lfm2-350m-webgpu — GPTQ-int4 weights

Custom GPTQ-int4 quants (group 32, byte-sliced nibbles, dedicated q4 lm_head, q8 embedding gather) of LiquidAI/LFM2-350M and LiquidAI/LFM2.5-350M, in the wire format of the lfm2-350m-webgpu browser runtime (hand-written WGSL kernels, no ONNX / transformers.js / GGUF).

• 269 MB per model; ~390 tok/s single-stream decode and ~1640 tok/s batched (64 streams) on an Apple M4 (10-core GPU) in Chrome.
• LFM2: ppl +3.2% vs f32 on held-out text.
• LFM2.5: GPTQ calibrated on chat-formatted passages (raw-prose calibration degrades an instruct checkpoint badly): ppl +4.9% vs f32 in its native chat format.
• The runtime is verified bit-exact against a PyTorch dequantized-weights reference, greedy token-for-token.

f16/q8 blobs are reproducible with the repo's tools/ scripts.

Try it: https://huggingface.co/spaces/borkiss/lfm2-350m-webgpu-demo · run the benchmark on your GPU and share the .log from the /web/bench.html page.