qwen35: force pre-norm hidden state for MTP

Tap the MTP/nextn hidden state before the final output norm in both the
trunk and MTP graphs of Qwen35 and Qwen35MoE, reverting the semantic core
of the post-norm migration (ggml-org#24025) while keeping the upstream `nextn`
naming and the masked (embeddings_nextn_masked) extraction path.

This restores the fork's pre-norm Qwen MTP optimization (TheTom#149) on top of
the rebased TurboQuant+ KV cache lineage, per explicit request to prefer
pre-norm over upstream's post-norm/nextn approach.

https://claude.ai/code/session_01HZZQMPA6D6KWgcuDsDarWa

Author: claude

src/models/qwen35.cpp

@@ -204,15 +204,17 @@ llama_model_qwen35::graph::graph(const llama_model & model, const llm_graph_para
     }
     cur = inpL;
 
-    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
-
+    // Pre-norm hidden state (before the final output norm) seeds the MTP/nextn draft loop.
     cb(cur, "h_nextn", -1);
     res->t_h_nextn = cur;
 
     if (!cparams.embeddings_nextn_masked && inp_out_ids) {
         cur = ggml_get_rows(ctx0, cur, inp_out_ids);
     }
 
+    // Final norm
+    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
+
     cb(cur, "result_norm", -1);
     res->t_embd = cur;
 
@@ -619,16 +621,17 @@ llama_model_qwen35::graph_mtp::graph_mtp(const llama_model & model, const llm_gr
     cur = ggml_add(ctx0, cur, ffn_residual);
     cb(cur, "mtp_post_ffn", il);
 
+    // Pre-norm hidden state: used by the AR draft loop to seed the next MTP step.
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+
     ggml_tensor * head_norm_w = layer.nextn.shared_head_norm
             ? layer.nextn.shared_head_norm
             : model.output_norm;
     GGML_ASSERT(head_norm_w && "QWEN35 MTP: missing both nextn.shared_head_norm and output_norm");
     cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1);
-
-    cb(cur, "h_nextn", -1);
-    res->t_h_nextn = cur;
-
-    cur = ggml_get_rows(ctx0, cur, inp_out_ids);
     cb(cur, "mtp_shared_head_norm", -1);
 
     ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output;

src/models/qwen35moe.cpp

@@ -227,16 +227,17 @@ llama_model_qwen35moe::graph::graph(const llama_model & model, const llm_graph_p
     }
     cur = inpL;
 
-    // post-norm hidden state feeds both the LM head and the MTP seed below
-    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
-
+    // Pre-norm hidden state (before the final output norm) seeds the MTP/nextn draft loop.
     cb(cur, "h_nextn", -1);
     res->t_h_nextn = cur;
 
     if (!cparams.embeddings_nextn_masked && inp_out_ids) {
         cur = ggml_get_rows(ctx0, cur, inp_out_ids);
     }
 
+    // Final norm
+    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
+
     cb(cur, "result_norm", -1);
     res->t_embd = cur;
 
@@ -716,16 +717,17 @@ llama_model_qwen35moe::graph_mtp::graph_mtp(const llama_model & model, const llm
     cur = ggml_add(ctx0, cur, ffn_residual);
     cb(cur, "mtp_post_ffn", il);
 
+    // Pre-norm hidden state: used by the AR draft loop to seed the next MTP step.
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+
     ggml_tensor * head_norm_w = layer.nextn.shared_head_norm
             ? layer.nextn.shared_head_norm
             : model.output_norm;
     GGML_ASSERT(head_norm_w && "QWEN35MOE MTP: missing both nextn.shared_head_norm and output_norm");
     cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1);
-
-    cb(cur, "h_nextn", -1);
-    res->t_h_nextn= cur;
-
-    cur = ggml_get_rows(ctx0, cur, inp_out_ids);
     cb(cur, "mtp_shared_head_norm", -1);
 
     ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output;