gemma.cpp
gemma.cpp copied to clipboard
Abort / Assertion-error when loading gemma-3/gemmaCpp/3.0-1b-it-sfp
With current dev (commit 684a0444e9bb6ddba53c361d87def055235ca387 (HEAD -> dev, origin/dev)) following the Readme.md using the bazel-based path:
Gemma2: Looks ok
https://www.kaggle.com/models/google/gemma-2/gemmaCpp/2.0-9b-it-sfp
bazel-bin/gemma --tokenizer tokenizer.spm --weights 2.0-2b-it-sfp.sbs --prompt_file /mnt/c/Users/XXX/Desktop/llm_prompt.txt
1180131840 blob bytes (36.83%) of bf16
2024275968 blob bytes (63.17%) of sfp
normalizer.cc(52) LOG(INFO) precompiled_charsmap is empty. use identity normalization.
[ Reading prompt ] ......................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................
[ Timing info ] Prefill: 24927 ms for 3829 prompt tokens (153.60 tokens / sec); Time to first token: 25039 ms
...
Gemma3: Assertion error
https://www.kaggle.com/models/google/gemma-3/gemmaCpp/3.0-1b-it-sfp
bazel-bin/gemma --tokenizer tokenizer.spm.GEMMA3 --weights 1b-it-sfp.sbs.GEMMA3 --prompt_file /mnt/c/Users/XXX/Desktop/llm_prompt.txt
604248320 blob bytes (46.41%) of bf16
697761792 blob bytes (53.59%) of sfp
normalizer.cc(52) LOG(INFO) precompiled_charsmap is empty. use identity normalization.
Abort at weights.cc:454: Assert offset == range.End():
Thanks for reporting. There is a known issue with Gemma3 1B which we have not addressed yet, but it's being bumped up the priority list. In the meantime, would the 4B version work for you instead?
Thanks for the confirmation.
Indeed 4b-it-sfp.sbs works fine and is a viable alternative for my experiments.