@fugood/llama.node 1.1.4 → 1.1.6

package/src/llama.cpp/src/llama-graph.h

@@ -39,6 +39,7 @@ enum llm_ffn_op_type {
     LLM_FFN_SWIGLU,
     LLM_FFN_GEGLU,
     LLM_FFN_REGLU,
+    LLM_FFN_SWIGLU_OAI_MOE,
 };
 
 enum llm_ffn_gate_type {
@@ -144,7 +145,7 @@ public:
 
     ggml_tensor * pos_bucket = nullptr; // I32 [n_batch, n_batch]
 
-    const llama_hparams & hparams;
+    const llama_hparams hparams;
 };
 
 class llm_graph_input_pos_bucket_kv : public llm_graph_input_i {
@@ -158,7 +159,7 @@ public:
 
     ggml_tensor * pos_bucket = nullptr; // I32 [n_kv, n_batch]
 
-    const llama_hparams & hparams;
+    const llama_hparams hparams;
 
     const llama_kv_cache_unified_context * mctx;
 };
@@ -177,8 +178,8 @@ public:
 
     ggml_tensor * out_ids; // I32 [n_outputs]
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
+    const llama_hparams hparams;
+    const llama_cparams cparams;
 
     const uint32_t n_outputs;
 };
@@ -192,7 +193,7 @@ public:
 
     ggml_tensor * mean; // F32 [n_batch, n_batch]
 
-    const llama_cparams & cparams;
+    const llama_cparams cparams;
 };
 
 class llm_graph_input_cls : public llm_graph_input_i {
@@ -204,7 +205,7 @@ public:
 
     ggml_tensor * cls; // I32 [n_batch]
 
-    const llama_cparams & cparams;
+    const llama_cparams cparams;
 };
 
 class llm_graph_input_rs : public llm_graph_input_i {
@@ -214,7 +215,12 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
-    ggml_tensor * s_copy; // I32 [kv_size]
+    ggml_tensor * s_copy;  // I32 [n_rs]
+
+    // views of s_copy, computed once per graph
+    // and shared across layers which use build_rs
+    ggml_tensor * s_copy_main;   // I32 [n_seqs]
+    ggml_tensor * s_copy_extra;  // I32 [n_rs - n_seqs]
 
     const llama_memory_recurrent_context * mctx;
 };
@@ -247,8 +253,8 @@ public:
     ggml_tensor * kq_mask     = nullptr; // F32 [n_tokens, n_batch, 1, 1]
     ggml_tensor * kq_mask_cnv = nullptr; //     [n_tokens, n_batch, 1, 1]
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
+    const llama_hparams hparams;
+    const llama_cparams cparams;
 };
 
 class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
@@ -278,8 +284,11 @@ public:
     ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
     ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
+    // note: these have to be copies because in order to be able to reuse a graph, its inputs
+    //       need to carry these parameters with them. otherwise, they can point to freed
+    //       llm_graph_params from a previous batch, causing stack-use-after-return
+    const llama_hparams hparams;
+    const llama_cparams cparams;
 
     const llama_kv_cache_unified_context * mctx;
 };
@@ -318,8 +327,8 @@ public:
     ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
     ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
+    const llama_hparams hparams;
+    const llama_cparams cparams;
 
     const llama_kv_cache_unified_iswa_context * mctx;
 };
@@ -415,7 +424,9 @@ struct llm_graph_params {
                 (!ubatch.embd  && !other.ubatch.embd)
             );
 
-        if (can_reuse_ubatch && !ubatch.equal_seqs()) {
+        // when we split the batch using "equal_seqs" we have to verify that the participating sequences are the same
+        //   the reason is because the set of attention streams would be different for different sequences
+        if (can_reuse_ubatch && ubatch.equal_seqs()) {
             if (!ubatch.data) {
                 // if the old ubatch does not own it's data, then we cannot guarantee that it is still alive, and
                 //   therefore we cannot perform the sequence id check. normally should never happen
@@ -609,6 +620,7 @@ struct llm_graph_context {
        llm_ffn_gate_type   type_gate,
                      int   il) const;
 
+    // build MoE FFN without bias tensors
     ggml_tensor * build_moe_ffn(
              ggml_tensor * cur,
              ggml_tensor * gate_inp,
@@ -623,19 +635,29 @@ struct llm_graph_context {
                     bool   scale_w,
                    float   w_scale,
             llama_expert_gating_func_type gating_op,
-                     int   il) const;
+                     int   il,
+             ggml_tensor * probs_in = nullptr) const;
 
