Mojo function

mla_decode_branch_fp8

mla_decode_branch_fp8[dtype: DType, fp8_dtype: DType, fp8_scale_dtype: DType, collection_t: KVCollectionT, //, 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: TileTensor[dtype, origin, linear_idx_type=linear_idx_type, element_shape_types=element_shape_types], q: TileTensor[dtype, origin, linear_idx_type=linear_idx_type, element_shape_types=element_shape_types], input_row_offsets: TileTensor[DType.uint32, origin, linear_idx_type=linear_idx_type, element_shape_types=element_shape_types], freqs_cis: TileTensor[dtype, origin, linear_idx_type=linear_idx_type, element_shape_types=element_shape_types], kv_norm_gamma: TileTensor[dtype, origin, linear_idx_type=linear_idx_type, element_shape_types=element_shape_types], kv_collection: collection_t, layer_idx: UInt32, scale: Float32, epsilon: Float32, w_uk: TileTensor[fp8_dtype, origin, linear_idx_type=linear_idx_type, element_shape_types=element_shape_types], w_uk_scale: TileTensor[fp8_scale_dtype, origin, linear_idx_type=linear_idx_type, element_shape_types=element_shape_types], w_uv: TileTensor[fp8_dtype, origin, linear_idx_type=linear_idx_type, element_shape_types=element_shape_types], w_uv_scale: TileTensor[fp8_scale_dtype, origin, linear_idx_type=linear_idx_type, element_shape_types=element_shape_types], ctx: DeviceContext)

This is a manually fused kernel that performs the following operations: - Apply RoPE to the query and the key cache (in-place). - Apply RMSNorm to the non-rope portion of the key cache (in-place). - Project q_nope to kv_latent_dim through a fp8 batched matmul: q_nope_proj = q_nope_t @ w_uk. - Concatenate q_nope_proj and q_rope: q_full = concat(q_nope_proj, q_rope, axis=2). - Perform MLA decode. - Project raw_output to v_head_dim through another fp8 batched matmul: output = raw_output_t @ w_uv.

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.
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 (TileTensor): Output tensor of shape [tot_seq_len, num_heads, v_head_dim].
q (TileTensor): Combined query tensor containing both nope and rope parts. Shape: [tot_seq_len, num_heads, qk_nope_head_dim + qk_rope_head_dim].
input_row_offsets (TileTensor): Indicates where each request starts and ends in q. Shape: [num_batches + 1].
freqs_cis (TileTensor): Precomputed RoPE frequency values for rotary position embeddings. Shape: [max_seq_len, qk_rope_head_dim].
kv_norm_gamma (TileTensor): RMSNorm gamma weights for normalizing the KV cache. Shape: [kv_lora_rank].
kv_collection (collection_t): Paged KV Cache object.
layer_idx (UInt32): Layer index.
scale (Float32): Scale for the attention calculation.
epsilon (Float32): Small constant for numerical stability in RMSNorm.
w_uk (TileTensor): Weight matrix for projecting the non-rope part of each query head to KV latent space. Shape: [num_heads, kv_latent_dim, qk_nope_head_dim].
w_uk_scale (TileTensor): The scale for the w_uk weight matrix. Shape varies depending on the float8_config.
w_uv (TileTensor): Weight matrix for projecting the output of the attention back to each head's original space. Shape: [num_heads, v_head_dim, kv_latent_dim].
w_uv_scale (TileTensor): The scale for the w_uv weight matrix. Shape varies depending on the float8_config.
ctx (DeviceContext): Device context.