Skip to content

Help wanted: published vindex hydration and Gemma 4 extraction sanity check #179

Description

@henry-p

Help wanted: published Gemma 3 vindex hydration + Gemma 4 extraction sanity check

Hi! I'm experimenting with LarQL and trying to reproduce the README happy path, then build an equivalent vindex for Gemma 4. I may be using the wrong workflow, but I'm seeing two confusing things.

1. Gemma 3 published vindex: inference files exist on HF, but local pull/run misses them

The README suggests:

larql pull hf://chrishayuk/gemma-3-4b-it-vindex
larql run gemma-3-4b-it-vindex "The capital of France is"

The HF repo appears to include inference files such as:

attn_weights_q4k.bin
interleaved_q4k.bin
lm_head_q4.bin
norms.bin
weight_manifest.json

But after larql pull, my local snapshot only has the browse/core files:

down_meta.bin
embeddings.bin
feature_labels.json
gate_vectors.bin
index.json
tokenizer.json

Then larql run ... fails with:

Error: parse error: attention k-quant weights not found
(looked for attn_weights_kquant.bin and legacy attn_weights_q4k.bin)

So my question is: should larql pull hydrate the inference files for a full vindex, or should larql run hf://... lazily call the weight-download path before loading?

2. Locally extracted Gemma 4 vindex loads, but inference looks implausible

I extracted google/gemma-4-12B-it from HF safetensors:

larql extract <hf-snapshot> \
  --output gemma-4-12b-it.q4k-inference.vindex \
  --level inference \
  --quant q4k \
  --drop-gate-vectors \
  --resume

The vindex verifies, serves, and reports inference loaded. But /v1/infer for:

The capital of France is

returns flat/unrelated top tokens like:

" f", "?", "en", "is", " in", " of"

The same model in LM Studio answers "The capital of France is Paris."

Questions

  1. What is the intended command sequence to reproduce the README INFER "The capital of France is" -> Paris example from the published HF vindex?
  2. Should larql pull download VINDEX_WEIGHT_FILES, or is there another command I should run before inference?
  3. Is the server preset intentionally FFN-only/no-attention, meant to pair with a client slice?
  4. Is dense google/gemma-4-12B-it expected to work with larql extract --quant q4k today?
  5. If Gemma 4 should work, what diagnostic path would you recommend for implausible logits?

Thanks! LarQL is a very interesting project; I'm trying to separate a docs/artifact/hydration issue from user error.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions