| --- |
| library_name: kernels |
| tags: |
| - cuda |
| - pytorch |
| - flashrt |
| - gated-delta |
| - linear-attention |
| - qwen3 |
| - transformer |
| --- |
| |
| # Gated Delta Attention |
|
|
| BF16 Gated DeltaNet recurrent/chunk/WY kernels from FlashRT, packaged for |
| Hugging Face Kernel Hub. The v2 public profile covers Qwen3.6-style |
| linear-attention decode recurrence and prefill WY building blocks. |
|
|
| ## Available functions |
|
|
| - `gated_delta_recurrent_bf16` |
| - `gated_delta_recurrent_inout_bf16` |
| - `gated_delta_recurrent_f32state_bf16io` |
| - `gated_delta_chunk_bf16` |
| - `gated_delta_chunk_smem_bf16` |
| - `lin_split_qkv_broadcast_bf16` |
| - `lin_split_qkv_gqa_bf16` |
| - `split_q_gate_bf16` |
| - `gdn_gating_bf16` |
| - `gdn_gating_strided_bf16` |
| - `gdn_chunk_from_conv_smem_bf16` |
| - `gdn_wy_norm_cumsum_pack_qk_bf16` |
| - `gdn_wy_kkt_b64_bf16` |
| - `gdn_wy_solve_tril_b64_f32` |
| - `gdn_wy_recompute_wu_b64_bf16` |
| - `gdn_wy_chunk_h_b64_bf16` |
| - `gdn_wy_output_o_b64_bf16` |
|
|
| ## Usage |
|
|
| ```python |
| from kernels import get_kernel |
| |
| gdn = get_kernel("flashrt/gated-delta-attention", version=2, trust_remote_code=True) |
| out = gdn.gated_delta_recurrent_bf16(q, k, v, g, beta, state) |
| ``` |
|
|
| The WY helpers use the Qwen3.6 profile: `conv_out=(S,10240)`, |
| Q/K heads `16`, value heads `48`, head dimension `128`, and 64-token WY |
| blocks. |
