From decfaf4dd32c8b14b2dec12352f4e559477f3cd1 Mon Sep 17 00:00:00 2001
From: Kawrakow <iwankawrakow@gmail.com>
Date: Wed, 10 Jun 2026 09:31:31 +0000
Subject: [PATCH] A few more named nodes

---
 src/graphs/build_gemma4.cpp | 4 +++-
 src/llama-build-context.cpp | 5 ++---
 2 files changed, 5 insertions(+), 4 deletions(-)

diff --git a/src/graphs/build_gemma4.cpp b/src/graphs/build_gemma4.cpp
index ffbfb3e6..a0c555df 100644
--- a/src/graphs/build_gemma4.cpp
+++ b/src/graphs/build_gemma4.cpp
@@ -690,10 +690,11 @@ static ggml_tensor * gemma4_project_per_layer_inputs(ggml_context * ctx0, const
     const float per_layer_input_scale      = 1.0f / sqrtf(2.0f);
 
     ggml_tensor * per_layer_proj = ggml_mul_mat(ctx0, model.per_layer_model_proj, inputs_embeds);
+    cb(per_layer_proj, "per_layer_proj", -1);
     per_layer_proj               = ggml_reshape_3d(ctx0, per_layer_proj, n_embd_per_layer, n_layer, n_tokens);
     per_layer_proj               = llm_build_context::llm_build_norm(ctx0, per_layer_proj, model.hparams,
             model.per_layer_proj_norm, nullptr, LLM_NORM_RMS, cb, -1, 1.0f*n_embd);  // [n_embd_per_layer, n_layer, n_tokens]
-    cb(per_layer_proj, "per_layer_proj", -1);
+    cb(per_layer_proj, "per_layer_proj_normed", -1);
 
     inp_per_layer = ggml_add(ctx0, per_layer_proj, inp_per_layer);
     inp_per_layer = ggml_scale(ctx0, inp_per_layer, per_layer_input_scale);
@@ -948,6 +949,7 @@ ggml_cgraph * llm_build_context::build_gemma4() {
     cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
 
     if (hparams.f_final_logit_softcapping > 0) {
+        cb(cur, "result_pre_softcap", -1);
         cur = ggml_softcap(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping, hparams.f_final_logit_softcapping);
     }
 
diff --git a/src/llama-build-context.cpp b/src/llama-build-context.cpp
index e7dfa966..d0b3f49b 100644
--- a/src/llama-build-context.cpp
+++ b/src/llama-build-context.cpp
@@ -843,16 +843,15 @@ ggml_tensor * llm_build_context::llm_build_ffn(
         }
         if (down) {
             cur = llm_build_lora_mm(lctx, ctx, down, cur);
+            cb(cur, "ffn_down", il);
             if (lctx.model.arch == LLM_ARCH_GLM4 || lctx.model.arch == LLM_ARCH_GLM4_MOE) {
                 // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
                 ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
             }
         }
-        if (down_b) {
-            cb(cur, "ffn_down", il);
-        }
         if (down_b) {
             cur = ggml_add(ctx, cur, down_b);
+            cb(cur, "ffn_down_b", il);
         }
         if (down_s) {
             cur = ggml_mul(ctx, cur, down_s);