Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	.devops/rocm.Dockerfile
#	docs/build-s390x.md
#	docs/development/HOWTO-add-model.md
#	docs/ops.md
#	docs/ops/CPU.csv
#	docs/ops/CUDA.csv
#	ggml/CMakeLists.txt
#	ggml/src/ggml-cann/acl_tensor.cpp
#	ggml/src/ggml-cann/aclnn_ops.cpp
#	ggml/src/ggml-cann/aclnn_ops.h
#	ggml/src/ggml-cann/ggml-cann.cpp
#	ggml/src/ggml-cpu/CMakeLists.txt
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-opencl/kernels/rms_norm.cl
#	scripts/create_ops_docs.py
#	tests/test-backend-ops.cpp
#	tools/export-lora/export-lora.cpp

src/llama-context.cpp

@@ -105,7 +105,7 @@ llama_context::llama_context(
 
     {
         const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
-        const bool supports_set_rows = LLAMA_SET_ROWS ? (atoi(LLAMA_SET_ROWS) != 0) : false;
+        supports_set_rows = LLAMA_SET_ROWS ? (atoi(LLAMA_SET_ROWS) != 0) : false;
 
         if (!supports_set_rows && !cparams.kv_unified) {
             LLAMA_LOG_WARN("%s: non-unified KV cache requires ggml_set_rows() - forcing unified KV cache\n", __func__);
@@ -899,6 +899,12 @@ int llama_context::encode(const llama_batch & batch_inp) {
         }
     }
 
+    if (!supports_set_rows) {
+        // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
+        // overlap with device computation.
+        ggml_backend_sched_reset(sched.get());
+    }
+
     // TODO: hacky solution
     if (model.arch == LLM_ARCH_T5 && t_embd) {
         //cross.t_embd = t_embd;
@@ -1229,6 +1235,12 @@ int llama_context::decode(const llama_batch & batch_inp) {
     // wait for the computation to finish (automatically done when obtaining the model output)
     //synchronize();
 
+    if (!supports_set_rows) {
+        // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
+        // overlap with device computation.
+        ggml_backend_sched_reset(sched.get());
+    }
+
     return 0;
 }
 

src/llama-context.h

@@ -287,6 +287,10 @@ private:
 
     bool has_evaluated_once = false;
 
+    // env: LLAMA_SET_ROWS (temporary)
+    // ref: https://github.com/ggml-org/llama.cpp/pull/14285
+    bool supports_set_rows = false;
+
     // perf
     mutable int64_t t_start_us  = 0;
     mutable int64_t t_load_us   = 0;

src/llama-hparams.h

@@ -98,7 +98,7 @@ struct llama_hparams {
     float    rope_freq_scale_train;
     float    rope_freq_scale_train_swa;
     uint32_t n_ctx_orig_yarn;
-    float    rope_yarn_log_mul;
+    float    rope_yarn_log_mul = 0.0f;
 
     std::array<int, 4> rope_sections;
 

src/llama-model.cpp

@@ -1374,7 +1374,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     // that have no expert_gating_func model parameter set
                     hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX;
                 }
-                ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul);
+                ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, false);
 
                 switch (hparams.n_layer) {
                     case 27: type = LLM_TYPE_16B; break;
@@ -16291,7 +16291,7 @@ private:
         {
             // PLaMo-2 uses combined QKV tensor
             ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur);
-            cb(qkv, "qkv", il);
+            cb(qkv, "wqkv", il);
 
             // split QKV tensor into Q, K, V
             const int64_t n_embd_head_q = hparams.n_embd_head_k;
@@ -16331,7 +16331,7 @@ private:
                     ext_factor, attn_factor, beta_fast, beta_slow
                     );
 
-            cur = build_attn(inp, model.layers[il].wo, NULL, Qcur, Kcur, Vcur, NULL, NULL, 1.0f, il);
+            cur = build_attn(inp, model.layers[il].wo, NULL, Qcur, Kcur, Vcur, NULL, NULL, 1.0f/sqrtf(float(n_embd_head_v)), il);
         }
 
         cb(cur, "attn_out", il);
@@ -16406,8 +16406,9 @@ private:
             ggml_build_forward_expand(gf,
                 ggml_cpy(ctx0, last_conv,
                     ggml_view_1d(ctx0, conv_states_all,
-                        (d_conv - 1)*(d_inner)*(n_seqs),
-                        kv_head*(d_conv - 1)*(d_inner)*ggml_element_size(conv_states_all))));
+                        (d_conv - 1)*(d_inner + 2*n_group*d_state)*(n_seqs),
+                        kv_head*(d_conv - 1)*(d_inner + 2*n_group*d_state)*ggml_element_size(conv_states_all))));
+            cb(conv_states_all, "mamba_conv1d_state", il);
 
             // 1D convolution
             x = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d);
@@ -16470,9 +16471,9 @@ private:
             // store last states
             ggml_build_forward_expand(gf,
                 ggml_cpy(ctx0,
-                    ggml_view_1d(ctx0, y_ssm, d_state*d_inner*n_seqs, x->nb[3]*x->ne[3]),
-                    ggml_view_1d(ctx0, ssm_states_all, d_state*d_inner*n_seqs,
-                            kv_head*d_state*d_inner*ggml_element_size(ssm_states_all))));
+                    ggml_view_1d(ctx0, y_ssm, n_heads*head_dim*d_state*n_seqs, n_heads*head_dim*n_seq_tokens*n_seqs*ggml_element_size(y_ssm)),
+                    ggml_view_1d(ctx0, ssm_states_all, n_heads*head_dim*d_state*n_seqs, kv_head*n_seqs*n_heads*head_dim*d_state*ggml_element_size(ssm_states_all))));
+            cb(ssm_states_all, "mamba_ssm_states", il);
 
             ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_heads, n_seq_tokens, n_seqs, head_dim * ggml_element_size(x), head_dim * n_heads * ggml_element_size(x), head_dim * n_heads * n_seq_tokens * ggml_element_size(x), 0);
             cb(y, "mamba_y_view", il);