llama: dynamic head_dim and n_rot for SWA (#20301)

* llama: dynamic head_dim and n_rot for SWA

* also add gguf_writer wrappers

* fix build

* build_rope_shift arg reorder

diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index 839c6e787fc0db9db0eb5ff94b9894dd04bea961..c5f5469506741330507b6585627e709f3db3cafe 100644 (file)
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -177,6 +177,8 @@ class Keys:
         TEMPERATURE_LENGTH           = "{arch}.attention.temperature_length"
         KEY_LENGTH_MLA               = "{arch}.attention.key_length_mla"
         VALUE_LENGTH_MLA             = "{arch}.attention.value_length_mla"
+        KEY_LENGTH_SWA               = "{arch}.attention.key_length_swa"
+        VALUE_LENGTH_SWA             = "{arch}.attention.value_length_swa"
         SHARED_KV_LAYERS             = "{arch}.attention.shared_kv_layers"
         SLIDING_WINDOW_PATTERN       = "{arch}.attention.sliding_window_pattern"
         TEMPERATURE_SCALE            = "{arch}.attention.temperature_scale"
@@ -188,6 +190,7 @@ class Keys:
 
     class Rope:
         DIMENSION_COUNT           = "{arch}.rope.dimension_count"
+        DIMENSION_COUNT_SWA       = "{arch}.rope.dimension_count_swa"
         DIMENSION_SECTIONS        = "{arch}.rope.dimension_sections"
         FREQ_BASE                 = "{arch}.rope.freq_base"
         FREQ_BASE_SWA             = "{arch}.rope.freq_base_swa"

diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py
index 9ee3ac9e8ffe7df7df4e2a9fd81823fc928c92ce..e790be95332fb4e28a6e3b8455bdae8c7b7891be 100644 (file)
--- a/gguf-py/gguf/gguf_writer.py
+++ b/gguf-py/gguf/gguf_writer.py
@@ -773,6 +773,12 @@ class GGUFWriter:
     def add_value_length_mla(self, length: int) -> None:
         self.add_uint32(Keys.Attention.VALUE_LENGTH_MLA.format(arch=self.arch), length)
 
+    def add_key_length_swa(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.KEY_LENGTH_SWA.format(arch=self.arch), length)
+
+    def add_value_length_swa(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.VALUE_LENGTH_SWA.format(arch=self.arch), length)
+
     def add_indexer_head_count(self, count: int) -> None:
         self.add_uint32(Keys.Attention.Indexer.HEAD_COUNT.format(arch=self.arch), count)
 
@@ -946,6 +952,9 @@ class GGUFWriter:
     def add_rope_dimension_count(self, count: int) -> None:
         self.add_uint32(Keys.Rope.DIMENSION_COUNT.format(arch=self.arch), count)
 
+    def add_rope_dimension_count_swa(self, count: int) -> None:
+        self.add_uint32(Keys.Rope.DIMENSION_COUNT_SWA.format(arch=self.arch), count)
+
     def add_rope_dimension_sections(self, dims: Sequence[int]) -> None:
         self.add_array(Keys.Rope.DIMENSION_SECTIONS.format(arch=self.arch), dims)
 

diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index 9d8eb88d0bc58f38affae81882e1cfaccba686cd..ce49bbd988e761a6d9c095d4c45fc4cb47d90bd7 100644 (file)
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -230,11 +230,14 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_TEMPERATURE_SCALE,            "%s.attention.temperature_scale"            },
     { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
     { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
+    { LLM_KV_ATTENTION_KEY_LENGTH_SWA,               "%s.attention.key_length_swa"               },
+    { LLM_KV_ATTENTION_VALUE_LENGTH_SWA,             "%s.attention.value_length_swa"             },
     { LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,           "%s.attention.indexer.head_count"           },
     { LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,           "%s.attention.indexer.key_length"           },
     { LLM_KV_ATTENTION_INDEXER_TOP_K,                "%s.attention.indexer.top_k"                },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,           "%s.rope.dimension_count"                 },
+    { LLM_KV_ROPE_DIMENSION_COUNT_SWA,       "%s.rope.dimension_count_swa"             },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,        "%s.rope.dimension_sections"              },
     { LLM_KV_ROPE_FREQ_BASE,                 "%s.rope.freq_base"                       },
     { LLM_KV_ROPE_FREQ_BASE_SWA,             "%s.rope.freq_base_swa"                   },

diff --git a/src/llama-arch.h b/src/llama-arch.h
index 07aac40aa11d35f5a04c9a4872c2e3d31a726e73..28dd1ffac77d30164c009beaeb37b873a1e10a6c 100644 (file)
--- a/src/llama-arch.h
+++ b/src/llama-arch.h
@@ -234,11 +234,14 @@ enum llm_kv {
     LLM_KV_ATTENTION_TEMPERATURE_SCALE,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
+    LLM_KV_ATTENTION_KEY_LENGTH_SWA,
+    LLM_KV_ATTENTION_VALUE_LENGTH_SWA,
     LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,
     LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,
     LLM_KV_ATTENTION_INDEXER_TOP_K,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
+    LLM_KV_ROPE_DIMENSION_COUNT_SWA,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
     LLM_KV_ROPE_FREQ_BASE,
     LLM_KV_ROPE_FREQ_BASE_SWA,

diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 009d07e00e396059062272cac970b8ac65516eac..ee2669c154ec83462880673aeb734102a0cbbab2 100644 (file)
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -2876,19 +2876,23 @@ llama_context * llama_init_from_model(
 
     if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO && ggml_is_quantized(params.type_k)) {
         const uint32_t blck_size = ggml_blck_size(params.type_k);
-        if (model->hparams.n_embd_head_k % blck_size != 0) {
-            LLAMA_LOG_ERROR("%s: K cache type %s with block size %u does not divide n_embd_head_k=%u\n",
-                __func__, ggml_type_name(params.type_k), blck_size, model->hparams.n_embd_head_k);
-            return nullptr;
+        for (uint32_t il = 0; il < model->hparams.n_layer; ++il) {
+            if (model->hparams.n_embd_head_k(il) % blck_size != 0) {
+                LLAMA_LOG_ERROR("%s: K cache type %s with block size %u does not divide n_embd_head_k=%u\n",
+                    __func__, ggml_type_name(params.type_k), blck_size, model->hparams.n_embd_head_k(il));
+                return nullptr;
+            }
         }
     }
 
     if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO && ggml_is_quantized(params.type_v)) {
         const uint32_t blck_size = ggml_blck_size(params.type_v);
-        if (model->hparams.n_embd_head_v % blck_size != 0) {
-            LLAMA_LOG_ERROR("%s: V cache type %s with block size %u does not divide n_embd_head_k=%u\n",
-                __func__, ggml_type_name(params.type_v), blck_size, model->hparams.n_embd_head_v);
-            return nullptr;
+        for (uint32_t il = 0; il < model->hparams.n_layer; ++il) {
+            if (model->hparams.n_embd_head_v(il) % blck_size != 0) {
+                LLAMA_LOG_ERROR("%s: V cache type %s with block size %u does not divide n_embd_head_v=%u\n",
+                    __func__, ggml_type_name(params.type_v), blck_size, model->hparams.n_embd_head_v(il));
+                return nullptr;
+            }
         }
     }
 

diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index f4cb7dce15e0869002d7ea61dd34d6bd18814344..5f875136a1c5691a4497f61d54f008b6674a1ecc 100644 (file)
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -849,13 +849,13 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     ubatch           (params.ubatch),
     n_embd           (hparams.n_embd),
     n_layer          (hparams.n_layer),
-    n_rot            (hparams.n_rot),
+    n_rot            (hparams.n_rot()),
     n_ctx            (cparams.n_ctx),
     n_head           (hparams.n_head()),
     n_head_kv        (hparams.n_head_kv()),
-    n_embd_head_k    (hparams.n_embd_head_k),
+    n_embd_head_k    (hparams.n_embd_head_k()),
     n_embd_k_gqa     (hparams.n_embd_k_gqa()),
-    n_embd_head_v    (hparams.n_embd_head_v),
+    n_embd_head_v    (hparams.n_embd_head_v()),
     n_embd_v_gqa     (hparams.n_embd_v_gqa()),
     n_expert         (hparams.n_expert),
     n_expert_used    (cparams.warmup ? hparams.n_expert : hparams.n_expert_used),

diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp
index 756dda1a7ab658f25f4b1e0095845ba3c32aeef8..002d15d415fe0897eeb9ac582f652e135a0e7fc6 100644 (file)
--- a/src/llama-hparams.cpp
+++ b/src/llama-hparams.cpp
@@ -62,6 +62,14 @@ uint32_t llama_hparams::n_gqa(uint32_t il) const {
     return n_head/n_head_kv;
 }
 
+uint32_t llama_hparams::n_rot(uint32_t il) const {
+    if (il < n_layer) {
+        return is_swa(il) ? n_rot_swa : n_rot_full;
+    }
+
+    GGML_ABORT("fatal error");
+}
+
 uint32_t llama_hparams::n_embd_inp() const {
     uint32_t n_embd_inp = n_embd;
 
@@ -76,16 +84,32 @@ uint32_t llama_hparams::n_embd_out() const {
     return n_embd_out_impl > 0 ? n_embd_out_impl : n_embd;
 }
 
+uint32_t llama_hparams::n_embd_head_k(uint32_t il) const {
+    if (il < n_layer) {
+        return is_swa(il) ? n_embd_head_k_swa : n_embd_head_k_full;
+    }
+
+    GGML_ABORT("fatal error");
+}
+
+uint32_t llama_hparams::n_embd_head_v(uint32_t il) const {
+    if (il < n_layer) {
+        return is_swa(il) ? n_embd_head_v_swa : n_embd_head_v_full;
+    }
+
+    GGML_ABORT("fatal error");
+}
+
 uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
     const uint32_t n_head_kv = this->n_head_kv(il);
 
-    return n_embd_head_k * n_head_kv;
+    return n_embd_head_k(il) * n_head_kv;
 }
 
 uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
     const uint32_t n_head_kv = this->n_head_kv(il);
 
