flashinfer.mla.supported_sparse_mla_sm120_configs

flashinfer.mla.supported_sparse_mla_sm120_configs() → dict[str, SparseMLASm120DecodeConfig]

Enumerate the instantiated SM120 sparse-MLA decode kernel configurations.

Lets callers validate a serving configuration at initialization time instead of discovering an uninstantiated (num_heads, topk) pair on the first decode-form request.

Returns:

Mapping from kernel family to its instantiated decode set, keyed by "dsv4" (d_qk=512), "dsv3_2" (d_qk=576, power-of-2 FP32 scales), "glm_nsa" (d_qk=576, arbitrary FP32 scales; shares the DSv3.2 decode instantiations), "glm53_nope" (GLM-5.3 native NoPE, d_qk=512, arbitrary FP32 scales), and "dots3_swa" (sliding-window family, d_qk=1088, UE8M0 scales).

Return type:

dict[str, SparseMLASm120DecodeConfig]

Examples

>>> import flashinfer
>>> configs = flashinfer.mla.supported_sparse_mla_sm120_configs()
>>> configs["dsv4"].supports_decode(num_heads=64, topk=256)
True