Mojo function

mla_prefill_branch_fp8

mla_prefill_branch_fp8[dtype: DType, fp8_dtype: DType, fp8_scale_dtype: DType, collection_t: KVCollectionT, //, qk_nope_head_dim: Int, m_scale_granularity: Int, n_scale_granularity: Int, k_scale_granularity: Int, mask_str: StringSlice[StaticConstantOrigin], score_mod_str: StringSlice[StaticConstantOrigin], target: StringSlice[StaticConstantOrigin] = "cpu"](output: LayoutTensor[dtype, layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], q: LayoutTensor[dtype, layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], input_row_offsets: LayoutTensor[DType.uint32, layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], kv_collection: collection_t, layer_idx: UInt32, scale: Float32, buffer_row_offsets: LayoutTensor[DType.uint32, layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], cache_offsets: LayoutTensor[DType.uint32, layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], buffer_length: Int, kv_b_proj: LayoutTensor[fp8_dtype, layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], kv_b_proj_scale: LayoutTensor[fp8_scale_dtype, layout, origin, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], ctx: DeviceContext)

This is a manually fused kernel that performs the following operations: - Copy the KV latent values from PagedKVCache to a contiguous buffer. - Quantize the KV latent values to fp8. - Up-project the latent KV values to full K and V through a matmul. - Split the concatenated KV into K and V. - Perform MLA prefill.

Parameters:

dtype (DType): Data type of the input and output tensors.
fp8_dtype (DType): Data type of the fp8 input and output tensors.
fp8_scale_dtype (DType): Data type of the fp8 scale input and output tensors.
collection_t (KVCollectionT): Type of the KV collection.
qk_nope_head_dim (Int): Dimension of non-rope parts of the Q/K heads.
m_scale_granularity (Int): Granularity of the scale for M dimension of the matrix multiplication.
n_scale_granularity (Int): Granularity of the scale for N dimension of the matrix multiplication.
k_scale_granularity (Int): Granularity of the scale for K dimension of the matrix multiplication.
mask_str (StringSlice): Mask variant.
score_mod_str (StringSlice): Positional encoding variant.
target (StringSlice): Target device.

Args:

output (LayoutTensor): Output tensor of shape [tot_seq_len, num_heads, v_head_dim].
q (LayoutTensor): Query tensor of shape [tot_seq_len, num_heads, qk_nope_head_dim + qk_rope_head_dim].
input_row_offsets (LayoutTensor): Indicates where each request starts and ends in q. Shape: [num_batches + 1].
kv_collection (collection_t): Paged KV Cache object.
layer_idx (UInt32): Layer index.
scale (Float32): Scale for the attention calculation.
buffer_row_offsets (LayoutTensor): Indicates where each request's KV latent values should be stored in the contiguous K buffer. This is a 1D tensor of shape [num_batches + 1].
cache_offsets (LayoutTensor): Indicates the starting token position in the KV cache from which to copy KV latent values for each request. This is a 1D tensor of shape [num_batches + 1].
buffer_length (Int): The total number of tokens in the KV cache. Scalar.
kv_b_proj (LayoutTensor): Weight matrix for up-projecting the KV latent values to full K and V. Shape: [num_heads * (qk_nope_head_dim + v_head_dim), kv_latent_dim].
kv_b_proj_scale (LayoutTensor): The scale for the weight matrix. Shape varies depending on the float8_config.
ctx (DeviceContext): Device context.