-    return n_embd_head_v * n_head_kv;
+    return n_embd_head_v(il) * n_head_kv;
 }
 
 bool llama_hparams::is_n_embd_k_gqa_variable() const {
@@ -197,11 +221,11 @@ bool llama_hparams::is_mla() const {
 }
 
 uint32_t llama_hparams::n_embd_head_k_mla() const {
-    return is_mla() ? n_embd_head_k_mla_impl : n_embd_head_k;
+    return is_mla() ? n_embd_head_k_mla_impl : n_embd_head_k();
 }
 
 uint32_t llama_hparams::n_embd_head_v_mla() const {
-    return is_mla() ? n_embd_head_v_mla_impl : n_embd_head_v;
+    return is_mla() ? n_embd_head_v_mla_impl : n_embd_head_v();
 }
 
 bool llama_hparams::has_kv(uint32_t il) const {

diff --git a/src/llama-hparams.h b/src/llama-hparams.h
index c4b2a99da5a98ac908237c647661bf50622b945f..abfd7f2c4b99e1888f6852595a49921a372d4a75 100644 (file)
--- a/src/llama-hparams.h
+++ b/src/llama-hparams.h
@@ -44,13 +44,20 @@ struct llama_hparams {
     uint32_t n_embd;
     uint32_t n_layer;
     int32_t n_layer_kv_from_start = -1; // if non-negative, the first n_layer_kv_from_start layers have KV cache
-    uint32_t n_rot;
-    uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
-    uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
     uint32_t n_expert = 0;
     uint32_t n_expert_used = 0;
     uint32_t n_rel_attn_bkts = 0;
 
+    // different head size for full_attention and SWA layers
+    uint32_t n_embd_head_k_full; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
+    uint32_t n_embd_head_v_full; // dimension of values (d_v) aka n_embd_head
+    uint32_t n_embd_head_k_swa;
+    uint32_t n_embd_head_v_swa;
+
+    // different RoPE dimensions for full_attention and SWA layers
+    uint32_t n_rot_full;
+    uint32_t n_rot_swa;
+
     // note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
     uint32_t n_embd_head_k_mla_impl = 0;
     uint32_t n_embd_head_v_mla_impl = 0;
@@ -247,12 +254,18 @@ struct llama_hparams {
 
     uint32_t n_gqa(uint32_t il = 0) const;
 
+    uint32_t n_rot(uint32_t il = 0) const;
+
     // dimension of main + auxiliary input embeddings
     uint32_t n_embd_inp() const;
 
     // dimension of output embeddings
     uint32_t n_embd_out() const;
 
+    // dimension of key/value embeddings for each head (per layer)
+    uint32_t n_embd_head_k(uint32_t il = 0) const;
+    uint32_t n_embd_head_v(uint32_t il = 0) const;
+
     // dimension of key embeddings across all k-v heads
     uint32_t n_embd_k_gqa(uint32_t il = 0) const;
 

diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
index d80e8a70bc277301539757d97d1206a233290760..82fe58fac46233505d51f72dd925324ae4a84afd 100644 (file)
--- a/src/llama-kv-cache.cpp
+++ b/src/llama-kv-cache.cpp
@@ -1033,8 +1033,8 @@ ggml_tensor * llama_kv_cache::get_k(ggml_context * ctx, int32_t il, uint32_t n_k
     const uint32_t ns = sinfo.s1 - sinfo.s0 + 1;
 
     return ggml_view_4d(ctx, k,
-            hparams.n_embd_head_k, hparams.n_head_kv(il), n_kv, ns,
-            ggml_row_size(k->type, hparams.n_embd_head_k),
+            hparams.n_embd_head_k(il), hparams.n_head_kv(il), n_kv, ns,
+            ggml_row_size(k->type, hparams.n_embd_head_k(il)),
             ggml_row_size(k->type, n_embd_k_gqa),
             ggml_row_size(k->type, n_embd_k_gqa*kv_size),
             ggml_row_size(k->type, n_embd_k_gqa*kv_size)*sinfo.s0);
@@ -1056,8 +1056,8 @@ ggml_tensor * llama_kv_cache::get_v(ggml_context * ctx, int32_t il, uint32_t n_k
     if (!v_trans) {
         // note: v->nb[1] <= v->nb[2]
         return ggml_view_4d(ctx, v,
-                hparams.n_embd_head_v, hparams.n_head_kv(il), n_kv, ns,
-                ggml_row_size(v->type, hparams.n_embd_head_v),          // v->nb[1]
+                hparams.n_embd_head_v(il), hparams.n_head_kv(il), n_kv, ns,
+                ggml_row_size(v->type, hparams.n_embd_head_v(il)),          // v->nb[1]
                 ggml_row_size(v->type, n_embd_v_gqa),                   // v->nb[2]
                 ggml_row_size(v->type, n_embd_v_gqa*kv_size),           // v->nb[3]
                 ggml_row_size(v->type, n_embd_v_gqa*kv_size)*sinfo.s0);
@@ -1065,8 +1065,8 @@ ggml_tensor * llama_kv_cache::get_v(ggml_context * ctx, int32_t il, uint32_t n_k
 
     // note: v->nb[1] > v->nb[2]
     return ggml_view_4d(ctx, v,
-            n_kv, hparams.n_head_kv(il), hparams.n_embd_head_v, ns,
-            ggml_row_size(v->type, kv_size*hparams.n_embd_head_v),  // v->nb[1]
+            n_kv, hparams.n_head_kv(il), hparams.n_embd_head_v(il), ns,
+            ggml_row_size(v->type, kv_size*hparams.n_embd_head_v(il)),  // v->nb[1]
             ggml_row_size(v->type, kv_size),                        // v->nb[2]
             ggml_row_size(v->type, kv_size*n_embd_v_gqa),           // v->nb[3]
             ggml_row_size(v->type, kv_size*n_embd_v_gqa)*sinfo.s0);
@@ -1544,7 +1544,8 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
                 ggml_tensor * shift,
                 ggml_tensor * factors,
                       float   freq_base,
-                      float   freq_scale) const {
+                      float   freq_scale,
+                   uint32_t   il) const {
     const auto & n_ctx_orig = cparams.n_ctx_orig_yarn;
 
     const auto & yarn_ext_factor  = cparams.yarn_ext_factor;
@@ -1552,7 +1553,7 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
     const auto & yarn_beta_slow   = cparams.yarn_beta_slow;
     const auto & yarn_attn_factor = cparams.yarn_attn_factor;
 
-    const auto & n_rot     = hparams.n_rot;
+    const auto & n_rot     = hparams.n_rot(il);
     const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE || hparams.rope_type == LLAMA_ROPE_TYPE_IMROPE
                                 // @ngxson : this is a workaround
                                 // for M-RoPE, we want to rotate the whole vector when doing KV shift
@@ -1606,13 +1607,6 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
     auto * ctx = res->get_ctx();
     auto * gf  = res->get_gf();
 
-    const auto & n_embd_head_k = hparams.n_embd_head_k;
-  //const auto & n_embd_head_v = hparams.n_embd_head_v;
-
-    const auto & n_rot = hparams.n_rot;
-
-    const auto n_embd_nope = hparams.n_lora_kv > 0 ? n_embd_head_k - n_rot : 0;
-
     auto inp = std::make_unique<llm_graph_input_k_shift>(this);
 
     inp->k_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, (int64_t) get_size()*n_stream);
@@ -1626,6 +1620,10 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
         const int64_t n_head_kv    = hparams.n_head_kv(il);
         const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
 
+        const auto n_rot         = hparams.n_rot(il);
+        const auto n_embd_head_k = hparams.n_embd_head_k(il);
+        const auto n_embd_nope   = hparams.n_lora_kv > 0 ? n_embd_head_k - n_rot : 0;
+
         const float freq_base_l  = model.get_rope_freq_base (cparams, il);
         const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
 
@@ -1638,7 +1636,7 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
                 ggml_row_size(layer.k->type, n_embd_k_gqa),
                 ggml_row_size(layer.k->type, n_embd_nope));
 
-        ggml_tensor * cur = build_rope_shift(cparams, ctx, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l);
+        ggml_tensor * cur = build_rope_shift(cparams, ctx, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l, il);
 
         ggml_build_forward_expand(gf, cur);
     }

diff --git a/src/llama-kv-cache.h b/src/llama-kv-cache.h
index e194bf3e26f3820eb95b78d7f47175b66f5fc031..33c78c5f2107ec7f48b213240a706e589dab8f8d 100644 (file)
--- a/src/llama-kv-cache.h
+++ b/src/llama-kv-cache.h
@@ -264,7 +264,8 @@ private:
                     ggml_tensor * shift,
                     ggml_tensor * factors,
                           float   freq_base,
-                          float   freq_scale) const;
+                          float   freq_scale,
+                       uint32_t   il) const;
 
     ggml_cgraph * build_graph_shift(
                llm_graph_result * res,

diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
index 2a6196eff3f8af8fff2017ffdfe96093738e996e..623a3455dd7d8a0b3b22dc145e16abe2546c54c2 100644 (file)
--- a/src/llama-model-loader.cpp
+++ b/src/llama-model-loader.cpp
@@ -918,7 +918,7 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
             } break;
         case GGML_OP_ROPE:
             {
-                const int n_embd_head = hparams.n_embd_head_v;
+                const int n_embd_head = hparams.n_embd_head_v();
                 const int n_head = hparams.n_head();
                 ggml_tensor * a = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_embd_head, n_head, 512);
                 ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 512);

diff --git a/src/llama-model-saver.cpp b/src/llama-model-saver.cpp
index 9f677b40cfc0d0b5664d105ce0afd3a23fbe45cf..6f6538aeccd2c0e649f5584fa5190223548499d5 100644 (file)
--- a/src/llama-model-saver.cpp
+++ b/src/llama-model-saver.cpp
@@ -186,8 +186,10 @@ void llama_model_saver::add_kv_from_model() {
     add_kv(LLM_KV_ATTENTION_HEAD_COUNT_KV,           hparams.n_head_kv_arr, true);
     add_kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS,          hparams.f_max_alibi_bias);
     add_kv(LLM_KV_ATTENTION_CLAMP_KQV,               hparams.f_clamp_kqv);
-    add_kv(LLM_KV_ATTENTION_KEY_LENGTH,              hparams.n_embd_head_k);
-    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH,            hparams.n_embd_head_v);
+    add_kv(LLM_KV_ATTENTION_KEY_LENGTH,              hparams.n_embd_head_k_full);
+    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH,            hparams.n_embd_head_v_full);
+    add_kv(LLM_KV_ATTENTION_KEY_LENGTH_SWA,          hparams.n_embd_head_k_swa);
+    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH_SWA,        hparams.n_embd_head_v_swa);
     add_kv(LLM_KV_ATTENTION_LAYERNORM_EPS,           hparams.f_norm_eps);
     add_kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
     add_kv(LLM_KV_ATTENTION_CAUSAL,                  hparams.causal_attn);
@@ -199,7 +201,8 @@ void llama_model_saver::add_kv_from_model() {
 
     const float rope_scaling_factor = hparams.rope_freq_scale_train == 1.0f ? 0.0f : 1.0f/hparams.rope_freq_scale_train;
 
