DAOS Parameter Configuration Workflow Guide

1. Scope

This guide is intended for users who need to use the DAOS Low-Bit Quantization Solution. DAOS is a joint low-bit solution for multimodal generative models. By chaining oasq (outlier-aware smoothing) with trainable_linear_quant (TLQ, trainable linear quantization), it targets extremely low-bit scenarios such as W4A4 / MXFP; when further accelerating Attention inference is needed, fa3_quant can additionally be configured for FA quantization.

This guide explains the chaining order, stage goals, and constraints. For each algorithm's recommended starting configuration and field tuning, see the OASQ Outlier-Aware Smooth Quantization Algorithm Usage Guide, the Trainable Linear Quantization Algorithm Usage Guide, and the FA3 Quant Parameter Configuration Workflow Guide respectively; they are not repeated here.

• Applicable: you need OASQ + TLQ to jointly recover low-bit accuracy; the task protocol is multimodal_sd_modelslim_v1 (or the corresponding multimodal generative task protocol).
• Not applicable: a single calibration pass is sufficient and OASQ/TLQ is not needed; the target dtype/scope/method has no corresponding TLQ kernel; the adapter cannot provide OASQ-fusible subgraphs and default probing does not apply either; or the device memory/duration cannot support block-level training.

Resource note: DAOS includes OASQ smoothing and TLQ block-level training, so quantization time and memory usage are significantly higher than for single-calibration algorithms; whether memory is sufficient usually cannot be confirmed until quantization is started, as described in the relevant step.

If the target model already has a complete, verified quantization recipe, prefer reusing that recipe.

2. Inputs and Deliverables

Type	Name	Source or Storage Location	Format or Constraints	Acceptance Criteria
Input	Floating-point model weight directory	Model download or a local path	HuggingFace format containing config.json and *.safetensors shards	Loadable by the target Transformers version
Input	Model adapter	User adaptation code, invoked via --model_type	Implements PipelineInterface and the dedicated interfaces for OASQ/TLQ (and optionally FA3)	Properly driven by the runner and processor
Input	Calibration dataset	Task dataset, a calibration set, or a model practice recipe	Text prompts in JSONL or JSON format; 50 samples recommended	Represents the real input distribution and is successfully encoded into forward tensors by the adapter's handle_dataset
Input	Quantization task configuration	The user's YAML or a lab_practice recipe	Complies with multimodal_sd_modelslim_v1 (or the corresponding task protocol) and can be chained into spec.process	Passes schema validation and can be loaded by the quantization workflow
Deliverable	DAOS configuration snippet	The process entries written into the above YAML	oasq first, then trainable_linear_quant; append fa3_quant as needed; each section's fields comply with the corresponding algorithm's configuration description	Quantization can be reproduced with the complete configuration

3. Process Overview

The overall workflow of DAOS is as follows:

flowchart LR
    A[Determine the quantization mode] --> B[Adapt and write the configuration]
    B --> C[Start quantization]
    C --> D[Evaluate and iterate on tuning]

Key details of each stage are as follows:

• Determine the quantization mode: fix the quantization mode DAOS will implement, including the target bit width and numerical format of the linear layers' W/A, whether to apply FA quantization to Attention, and TLQ's preferred OP path.
• Adapt and write the configuration: complete subgraph adaptation in the order OASQ → TLQ → optional FA3, and write a loadable DAOS chained configuration into spec.process.
• Start quantization: start the whole-chain quantization and get it running; if memory is insufficient, adjust the resource-related parameters first and retry.
• Evaluate and iterate on tuning: complete the end-to-end evaluation; if it falls short, iterate on single-variable tuning according to the corresponding algorithm's usage guide until the target is met or a fallback strategy is determined.

4. Procedure

Step 1: Determine the Quantization Mode

Goal: Fix the quantization mode DAOS will implement, as the precondition for subsequent adaptation and configuration.

Action:

1. Determine the target bit width and numerical format of the linear layers' W/A (such as MXFP4 / INT4).
2. Determine whether FA quantization of Attention is needed to accelerate inference (whether to include fa3_quant).
3. Determine TLQ's preferred OP path: usually minmax_tune + round_tune; you may also enable trainable_smooth alone according to the TLQ usage guide (it cannot be combined with the former two).

The required memory and time are hard to estimate accurately before configuration, so there is no need to quantify resources in this step; rely on the actual run in Step 3.

Output: A clear statement of the quantization mode (W/A target, whether FA is enabled, TLQ OP path).

Step 2: Complete Model Adaptation and Configuration Authoring

Goal: Complete adaptation according to the corresponding algorithm documents, and write a loadable DAOS chained configuration into spec.process.

