Falcon-H1-1.5B-Deep-Instruct — LiteRT-LM
tiiuae/Falcon-H1-1.5B-Deep-Instruct converted to the
LiteRT-LM (
.litertlm) format for on-device inference with Google's
LiteRT-LM runtime.
Requires litert-lm ≥ 0.15. Sibling of
litert-community/Falcon-H1-0.5B-Instruct,
litert-community/Falcon-H1-1.5B-Instruct and
litert-community/Falcon-H1-3B-Instruct — same conversion, same patch.
Falcon-H1 is TII's fully-hybrid design: every layer runs a grouped-query attention branch and a Mamba2 selective-scan branch in parallel on the same input and sums them. Each layer therefore carries both a KV cache and constant-size conv + SSM recurrent state. The Deep variant trades width for depth — 66 layers at hidden size 1280 — so the bundle carries 264 state buffers (132 conv/SSM + 132 KV), the most of any Falcon-H1 size.
| File | Recipe | Size |
|---|
Falcon-H1-1.5B-Deep-Instruct_int8.litertlm | int8 dynamic on linears + embedding (convs and the scan stay float); fp32 activations declared for GPU | 1.83 GB |
Correctness
- Logits parity vs PyTorch: the float export matches the HF model teacher-forced across 48 decode positions — top-1 and top-5 identical at every position, mean per-position logit correlation 1.0000, mean KL ≈ 0.
- 8-question sanity gate: 8/8 on every lane — GPU and CPU, litert-lm 0.15.0 and 0.16.0, no degeneration.
- Prompt-length robustness: hermetic prefill-chunk sweep (fresh engine per length) — CPU fills 12–51 and GPU fills 12–31 all clean.
- iPhone 17 Pro (Metal): the 8-item composite probe answers 8/8 on GPU and 8/8 on CPU. The two backends produce the same answers word for word, apart from one function word ("8 times 7 equals 56" on GPU, "is 56" on CPU).
Usage
1litert-lm run ./Falcon-H1-1.5B-Deep-Instruct_int8.litertlm --prompt "What is the capital of France? Answer in one word."
2
3# GPU
4litert-lm run ./Falcon-H1-1.5B-Deep-Instruct_int8.litertlm --backend gpu --cache no --prompt "..."
Multi-length prefill signatures (1–1024) are exported so the runtime picks tight chunks. The bundle carries the tokenizer and the stock ChatML-style Falcon-H1 chat template.
Performance
litert-lm benchmark (litert-lm 0.16.0), Apple M4 Max, -p 256 -d 256 --runs 3 --cache no, quiet machine:
| Backend | Prefill (256) | Decode | TTFT |
|---|
| GPU | 1134 tok/s | 51.7 tok/s | 0.25 s |
| CPU | 192 tok/s | 22.7 tok/s | 1.38 s |
On device (cold start, single runs, 146-token composite prompt, quality harness):
| Device | Backend | Prefill | Decode | TTFT | Peak memory |
|---|
| iPhone 17 Pro | GPU (Metal) | 140.3 tok/s | 16.5 tok/s | 1.19 s | 4.89 GB |
| iPhone 17 Pro | CPU | 103.2 tok/s | 9.9 tok/s | 1.53 s | 1.74 GB |
Honest notes:
- GPU runs with fp32 activations (declared in the bundle) — expect a corresponding memory multiple over CPU.
- Depth costs decode: 66 layers means 66 sequential attention+scan blocks per token. Decode lands below the shallower 3B sibling (65.3 tok/s on the same machine and settings) despite this model being roughly half its size, while prefill — which batches over the sequence — comes out ahead (1134 vs 979 tok/s).
- On GPU the runtime logs
TopK requires src tensor C dimension to be divisible by 4 (vocab 65537). It is family-normal on every Falcon-H1 size — a fallback path, not a failure; generation and the quality gates are unaffected.
- The full prefill ladder loads on a 12 GB phone, all 264 state buffers included: GPU peak is 4.89 GB, 2.8× the file. Load takes about a minute on GPU (8 s on CPU) — the Metal kernels for every signature are built at engine creation.
Galaxy S26 — GPU backend
The published bundle runs on the Android GPU backend: LiteRT takes the whole graph and the model generates.
| file | GPU backend | delegation | peak |
|---|
Falcon-H1-1.5B-Deep-Instruct_int8.litertlm | runs | 115393 / 115393 ops across 12 subgraphs on LiteRT GPU | 3745 MB |
Measured on a Samsung Galaxy S26 (SM-S942Q / SM8850, Android 16) with litert_lm_advanced_main from litert-lm 0.16.0, --backend=gpu --sampler_backend=cpu, prompt What is the capital of France?. Peak is the process high-water mark (VmHWM) sampled during that same run. Gated 2026-08-24.
No speed rows, on purpose. On this handset the GPU backend wins prefill and does not win decode, so a GPU throughput figure only means something beside a CPU row from the same handset, and no S26 CPU row exists for this model yet.
GPU wiring, including the Gallery import toggle:
GPU guide.
Conversion notes
Converted with
litert-torch plus a hybrid-cache patch (reproduction script + patch:
hf-to-litertlm falcon_h1_work/):
- Composite hybrid cache layer: every layer holds KV + conv + recurrent state at ONE layer index — a cache layer class that is full-attention and Mamba2 at the same time (the runtime binds states by tensor name, so co-residency is just packaging).
- Folded selective scan: the Mamba2 scan is re-expressed as batched matmuls with chunk and head axes folded into the batch axis (all tensors rank ≤ 4, no
BROADCAST_TO, no int64 index math) — this is what makes the graph fully delegable on GPU.
- Falcon-specific wiring: the µP multiplier vector (
mup_vector, a non-persistent model-level buffer) and ssm_in_multiplier are preserved in the traced scan; the exporter's timestamp-index kwargs are re-injected at the attention layer (FalconH1's layer loop drops kwargs).
- Prefill-pad guard: the runtime runs partially-filled prefill chunks; pad positions are made exact identity steps for the SSM and the stored conv window is gathered at the last valid column.
- Quantization: post-hoc dynamic int8 over linears + embedding only; convs and the scan stay float.
Raspberry Pi 5 (CPU)
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 |
|---|
Falcon-H1-1.5B-Deep-Instruct_int8.litertlm | 24.6 (24.6–24.7) | 2.0 (2.0–2.0) | 10.9 s | 3.3 GB |
License and changes
Distributed under the Falcon LLM License (inherited from the base model — see the license link). Changes from the original work: weights converted from safetensors bf16 to LiteRT flatbuffers and quantized as described above; tokenizer and chat template repackaged unmodified. This repository is a community conversion and is not affiliated with TII.