-    add_kv(LLM_KV_ROPE_DIMENSION_COUNT,              hparams.n_rot);
+    add_kv(LLM_KV_ROPE_DIMENSION_COUNT,              hparams.n_rot_full);
+    add_kv(LLM_KV_ROPE_DIMENSION_COUNT_SWA,          hparams.n_rot_swa);
     add_kv(LLM_KV_ROPE_FREQ_BASE,                    hparams.rope_freq_base_train);
     // add_kv(LLM_KV_ROPE_SCALE_LINEAR,                 rope_scaling_factor); // old name
     add_kv(LLM_KV_ROPE_SCALING_TYPE,                 llama_rope_scaling_type_name(hparams.rope_scaling_type_train));

diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index e18cca0524e1ae713b064582d94f324dbebe6f18..349f70a611ac8a2ac1745b55648bb2d45aae971c 100644 (file)
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -459,26 +459,37 @@ void llama_model::load_hparams(llama_model_loader & ml) {
         // gpt-neox n_rot = rotary_pct * (n_embd / n_head)
         // gpt-j n_rot = rotary_dim
 
-        hparams.n_embd_head_k = hparams.n_embd / hparams.n_head();
-        ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k, false);
+        hparams.n_embd_head_k_full = hparams.n_embd / hparams.n_head();
+        ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k_full, false);
 
-        hparams.n_embd_head_v = hparams.n_embd / hparams.n_head();
-        ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false);
+        hparams.n_embd_head_v_full = hparams.n_embd / hparams.n_head();
+        ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v_full, false);
 
         // sanity check for n_rot (optional)
-        hparams.n_rot = hparams.n_embd_head_k;
+        hparams.n_rot_full = hparams.n_embd_head_k_full;
 
-        ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
+        ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot_full, false);
 
         if (arch == LLM_ARCH_LLAMA || arch == LLM_ARCH_DECI || arch == LLM_ARCH_FALCON || arch == LLM_ARCH_LLAMA_EMBED) {
-            if (hparams.n_rot != hparams.n_embd_head_k) {
-                throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k));
+            if (hparams.n_rot_full != hparams.n_embd_head_k_full) {
+                throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot_full, hparams.n_embd_head_k_full));
             }
         }
     } else {
-        hparams.n_rot = 0;
-        hparams.n_embd_head_k = 0;
-        hparams.n_embd_head_v = 0;
+        hparams.n_rot_full = 0;
+        hparams.n_embd_head_k_full = 0;
+        hparams.n_embd_head_v_full = 0;
+    }
+
+    // head size and n_rot for SWA layers
+    {
+        hparams.n_embd_head_k_swa = hparams.n_embd_head_k_full;
+        hparams.n_embd_head_v_swa = hparams.n_embd_head_v_full;
+        ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_SWA, hparams.n_embd_head_k_swa, false);
+        ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_SWA, hparams.n_embd_head_v_swa, false);
+
+        hparams.n_rot_swa = hparams.n_rot_full;
+        ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT_SWA, hparams.n_rot_swa, false);
     }
 
     // for differentiating model types
@@ -1114,10 +1125,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                         break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
-
-                // Load attention parameters
-                ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH,   hparams.n_embd_head_k, false);
-                ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false);
             } break;
         case LLM_ARCH_PLAMO3:
             {
@@ -1212,7 +1219,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/config.py#L173
                 hparams.f_attention_scale = type == LLM_TYPE_27B
                     ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0)))
-                    : 1.0f / std::sqrt(float(hparams.n_embd_head_k));
+                    : 1.0f / std::sqrt(float(hparams.n_embd_head_k()));
             } break;
         case LLM_ARCH_GEMMA3:
             {
@@ -1245,7 +1252,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/config.py#L289
                 hparams.f_attention_scale = type == LLM_TYPE_27B
                     ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0)))
-                    : 1.0f / std::sqrt(float(hparams.n_embd_head_k));
+                    : 1.0f / std::sqrt(float(hparams.n_embd_head_k()));
             } break;
         case LLM_ARCH_GEMMA3N:
             {
@@ -1294,7 +1301,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     case 24: type = LLM_TYPE_0_3B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
-                hparams.f_attention_scale = 1.0f / std::sqrt(float(hparams.n_embd_head_k));
+                hparams.f_attention_scale = 1.0f / std::sqrt(float(hparams.n_embd_head_k()));
 
             } break;
         case LLM_ARCH_STARCODER2:
@@ -2487,7 +2494,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,    hparams.n_embd_head_k_mla_impl);
                 ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA,  hparams.n_embd_head_v_mla_impl);
                 ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,      hparams.n_lora_kv);
-                ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT,        hparams.n_rot);
                 ml.get_key(LLM_KV_SSM_CONV_KERNEL,             hparams.ssm_d_conv);
                 ml.get_key(LLM_KV_KDA_HEAD_DIM,                hparams.n_embd_head_kda);
 
@@ -2518,6 +2524,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
                 hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
 
+                // full_attention layer only use half of the RoPE dimensions
+                hparams.n_rot_full = hparams.n_rot_full / 2;
+
                 // MoE + SWA parameters
                 ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
                 ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
@@ -2661,13 +2670,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         const int64_t n_embd        = hparams.n_embd;
         const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa();
         const int64_t n_embd_v_gqa  = hparams.n_embd_v_gqa();
-        const int64_t n_embd_head_k = hparams.n_embd_head_k;
-        const int64_t n_embd_head_v = hparams.n_embd_head_v;
+        const int64_t n_embd_head_k = hparams.n_embd_head_k();
+        const int64_t n_embd_head_v = hparams.n_embd_head_v();
         const int64_t n_ff          = hparams.n_ff();
         const int64_t n_embd_gqa    = n_embd_v_gqa;
         const int64_t n_vocab       = vocab.n_tokens();
         const int64_t n_token_types = vocab.n_token_types();
-        const int64_t n_rot         = hparams.n_rot;
+        const int64_t n_rot         = hparams.n_rot();
         const int64_t n_expert      = hparams.n_expert;
         const int64_t n_expert_used = hparams.n_expert_used;
         const int64_t n_ctx_train   = hparams.n_ctx_train;
@@ -2967,8 +2976,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_MINICPM3:
                 {
-                    const int64_t n_embd_head_qk_rope = hparams.n_rot;
-                    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+                    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+                    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k() - hparams.n_rot();
 
                     const int64_t q_lora_rank  = hparams.n_lora_q;
                     const int64_t kv_lora_rank = hparams.n_lora_kv;
@@ -3840,8 +3849,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     const int64_t dt_dim              = std::max(64, int(hparams.n_embd / 16));
 
                     // attention parameters
-                    const uint32_t qk_dim = hparams.n_embd_head_k;
-                    const uint32_t v_dim  = hparams.n_embd_head_v;
+                    const uint32_t qk_dim = hparams.n_embd_head_k();
+                    const uint32_t v_dim  = hparams.n_embd_head_v();
 
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -3901,8 +3910,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_PLAMO3:
                 {
-                    const int64_t head_dim_q = hparams.n_embd_head_k;
-                    const int64_t head_dim_v = hparams.n_embd_head_v;
+                    const int64_t head_dim_q = hparams.n_embd_head_k();
+                    const int64_t head_dim_v = hparams.n_embd_head_v();
 
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -4649,7 +4658,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_SEED_OSS:
                 {
-                    const uint32_t head_dim             = hparams.n_embd_head_k;
+                    const uint32_t head_dim             = hparams.n_embd_head_k();
                     const int64_t n_qo_dim              = n_head * head_dim;
                     const int64_t n_kv_dim              = n_head_kv * head_dim;
 
@@ -4878,7 +4887,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();
                     const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();
 
-                    const int64_t n_embd_head_qk_rope = hparams.n_rot;
+                    const int64_t n_embd_head_qk_rope = hparams.n_rot();
                     const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
                     GGML_ASSERT(n_embd_head_qk_nope >= 1);
 
@@ -4957,8 +4966,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_PLM:
                 {
-                    const int64_t n_embd_head_qk_rope = hparams.n_rot;
-                    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+                    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+                    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k() - hparams.n_rot();
                     const int64_t kv_lora_rank = hparams.n_lora_kv;
 
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -5396,7 +5405,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();
                     const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();
 
-                    const int64_t n_embd_head_qk_rope = hparams.n_rot;
+                    const int64_t n_embd_head_qk_rope = hparams.n_rot();
                     const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
 
                     const int64_t q_lora_rank  = hparams.n_lora_q;
@@ -5680,7 +5689,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     const int64_t n_expert       = hparams.n_expert;
                     const int64_t n_expert_used  = hparams.n_expert_used;
                     const int64_t n_ff_shexp     = hparams.n_ff_shexp > 0 ? hparams.n_ff_shexp : n_ff_exp;
-                    const int64_t head_dim       = hparams.n_embd_head_k;
+                    const int64_t head_dim       = hparams.n_embd_head_k();
                     const int64_t n_qo_dim       = n_head * head_dim;
                     const int64_t n_kv_dim       = n_head_kv * head_dim;
 
