LLM inference in C/C++
C++ ggml
llama : share the k-pool input views between layers (#30613)
Aman Gupta6afa9a3134dc
diff --git a/src/models/glm5-next.cpp b/src/models/glm5-next.cppindex a1b7229507..90870bd967 100644--- a/src/models/glm5-next.cpp+++ b/src/models/glm5-next.cpp@@ -271,8 +271,10 @@ public:ggml_tensor * pool_mask = nullptr; // F32/F16 [n_pool, n_tokens]ggml_tensor * tail_idxs = nullptr; // I32 [kpool - 1, n_tokens]ggml_tensor * sel_mask = nullptr; // F32 [n_sel, 1, 1, n_tokens] 0 for live selection slots, -inf for dead ones+ ggml_tensor * sel_mask_2d = nullptr; // F32 [n_sel, n_tokens] view of sel_mask, shared by the layers// n_new is never below 1, see build_inp_kpoolggml_tensor * new_pool_idxs = nullptr; // I32 [kpool, n_new] members of the pools completed this ubatch+ ggml_tensor * new_pool_flat = nullptr; // I32 [kpool*n_new] view of new_pool_idxs, shared by the layersggml_tensor * new_pool_rep = nullptr; // I64 [n_new] cell to write each new pooled key intoconst llama_memory_hybrid_idx_context * mctx;@@ -326,12 +328,15 @@ llama_model_glm5_next::llm_graph_input_kpool * llama_model_glm5_next::graph::buiggml_set_input(inp->sel_mask);// set_input_kpool always fills the mask, so it stays allocated in every graphggml_build_forward_expand(gf, inp->sel_mask);+ // one view for all layers: each view of a host input is a separate upload+ inp->sel_mask_2d = ggml_reshape_2d(ctx0, inp->sel_mask, n_sel, n_tokens);}inp->n_new = n_new;inp->new_pool_idxs = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, kpool, n_new);ggml_set_input(inp->new_pool_idxs);+ inp->new_pool_flat = ggml_reshape_1d(ctx0, inp->new_pool_idxs, kpool*n_new);// one scatter row per new pool, each a distinct rep row (see kpool_build_state)inp->new_pool_rep = ggml_new_tensor_1d(ctx0, GGML_TYPE_I64, n_new);ggml_set_input(inp->new_pool_rep);@@ -917,7 +922,7 @@ ggml_tensor * llama_model_glm5_next::graph::build_kpool_select(const int64_t n_kv = mctx_lid->get_n_kv();// Pool the entries completed by this ubatch. The last one is a dummy when the ubatch completes none.- ggml_tensor * rows = kpool_cache.gather_key_gate(ggml_reshape_1d(ctx0, inp_kpool->new_pool_idxs, kpool*n_new));+ ggml_tensor * rows = kpool_cache.gather_key_gate(inp_kpool->new_pool_flat);rows = ggml_reshape_3d(ctx0, rows, 2*n_embd_indexer, kpool, n_new);ggml_tensor * pk = ggml_view_3d(ctx0, rows, n_embd_indexer, kpool, n_new, rows->nb[1], rows->nb[2], 0);@@ -1005,7 +1010,7 @@ ggml_tensor * llama_model_glm5_next::graph::build_kpool_select(// Live slots (visible pools, real tail cells) address disjoint cells. Each dead slot writes its own dump row// n_kv + slot, so the scatter indices of a token are unique: idx = dump + live*(idx - dump), live = exp(mask).GGML_ASSERT(inp_kpool->sel_mask->ne[0] == n_sel && inp_kpool->sel_mask->ne[3] == n_tokens);- ggml_tensor * live = ggml_exp(ctx0, ggml_reshape_2d(ctx0, inp_kpool->sel_mask, n_sel, n_tokens));+ ggml_tensor * live = ggml_exp(ctx0, inp_kpool->sel_mask_2d);ggml_tensor * dump = ggml_arange(ctx0, (float) n_kv, (float) (n_kv + n_sel), 1.0f);ggml_tensor * idx_f = ggml_cast(ctx0, sel_idx, GGML_TYPE_F32);idx_f = ggml_add(ctx0, ggml_mul(ctx0, ggml_sub(ctx0, idx_f, dump), live), dump);diff --git a/src/models/qwen4exp.cpp b/src/models/qwen4exp.cppindex e055abdf55..77430e7afd 100644--- a/src/models/qwen4exp.cpp+++ b/src/models/qwen4exp.cpp@@ -682,6 +682,7 @@ public:ggml_tensor * pool_mask = nullptr; // F16 [n_pool, n_tokens]ggml_tensor * tail_idxs = nullptr; // I32 [kpool - 1, n_tokens]ggml_tensor * new_pool_idxs = nullptr; // I32 [kpool, n_new] members of the blocks to re-pool this ubatch+ ggml_tensor * new_pool_flat = nullptr; // I32 [kpool*n_new] view of new_pool_idxs, shared by the QSA layersggml_tensor * new_pool_rep = nullptr; // I64 [n_new] cell to write each new pooled key intoggml_tensor * new_pool_pos = nullptr; // I32 [4*n_new] M-RoPE position of each new block's first member@@ -724,6 +725,8 @@ llama_model_qwen4exp::llm_graph_input_kpool * llama_model_qwen4exp::graph::buildinp->new_pool_idxs = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, kpool, inp->n_new);ggml_set_input(inp->new_pool_idxs);+ // one view for all layers: each view of a host input is a separate upload+ inp->new_pool_flat = ggml_reshape_1d(ctx0, inp->new_pool_idxs, kpool*inp->n_new);// one scatter row per new pool, each a distinct rep row (see kpool_build_state)inp->new_pool_rep = ggml_new_tensor_1d(ctx0, GGML_TYPE_I64, inp->n_new);ggml_set_input(inp->new_pool_rep);@@ -765,7 +768,7 @@ ggml_tensor * llama_model_qwen4exp::graph::build_qsa_sel(auto kpool_cache = mctx_hyb->get_kpool_access(ctx0, il, idx_dim);// pool only the blocks this ubatch completes or regroups- ggml_tensor * rows = kpool_cache.gather_key_gate(ggml_reshape_1d(ctx0, inp_kpool->new_pool_idxs, kpool*n_new));+ ggml_tensor * rows = kpool_cache.gather_key_gate(inp_kpool->new_pool_flat);rows = ggml_reshape_3d(ctx0, rows, idx_dim, kpool, n_new);// mean over the members; kpool is small, so summing slices beats a transpose plus sum_rows