Training
MXFP4 quantization
FP4 microscalingReferenceOCP microscaling FP4, block size 32 — the open cross-vendor standard.
- Relieves
- CapacityBandwidthCompute
- Targets
- Weights · Activations
- Format
- MXFP4 (E2M1, block 32)
- Granularity
- Microscaling — block size 32
- Lifecycle
- PTQ · no gradients
- Calibration
- Small calibration set
- Compression
- ~4× vs BF16
- Quality
- Good; block 32 adapts less locally than NVFP4
- Hardware
- Blackwell / MX-capable tensor cores for speedup
- Runtimes
- vLLMTensorRT-LLM
A worked Spark recipe for this method hasn't been written yet — it lives here as a reference point in the ontology.