@@ -6968,7 +6977,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                              // Kimi: qk_rope_head_dim = 64 (actual RoPE dimension for MLA)
                              // Note: hparams.n_rot may be 72 (from conversion) but actual is 64
-                             const int64_t qk_rope_head_dim = hparams.n_rot;  // From config: qk_rope_head_dim
+                             const int64_t qk_rope_head_dim = hparams.n_rot();  // From config: qk_rope_head_dim
                              layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + qk_rope_head_dim}, 0);
                              // Support Legacy GGUFs that don't split wkv_b (MLA KV cache disabled)
                              layer.wkv_b = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i),
@@ -7339,7 +7348,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     // ("rope_freqs.weight") and ggml uses only the first (n_rot_l/2) entries per layer.
                     uint32_t n_rot_max = 0;
                     for (int i = 0; i < n_layer; ++i) {
-                        n_rot_max = std::max(n_rot_max, hparams.n_rot);
+                        n_rot_max = std::max(n_rot_max, hparams.n_rot());
                     }
                     if (n_rot_max == 0) {
                         n_rot_max = n_rot;
@@ -7674,11 +7683,11 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_layer               = %u\n",     __func__, hparams.n_layer);
         LLAMA_LOG_INFO("%s: n_head                = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_head(il);    }, hparams.n_layer).c_str());
         LLAMA_LOG_INFO("%s: n_head_kv             = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str());
-        LLAMA_LOG_INFO("%s: n_rot                 = %u\n",     __func__, hparams.n_rot);
+        LLAMA_LOG_INFO("%s: n_rot                 = %u\n",     __func__, hparams.n_rot_full);
         LLAMA_LOG_INFO("%s: n_swa                 = %u\n",     __func__, hparams.n_swa);
         LLAMA_LOG_INFO("%s: is_swa_any            = %u\n",     __func__, hparams.is_swa_any());
-        LLAMA_LOG_INFO("%s: n_embd_head_k         = %u\n",     __func__, hparams.n_embd_head_k);
-        LLAMA_LOG_INFO("%s: n_embd_head_v         = %u\n",     __func__, hparams.n_embd_head_v);
+        LLAMA_LOG_INFO("%s: n_embd_head_k         = %u\n",     __func__, hparams.n_embd_head_k_full);
+        LLAMA_LOG_INFO("%s: n_embd_head_v         = %u\n",     __func__, hparams.n_embd_head_v_full);
         LLAMA_LOG_INFO("%s: n_gqa                 = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_gqa(il);        }, hparams.n_layer).c_str());
         LLAMA_LOG_INFO("%s: n_embd_k_gqa          = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_embd_k_gqa(il); }, hparams.n_layer).c_str());
         LLAMA_LOG_INFO("%s: n_embd_v_gqa          = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_embd_v_gqa(il); }, hparams.n_layer).c_str());
@@ -7702,6 +7711,9 @@ void llama_model::print_info() const {
         if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
             LLAMA_LOG_INFO("%s: freq_base_swa         = %.1f\n",   __func__, hparams.rope_freq_base_train_swa);
             LLAMA_LOG_INFO("%s: freq_scale_swa        = %g\n",     __func__, hparams.rope_freq_scale_train_swa);
+            LLAMA_LOG_INFO("%s: n_embd_head_k_swa     = %u\n",     __func__, hparams.n_embd_head_k_swa);
+            LLAMA_LOG_INFO("%s: n_embd_head_v_swa     = %u\n",     __func__, hparams.n_embd_head_v_swa);
+            LLAMA_LOG_INFO("%s: n_rot_swa             = %u\n",     __func__, hparams.n_rot_swa);
         }
         LLAMA_LOG_INFO("%s: n_ctx_orig_yarn       = %u\n",     __func__, hparams.n_ctx_orig_yarn);
         LLAMA_LOG_INFO("%s: rope_yarn_log_mul     = %.4f\n",   __func__, hparams.rope_yarn_log_mul);

diff --git a/src/models/afmoe.cpp b/src/models/afmoe.cpp
index 11f5ea2c27178ac0e49c41a9f5c2a6a4590a1868..9aabe25c96553f98087d4ff8650551a1c0ae1947 100644 (file)
--- a/src/models/afmoe.cpp
+++ b/src/models/afmoe.cpp
@@ -1,8 +1,8 @@
 #include "models.h"
 
 llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/apertus.cpp b/src/models/apertus.cpp
index 9af19c1bfe800994f5ba62a551af1582777dd04f..4d65614e466170bc23749ddf10beb76126dc4f8e 100644 (file)
--- a/src/models/apertus.cpp
+++ b/src/models/apertus.cpp
@@ -3,10 +3,10 @@
 
 
 llm_build_apertus::llm_build_apertus(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/arcee.cpp b/src/models/arcee.cpp
index aa6167dba1e7e617b394c23c449de86c1f43098c..20b9ffd49ebd18fac688ba1bb6158695ff5d6752 100644 (file)
--- a/src/models/arcee.cpp
+++ b/src/models/arcee.cpp
@@ -2,10 +2,10 @@
 
 
 llm_build_arcee::llm_build_arcee(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/arctic.cpp b/src/models/arctic.cpp
index d7db06de1d0a25eab2966849f9d56d96ad8dddf2..b712e08cbd32c233fb23658d4ce8879c45cf2d59 100644 (file)
--- a/src/models/arctic.cpp
+++ b/src/models/arctic.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_arctic::llm_build_arctic(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/baichuan.cpp b/src/models/baichuan.cpp
index d5c6528531e72ac0bc34f22fe0cc97feee8f508e..abd03cd0b9711e00ce7b99140e149d3688ab1b48 100644 (file)
--- a/src/models/baichuan.cpp
+++ b/src/models/baichuan.cpp
@@ -2,10 +2,10 @@
 
 
 llm_build_baichuan::llm_build_baichuan(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/bailingmoe2.cpp b/src/models/bailingmoe2.cpp
index 81906cecb53df8372a8acdf9a250d119a3e43007..42098624663b6e85f48cbc922d77fae4b9f0df63 100644 (file)
--- a/src/models/bailingmoe2.cpp
+++ b/src/models/bailingmoe2.cpp
@@ -2,10 +2,10 @@
 
 llm_build_bailingmoe2::llm_build_bailingmoe2(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/bert.cpp b/src/models/bert.cpp
index 17efdafec3f5a269b6bfede2c3bc2ca05e54f950..8733179141885ce7f1a8f9032d45a13a93cd743e 100644 (file)
--- a/src/models/bert.cpp
+++ b/src/models/bert.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/bitnet.cpp b/src/models/bitnet.cpp
index 331a3f111979575ffacd144a90a585465a5e3d07..d47638498dfabfb6a9fd13358e1681e104ebb4b1 100644 (file)
--- a/src/models/bitnet.cpp
+++ b/src/models/bitnet.cpp
@@ -2,9 +2,9 @@
 
 
 llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/bloom.cpp b/src/models/bloom.cpp
index 2c552d1d15ea6d8a29bebd08dcdda96cf15b9ee8..b1c19bb58a2e07dbe5db557b6d54110bbf6b3378 100644 (file)
--- a/src/models/bloom.cpp
+++ b/src/models/bloom.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_bloom::llm_build_bloom(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/chameleon.cpp b/src/models/chameleon.cpp
index 184511aed4c304002fa2b2636a7650058aa8aa98..2f24105fa14ea6dd73c33d11c88ce0d8fae2a4e8 100644 (file)
--- a/src/models/chameleon.cpp
+++ b/src/models/chameleon.cpp
@@ -3,10 +3,10 @@
 #include <float.h>
 
 llm_build_chameleon::llm_build_chameleon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/chatglm.cpp b/src/models/chatglm.cpp
index 2685d4fbcbee81378c380d1f225c2fdc085bd8f8..5887ed22e7eeddba1d0c86ee90a09837f589f2af 100644 (file)
--- a/src/models/chatglm.cpp
+++ b/src/models/chatglm.cpp
@@ -2,10 +2,10 @@
 
 
 llm_build_chatglm::llm_build_chatglm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/codeshell.cpp b/src/models/codeshell.cpp
index 0b3bdbff529ea613c3511ee110593c24c12240fd..e8e13e143f262d6dada8baaa757b630f274d71ea 100644 (file)
--- a/src/models/codeshell.cpp
+++ b/src/models/codeshell.cpp
@@ -1,11 +1,11 @@
 #include "models.h"
 
 llm_build_codeshell::llm_build_codeshell(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/cogvlm.cpp b/src/models/cogvlm.cpp
index 0ceae3aaeb550cc812fca7735166b6bbcb1ad7be..2ef2b6e389b4d243248295f704ff0228d2a47add 100644 (file)
--- a/src/models/cogvlm.cpp
+++ b/src/models/cogvlm.cpp
@@ -2,11 +2,11 @@
 
 llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const float   kq_scale    = 1.0f / sqrtf(float(n_embd_head));
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * inpL;
     ggml_tensor * cur;

diff --git a/src/models/cohere2-iswa.cpp b/src/models/cohere2-iswa.cpp
index 9334b5e42634f4905781bd2650d91d92a529dfae..7c71a59ae7f0f292c4901b869a6e4e2b40226690 100644 (file)
--- a/src/models/cohere2-iswa.cpp
+++ b/src/models/cohere2-iswa.cpp
@@ -1,9 +1,9 @@
 #include "models.h"
 
 llm_build_cohere2_iswa::llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     const float f_logit_scale = hparams.f_logit_scale;
 

diff --git a/src/models/command-r.cpp b/src/models/command-r.cpp
index 4d3b643b444ae54e788b2932b6ebaecf60b3c1d2..ba1230f0419358cd684231d560bcf98a5cf0256a 100644 (file)
--- a/src/models/command-r.cpp
+++ b/src/models/command-r.cpp
@@ -4,9 +4,9 @@
 
 llm_build_command_r::llm_build_command_r(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     const float f_logit_scale = hparams.f_logit_scale;
 

diff --git a/src/models/dbrx.cpp b/src/models/dbrx.cpp
index 5c7f108437e36b06011fa8c88ef91a6d803d71be..73eb5cd24e7de114e6b05ef8e5912b354558b140 100644 (file)
--- a/src/models/dbrx.cpp
+++ b/src/models/dbrx.cpp
@@ -1,11 +1,11 @@
 #include "models.h"
 
 llm_build_dbrx::llm_build_dbrx(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/deci.cpp b/src/models/deci.cpp
index 7410a3a46d93b9a9d83ff0a13cbbfcc0da493309..ac448bfcaa8f26149943faab9fb0d1b757385f93 100644 (file)
--- a/src/models/deci.cpp
+++ b/src/models/deci.cpp
@@ -3,10 +3,10 @@
 
 
 llm_build_deci::llm_build_deci(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/deepseek.cpp b/src/models/deepseek.cpp
index 57cb1724f259f7a9daf881f5cba390179d20352d..3432359e03a0a9e73dcac3a6bad55d693e58b04a 100644 (file)
--- a/src/models/deepseek.cpp
+++ b/src/models/deepseek.cpp
@@ -2,10 +2,10 @@
 
 llm_build_deepseek::llm_build_deepseek(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/deepseek2.cpp b/src/models/deepseek2.cpp
index abd54b7656f6a975cdfca45a9169667c9aeee6c3..d437fe29e71b3a7d4da716885d6ab04805acddd2 100644 (file)
--- a/src/models/deepseek2.cpp
+++ b/src/models/deepseek2.cpp
@@ -8,7 +8,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
     const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
     const int64_t n_embd_head_v = hparams.n_embd_head_v_mla();
 
