models : fix graph splits (#19866)

src/models/qwen35moe.cpp

@@ -29,6 +29,8 @@ llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_gr
         cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
         cb(cur, "attn_norm", il);
 
+        ggml_build_forward_expand(gf, cur);
+
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
@@ -269,7 +271,6 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
     cb(state_update_target, "state_update_target", il);
 
     ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
-    cb(conv_states_all, "conv_states_updated", il);
 
     ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
     state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);