Rebuild bf16 vmfbs for Torq v2.3.0
Rebuild of the bf16 LFM2-VL-450M bundle for the upcoming Torq v2.3.0 (compiler/runtime snapshot-v2.3, 639d1ed).
Built by torq-tools wip/liquid-vl-vision-fixes@f8d397a (synaptics-torq/torq-tools#50) with:
torq-export-model liquid-vl --convert-dtypes --skip-validation --vision-res 256 \
--image-decoder-parts --batch-prefill 0 --use-binary --compiler-path <v2.3 torq-compile>
decoder_nolm.vmfb is the exporter's transformer.vmfb (same I/O signature as the previous decoder_nolm.vmfb); all five entry points/signatures are unchanged. No NSS-program caps are needed any more (every component fits the compiler's 8 MB default).
Requires the v2.3 runtime. The W8 files are not touched by this PR.
Board validation (SL2619, fresh boot + cpu.sh per run, 3 runs, same image/prompt, bf16 load/run/release runner):
| current main (Aug-26 runtime) | this PR (v2.3 runtime) | |
|---|---|---|
| vision encode | 1274 ms | 1242 ms |
| image prefill | 2840 ms | 1249 ms |
| TTFT | 2371 ms | 2055 ms |
| decode | 4.2 tok/s | 5.2 tok/s |
Caption correct ("The picture shows a small beagle puppy and a gray tabby kitten sitting together in a basket ..."); vision features 0.9984 per-token cosine vs fp32 ORT.