-    const int64_t n_embd_head_qk_rope = hparams.n_rot;
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();
     const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
 
     const uint32_t kv_lora_rank = hparams.n_lora_kv;

diff --git a/src/models/dots1.cpp b/src/models/dots1.cpp
index 0bcf3fe0de5bf8c20b47dd14070bc97cba7f0ab3..07236dd27c9d3cdad9ee427fff432d8ef18f4a4f 100644 (file)
--- a/src/models/dots1.cpp
+++ b/src/models/dots1.cpp
@@ -2,10 +2,10 @@
 
 llm_build_dots1::llm_build_dots1(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/dream.cpp b/src/models/dream.cpp
index 2aafbae1397fb8f320a66287fac56cb9686af113..4edc8530cb3cfede88dd4b07ec0a7ea4521c376e 100644 (file)
--- a/src/models/dream.cpp
+++ b/src/models/dream.cpp
@@ -5,10 +5,10 @@
 llm_build_dream::llm_build_dream(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     //copied from qwen2
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/ernie4-5-moe.cpp b/src/models/ernie4-5-moe.cpp
index ed781d5076cff5b3bc10f540d7868ff6e7e5a23a..63baf152c40771a75ea74533d3da386719fbfa92 100644 (file)
--- a/src/models/ernie4-5-moe.cpp
+++ b/src/models/ernie4-5-moe.cpp
@@ -2,10 +2,10 @@
 
 llm_build_ernie4_5_moe::llm_build_ernie4_5_moe(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/ernie4-5.cpp b/src/models/ernie4-5.cpp
index 99aead53283f7e48f8af17a63ca4e24740e7c014..d548de0547b8d8f4127f7462d039d86f8ff6b517 100644 (file)
--- a/src/models/ernie4-5.cpp
+++ b/src/models/ernie4-5.cpp
@@ -2,10 +2,10 @@
 
 llm_build_ernie4_5::llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/eurobert.cpp b/src/models/eurobert.cpp
index 86e3176edc002b48fe8b34a4734ec1a34a799a04..e8628d165d02ddbcfa001833a6c1b1e78d0d02cf 100644 (file)
--- a/src/models/eurobert.cpp
+++ b/src/models/eurobert.cpp
@@ -1,9 +1,9 @@
 #include "models.h"
 
 llm_build_eurobert::llm_build_eurobert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/exaone-moe.cpp b/src/models/exaone-moe.cpp
index a7396829ca8a89c44859bd8b19b3e4e285d3373e..ea75701c528c3b8d165360d40b91e013f1051cc2 100644 (file)
--- a/src/models/exaone-moe.cpp
+++ b/src/models/exaone-moe.cpp
@@ -2,10 +2,10 @@
 
 llm_build_exaone_moe::llm_build_exaone_moe(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_k;
+    const int64_t n_embd_head = hparams.n_embd_head_k();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/exaone.cpp b/src/models/exaone.cpp
index 62602b284ded8a564b95b65979fe86e9eab7e283..d4eea58e2f1d6513955b998f6c2154af49026b50 100644 (file)
--- a/src/models/exaone.cpp
+++ b/src/models/exaone.cpp
@@ -4,10 +4,10 @@
 
 llm_build_exaone::llm_build_exaone(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/exaone4.cpp b/src/models/exaone4.cpp
index 8b7e3dc06e5cb0ad1e6d7c523a729de150df9e84..755af3b747b0ef539d648987cae9f30937bf05b3 100644 (file)
--- a/src/models/exaone4.cpp
+++ b/src/models/exaone4.cpp
@@ -4,10 +4,10 @@
 template <bool iswa>
 llm_build_exaone4<iswa>::llm_build_exaone4(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_k;
+    const int64_t n_embd_head = hparams.n_embd_head_k();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/falcon-h1.cpp b/src/models/falcon-h1.cpp
index 785a7e5e6629cb1190ee8d2bb5e2901550112fb7..ff842d93a41dc8761988fcaf461941ba7523385f 100644 (file)
--- a/src/models/falcon-h1.cpp
+++ b/src/models/falcon-h1.cpp
@@ -2,7 +2,7 @@
 
 llm_build_falcon_h1::llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params) :
     llm_build_mamba_base(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/falcon.cpp b/src/models/falcon.cpp
index db1ccdb50085ff0f5a2388935cc77823843c9c66..9fcba508878a209793a62dc7bfa4993dd89a3ca8 100644 (file)
--- a/src/models/falcon.cpp
+++ b/src/models/falcon.cpp
@@ -2,11 +2,11 @@
 
 
 llm_build_falcon::llm_build_falcon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/gemma-embedding.cpp b/src/models/gemma-embedding.cpp
index 944c198bf9502459101d966c58c0ef670a470474..98110d45e3b8d5385d3a9cdc6e146fa7eab57a27 100644 (file)
--- a/src/models/gemma-embedding.cpp
+++ b/src/models/gemma-embedding.cpp
@@ -2,7 +2,7 @@
 
 llm_build_gemma_embedding::llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_k;
+    const int64_t n_embd_head = hparams.n_embd_head_k();
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/gemma.cpp b/src/models/gemma.cpp
index 4893d9af4b8c8cee647b6d66ddc4db490fe41c62..1869efd389a9121b31d50e124b69a49955ede988 100644 (file)
--- a/src/models/gemma.cpp
+++ b/src/models/gemma.cpp
@@ -2,7 +2,7 @@
 
 
 llm_build_gemma::llm_build_gemma(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/gemma2-iswa.cpp b/src/models/gemma2-iswa.cpp
index 7a9198193acc50b4f10f5fedae0caa0a2d5bae66..3927ddd297b2754ea6f700ab16f4874ef3a12655 100644 (file)
--- a/src/models/gemma2-iswa.cpp
+++ b/src/models/gemma2-iswa.cpp
@@ -1,7 +1,7 @@
 #include "models.h"
 
 llm_build_gemma2_iswa::llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_k;
+    const int64_t n_embd_head = hparams.n_embd_head_k();
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/gemma3.cpp b/src/models/gemma3.cpp
index dec3fc4b8bc3962217d4f5ecb99c6cb3260034bb..bbb4d9a81e825dc8225f845dda040a31bba485a9 100644 (file)
--- a/src/models/gemma3.cpp
+++ b/src/models/gemma3.cpp
@@ -2,7 +2,7 @@
 
 template <bool iswa>
 llm_build_gemma3<iswa>::llm_build_gemma3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_k;
+    const int64_t n_embd_head = hparams.n_embd_head_k();
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/gemma3n-iswa.cpp b/src/models/gemma3n-iswa.cpp
index 7db6d3bf4ec5c9bf1a8b39d5b2004a1a077acff0..8ce2ae39c2f4c07f30f37835161b1432b1b840ef 100644 (file)
--- a/src/models/gemma3n-iswa.cpp
+++ b/src/models/gemma3n-iswa.cpp
@@ -3,7 +3,7 @@
 llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params),
     model(model),
-    n_embd_head(model.hparams.n_embd_head_k),
+    n_embd_head(model.hparams.n_embd_head_k()),
     n_embd_altup(model.hparams.n_embd_altup),
     n_altup(model.hparams.n_altup),
     i_altup_act(model.hparams.i_altup_act) {

diff --git a/src/models/glm4-moe.cpp b/src/models/glm4-moe.cpp
index 97a65f4116602db6ac05c7f28fec36bafe8d7caf..7938545ed8a416623c135c71fcb4c0fe1c82eb84 100644 (file)
--- a/src/models/glm4-moe.cpp
+++ b/src/models/glm4-moe.cpp
@@ -1,9 +1,9 @@
 #include "models.h"
 
 llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     int sections[4];
     std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);

diff --git a/src/models/glm4.cpp b/src/models/glm4.cpp
index bcd837b30d6220f4d109d9bb70a863a2f9a43a18..b6ad8febed3070d74eeba89be47183dea96fcd10 100644 (file)
--- a/src/models/glm4.cpp
+++ b/src/models/glm4.cpp
@@ -3,10 +3,10 @@
 
 
 llm_build_glm4::llm_build_glm4(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     int sections[4];
     std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);

diff --git a/src/models/gpt2.cpp b/src/models/gpt2.cpp
index 60761c8e76521ddc47a8d6db20503e3fef2506d8..cb1238f2d3440638f2fe3f6ec21a27cc80867fc9 100644 (file)
--- a/src/models/gpt2.cpp
+++ b/src/models/gpt2.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_gpt2::llm_build_gpt2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * pos;

diff --git a/src/models/gptneox.cpp b/src/models/gptneox.cpp
index 2151b14e9394f61829fc913c73fb6b5814f7f59f..1c8fe6c836d183c0c22becf5d191212e54669eba 100644 (file)
--- a/src/models/gptneox.cpp
+++ b/src/models/gptneox.cpp
@@ -2,10 +2,10 @@
 
 
 llm_build_gptneox::llm_build_gptneox(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/granite-hybrid.cpp b/src/models/granite-hybrid.cpp
index d9b1c06da82e5a65667cf1b433ef1431981c0cbd..9b54a38c386b3b5eb305b4003d4413ed7817ab1a 100644 (file)
--- a/src/models/granite-hybrid.cpp
+++ b/src/models/granite-hybrid.cpp
@@ -2,8 +2,8 @@
 
 llm_build_granite_hybrid::llm_build_granite_hybrid(const llama_model & model, const llm_graph_params & params) :
     llm_build_mamba_base(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/granite.cpp b/src/models/granite.cpp
index fd97116ed3f8d33b59bcc89cfdb4d71558fbe261..7a7e1664c296820a0cc530bf5c7869cd64dcb302 100644 (file)
--- a/src/models/granite.cpp
+++ b/src/models/granite.cpp
@@ -5,10 +5,10 @@ llm_build_granite::llm_build_granite(
     const llm_graph_params & params)
     : llm_graph_context(params) {
 
