Remove padding and multiple D2D copies for MTP (#24086)

* Make ggml_gated_delta_net take only the initial recurrent state (D, 1, n_seqs) and passes the snapshot count K as an op parameter instead of inferring it from state->ne[1].

Remove the padding hack and copy all emitted snapshots into the recurrent cache with a single strided ggml_cpy

* Make GDN changes in all backends. Address review comments.

* Fix CI build errors

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -11528,7 +11528,6 @@ static void ggml_vk_gated_delta_net(ggml_backend_vk_context * ctx, vk_context& s
     const ggml_tensor * src_q     = dst->src[0];
     const ggml_tensor * src_v     = dst->src[2];
     const ggml_tensor * src_beta  = dst->src[4];
-    const ggml_tensor * src_state = dst->src[5];
 
     GGML_ASSERT(dst->buffer != nullptr);
 
@@ -11537,8 +11536,8 @@ static void ggml_vk_gated_delta_net(ggml_backend_vk_context * ctx, vk_context& s
     const uint32_t n_tokens = (uint32_t)src_v->ne[2];
     const uint32_t n_seqs   = (uint32_t)src_v->ne[3];
 
-    // state is 3D (S_v*S_v*H, K, n_seqs); K is the snapshot slot count.
-    const uint32_t K = (uint32_t)src_state->ne[1];
+    // K (snapshot slot count) is an op param; state holds s0 only [S_v, S_v, H, n_seqs].
+    const uint32_t K = (uint32_t)ggml_get_op_params_i32(dst, 0);
 
     const uint32_t s_off = S_v * H * n_tokens * n_seqs;
 
@@ -17954,7 +17953,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
             src_clone[4], src_clone[5], src_clone[6]);
         } else if (tensor->op == GGML_OP_GATED_DELTA_NET) {
             tensor_clone = ggml_gated_delta_net(ggml_ctx, src_clone[0], src_clone[1],
-            src_clone[2], src_clone[3], src_clone[4], src_clone[5]);
+            src_clone[2], src_clone[3], src_clone[4], src_clone[5],
+            ggml_get_op_params_i32(tensor, 0));
         } else if (tensor->op == GGML_OP_OPT_STEP_ADAMW) {
             src_clone[0]->flags = tensor->src[0]->flags;
             tensor_clone = ggml_opt_step_adamw(ggml_ctx, src_clone[0], src_clone[1],

ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp

@@ -102,8 +102,8 @@ void main() {
     const uint iq3 = seq_id / rq3;
 
     const uint state_size = S_V * S_V;
-    // input state layout (D, K, n_seqs): per-seq stride is K*H*D; we read slot 0.
-    const uint state_in_base       = (seq_id * K * H + head_id) * state_size;
+    // input state holds s0 only [S_v, S_v, H, n_seqs]: per-seq stride is H*D.
+    const uint state_in_base       = (seq_id * H + head_id) * state_size;
     // output state layout per slot: same per-(seq,head) offset as the single-slot case.
     const uint state_out_base      = (seq_id * H + head_id) * state_size;
     const uint state_size_per_snap = state_size * H * n_seqs;
@@ -113,9 +113,8 @@ void main() {
         s_shard[r] = FLOAT_TYPE(data_state[state_in_base + col * S_V + r * LANES_PER_COLUMN + lane]);
     }
 
-    // snapshot slot mapping: target_slot = t - shift. When n_tokens < K, only the last
-    // n_tokens slots are written; earlier slots are left untouched (caller-owned).
-    const int shift = int(n_tokens) - int(K);
+    // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+    // When n_tokens < K, only slots 0..n_tokens-1 are written; older slots are caller-owned.
 
     uint attn_off = (seq_id * n_tokens * H + head_id) * S_V;
 
@@ -172,7 +171,7 @@ void main() {
         attn_off += S_V * H;
 
         if (K > 1u) {
-            const int target_slot = int(t) - shift;
+            const int target_slot = int(n_tokens) - 1 - int(t);
             if (target_slot >= 0 && target_slot < int(K)) {
                 const uint slot_base = s_off + uint(target_slot) * state_size_per_snap + state_out_base;
                 [[unroll]] for (uint r = 0; r < ROWS_PER_LANE; r++) {