For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo module

attention

`comptime` values

`logger`

comptime logger = Logger(stdout, prefix=String(""), source_location=False)

Structs

FlashAttentionGPU:
MaskedFlashAttentionGPU:
MLAIndexerRaggedFloat8Paged:
NoMaskFlashAttentionCPU:
PaddedFlashAttentionGPU:
RaggedFlashAttentionGPU:
Struct_cross_attention_ragged_paged:
Struct_fused_qk_rope_padded_paged:
Struct_fused_qk_rope_ragged_paged:
Struct_fused_qk_rope_ragged_paged_with_position_id:
Struct_fused_qkv_matmul_padded_paged:
Struct_fused_qkv_matmul_padded_ragged:
Struct_fused_qkv_matmul_padded_ragged_bias:
Struct_fused_qkv_matmul_padded_ragged_bias_quantized:
Struct_fused_qkv_matmul_padded_ragged_quantized:
Struct_fused_qkv_matmul_padded_ragged_scale:
Struct_fused_qkv_matmul_padded_ragged_scale_bias:
Struct_fused_qkv_matmul_padded_ragged_scale_float4:
Struct_mha_decode_num_partitions:
Struct_mha_padded_paged:
Struct_mha_ragged_paged_fp8_kv: MHA with bf16 Q and fp8_e4m3fn paged KV cache (dequant-staging path).
Struct_mha_ragged_paged_scalar_args:
Struct_mha_ragged_paged_sink_weights_scalar_args:
Struct_mla_compute_dispatch_args_scalar:
Struct_mla_decode_graph_bf16_paged:
Struct_mla_decode_graph_paged_fp8:
Struct_mla_decode_graph_paged_fp8_sparse:
Struct_mla_decode_ragged_paged:
Struct_mla_decode_ragged_paged_scaled:
Struct_mla_decompress_k_cache_ragged_paged:
Struct_mla_prefill_graph_bf16_paged:
Struct_mla_prefill_graph_decode_bf16_paged:
Struct_mla_prefill_graph_decode_bf16_paged_quantized:
Struct_mla_prefill_graph_decode_paged_fp8:
Struct_mla_prefill_graph_decode_paged_fp8_sparse:
Struct_mla_prefill_graph_paged:
Struct_mla_prefill_ragged_paged:
Struct_mla_prefill_ragged_plan:
Struct_mla_prefill_sparse_paged:
Struct_mla_prefill_sparse_paged_fp8:
WithMaskFlashAttentionCPU:
WithMaskFlashAttentionSplitKVCPU:

comptime values​

logger​

Structs​

`comptime` values

`logger`

Structs