-    ggml_tensor * build_moe_ffn_from_probs(
+    ggml_tensor * build_moe_ffn(
              ggml_tensor * cur,
-             ggml_tensor * probs,
+             ggml_tensor * gate_inp,
+             ggml_tensor * gate_inp_b,
              ggml_tensor * up_exps,
+             ggml_tensor * up_exps_b,
              ggml_tensor * gate_exps,
+             ggml_tensor * gate_exps_b,
              ggml_tensor * down_exps,
+             ggml_tensor * down_exps_b,
              ggml_tensor * exp_probs_b,
                  int64_t   n_expert,
                  int64_t   n_expert_used,
+         llm_ffn_op_type   type_op,
+                    bool   norm_w,
+                    bool   scale_w,
+                   float   w_scale,
             llama_expert_gating_func_type gating_op,
-                     int   il) const;
+                     int   il,
+             ggml_tensor * probs_in = nullptr) const;
 
     //
     // inputs
@@ -663,6 +685,7 @@ struct llm_graph_context {
              ggml_tensor * v,       // [n_embd_head_v, n_head_v, n_tokens] (v_trans == false)
              ggml_tensor * kq_b,
              ggml_tensor * kq_mask,
+             ggml_tensor * sinks,
              ggml_tensor * v_mla,   // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                    float   kq_scale) const;
 
@@ -709,6 +732,20 @@ struct llm_graph_context {
                   float   kq_scale,
                     int   il) const;
 
+    // TODO: temporary to keep the diff small. after the code is public will refactor to simplify this
+    ggml_tensor * build_attn_with_sinks(
+            llm_graph_input_attn_kv_unified_iswa * inp,
+            ggml_tensor * wo,
+            ggml_tensor * wo_b,
+            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
+            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] optional
+            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional
+            ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+            ggml_tensor * sinks, // [n_head_q]
+                  float   kq_scale,
+                    int   il) const;
+
     llm_graph_input_attn_cross * build_attn_inp_cross() const;
 
     ggml_tensor * build_attn(
@@ -727,7 +764,6 @@ struct llm_graph_context {
     // recurrent
     //
 
-    // TODO: avoid notion of "kv"
     // TODO: move this implementation to llama_memory_recurrent.
     //       this is analogous to llama_kv_cache_unified::cpy_k / cpy_v
     //       when moving, avoid passing `ggml_cgraph` - only pass `ggml_context`. would likely need to split the
@@ -735,12 +771,13 @@ struct llm_graph_context {
     //         `llama_memory_recurrent`
     ggml_tensor * build_rs(
             ggml_tensor * s,
-            ggml_tensor * state_copy,
+            ggml_tensor * state_copy_main,
+            ggml_tensor * state_copy_extra,
                 int32_t   state_size,
                 int32_t   n_seqs,
-               uint32_t   n_kv,
-               uint32_t   kv_head,
-               uint32_t   kv_size,
+               uint32_t   n_rs,
+               uint32_t   rs_head,
+               uint32_t   rs_size,
                 int32_t   rs_zero,
             const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) const;
 

package/src/llama.cpp/src/llama-hparams.h

@@ -9,9 +9,10 @@
 #define LLAMA_MAX_EXPERTS 384  // Kimi-K2
 
 enum llama_expert_gating_func_type {
-    LLAMA_EXPERT_GATING_FUNC_TYPE_NONE    = 0,
-    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX = 1,
-    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID = 2,
+    LLAMA_EXPERT_GATING_FUNC_TYPE_NONE           = 0,
+    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX        = 1,
+    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID        = 2,
+    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT = 3, // applied to the router weights instead of the logits
 };
 
 enum llama_swa_type {
@@ -73,6 +74,7 @@ struct llama_hparams {
     bool     expert_weights_norm  = false;
     uint32_t expert_gating_func   = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
     uint32_t moe_every_n_layers   = 0;
+    uint32_t nextn_predict_layers = 0;
 
     float f_norm_eps;
     float f_norm_rms_eps;

package/src/llama.cpp/src/llama-kv-cache-unified.cpp

@@ -39,6 +39,10 @@ llama_kv_cache_unified::llama_kv_cache_unified(
     if (model.arch == LLM_ARCH_GEMMA3N) {
         n_layer_cache = 20;
     }
+    if (model.arch == LLM_ARCH_GLM4_MOE) {
+        // GLM-4.5: Only process up to last layer, skip final NextN layer
+        n_layer_cache = hparams.n_layer - hparams.nextn_predict_layers;
+    }
 
