2.5-2.6B-heretic, distributed under the same terms (lfm1.0).

The mtk/ folders hold model binaries compiled with the MediaTek NeuroPilot Express SDK 8.0.8-build20250925 (mtk_converter 8.13.0+public, mtk_neuron 8.2.23) from MediaTek Inc., used under MediaTek's license terms for that SDK. No MediaTek SDK or runtime library is included. They run on MediaTek's LLM runner from ExecuTorch (examples/mediatek/executor_runner) with the device's NeuroPilot runtime; the settings it needs are in each folder's config.json, under each variant's runner.

Variants whose runner says state_layout: per-layer give each short-convolution layer its own two-position state, and take two small inputs that keep the runner's padding out of the convolution. MediaTek's runner as ExecuTorch 1.4.0 and 1.5.1 ship it cannot load them; the one that can is ExecuTorch's examples/mediatek/executor_runner/llama_runner with OpenWeights' patch, which tells caches from states by shape.