Skip to content

Quantization support

A check mark means that the platform class declares the method. NVIDIA CUDA accepts all registered methods, but each method can require a newer compute capability or an optional kernel package. Test your model on the target device before production use.

MethodMinimum NVIDIA capabilityAMD ROCmIntel XPU
awqSM 7.5
auto_awqSM 7.5
fp8SM 7.5
fbgemm_fp8SM 8.0
fp_quantSM 10.0
modeloptSM 8.9
modelopt_fp4SM 7.5
modelopt_mxfp8SM 8.0
modelopt_mixedSM 7.5
auto_gptqSM 6.0
gptqSM 6.0
gptq_marlinSM 6.0
awq_marlinSM 7.5
hummingSM 7.5
compressed-tensorsSM 7.0
bitsandbytesSM 7.0
experts_int8SM 8.0
exl3SM 8.0
quarkSM 7.0
moe_wna16SM 7.0
torchaoSM 7.5
incSM 6.0
mxfp4SM 8.0
gpt_oss_mxfp4SM 8.0
deepseek_v4_fp8SM 7.5
onlineSM 7.5
fp8_per_tensorSM 7.5
fp8_per_blockSM 7.5
fp8_per_channelSM 7.5
int8_per_channel_weight_onlySM 7.5
nvfp4_per_tokenSM 7.5
mxfp8SM 8.0

Read the table

  • means that the platform registry permits the method.
  • means that the platform does not declare the method.
  • A permitted method can still require a specific GPU generation.
  • CPU, Metal, and TPU do not yet publish a complete method list in their platform classes. Their columns will be added when that metadata becomes available.