Skip to content

LLM inference in C/C++

C++ ggml

models : Record layer inputs for all granite models (#30957)

Gabe Goodhartbf039c84969b
diff --git a/src/models/granite-hybrid.cpp b/src/models/granite-hybrid.cpp
index fb3bf7024b..a752a365cf 100644
--- a/src/models/granite-hybrid.cpp
+++ b/src/models/granite-hybrid.cpp
@@ -153,6 +153,8 @@ llama_model_granite_hybrid::graph::graph(const llama_model & model, const llm_gr
}
for (int il = 0; il < n_layer; ++il) {
+ res->t_layer_inp[il] = inpL;
+
struct ggml_tensor * inpSA = inpL;
// norm
diff --git a/src/models/granite-swa.cpp b/src/models/granite-swa.cpp
index f683ef6e8e..60c1eb8ba8 100644
--- a/src/models/granite-swa.cpp
+++ b/src/models/granite-swa.cpp
@@ -154,6 +154,8 @@ llama_model_granite_swa::graph::graph(
cb(inpL, "deepstack_in", il);
}
+ res->t_layer_inp[il] = inpL;
+
ggml_tensor * inpSA = inpL;
// norm
diff --git a/src/models/granite-switch.cpp b/src/models/granite-switch.cpp
index a73fcdaff5..f67bf6055f 100644
--- a/src/models/granite-switch.cpp
+++ b/src/models/granite-switch.cpp
@@ -293,6 +293,8 @@ llama_model_granite_switch::graph::graph(
ggml_tensor * cur;
for (int il = 0; il < n_layer; ++il) {
+ res->t_layer_inp[il] = inpL;
+
ggml_tensor * inpSA = inpL;
cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
diff --git a/src/models/granite.cpp b/src/models/granite.cpp
index 4cac495953..aa4d010956 100644
--- a/src/models/granite.cpp
+++ b/src/models/granite.cpp
@@ -159,6 +159,8 @@ llama_model_granite::graph::graph(
cb(inpL, "deepstack_in", il);
}
+ res->t_layer_inp[il] = inpL;
+
ggml_tensor * inpSA = inpL;
// norm