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 struct
Struct_mha_ragged_paged_fp8_kv
struct Struct_mha_ragged_paged_fp8_kv
MHA with bf16 Q and fp8_e4m3fn paged KV cache (dequant-staging path).
Dequantizes the fp8 KV blocks to a bf16 staging buffer using per-block float32 scales, then calls the standard bf16 flash attention kernel on the staging buffer.
Implemented traits
AnyType,
ImplicitlyDestructible
Methods
execute
static def execute[scale_dtype: DType, //, quantization_granularity: Int, target: StringSlice[StaticConstantOrigin], mask_str: StringSlice[StaticConstantOrigin], local_window_size: Int = -1](output: ManagedTensorSlice[Output, static_spec=output.static_spec], q: ManagedTensorSlice[Input, static_spec=q.static_spec], input_row_offsets: ManagedTensorSlice[Input, static_spec=input_row_offsets.static_spec], kv_blocks: ManagedTensorSlice[MutableInput, static_spec=kv_blocks.static_spec], cache_lengths: ManagedTensorSlice[Input, static_spec=cache_lengths.static_spec], kv_lookup_table: ManagedTensorSlice[Input, static_spec=kv_lookup_table.static_spec], max_lengths: ManagedTensorSlice[Input, static_spec=max_lengths.static_spec], kv_scales: ManagedTensorSlice[MutableInput, static_spec=kv_scales.static_spec], layer_idx: UInt32, scale: Float32, mha_decode_dispatch_metadata: ManagedTensorSlice[Input, static_spec=mha_decode_dispatch_metadata.static_spec], ctx: DeviceContext)
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!