qwen3next: add decode-only fused delta mode

docs/development/qwen3next_bench_16k_pp16384_tg128.md

@@ -113,3 +113,32 @@ Outcome:
 - No stable speed win in our setup after repeated runs.
 - Autoregressive rewrite specifically hurt TG throughput in non-fused mode and was reverted.
 - Final code keeps only the fused-default safety fix (non-fused by default).
+
+## Decode-Only Fused Mode Trial (`LLAMA_QWEN3NEXT_FUSED_DELTA=2`)
+
+Date: 2026-02-08
+
+Code change:
+
+- Added mode `2` for `LLAMA_QWEN3NEXT_FUSED_DELTA`:
+  - prompt / multi-token path: non-fused
+  - single-token decode path: fused
+
+Perplexity validation (`-c 2048`, GPU config as above):
+
+| Model | `=0` non-fused | `=2` decode-only fused |
+|---|---:|---:|
+| `/models/qwen3-next-coder.gguf` | `3.9378` | `3.9378` |
+| `/models/qwen-3-coder-next-mxfp4.gguf` | `3.9860` | `3.9860` |
+
+`llama-bench` at `-p 8192 -n 128 -b 2048 -ub 512 -r 3 -rtr 1`:
+
+| Mode | PP 8192 (tok/s) | TG 128 (tok/s) |
+|---|---:|---:|
+| `LLAMA_QWEN3NEXT_FUSED_DELTA=0` | `170.090` | `25.465` |
+| `LLAMA_QWEN3NEXT_FUSED_DELTA=2` | `166.212` | `29.599` |
+
+Notes:
+
+- Decode-only fused mode preserves prompt-quality metrics in this test.
+- TG improved significantly in this run; PP variance was higher, so PP delta should be treated as noisy.

src/llama-build-context.cpp

@@ -4178,12 +4178,21 @@ ggml_cgraph * llm_build_context::build_qwen3next() {
 
     const bool reset_state = batch.pos != nullptr && batch.pos[0] == 0;
 
-    // Keep legacy DeltaNet path as the default for correctness; enable fused path explicitly
-    // with LLAMA_QWEN3NEXT_FUSED_DELTA=1 for controlled testing.
-    const bool use_fused_delta_net = []() {
+    enum class qwen3next_fused_delta_mode {
+        off,
+        on,
+        tok1_only,
+    };
+
+    // Keep legacy DeltaNet path as default for correctness.
+    // LLAMA_QWEN3NEXT_FUSED_DELTA values:
+    //   unset / 0 : off
+    //   1         : fused for all token counts
+    //   2         : fused only for single-token decode steps
+    const qwen3next_fused_delta_mode fused_delta_mode = []() {
         const char * env = std::getenv("LLAMA_QWEN3NEXT_FUSED_DELTA");
         if (env == nullptr || env[0] == '\0') {
-            return false;
+            return qwen3next_fused_delta_mode::off;
         }
 
         switch (env[0]) {
@@ -4192,11 +4201,14 @@ ggml_cgraph * llm_build_context::build_qwen3next() {
             case 'Y':
             case 't':
             case 'T':
-                return true;
+                return qwen3next_fused_delta_mode::on;
+            case '2':
+                return qwen3next_fused_delta_mode::tok1_only;
             default:
-                return false;
+                return qwen3next_fused_delta_mode::off;
         }
     }();
+    const bool use_fused_delta_net_full = fused_delta_mode == qwen3next_fused_delta_mode::on;
 
     auto get_slice_2d = [&](ggml_tensor * t, int64_t c) -> ggml_tensor * {
         return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3],
@@ -4840,6 +4852,10 @@ ggml_cgraph * llm_build_context::build_qwen3next() {
         cb(v_conv, "v_conv_predelta", il);
 
         std::pair<ggml_tensor *, ggml_tensor *> attn_out;
+        const bool use_fused_delta_net =
+            use_fused_delta_net_full ||
+            (fused_delta_mode == qwen3next_fused_delta_mode::tok1_only && n_tok == 1);
+
         if (use_fused_delta_net) {
             attn_out = build_delta_net_fused(q_conv, k_conv, v_conv, gate, beta, state, il);
         } else {
@@ -4922,7 +4938,7 @@ ggml_cgraph * llm_build_context::build_qwen3next() {
     ggml_tensor * causal_mask = nullptr;
     ggml_tensor * identity    = nullptr;
     ggml_tensor * diag_mask   = nullptr;
-    if (!use_fused_delta_net) {
+    if (!use_fused_delta_net_full) {
         causal_mask = ggml_tri(ctx0,
                 ggml_fill_inplace(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, QWEN3NEXT_CHUNK_SIZE, QWEN3NEXT_CHUNK_SIZE), 1.0f),
                 GGML_TRI_TYPE_LOWER);