graph : constant topology for tokens/embeddings inputs

src/llama-graph.cpp

@@ -21,7 +21,8 @@ void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
     }
 
     if (ubatch->embd) {
-        const int64_t n_embd   = embd->ne[0];
+        GGML_ASSERT(n_embd == embd->ne[0]);
+
         const int64_t n_tokens = ubatch->n_tokens;
 
         ggml_backend_tensor_set(embd, ubatch->embd, 0, n_tokens*n_embd*ggml_element_size(embd));
@@ -1206,17 +1207,21 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
 ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
     const int64_t n_embd = hparams.n_embd_inp();
 
-    auto inp = std::make_unique<llm_graph_input_embd>();
+    auto inp = std::make_unique<llm_graph_input_embd>(n_embd);
+
+    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
+    cb(inp->tokens, "inp_tokens", -1);
+    ggml_set_input(inp->tokens);
+    res->t_tokens = inp->tokens;
+
+    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, ubatch.n_tokens);
+    ggml_set_input(inp->embd);
 
     ggml_tensor * cur = nullptr;
 
-    if (ubatch.token) {
-        inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
-        //cb(inp->tokens, "inp_tokens", -1);
-        ggml_set_input(inp->tokens);
-        res->t_tokens = inp->tokens;
-
+    {
         cur = ggml_get_rows(ctx0, tok_embd, inp->tokens);
+        cur = ggml_scale(ctx0, cur, ubatch.token ? 1.0f : 0.0f);
 
         // apply lora for embedding tokens if needed
         for (const auto & lora : *loras) {
@@ -1235,13 +1240,10 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
 
             cur = ggml_add(ctx0, cur, inpL_delta);
         }
-    } else {
-        inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, ubatch.n_tokens);
-        ggml_set_input(inp->embd);
-
-        cur = inp->embd;
     }
 
+    cur = ggml_add(ctx0, cur, ggml_scale(ctx0, inp->embd, ubatch.embd ? 1.0f : 0.0f));
+
     // For Granite architecture
     if (hparams.f_embedding_scale != 0.0f) {
         cur = ggml_scale(ctx0, cur, hparams.f_embedding_scale);

src/llama-graph.h

@@ -104,7 +104,7 @@ using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;
 
 class llm_graph_input_embd : public llm_graph_input_i {
 public:
-    llm_graph_input_embd()          = default;
+    llm_graph_input_embd(int64_t n_embd) : n_embd(n_embd) {}
     virtual ~llm_graph_input_embd() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
@@ -113,6 +113,8 @@ public:
 
     ggml_tensor * tokens = nullptr; // I32 [n_batch]
     ggml_tensor * embd   = nullptr; // F32 [n_embd, n_batch]
+
+    const int64_t n_embd = 0;
 };
 
 class llm_graph_input_pos : public llm_graph_input_i {

src/models/cogvlm.cpp

@@ -3,12 +3,14 @@
 llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v;
-    float         kq_scale    = 1.0f / sqrtf(float(n_embd_head));
+    const float   kq_scale    = 1.0f / sqrtf(float(n_embd_head));
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
     GGML_ASSERT(n_embd_head == hparams.n_rot);
 
-    ggml_tensor *inpL, *cur;
+    ggml_tensor * inpL;
+    ggml_tensor * cur;
+
     inpL = build_inp_embd(model.tok_embd);
 
     ggml_tensor * inp_pos = build_inp_pos();
@@ -44,7 +46,7 @@ llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_pa
         }
 
         ggml_tensor * inpSA = inpL;
-        cur                 = build_norm(inpSA, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cur = build_norm(inpSA, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
 
         // build self attention
         {

src/models/gemma-embedding.cpp

@@ -1,7 +1,5 @@
 #include "models.h"
 
-
-
 llm_build_gemma_embedding::llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_k;
@@ -12,10 +10,8 @@ llm_build_gemma_embedding::llm_build_gemma_embedding(const llama_model & model,
     inpL = build_inp_embd(model.tok_embd);
 
     // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
-    if (ubatch.token) {
-        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-        cb(inpL, "inp_scaled", -1);
-    }
+    inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);
+    cb(inpL, "inp_scaled", -1);
 
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();

src/models/gemma3.cpp

@@ -10,10 +10,9 @@ llm_build_gemma3<iswa>::llm_build_gemma3(const llama_model & model, const llm_gr
     inpL = build_inp_embd(model.tok_embd);
 
     // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
-    if (ubatch.token) {
-        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-        cb(inpL, "inp_scaled", -1);
-    }
+    inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);
+    cb(inpL, "inp_scaled", -1);
+
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 

src/models/gemma3n-iswa.cpp

@@ -1,7 +1,5 @@
 #include "models.h"
 
-
-
 llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params),
     model(model),
@@ -15,10 +13,9 @@ llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const
     inpL = build_inp_embd(model.tok_embd);
 
     // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
-    if (ubatch.token) {
-        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-        cb(inpL, "inp_scaled", -1);
-    }
+    inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);
+    cb(inpL, "inp_scaled", -1);
+
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -248,7 +245,7 @@ ggml_tensor * llm_build_gemma3n_iswa::view_2d_slice(ggml_tensor * x, int idx) {
 // equivalent to get_per_layer_inputs() in python code
 // output shape: [n_embd_altup, n_layer, n_tokens]
 ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() {
-    auto          inp = std::make_unique<llm_graph_input_embd>();
+    auto inp = std::make_unique<llm_graph_input_embd>(n_embd);
     ggml_tensor * inp_per_layer;
     if (ubatch.token) {
         inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);