models : fix Phi-3 SWA parameters

ggml-ci

Author: ggerganov

src/llama-graph.cpp

@@ -1316,9 +1316,8 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
     }
 
-    {
-        GGML_ASSERT(hparams.n_swa_pattern > 1 && "Use llama_kv_cache_unified for non-SWA");
-        GGML_ASSERT(hparams.n_swa > 0         && "Use llama_kv_cache_unified for non-SWA");
+    if (hparams.n_swa_pattern > 1) {
+        GGML_ASSERT(hparams.n_swa > 0          && "Use llama_kv_cache_unified for non-SWA");
 
         const auto n_kv = kv_self->get_kv_swa()->get_n();
 

src/llama-graph.h

@@ -256,10 +256,10 @@ class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
 
     void set_input(const llama_ubatch * ubatch) override;
 
-    ggml_tensor * get_kq_mask()     const { return self_kq_mask_cnv; }
+    ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
 
-    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch]
-    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch]
+    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch]
+    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;

src/llama-model.cpp

@@ -856,20 +856,34 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 // for backward compatibility ; see: https://github.com/ggerganov/llama.cpp/pull/8931
                 if ((hparams.n_layer == 32 || hparams.n_layer == 40) && hparams.n_ctx_train == 4096) {
                     // default value for Phi-3-mini-4k-instruct and Phi-3-medium-4k-instruct
+                    LLAMA_LOG_WARN("%s: assuming n_swa = 2047 for Phi-3-mini-4k-instruct and Phi-3-medium-4k-instruct\n", __func__);
+
                     hparams.n_swa = 2047;
                 } else if (hparams.n_layer == 32 && hparams.n_head_kv(0) == 32 && hparams.n_ctx_train == 131072) {
                     // default value for Phi-3-mini-128k-instruct
-                    // note: this seems incorrect because the window is bigger than the train context?
-                    hparams.n_swa = 262144;
+                    LLAMA_LOG_WARN("%s: assuming n_swa = n_ctx_train for Phi-3-mini-128k-instruct\n", __func__);
+
+                    hparams.n_swa = hparams.n_ctx_train;
+                    hparams.n_swa_pattern = 1;
                 } else if (hparams.n_layer == 40 && hparams.n_ctx_train == 131072) {
                     // default value for Phi-3-medium-128k-instruct
-                    // note: this seems incorrect because the window is equal to the train context?
-                    hparams.n_swa = 131072;
+                    LLAMA_LOG_WARN("%s: assuming n_swa = n_ctx_train for Phi-3-medium-128k-instruct\n", __func__);
+
+                    hparams.n_swa = hparams.n_ctx_train;
+                    hparams.n_swa_pattern = 1;
                 }
+
                 bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
                 if (!found_swa && hparams.n_swa == 0) {
                     throw std::runtime_error("invalid value for sliding_window");
                 }
+
+                if (hparams.n_swa > hparams.n_ctx_train) {
+                    LLAMA_LOG_WARN("%s: unexpected n_swa: %d >= %d, setting to 0\n", __func__, hparams.n_swa, hparams.n_ctx_train);
+
+                    hparams.n_swa = hparams.n_ctx_train;
+                    hparams.n_swa_pattern = 1;
+                }
             } break;
         case LLM_ARCH_PHIMOE:
             {