     // create a context for each buffer type
     std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
@@ -183,7 +187,7 @@ llama_kv_cache_unified::llama_kv_cache_unified(
         const size_t memory_size_k = size_k_bytes();
         const size_t memory_size_v = size_v_bytes();
 
-        LLAMA_LOG_INFO("%s: size = %7.2f MiB (%6u cells, %3d layers, %2u/%2u seqs), K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__,
+        LLAMA_LOG_INFO("%s: size = %7.2f MiB (%6u cells, %3d layers, %2u/%u seqs), K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__,
                 (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f), kv_size, (int) layers.size(), n_seq_max, n_stream,
                 ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f),
                 ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f));
@@ -193,7 +197,7 @@ llama_kv_cache_unified::llama_kv_cache_unified(
     debug = LLAMA_KV_CACHE_DEBUG ? atoi(LLAMA_KV_CACHE_DEBUG) : 0;
 
     const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
-    supports_set_rows = LLAMA_SET_ROWS ? atoi(LLAMA_SET_ROWS) != 0 : 0;
+    supports_set_rows = LLAMA_SET_ROWS ? atoi(LLAMA_SET_ROWS) != 0 : supports_set_rows;
 
     if (!supports_set_rows) {
         // ref: https://github.com/ggml-org/llama.cpp/pull/14363

package/src/llama.cpp/src/llama-kv-cache-unified.h

@@ -230,7 +230,7 @@ private:
 
     // env: LLAMA_SET_ROWS (temporary)
     // ref: https://github.com/ggml-org/llama.cpp/pull/14285
-    bool supports_set_rows = false;
+    bool supports_set_rows = true;
 
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 

package/src/llama.cpp/src/llama-memory-hybrid.cpp

@@ -25,6 +25,7 @@ llama_memory_hybrid::llama_memory_hybrid(
                          /* common */
              uint32_t    n_seq_max,
                  bool    offload,
+                 bool    unified,
                          /* layer filters */
       layer_filter_cb && filter_attn,
       layer_filter_cb && filter_recr) :
@@ -38,7 +39,7 @@ llama_memory_hybrid::llama_memory_hybrid(
         type_v,
         v_trans,
         offload,
-        1,
+        unified,
         kv_size,
         n_seq_max,
         n_pad,

package/src/llama.cpp/src/llama-memory-hybrid.h

@@ -39,6 +39,7 @@ public:
                              /* common */
                  uint32_t    n_seq_max,
                      bool    offload,
+                     bool    unified,
                              /* layer filters */
           layer_filter_cb && filter_attn = nullptr,
           layer_filter_cb && filter_recr = nullptr);

package/src/llama.cpp/src/llama-model-loader.cpp

@@ -35,6 +35,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q5_0:     return "Q5_0";
         case LLAMA_FTYPE_MOSTLY_Q5_1:     return "Q5_1";
         case LLAMA_FTYPE_MOSTLY_Q8_0:     return "Q8_0";
+        case LLAMA_FTYPE_MOSTLY_MXFP4_MOE: return "MXFP4 MoE";
         case LLAMA_FTYPE_MOSTLY_Q2_K:     return "Q2_K - Medium";
         case LLAMA_FTYPE_MOSTLY_Q2_K_S:   return "Q2_K - Small";
         case LLAMA_FTYPE_MOSTLY_Q3_K_S:   return "Q3_K - Small";

package/src/llama.cpp/src/llama-model-loader.h

@@ -58,8 +58,9 @@ struct llama_model_loader {
         }
     };
 
-    static const int TENSOR_NOT_REQUIRED = 1;
-    static const int TENSOR_DUPLICATED   = 2;
+    static const int TENSOR_NOT_REQUIRED = 1 << 0;
+    static const int TENSOR_DUPLICATED   = 1 << 1;
+    static const int TENSOR_SKIP         = 1 << 2;
 
     int n_kv      = 0;
     int n_tensors = 0;