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/grok.cpp b/src/models/grok.cpp
index 24232604b37aea68a4716200f625837b1f1b49b3..580d63e36ae5c6a82ee3cd909c61554e6d784e10 100644 (file)
--- a/src/models/grok.cpp
+++ b/src/models/grok.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_grok::llm_build_grok(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/grovemoe.cpp b/src/models/grovemoe.cpp
index 2081f068569d1a73d2cc74e725b91f4cc0d070bc..aa60d3e9388d24062eaf022af11b21c8e5d1a3c3 100644 (file)
--- a/src/models/grovemoe.cpp
+++ b/src/models/grovemoe.cpp
@@ -2,11 +2,11 @@
 
 llm_build_grovemoe::llm_build_grovemoe(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t n_embd_head    = hparams.n_embd_head_v;
+    const int64_t n_embd_head    = hparams.n_embd_head_v();
     const int64_t n_chunk_expert = n_expert / hparams.n_group_experts;
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/hunyuan-dense.cpp b/src/models/hunyuan-dense.cpp
index 7d5dcc7828b6f1b3331de3ad84b23eb2c03ae8ac..6a51707c85ba1edad34dc047059a038c021b1e9e 100644 (file)
--- a/src/models/hunyuan-dense.cpp
+++ b/src/models/hunyuan-dense.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_hunyuan_dense::llm_build_hunyuan_dense(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/hunyuan-moe.cpp b/src/models/hunyuan-moe.cpp
index cee2b541b73ecb8b76d3b7e29af8c9fc700ac465..806c30b36673e4085cb4ce491df88418878f50f4 100644 (file)
--- a/src/models/hunyuan-moe.cpp
+++ b/src/models/hunyuan-moe.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_hunyuan_moe::llm_build_hunyuan_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/internlm2.cpp b/src/models/internlm2.cpp
index 387e8211270d798a98a7f15c2996fdc5ffad3c07..441d250268ee2b0b2da0eb9eed935de9f609ab85 100644 (file)
--- a/src/models/internlm2.cpp
+++ b/src/models/internlm2.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_internlm2::llm_build_internlm2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/jais.cpp b/src/models/jais.cpp
index 3e3376e6a62431f15e269ee25da86afd976591f4..135bf288ba1b5e9375a55ec365d0a59d551a4857 100644 (file)
--- a/src/models/jais.cpp
+++ b/src/models/jais.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_jais::llm_build_jais(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/jais2.cpp b/src/models/jais2.cpp
index a69fcaa3bb3a19f641876356bfb4ec40bc45fd09..2cfe484eb52d510219086d722f40e36b04106f0e 100644 (file)
--- a/src/models/jais2.cpp
+++ b/src/models/jais2.cpp
@@ -3,10 +3,10 @@
 // JAIS-2 model graph builder
 // Uses: LayerNorm (not RMSNorm), relu2 activation, separate Q/K/V, RoPE embeddings
 llm_build_jais2::llm_build_jais2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/jamba.cpp b/src/models/jamba.cpp
index 1d482e425a62718b690d79bcbc3f1f8f9bf633b1..c0c89de187a6464da394aeac762012679dee344b 100644 (file)
--- a/src/models/jamba.cpp
+++ b/src/models/jamba.cpp
@@ -1,7 +1,7 @@
 #include "models.h"
 
 llm_build_jamba::llm_build_jamba(const llama_model & model, const llm_graph_params & params) : llm_build_mamba_base(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/kimi-linear.cpp b/src/models/kimi-linear.cpp
index 8d4b95e5192ea63770850e771060ee64b49e5014..063b17a2f6a20a3f75370f77253ca3aaaa4511e0 100644 (file)
--- a/src/models/kimi-linear.cpp
+++ b/src/models/kimi-linear.cpp
@@ -102,7 +102,7 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
     const int64_t kv_lora_rank = hparams.n_lora_kv;
     // qk_rope_head_dim = 64 (from Kimi config) which is hparams.n_rot
     // Confirmed from tensor shape: wkv_a_mqa [2304, 576] = [n_embd, kv_lora_rank + qk_rope_head_dim]
-    const int64_t n_embd_head_qk_rope = hparams.n_rot;  // config.qk_rope_head_dim
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();  // config.qk_rope_head_dim
     const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;  // 192 - 64 = 128
     // Attention scale for MLA
     const float kq_scale_mla = 1.0f / sqrtf((float)n_embd_head_k_mla);

diff --git a/src/models/lfm2.cpp b/src/models/lfm2.cpp
index 8ca8e6c8e2ce24a3a279a17af6ef1bbd5bf99063..dfa322166b109301be83f04b405d3c74b054fc64 100644 (file)
--- a/src/models/lfm2.cpp
+++ b/src/models/lfm2.cpp
@@ -39,7 +39,7 @@ llm_build_lfm2<iswa>::llm_build_lfm2(const llama_model & model, const llm_graph_
                                            inp_attn_type * inp_attn,
                                            int             il) -> ggml_tensor * {
         GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il));
-        const auto n_embd_head = hparams.n_embd_head_v;
+        const auto n_embd_head = hparams.n_embd_head_v();
         const auto n_head_kv   = hparams.n_head_kv(il);
 
         auto * q = build_lora_mm(model.layers[il].wq, cur);

diff --git a/src/models/llada-moe.cpp b/src/models/llada-moe.cpp
index 3bb9943f49d9a3e9d57a866b9399ea09bab45fee..18de88fde1f338264bbb08f7d3c262192e7cf54d 100644 (file)
--- a/src/models/llada-moe.cpp
+++ b/src/models/llada-moe.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_llada_moe::llm_build_llada_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/llada.cpp b/src/models/llada.cpp
index 857033660a04e5ba1d12633aeead51fa94bb3182..0dac9d616ae84faf302e5630aab38d258958b232 100644 (file)
--- a/src/models/llada.cpp
+++ b/src/models/llada.cpp
@@ -2,10 +2,10 @@
 
 llm_build_llada::llm_build_llada(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     // LLaDA is similar to LLaMA but uses non-causal attention for diffusion
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/llama-iswa.cpp b/src/models/llama-iswa.cpp
index 40dc2427a8d79cb1926a5da428284895d3d10701..67cb9a10ec5c99e4f412d2877c3a94d2dd6c50a1 100644 (file)
--- a/src/models/llama-iswa.cpp
+++ b/src/models/llama-iswa.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_llama_iswa::llm_build_llama_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/llama.cpp b/src/models/llama.cpp
index 279f2e301aee9ffc33d198f3f23b5b2cfa0d9324..ca4beac51f452e44de620619c7cf42300af028e4 100644 (file)
--- a/src/models/llama.cpp
+++ b/src/models/llama.cpp
@@ -2,10 +2,10 @@
 
 template <bool embed>
 llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/maincoder.cpp b/src/models/maincoder.cpp
index da57308167e8d4bc2aa9593968fe3a23ac8ae7a1..a72b7790a1f706655471f04b7a10fb298c59f8d2 100644 (file)
--- a/src/models/maincoder.cpp
+++ b/src/models/maincoder.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_maincoder::llm_build_maincoder(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/minicpm3.cpp b/src/models/minicpm3.cpp
index 297cc34ba58b738de4ec835613e711c6c7f3e358..89dd71051577e80eb07d9c1ddd81e7485bee580c 100644 (file)
--- a/src/models/minicpm3.cpp
+++ b/src/models/minicpm3.cpp
@@ -5,10 +5,10 @@ llm_build_minicpm3::llm_build_minicpm3(const llama_model & model, const llm_grap
     const int64_t n_embd_base = 256;
     const float scale_embd  = 12.0f;
     const float scale_depth = 1.4f;
-    const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k));
+    const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k()));
 
-    const uint32_t n_embd_head_qk_rope = hparams.n_rot;
-    const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+    const uint32_t n_embd_head_qk_rope = hparams.n_rot();
+    const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k() - hparams.n_rot();
 
     const uint32_t kv_lora_rank = hparams.n_lora_kv;
 
@@ -51,21 +51,21 @@ llm_build_minicpm3::llm_build_minicpm3(const llama_model & model, const llm_grap
                     LLM_NORM_RMS, il);
             cb(q, "q", il);
 
-            // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens}
+            // {q_lora_rank, n_head * hparams.n_embd_head_k()} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k(), n_tokens}
             q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
             cb(q, "q", il);
 
             // split into {n_head * n_embd_head_qk_nope, n_tokens}
             ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
-                    ggml_row_size(q->type, hparams.n_embd_head_k),
-                    ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                    ggml_row_size(q->type, hparams.n_embd_head_k()),
+                    ggml_row_size(q->type, hparams.n_embd_head_k() * n_head),
                     0);
             cb(q_nope, "q_nope", il);
 
             // and {n_head * n_embd_head_qk_rope, n_tokens}
             ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
-                    ggml_row_size(q->type, hparams.n_embd_head_k),
-                    ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                    ggml_row_size(q->type, hparams.n_embd_head_k()),
+                    ggml_row_size(q->type, hparams.n_embd_head_k() * n_head),
                     ggml_row_size(q->type, n_embd_head_qk_nope));
             cb(q_pe, "q_pe", il);
 
@@ -97,15 +97,15 @@ llm_build_minicpm3::llm_build_minicpm3(const llama_model & model, const llm_grap
 
             // split into {n_head * n_embd_head_qk_nope, n_tokens}
             ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
-                    ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
-                    ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                    ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v()),
+                    ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v())),
                     0);
             cb(k_nope, "k_nope", il);
 
