CLI¶
Metadata Generation¶
Script:
Arguments:
| Argument | Meaning |
|---|---|
--model |
Target HuggingFace model. |
--calibration-model |
Optional model used for activation collection. |
--kv-cache-dtype |
turboquant25 or turboquant35. |
--prompts-file |
Text file with one prompt per line. |
--output |
Metadata JSON path. |
--trust-remote-code |
Passed to HuggingFace loaders. |
--layer-pattern |
Layer-name template, default uses {i}. |
--max-prompts |
Prompt limit. |
--max-seq-len |
Token truncation limit. |
--batch-size |
Calibration batch size. |
--dtype |
auto, float32, float16, or bfloat16. |
--device |
PyTorch device string. |
The prompts file is hashed and the hash is stored in metadata. Empty lines are
ignored. --max-prompts can be used to bound calibration cost while preserving
deterministic prompt ordering.
The script rejects using the same quantized checkpoint as the calibration source when the target config reports quantization metadata. Use a non-quantized calibration checkpoint when available.
vLLM/HF Verification¶
Script:
Arguments:
| Argument | Meaning |
|---|---|
--model |
HuggingFace model id or path. |
--bits |
Shared K/V bit width. |
--k-bits and --v-bits |
Per-component bits; must appear together. |
--kv-cache-dtype |
tq4, k4v4, turboquant25, or turboquant35. |
--metadata-path |
Required for recipe modes. |
--threshold |
Minimum cosine for pass status. |
--json |
Write JSON to stdout and text to stderr. |
--trust-remote-code |
Passed to HuggingFace loaders. |
--token |
Passed to HuggingFace loaders. |
Argument rules:
--bitscannot be mixed with--k-bitsor--v-bits.--k-bitsand--v-bitsmust appear together.- recipe cache dtypes require
--metadata-path. - JSON mode returns nonzero when the verification status is
FAIL.
Qwen A/B Evaluation¶
Script:
This script runs baseline and TurboQuant diagnostic scenarios and writes separate artifacts for each run.
The script is designed for controlled A/B runs: same model, same prompts, same generation settings, separate artifact directories.
Phase Check Scripts¶
The repository includes scripts named:
phase2_real_data_check.py
phase3_real_data_check.py
phase4_real_data_check.py
phase5_real_data_check.py
phase6_real_data_check.py
phase7_real_data_check.py
These are development validation scripts tied to the project phases.
They are not a replacement for end-to-end serving benchmarks, but they are useful for finding regressions in layout, memory accounting, and attention reference behavior.