Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	src/llama-model.cpp

Author: LostRuins

src/llama-context.cpp

@@ -537,16 +537,12 @@ llm_graph_result_ptr llama_context::build_kv_self_shift(
         const int64_t n_head_kv    = hparams.n_head_kv(il);
         const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
 
-        float freq_base_l  = cparams.rope_freq_base;
-        float freq_scale_l = cparams.rope_freq_scale;
+        const bool is_swa = hparams.is_swa(il);
 
-        // TODO: improve
-        if (model.arch == LLM_ARCH_GEMMA3) {
-            const bool is_sliding = hparams.is_sliding(il);
-
-            freq_base_l  = is_sliding ? 10000.0f : cparams.rope_freq_base;
-            freq_scale_l = is_sliding ? 1.0f     : cparams.rope_freq_scale;
-        }
+        // note: the swa rope params could become part of the cparams in the future
+        //       if we decide to make them configurable, like the non-sliding ones
+        const float freq_base_l  = is_swa ? hparams.rope_freq_base_train_swa  : cparams.rope_freq_base;
+        const float freq_scale_l = is_swa ? hparams.rope_freq_scale_train_swa : cparams.rope_freq_scale;
 
         ggml_tensor * rope_factors = kv_self->cbs.get_rope_factors(n_ctx_per_seq(), il);
 

src/llama-graph.cpp

@@ -1311,29 +1311,23 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
-llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified(
-                bool   causal,
-                bool   swa) const {
+llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() const {
     const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
 
     auto inp = std::make_unique<llm_graph_input_attn_kv_unified>(hparams, cparams, kv_self);
 
     const auto n_kv = kv_self->n;
 
-    inp->self_kq_mask = causal
-        ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv,     GGML_PAD(n_tokens, GGML_KQ_MASK_PAD))
-        : ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+    inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
     //cb(inp->self_kq_mask, "KQ_mask", -1);
     ggml_set_input(inp->self_kq_mask);
 
     inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
 
-    if (swa) {
+    if (hparams.n_swa_pattern > 1) {
         GGML_ASSERT(hparams.n_swa > 0);
 
-        inp->self_kq_mask_swa = causal
-            ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv,     GGML_PAD(n_tokens, GGML_KQ_MASK_PAD))
-            : ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        inp->self_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
         //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1);
         ggml_set_input(inp->self_kq_mask_swa);
 
@@ -1403,9 +1397,9 @@ ggml_tensor * llm_graph_context::build_attn(
         ggml_build_forward_expand(gf, ggml_cpy(ctx0, v_cur, v_cache_view));
     }
 
-    const bool is_sliding = hparams.is_sliding(il);
+    const bool is_swa = hparams.is_swa(il);
 
-    const auto & kq_mask = is_sliding ? inp->get_kq_mask_swa() : inp->get_kq_mask();
+    const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask();
 
     const auto n_kv = kv_self->n;
 

src/llama-graph.h

@@ -509,9 +509,7 @@ struct llm_graph_context {
                   float   kq_scale,
                     int   il) const;
 
-    llm_graph_input_attn_kv_unified * build_attn_inp_kv_unified(
-            bool causal,
-            bool swa) const;
+    llm_graph_input_attn_kv_unified * build_attn_inp_kv_unified() const;
 
     ggml_tensor * build_attn(
             llm_graph_input_attn_kv_unified * inp,

src/llama-hparams.cpp

@@ -70,7 +70,7 @@ uint32_t llama_hparams::n_embd_v_s() const {
     return ssm_d_state * ssm_d_inner;
 }
 
-bool llama_hparams::is_sliding(uint32_t il) const {
+bool llama_hparams::is_swa(uint32_t il) const {
     if (il < n_layer) {
         return n_swa > 0 && n_swa_pattern > 0 && il % n_swa_pattern < (n_swa_pattern - 1);
     }

src/llama-hparams.h

@@ -79,7 +79,9 @@ struct llama_hparams {
 
     float    rope_attn_factor = 1.0f;
     float    rope_freq_base_train;
+    float    rope_freq_base_train_swa;
     float    rope_freq_scale_train;
+    float    rope_freq_scale_train_swa;
     uint32_t n_ctx_orig_yarn;
     float    rope_yarn_log_mul;
 
@@ -135,7 +137,7 @@ struct llama_hparams {
     // dimension of the recurrent state embeddings
     uint32_t n_embd_v_s() const;
 
-    bool is_sliding(uint32_t il) const;
+    bool is_swa(uint32_t il) const;
 };
 
 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");