Mix Precision Auto-Search

The Mix Precision Auto-Search workflow automatically finds the optimal mixed-precision quantization configuration for large language models running on AMD Instinct GPUs via vLLM.

Given an accuracy loss budget—for example, GSM8K must not drop by more than 2%—it uses a hardware-aware decode Roofline to walk adjacent performance candidates and selects the fastest evaluated configuration whose accuracy stays within the threshold.

Supported Quantization Modes

Hardware	Supported Modes
MI300	native, fp8, ptpc_fp8
MI325	native, fp8, ptpc_fp8
MI355	native, fp8, ptpc_fp8, mxfp4, mxfp4_fp8, mxfp6_e2m3

Workflow Overview

1. Load model on meta device — builds the layer graph for quantization config generation with no GPU memory cost.

2. Generate candidate configs — constrained by the target hardware and requested mode subset. With --file2file_quantization, candidates containing calibration-dependent fp8 or mxfp4_fp8 (W4A8) modes are reported in a warning and removed before evaluation.

3. Compute the Roofline — walk the meta model’s Linear shapes and score candidates with per-op GEMM, FusedMoE, and SDPA compute/memory ceilings on the target GPU, with aggregate memory as a fallback. The default workload is ISL=8192 / OSL=1024; a separate prefill Roofline is reported but does not affect ordering.

4. Evaluate baseline — run GSM8K on the original (unquantized) vLLM model.

5. Search loop — start from the hardware anchor, walk adjacent Roofline candidates, re-quantize, evaluate, check the threshold, and reset.

6. Select best config — the highest-scoring evaluated config under the Roofline model that passes.

7. Export (optional) — apply best config and save as safetensors. With --file2file_quantization, the best calibration-free config is exported one checkpoint shard at a time without loading the full model. Hugging Face model IDs are resolved to the local cache automatically; local inputs must contain config.json and at least one .safetensors shard.

MI300 and MI325 use mlp=ptpc_fp8 as the preferred anchor. MI355 uses mlp=mxfp4. All other layer partitions start as native.

Quick Start

Note

This workflow requires the vLLM ROCm container. See the example README for full environment setup instructions.

cd /workspace/quark/examples/torch/experimental/mix_precision

# Search all modes for MI300
python mix_precision.py \
    --model_dir /models/Qwen3.5-397B-A17B \
    --hardware mi300 \
    -tp 8 \
    --gpu-memory-utilization 0.8 \
    --export_best_model \
    --file2file_quantization

Further Reading

• Example script and README

• API design document