             // and {n_head * n_embd_head_v, n_tokens}
-            ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
-                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
+            ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v(), n_head, n_tokens,
+                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v())),
+                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v())*n_head),
                     ggml_row_size(kv->type, (n_embd_head_qk_nope)));
             cb(v_states, "v_states", il);
 

diff --git a/src/models/minimax-m2.cpp b/src/models/minimax-m2.cpp
index fbeed8eab9b0644b55dbae0451592d7119b02d02..83d0916c08c1888ab418772c7702ad37d09e62a9 100644 (file)
--- a/src/models/minimax-m2.cpp
+++ b/src/models/minimax-m2.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_minimax_m2::llm_build_minimax_m2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    // GGML_ASSERT(n_embd_head == hparams.n_rot); this is wrong in case of minimax, head_dim = 128, n_rot = 64
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    // GGML_ASSERT(n_embd_head == n_rot); this is wrong in case of minimax, head_dim = 128, n_rot = 64
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/mistral3.cpp b/src/models/mistral3.cpp
index 49734989df24c298381ace7c6c502b4c179c2052..42a5117ff020253754a6ecca26dce9b60291af2b 100644 (file)
--- a/src/models/mistral3.cpp
+++ b/src/models/mistral3.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_mistral3::llm_build_mistral3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/modern-bert.cpp b/src/models/modern-bert.cpp
index 32066c712b494d29ae5bf0e22d7e3d623a103dbb..26020584c6dbff95a7dc0da2ba58eb067fc2cae6 100644 (file)
--- a/src/models/modern-bert.cpp
+++ b/src/models/modern-bert.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_modern_bert::llm_build_modern_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/mpt.cpp b/src/models/mpt.cpp
index 2328e027a74102876b8d60139e752d1ba7a92015..ce44a805f5c2d849e4d25bc913d5d43a618dea3e 100644 (file)
--- a/src/models/mpt.cpp
+++ b/src/models/mpt.cpp
@@ -3,10 +3,10 @@
 
 
 llm_build_mpt::llm_build_mpt(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * pos;

diff --git a/src/models/nemotron-h.cpp b/src/models/nemotron-h.cpp
index fa2b55a28417d2d5fb64c7db8103ea344641db2c..635821505006e2101c842f1c47ff82c4828a45f5 100644 (file)
--- a/src/models/nemotron-h.cpp
+++ b/src/models/nemotron-h.cpp
@@ -2,8 +2,8 @@
 
 llm_build_nemotron_h::llm_build_nemotron_h(const llama_model & model, const llm_graph_params & params) :
     llm_build_mamba_base(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/nemotron.cpp b/src/models/nemotron.cpp
index fcead041f0a2561694ab3f84795d79d8110fd08c..34aa6fa5ec4cf82d6d7e8856ca65d8486d70949d 100644 (file)
--- a/src/models/nemotron.cpp
+++ b/src/models/nemotron.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_nemotron::llm_build_nemotron(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    //GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    //GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/neo-bert.cpp b/src/models/neo-bert.cpp
index 7c32bfca5f5604f20a8867c778fe46f064aa5402..2fdf4a3692f649b685b0fd805920707dea22b30a 100644 (file)
--- a/src/models/neo-bert.cpp
+++ b/src/models/neo-bert.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_neo_bert::llm_build_neo_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/olmo.cpp b/src/models/olmo.cpp
index bbd623f1112a1969603d10dbc2aeecfe0e12b64e..26f4b6ee628523eee8c2f7bd0fb68958cd290fe9 100644 (file)
--- a/src/models/olmo.cpp
+++ b/src/models/olmo.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_olmo::llm_build_olmo(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/olmo2.cpp b/src/models/olmo2.cpp
index 713552dab89732d01bdb874359baea51e888a640..5076359e3f9aa887e6ee80bcf342f6fec1dad006 100644 (file)
--- a/src/models/olmo2.cpp
+++ b/src/models/olmo2.cpp
@@ -2,10 +2,10 @@
 
 template <bool iswa>
 llm_build_olmo2<iswa>::llm_build_olmo2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/olmoe.cpp b/src/models/olmoe.cpp
index a05b892c70810e6d7e2f66976f1b3a8de05db83c..83a56a0b3b6a2f8e76e2d368aa1bc117aa964be5 100644 (file)
--- a/src/models/olmoe.cpp
+++ b/src/models/olmoe.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_olmoe::llm_build_olmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/openelm.cpp b/src/models/openelm.cpp
index fbf682ec8352e790f8810014d952315a8f8837c1..5df6fe3e3ce8f71f36457a64dd2579f053be04c6 100644 (file)
--- a/src/models/openelm.cpp
+++ b/src/models/openelm.cpp
@@ -1,9 +1,9 @@
 #include "models.h"
 
 llm_build_openelm::llm_build_openelm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/orion.cpp b/src/models/orion.cpp
index bb02273bfe74b88b5908dd1ef69c8ed3d18b5f97..48c01efe368111d5dd861b97ee7f1a938e7fd9f2 100644 (file)
--- a/src/models/orion.cpp
+++ b/src/models/orion.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_orion::llm_build_orion(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/paddleocr.cpp b/src/models/paddleocr.cpp
index 39a368df53ba0675fb9765a456be3a5a29c6c045..340455c2d5f07614e4e786043b36a76073225b04 100644 (file)
--- a/src/models/paddleocr.cpp
+++ b/src/models/paddleocr.cpp
@@ -5,10 +5,10 @@ llm_build_paddleocr::llm_build_paddleocr(const llama_model & model, const llm_gr
 
     // NOTE: same with qwen2vl.cpp, but bias tensors are optional
 
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/pangu-embedded.cpp b/src/models/pangu-embedded.cpp
index 664572a5001467406f05d13a6fab8ae2cb9aece9..1cf0938e68f9cca814c13e9cad87c6f4246c2d4c 100644 (file)
--- a/src/models/pangu-embedded.cpp
+++ b/src/models/pangu-embedded.cpp
@@ -2,10 +2,10 @@
 
 
 llm_build_pangu_embedded::llm_build_pangu_embedded(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/phi2.cpp b/src/models/phi2.cpp
index 22dbf610767d24c7d05000cd690bfa6f9c8989c8..32d40d71fb71fb777883417a6e111685094288de 100644 (file)
--- a/src/models/phi2.cpp
+++ b/src/models/phi2.cpp
@@ -2,10 +2,10 @@
 
 
 llm_build_phi2::llm_build_phi2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * attn_norm_output;

diff --git a/src/models/phi3.cpp b/src/models/phi3.cpp
index 803e374aa55e3e1b274174ea2fe4bae9535280bd..3d11a9459c4caf0d93ed73434767936438b5d080 100644 (file)
--- a/src/models/phi3.cpp
+++ b/src/models/phi3.cpp
@@ -2,10 +2,10 @@
 
 template<bool iswa>
 llm_build_phi3<iswa>::llm_build_phi3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/plamo.cpp b/src/models/plamo.cpp
index 04ff709f9c6ffcd4e4f8d5fffb5f1e9a6703e643..b7a7121104258e29b047736a8a052ecebffb6616 100644 (file)
--- a/src/models/plamo.cpp
+++ b/src/models/plamo.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_plamo::llm_build_plamo(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/plamo2.cpp b/src/models/plamo2.cpp
index 276d3829b1f3909ed46962f15fb4aa590b131c4d..f02acbc18699cddc0b229a56475845ea220f6970 100644 (file)
--- a/src/models/plamo2.cpp
+++ b/src/models/plamo2.cpp
@@ -106,9 +106,9 @@ ggml_tensor * llm_build_plamo2::build_plamo2_attn_layer(llm_graph_input_attn_kv
         cb(qkv, "wqkv", il);
 
         // split QKV tensor into Q, K, V
-        const int64_t n_embd_head_q = hparams.n_embd_head_k;
-        const int64_t n_embd_head_k = hparams.n_embd_head_k;
-        const int64_t n_embd_head_v = hparams.n_embd_head_v;
+        const int64_t n_embd_head_q = hparams.n_embd_head_k();
+        const int64_t n_embd_head_k = hparams.n_embd_head_k();
+        const int64_t n_embd_head_v = hparams.n_embd_head_v();
         int32_t       n_head        = hparams.n_head(il);
         int32_t       n_head_kv     = hparams.n_head_kv(il);
 

diff --git a/src/models/plamo3.cpp b/src/models/plamo3.cpp
index 55c8064679e27e464adfe93ddc412c800b5aaa23..32af6e04663ea54a45238b915353d22fe31d716e 100644 (file)
--- a/src/models/plamo3.cpp
+++ b/src/models/plamo3.cpp
@@ -3,8 +3,8 @@
 template <bool iswa>
 llm_build_plamo3<iswa>::llm_build_plamo3(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    const int64_t head_dim_q = hparams.n_embd_head_k;
-    const int64_t head_dim_v = hparams.n_embd_head_v;
+    const int64_t head_dim_q = hparams.n_embd_head_k();
+    const int64_t head_dim_v = hparams.n_embd_head_v();
 
     ggml_tensor * cur;
     ggml_tensor * inpL = build_inp_embd(model.tok_embd);

diff --git a/src/models/plm.cpp b/src/models/plm.cpp
index 612a487c564c45c660ad1b1ac25256fac823504b..bcb651ce543f328d10b301dd0979d9dc180b39cd 100644 (file)
--- a/src/models/plm.cpp
+++ b/src/models/plm.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_plm::llm_build_plm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const float kq_scale = 1.0f/sqrtf(float(hparams.n_embd_head_k));
+    const float kq_scale = 1.0f/sqrtf(float(hparams.n_embd_head_k()));
 
-    const uint32_t n_embd_head_qk_rope = hparams.n_rot;
-    const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+    const uint32_t n_embd_head_qk_rope = hparams.n_rot();
+    const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k() - hparams.n_rot();
 
     const uint32_t kv_lora_rank = hparams.n_lora_kv;
 
@@ -38,15 +38,15 @@ llm_build_plm::llm_build_plm(const llama_model & model, const llm_graph_params &
 
             // split into {n_head * n_embd_head_qk_nope, n_tokens}
             ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
-                    ggml_row_size(q->type, hparams.n_embd_head_k),
-                    ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                    ggml_row_size(q->type, hparams.n_embd_head_k()),
+                    ggml_row_size(q->type, hparams.n_embd_head_k() * n_head),
                     0);
             cb(q_nope, "q_nope", il);
 
             // and {n_head * n_embd_head_qk_rope, n_tokens}
             ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
-                    ggml_row_size(q->type, hparams.n_embd_head_k),
-                    ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                    ggml_row_size(q->type, hparams.n_embd_head_k()),
+                    ggml_row_size(q->type, hparams.n_embd_head_k() * n_head),
                     ggml_row_size(q->type, n_embd_head_qk_nope));
             cb(q_pe, "q_pe", il);
 
@@ -78,23 +78,23 @@ llm_build_plm::llm_build_plm(const llama_model & model, const llm_graph_params &
 
             // split into {n_head * n_embd_head_qk_nope, n_tokens}
             ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
-                    ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
-                    ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                    ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v()),
+                    ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v())),
                     0);
             cb(k_nope, "k_nope", il);
 
             // and {n_head * n_embd_head_v, n_tokens}
-            ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
-                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
+            ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v(), n_head, n_tokens,
+                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v())),
+                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v())*n_head),
                     ggml_row_size(kv->type, (n_embd_head_qk_nope)));
             cb(v_states, "v_states", il);
 
             v_states = ggml_cont(ctx0, v_states);
             cb(v_states, "v_states", il);
 
