Skip to content

LLM inference in C/C++

C++ ggml

llama : build the k-pool selection in the causal mask layout (#30480)

Georgi Gerganove7268bf6a996
diff --git a/src/models/glm5-next.cpp b/src/models/glm5-next.cpp
index 32199a2603..a1b7229507 100644
--- a/src/models/glm5-next.cpp
+++ b/src/models/glm5-next.cpp
@@ -988,6 +988,7 @@ ggml_tensor * llama_model_glm5_next::graph::build_kpool_select(
}
}
const int64_t n_sel = sel_idx->ne[0];
+ GGML_ASSERT(n_sel == (int64_t) inp_kpool->n_sel);
ggml_build_forward_expand(gf, sel_idx);
@@ -1011,11 +1012,13 @@ ggml_tensor * llama_model_glm5_next::graph::build_kpool_select(
sel_idx = ggml_cast(ctx0, idx_f, GGML_TYPE_I32);
ggml_tensor * sel = ggml_set_rows(ctx0, mask_all, zeros, ggml_reshape_3d(ctx0, sel_idx, n_sel, n_tokens, 1));
- sel = ggml_view_2d(ctx0, sel, n_kv, n_tokens, sel->nb[2], 0);
// Fold causal visibility before shared-indexer reuse.
GGML_ASSERT(kq_mask->ne[0] == n_kv && kq_mask->ne[1]*kq_mask->ne[2]*kq_mask->ne[3] == n_tokens);
- sel = ggml_add(ctx0, sel, ggml_reshape_2d(ctx0, kq_mask, n_kv, n_tokens));
+ const size_t row = sel->nb[2];
+ sel = ggml_view_4d(ctx0, sel, n_kv, kq_mask->ne[1], kq_mask->ne[2], kq_mask->ne[3],
+ row, row*kq_mask->ne[1], row*kq_mask->ne[1]*kq_mask->ne[2], 0);
+ sel = ggml_add(ctx0, sel, kq_mask);
cb(sel, "indexer_sel", il);
return sel;
@@ -1070,14 +1073,14 @@ ggml_tensor * llama_model_glm5_next::graph::build_dsa_layer(
ggml_build_forward_expand(gf, kv_cmpr);
ggml_build_forward_expand(gf, mctx_mla->cpy_k(ctx0, kv_cmpr, inp_attn->get_k_idxs(), il));
- // The scatter selection already includes the causal mask.
- ggml_tensor * mask = ggml_reshape_4d(ctx0, sel, kq_mask->ne[0], kq_mask->ne[1], kq_mask->ne[2], kq_mask->ne[3]);
- cb(mask, "kq_mask_dsa", il);
+ // The scatter selection already includes the causal mask, and it comes in the mask layout
+ GGML_ASSERT(ggml_are_same_shape(sel, kq_mask));
+ cb(sel, "kq_mask_dsa", il);
ggml_tensor * k = mctx_mla->get_k(ctx0, il);
ggml_tensor * v = ggml_view_4d(ctx0, k, kv_lora_rank, k->ne[1], k->ne[2], k->ne[3], k->nb[1], k->nb[2], k->nb[3], 0);
- ggml_tensor * out = build_attn_mha(q_absorbed, k, v, nullptr, mask, nullptr, layer.wv_b, inp_kpool->n_sel, kq_scale, il);
+ ggml_tensor * out = build_attn_mha(q_absorbed, k, v, nullptr, sel, nullptr, layer.wv_b, inp_kpool->n_sel, kq_scale, il);
cb(out, "kqv_out", il);
out = ggml_mul_mat(ctx0, layer.wo, out);
diff --git a/src/models/qwen4exp.cpp b/src/models/qwen4exp.cpp
index f68a8622c7..e055abdf55 100644
--- a/src/models/qwen4exp.cpp
+++ b/src/models/qwen4exp.cpp
@@ -905,16 +905,16 @@ ggml_tensor * llama_model_qwen4exp::graph::build_attn_qsa(
ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il));
}
- // the selection mask already carries the causal mask
+ // the selection already carries the causal mask, and it comes in the mask layout
ggml_tensor * kq_mask = inp->get_kq_mask();
- ggml_tensor * mask = ggml_reshape_4d(ctx0, sel, kq_mask->ne[0], kq_mask->ne[1], kq_mask->ne[2], kq_mask->ne[3]);
- cb(mask, "kq_mask_qsa", il);
+ GGML_ASSERT(ggml_are_same_shape(sel, kq_mask));
+ cb(sel, "kq_mask_qsa", il);
ggml_tensor * q = q_cur;
ggml_tensor * k = mctx_cur->get_k(ctx0, il);
ggml_tensor * v = mctx_cur->get_v(ctx0, il);
- ggml_tensor * cur = build_attn_mha(q, k, v, nullptr, mask, nullptr, nullptr, n_sel, kq_scale, il);
+ ggml_tensor * cur = build_attn_mha(q, k, v, nullptr, sel, nullptr, nullptr, n_sel, kq_scale, il);
cb(cur, "kqv_out", il);
// the rotation is its own inverse, so undo it on the value side of the output