qwen35 : in-place recurrent state rows for plain decode (GGML_GDN_ROWS_PLAIN=1)

src/llama-context.cpp

@@ -182,6 +182,12 @@ llama_context::llama_context(
         throw std::runtime_error("n_seq_max must be <= " + std::to_string(LLAMA_MAX_SEQ));
     }
 
+    {
+        const char * rows_plain = getenv("GGML_GDN_ROWS_PLAIN");
+        const char * rows_plain_max = getenv("GGML_GDN_ROWS_PLAIN_MAX_TOKENS");
+        cparams.gdn_rows_plain            = rows_plain && atoi(rows_plain) == 1;
+        cparams.gdn_rows_plain_max_tokens = rows_plain_max ? std::max(0, atoi(rows_plain_max)) : 0;
+    }
     cparams.n_rs_seq = params.n_rs_seq;
     if (cparams.n_rs_seq > 0 && !llm_arch_supports_rs_rollback(model.arch)) {
         LLAMA_LOG_DEBUG("%s: n_rs_seq=%u requested but model does not support recurrent partial rollback; clamping to 0\n",

src/llama-cparams.h

@@ -51,6 +51,8 @@ struct llama_cparams {
     bool offload_kqv;
     bool flash_attn;
     bool auto_fa;
+    bool gdn_rows_plain            = false; // GGML_GDN_ROWS_PLAIN=1: in-place recurrent state rows for plain decode
+    int  gdn_rows_plain_max_tokens = 0;     // GGML_GDN_ROWS_PLAIN_MAX_TOKENS: only up to this many tokens per sequence (0: any)
     bool fused_gdn_ar;       // use fused gated delta net (autoregressive)
     bool fused_gdn_ch;       // use fused gated delta net (chunked)
     bool auto_fgdn;

src/models/delta-net-base.cpp

@@ -551,9 +551,8 @@ ggml_tensor * llm_build_delta_net_base::build_recurrent_attn(
     const int64_t n_seqs       = v->ne[3];
     const int64_t n_seq_tokens = q->ne[2];
 
-    const bool keep = cparams.n_rs_seq > 0;
-
-    GGML_ASSERT(state_rows == nullptr || keep); // rows mode is a ring-path optimization
+    // rows mode also serves plain decode (GGML_GDN_ROWS_PLAIN): one snapshot slot, K = 1
+    const bool keep = cparams.n_rs_seq > 0 || state_rows != nullptr;
 
     if (!keep) {
         auto attn_out = build_delta_net(q, k, v, g, b, s, il);

src/models/qwen35.cpp

@@ -479,7 +479,12 @@ ggml_tensor * llama_model_qwen35::graph::build_layer_attn_linear(
     // GPU device in the model is Metal.
     static const bool gdn_state_rows_env = getenv("GGML_GDN_STATE_GATHER") == nullptr;
 
-    const bool gdn_state_rows = gdn_state_rows_env && gdn_state_rows_dev_ok && cparams.n_rs_seq > 0;
+    // GGML_GDN_ROWS_PLAIN=1: plain decode also updates the state rows in place (bitwise identical), but only when no extra cells are relocated, as that relocation runs before the in-place read
+    // GGML_GDN_ROWS_PLAIN_MAX_TOKENS caps it by tokens per sequence (A19: the in-place recurrence is ~18% slower at 512-token prefill)
+    const bool gdn_rows_plain_ok = cparams.gdn_rows_plain && mctx_cur->get_n_rs() == (uint32_t) n_seqs &&
+        (cparams.gdn_rows_plain_max_tokens <= 0 || n_seq_tokens <= cparams.gdn_rows_plain_max_tokens);
+
+    const bool gdn_state_rows = gdn_state_rows_env && gdn_state_rows_dev_ok && (cparams.n_rs_seq > 0 || gdn_rows_plain_ok);
 
     ggml_tensor * state;
     if (gdn_state_rows) {