Troubleshooting¶
Use this page to identify whether a failure is an installation issue, a shape contract issue, a metadata issue, or a claim-boundary issue.
Import Errors¶
If transformers, triton, or vllm is missing, install the needed package in
the active environment.
For docs:
Metadata Errors¶
Recipe modes require metadata. If resolve_metadata raises a file error, pass
metadata_path or place turboquant_kv.json beside the model directory.
If metadata head dimension differs from runtime head dimension, regenerate metadata for that model/config pair.
If group indices fail validation, check that each KV head has the correct recipe count:
| Recipe | High-group count at head dim 128 |
|---|---|
turboquant25 |
32 |
turboquant35 |
64 |
Shape Errors¶
Most KV tensors use:
The packed-page TQ helpers use:
Check the function docstring before calling a path.
Slot Mapping Errors¶
Negative slots are ignored. Non-negative slots outside allocated pages raise
IndexError.
For block size B:
If block >= num_blocks, allocate more pages or fix the caller's slot mapping.
CUDA Errors¶
Some tests skip when CUDA or Triton kernels are unavailable. CPU reference tests still validate layout and math contracts.
Quality Drops¶
Try:
- higher key precision,
- K16V4 or K8V4 before K4V4,
- larger recent dense window,
- boundary-token preservation,
- real logits and retrieval checks,
- per-layer analysis.
Also check whether the route is compressing full-attention layers in a
sliding/global architecture. The HF wrapper can bypass full-attention layers
when model_config.layer_types exposes those labels.
Documentation Build Issues¶
If MkDocs cannot find modules, run from the repository root. mkdocs.yml sets
mkdocstrings search paths to ".".