[Kernel][Perf] Add fused CUDA post-conv MTP decode kernel for Qwen3.5 GDN (#51674)

Signed-off-by: Jie Fang <jief@nvidia.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Jie Fang
2026-08-14 11:44:45 +08:00
committed by GitHub
co-authored by OpenAI Codex Claude Fable 5
parent d18bb7b390
commit 1be3628367
12 changed files with 1456 additions and 7 deletions
+14
View File
@@ -534,6 +534,15 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
"Tensor? norm_weight=None, float norm_eps=1e-5) -> ()");
#endif
#ifdef VLLM_ENABLE_FUSED_GDN_DECODE
ops.def(
"fused_gdn_decode_post_conv_mtp("
"Tensor mixed_qkv, Tensor a, Tensor b, Tensor A_log, Tensor dt_bias, "
"Tensor state_indices, Tensor cu_seqlens, Tensor num_accepted_tokens, "
"Tensor! state, Tensor output_gate, Tensor norm_weight, Tensor! out, "
"float scale, float norm_eps=1e-5) -> ()");
#endif
#ifdef VLLM_ENABLE_KIMI_K3_ATTN_RES
ops.def(
"kimi_k3_attn_res("
@@ -797,6 +806,11 @@ STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, ops) {
ops.impl("fused_kda_decode", TORCH_BOX(&fused_kda_decode));
#endif
#ifdef VLLM_ENABLE_FUSED_GDN_DECODE
ops.impl("fused_gdn_decode_post_conv_mtp",
TORCH_BOX(&fused_gdn_decode_post_conv_mtp));
#endif
#ifdef VLLM_ENABLE_KIMI_K3_ATTN_RES
ops.impl("kimi_k3_attn_res", TORCH_BOX(&kimi_k3_attn_res));
#endif