Action:

1. If the target model already has a lab_practice recipe containing the complete DAOS chain, prefer reusing it.
2. Otherwise, complete adaptation and configuration authoring in the order OASQ → TLQ → optional FA3:
3. OASQ: see the OASQ Outlier-Aware Smooth Quantization Algorithm Usage Guide to complete subgraph adaptation and write oasq.
4. TLQ: see the Trainable Linear Quantization Algorithm Usage Guide to complete adaptation and write trainable_linear_quant; configure operations according to the path chosen in Step 1; keep strategies[].qconfig consistent with the final deployment bit width. If trainable_smooth is selected, confirm the subgraph capability according to that guide.
5. FA3: if Step 1 decided to enable FA, see the FA3 Quant Parameter Configuration Workflow Guide to complete adaptation, and append fa3_quant after TLQ.
6. For field types and constraints, see the related documents in Chapter 6.
7. Verify the chaining constraints:
8. The order is OASQ → TLQ → optional FA3.
9. OASQ's symmetric should be consistent with the symmetry assumption of the subsequent quantization.
10. Do not mix minmax_tune / round_tune with trainable_smooth.

Output: The DAOS process chained entries written into the YAML.

Step 3: Start Quantization

Goal: Start the whole-chain quantization and get it running; if memory is insufficient, adjust the resource-related parameters below first and retry.

Action:

1. Start quantization using the configuration from Step 2.
2. If OASQ appears to be a no-op, first troubleshoot the adaptation and include/exclude according to the OASQ Outlier-Aware Smooth Quantization Algorithm Usage Guide, then continue with the subsequent training-related parameter adjustments.
3. When memory is insufficient, try the following adjustments first, changing only one item at a time; for details see the Trainable Linear Quantization Algorithm Usage Guide:
4. Increase train_config.gradient_accumulate_steps to reduce peak memory; this change alters the effective update frequency.
5. Narrow the low-bit range that needs training in the TLQ strategies, for example by adding exclude entries or moving some modules to a higher-bit-width strategy.
6. If necessary, temporarily lower train_config.iters to first verify the chain runs; restore a reasonable number of iterations for formal accuracy runs.
7. Once it runs, the quantization artifacts are produced; proceed to the evaluation in Step 4.

Output: A reproducible quantization run result; if the configuration was adjusted due to insufficient memory, keep the final working configuration.

Step 4: Evaluation and Iterative Tuning

Goal: After quantization, perform an end-to-end evaluation; if it falls short, iterate on single-variable tuning until the target is met or a fallback strategy is determined.

Action:

1. Keep the evaluation conditions unchanged and compare the end-to-end metrics before and after quantization (or against a single-calibration baseline).
2. When accuracy falls short, make single-variable adjustments according to the corresponding algorithm's usage guide and re-quantize, for example:
3. OASQ: scope, enable_subgraph_type, max_iters, and so on.
4. TLQ: strategies mixed bit widths, train_config (iters / lr, etc.), and OP-related switches.
5. FA3: qconfig and the Attention scope (if enabled).
6. After the upstream OASQ or the main target bit width changes, re-run the whole chain rather than splicing together old results.
7. Change only one item at a time and keep a record of adjustments relative to the baseline.

Output: The final DAOS snippet that can be merged into the task configuration, along with the record of adjustments relative to the baseline.

5. Terminology

Term	Brief Description	Link
DAOS Low-Bit Quantization Solution	The solution's definition, principle, properties, and limitations.	DAOS Low-Bit Quantization Solution Terminology Entry
OASQ	DAOS's upstream outlier suppression.	OASQ Outlier-Aware Smooth Quantization Algorithm Terminology Entry
Trainable Linear Quant	DAOS's trainable linear quantization stage.	Trainable Linear Quantization Algorithm Terminology Entry
FA3 Quant	Optional Attention FA quantization.	FA3 Quant Attention Activation Quantization Algorithm Terminology Entry

6. Related Documents

API or Document	Brief Description	Link
oasq Configuration Description	Field types, default values, and constraints.	oasq Configuration Description
trainable_linear_quant Configuration Description	Field types, default values, and constraints.	trainable_linear_quant Configuration Description
fa3_quant Configuration Description	Field types, default values, and constraints.	fa3_quant Configuration Description
multimodal_sd_modelslim_v1	Multimodal generative task-level configuration.	multimodal_sd_modelslim_v1 Configuration Description
Weight Quantization Usage Guide	User guide: quantization command parameters and complete usage instructions.	Weight Quantization Usage Guide