LiteRT LLM Calibration & Quantization Skill

A skill to guide the calibration, alignment, precision protection, and static range quantization of LiteRT LLMs (such as Gemma 3 1B) for high-performance NPU deployments.

Triggers and Anti-Triggers

• Use when:

  • The user wants to calibrate a LiteRT model against conversational prompt datasets.
  • The user needs to merge calibration JSON files from multiple different tasks.
  • The user wants to statically quantize a model suite (e.g. A8W8 or A16W8 configurations).
  • The user needs to protect RMS Norm or residual paths from quantization noise in Float32.
  • The user needs to align KV Cache quantization scaling parameters across main and auxiliary models.
  • The user wants to run CPU/NPU inference testing on unquantized or quantized LiteRT models.

• Don't use for:

  • Quantizing non-LiteRT models (e.g. PyTorch native, JAX native, or ONNX models).
  • General quantizations that do not involve LLM prefill/decode subgraphs or KV caches.

Principles & Mathematical Gotchas

Every LLM NPU quantization task has highly sensitive numerical paths. You must strictly adhere to these verified gotchas to prevent quality degradation or loop crashes:

1. Mandatory BOS (Beginning of Sequence) Token

• Gotcha: Tokenizing prompts without the BOS token (2) will completely corrupt the model's attention states and positional encodings at step 0. This causes the model to generate garbage loop outputs (e.g. does France does France infinitely) during both calibration and inference.
• Rule: You must ensure that prepend_bos=True is active during all tokenization operations. If using the open-source SentencePiece or Transformers tokenizers, verify that the BOS token ID (2) is explicitly prepended to the start of the token IDs list.

2. Precision Protection in Float32 (allow_float_operations=True)

• Gotcha: Statically quantizing RMS Norm operations, residual layer additions, or skip scale multipliers down to Int8/Int16 introduces severe quantization noise that ruins text generation quality.
• Rule: Keep these three mathematical paths protected in Float32:
  1. RMS Norm / Layer Norm operations (typically the StableHLO composite or norm nodes): Skip from quantization to prevent numerical accuracy loss.
  2. Residual connection additions (typically the element-wise add nodes immediately following attention or MLP projection blocks): Keep in Float32.
  3. Residual connection scaling multipliers (typically the element-wise multiplication nodes matching the skip connection scaling factors): Keep in Float32.

3. Key-Value (KV) Cache Parameter Alignment (align_kv_cache=True)

• Gotcha: Key and Value cache tensors across the main model (prefill/decode subgraphs) and auxiliary model must use identical scaling factors (scale and zero-point) so the hardware NPU can reuse slices dynamically without expensive runtime rescalings.
• Rule: Always execute align_kv_cache_params() to search and align all K and V cache layers across the model suite before calling the quantizer.

4. Keep embedder.tflite in Float32 (Never Quantize)

• Gotcha: The embedder model contains a massive lookup table. Quantizing its weights leads to massive accuracy degradation. Since the embedder only executes once at step 0 (prefill), it has zero speed impact.
• Rule: Do not quantize embedder.tflite. Copy the original Float32 version directly into the final quantized deployment folder.

5. Profiler-Based Calibration (use_profiler_based_calibration=True)

• Gotcha: Standard calibration runs inference and collects activation statistics from outputs. However, some internal activation bounds might be missed or underestimated.
• Rule: Use --use_profiler_based_calibration=True to run the calibration interpreter in profiler mode. This extracts precise tensor bounds directly from the interpreter's internal profiling metrics. This mode requires --enable_min_max_calibration_update=True to correctly update the min/max calibration parameters.

6. Skipping MLIR Passes and Scaling Ranges

• Gotcha: Standard post-quantization MLIR graph surgery passes can sometimes fail or crash on custom architectures (like Kanana's swapped KV cache). Additionally, tight min/max calibration bounds can cause clipping of activation outliers, degrading model quality.
• Rule:
  • Use --skip_mlir_passes=True to bypass failing MLIR optimizations for sensitive custom model architectures.
  • Use --calibration_range_scale=1.15 (or a similar factor) to slightly scale up the calibration bounds, providing headroom for outliers and reducing clipping distortion.

Core Command Patterns

1. Open-Source Calibration Run

Generates the calibration range JSONs under a task-specific sub-directory:

python3 -m litert_torch.generative.export_hf.experimental.calib.calibrate \
  --model_path={model_path} \
  --embedder_model_path={embedder_path} \
  --auxiliary_model_path={aux_path} \
  --spm_path={tokenizer_spm_path} \
  --eval_task_names="{task_name}" \
  --dataset_dir={dataset_dir} \
  --dataset_format=json \
  --calibration_result_save_dir={output_save_dir}/{task_name} \
  --max_decode_steps=128 \
  --use_profiler_based_calibration=True \
  --enable_min_max_calibration_update=True

2. Merging Calibration Results

Discovers and merges QSV files across task sub-directories into a single aligned calibration file:

python3 -m litert_torch.generative.export_hf.experimental.calib.merge_calibration_results \
  --input_dir={calibration_dir} \
  --output_dir={merged_output_dir}

Note: The merger expects a directory structure where task files are grouped inside sub-folders (e.g. input_dir/task_name/model.tflite.json). Flat folders will result in no tasks discovered.

3. Advanced Multi-Model Quantization (A8W8 or A16W8)

Statically quantizes both the main model and the auxiliary model with KV Cache Alignment and Float Protection active:

python3 -m litert_torch.generative.export_hf.experimental.calib.quantize \
  --model_path={model_path} \
  --calibration_path={merged_dir}/model.tflite.json \
  --output_path={quantized_dir}/model.tflite \
  --aux_model_path={aux_path} \
  --aux_calibration_path={merged_dir}/aux.tflite.json \
  --aux_output_path={quantized_dir}/aux.tflite \
  --a16w8={true_or_false} \
  --align_kv_cache=True \
  --allow_float_operations=True \
  --skip_mlir_passes=True \
  --calibration_range_scale=1.15

4. Quantized Inference Testing

Executes text generation using the quantized model suite:

python3 -m litert_torch.generative.export_hf.experimental.calib.sampling_executor_main \
  --model_path={quantized_dir}/model.tflite \
  --embedder_model_path={embedder_path} \
  --auxiliary_model_path={quantized_dir}/aux.tflite \
  --spm_path={tokenizer_spm_path} \
  --prompt="{prompt_text}" \
  --enable_formatting=False \
  --max_decode_steps=16 \
  --stop_token={stop_token_id} \
  --stream_output=True

Note: For Gemma 3, set --stop_token=106 to ensure the executor cleanly halts immediately when the model generates its end-of-turn sequence.

The E2E Interactive Driver tool

For the easiest, most robust end-to-end execution, you can use the interactive bash script inside the repository: bash cd litert_torch/generative/export_hf/experimental/calib/ ./run_quantize_and_inference.sh This script provides interactive path configuration, parallel calibration task management, automatic task grouping, merging, and A16W8/A8W8 recipe quantization!