-            v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens,
-                    ggml_row_size(kv->type, hparams.n_embd_head_v * n_head),
+            v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v() * n_head, n_tokens,
+                    ggml_row_size(kv->type, hparams.n_embd_head_v() * n_head),
                     0);
             cb(v_states, "v_states", il);
 

diff --git a/src/models/qwen.cpp b/src/models/qwen.cpp
index 31fd9b73763dedb50e211274b82bfc99a241db87..7390f1320bf5d0071012150c779a80c0d72c326d 100644 (file)
--- a/src/models/qwen.cpp
+++ b/src/models/qwen.cpp
@@ -2,9 +2,9 @@
 
 
 llm_build_qwen::llm_build_qwen(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/qwen2.cpp b/src/models/qwen2.cpp
index 3da4dea3c16752b35040398d6fd6d5b293b9e928..58c10622508422a786bd9c4e90edf81d680e103e 100644 (file)
--- a/src/models/qwen2.cpp
+++ b/src/models/qwen2.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_qwen2::llm_build_qwen2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/qwen2moe.cpp b/src/models/qwen2moe.cpp
index e19061334c01ddc98f50839e41ba1287c965c007..60761789dc9c378e2ffa3f473b3593b91d222a1c 100644 (file)
--- a/src/models/qwen2moe.cpp
+++ b/src/models/qwen2moe.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_qwen2moe::llm_build_qwen2moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/qwen2vl.cpp b/src/models/qwen2vl.cpp
index 9be38675cf7e2cbb05564d1d6eb7663d03b9eb5c..9004bab9db12a433132248f8092b9adc56f5d1b8 100644 (file)
--- a/src/models/qwen2vl.cpp
+++ b/src/models/qwen2vl.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_qwen2vl::llm_build_qwen2vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/qwen3.cpp b/src/models/qwen3.cpp
index a5cfffa531491a324cf7d15491b0465057cb82c5..be4811aba1e96b15d29d975267023817925d2f89 100644 (file)
--- a/src/models/qwen3.cpp
+++ b/src/models/qwen3.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_qwen3::llm_build_qwen3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/qwen35.cpp b/src/models/qwen35.cpp
index afc5a1aad7753a3b9b5b87dd20dda0ac2e78b108..ba096a5a7ba14cac306453230ed76f17d3370b22 100644 (file)
--- a/src/models/qwen35.cpp
+++ b/src/models/qwen35.cpp
@@ -4,9 +4,9 @@
 
 llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_params & params) :
     llm_build_delta_net_base(params), model(model) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     int sections[4];
     std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
@@ -117,8 +117,8 @@ ggml_tensor * llm_build_qwen35::build_layer_attn(
         ggml_tensor *             inp_pos,
         int *                     sections,
         int                       il) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention
 

diff --git a/src/models/qwen35moe.cpp b/src/models/qwen35moe.cpp
index fe081af212f50268fbb8b370de04535d364d8271..fe382286e984d6ec9ca48aea8b12b64c888cc5f1 100644 (file)
--- a/src/models/qwen35moe.cpp
+++ b/src/models/qwen35moe.cpp
@@ -4,9 +4,9 @@
 
 llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_graph_params & params) :
     llm_build_delta_net_base(params), model(model) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     int sections[4];
     std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
@@ -117,8 +117,8 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn(
         ggml_tensor *             inp_pos,
         int *                     sections,
         int                       il) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention
 

diff --git a/src/models/qwen3moe.cpp b/src/models/qwen3moe.cpp
index 9cf1ec03c65dbe896a0f80d4b4f7b5c10b8b8c09..5912a715822cf541feefd52e1b0b6d22f1f2ac77 100644 (file)
--- a/src/models/qwen3moe.cpp
+++ b/src/models/qwen3moe.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_qwen3moe::llm_build_qwen3moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/qwen3next.cpp b/src/models/qwen3next.cpp
index 9b8164ddfab7f2f0700209a89215c4d29b08cc60..30912fd5e363b3b6324ad65b7be2a8a28b78f18b 100644 (file)
--- a/src/models/qwen3next.cpp
+++ b/src/models/qwen3next.cpp
@@ -100,8 +100,8 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn(
         ggml_tensor *             cur,
         ggml_tensor *             inp_pos,
         int                       il) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention
 

diff --git a/src/models/qwen3vl-moe.cpp b/src/models/qwen3vl-moe.cpp
index f347c5d6fb035da371d5c89a3dfcf955d8b8ec22..195daea66c988949c8a6a945037b3c3038552610 100644 (file)
--- a/src/models/qwen3vl-moe.cpp
+++ b/src/models/qwen3vl-moe.cpp
@@ -4,10 +4,10 @@ llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_
     const size_t n_deepstack_layers = hparams.n_deepstack_layers;
 
     const int64_t n_embd      = hparams.n_embd;
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/qwen3vl.cpp b/src/models/qwen3vl.cpp
index 0f8315b32404b5d9ba521f130eb4687dc7806a10..bbd5f42ba5bd8db77a16eee2982185ca469fe81b 100644 (file)
--- a/src/models/qwen3vl.cpp
+++ b/src/models/qwen3vl.cpp
@@ -4,10 +4,10 @@ llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_
     const size_t n_deepstack_layers = hparams.n_deepstack_layers;
 
     const int64_t n_embd      = hparams.n_embd;
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/refact.cpp b/src/models/refact.cpp
index ff5eb2841db9342ea0d2928d03aa847710132421..140700d9e2d8b46b4325ce331fadd2080fc25c26 100644 (file)
--- a/src/models/refact.cpp
+++ b/src/models/refact.cpp
@@ -1,9 +1,9 @@
 #include "models.h"
 
 llm_build_refact::llm_build_refact(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/rnd1.cpp b/src/models/rnd1.cpp
index de9ab1c652558d3cc2ffacb1b1e86449e8f11e5f..c8e1f43400fffbed70f618c372342fa319c1148a 100644 (file)
--- a/src/models/rnd1.cpp
+++ b/src/models/rnd1.cpp
@@ -2,10 +2,10 @@
 
 // RND1 is a Qwen3Moe AR model converted to diffusion model.
 llm_build_rnd1::llm_build_rnd1(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/seed-oss.cpp b/src/models/seed-oss.cpp
index 0dc33c50ba33f67c41afdc622dd689ec4029bc57..a4d0b75d84699020307e5c53361055068dbbbef1 100644 (file)
--- a/src/models/seed-oss.cpp
+++ b/src/models/seed-oss.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_seed_oss::llm_build_seed_oss(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/smallthinker.cpp b/src/models/smallthinker.cpp
index 8723905e839039b9ce191bcb41c72c0b316bba33..e2155aacef41a0c71488691f9bbb96e2fa52e391 100644 (file)
--- a/src/models/smallthinker.cpp
+++ b/src/models/smallthinker.cpp
@@ -2,10 +2,10 @@
 
 template <bool iswa>
 llm_build_smallthinker<iswa>::llm_build_smallthinker(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params){
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/smollm3.cpp b/src/models/smollm3.cpp
index 97c30deed54e6c158279c6594c370ecbcef23ffc..e267fd8f32f5cb3bec368dc1d5e51cccbee60049 100644 (file)
--- a/src/models/smollm3.cpp
+++ b/src/models/smollm3.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_smollm3::llm_build_smollm3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/stablelm.cpp b/src/models/stablelm.cpp
index bed1915c0067686727a55d1f15cbc1f388603924..ff5aced93b3f4e7648addfda066155eaa3c121d6 100644 (file)
--- a/src/models/stablelm.cpp
+++ b/src/models/stablelm.cpp
@@ -1,9 +1,9 @@
 #include "models.h"
 
 llm_build_stablelm::llm_build_stablelm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/starcoder.cpp b/src/models/starcoder.cpp
index e197af4a8c63f0f521826f542b9e6ca2e372c999..941cee982194abd4426ee2e93d79cf42b7c3508f 100644 (file)
--- a/src/models/starcoder.cpp
+++ b/src/models/starcoder.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_starcoder::llm_build_starcoder(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/starcoder2.cpp b/src/models/starcoder2.cpp
index e40ef2cb7493a20549ac33618026c58a2bb6d8d9..a5965aceb3b4ac4b12923dafbf76a915985e4b72 100644 (file)
--- a/src/models/starcoder2.cpp
+++ b/src/models/starcoder2.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_starcoder2::llm_build_starcoder2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/step35-iswa.cpp b/src/models/step35-iswa.cpp
index aa8e98b737015f597d6493f65a36714f102260b1..176209cd93e15b7b60b91b74ebee0ce2517547d1 100644 (file)
--- a/src/models/step35-iswa.cpp
+++ b/src/models/step35-iswa.cpp
@@ -52,7 +52,7 @@ llm_build_step35_iswa::llm_build_step35_iswa(const llama_model & model, const ll
             // RoPE (partial rotary factors per layer)
             const bool is_swa = hparams.is_swa(il);
             ggml_tensor * rope_factors = is_swa ? nullptr : model.get_rope_factors(cparams, il);
-            const int64_t n_rot_l = is_swa ? hparams.n_rot : (hparams.n_rot / 2);
+            const int64_t n_rot_l = hparams.n_rot(il);
             Qcur = ggml_rope_ext(
                 ctx0, Qcur, inp_pos, rope_factors,
                 n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,

diff --git a/src/models/t5-dec.cpp b/src/models/t5-dec.cpp
index 297e450de76fe60073d459b50a3a40a672ec86c5..8ca8372bd4ccd31864dcf06709b06d90602db2f4 100644 (file)
--- a/src/models/t5-dec.cpp
+++ b/src/models/t5-dec.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_t5_dec::llm_build_t5_dec(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
     //const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/t5-enc.cpp b/src/models/t5-enc.cpp
index 70e1d80dcddc7f6b1e13658d53305046456c1512..395dfb510420f02d497ba6b2c0838035a0264844 100644 (file)
--- a/src/models/t5-enc.cpp
+++ b/src/models/t5-enc.cpp
@@ -1,9 +1,9 @@
 #include "models.h"
 
 llm_build_t5_enc::llm_build_t5_enc(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

diff --git a/src/models/xverse.cpp b/src/models/xverse.cpp
index 364797dd31b883ef36c94ee79a0c5466627aa79f..3a8dfafccebe1f0315204dd8b636d80958056ff5 100644 (file)
--- a/src/models/xverse.cpp
+++ b/src/models/xverse.cpp
@@ -1,10 +1,10 @@
 #include "models.h"
 
 llm_build_xverse::llm_build_xverse(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_head = hparams.n_embd_head_v();
 
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-    GGML_ASSERT(n_embd_head == hparams.n_rot);
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
 
     ggml_tensor * cur;
     ggml_tensor * inpL;