cui-llama.rn 1.3.0 → 1.3.4

package/cpp/ggml-cpu-quants.h

@@ -1,63 +1,63 @@
-#pragma once
-
-#define LM_GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
-#include "ggml.h"
-
-// GGML CPU internal header
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// Quantization
-void quantize_row_q4_0(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q2_K(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-
-void quantize_row_tq1_0(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_tq2_0(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-
-void quantize_row_iq4_nl (const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs (const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-
-// Dot product
-void lm_ggml_vec_dot_q4_0_q8_0(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_q4_1_q8_1(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_q5_0_q8_0(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_q5_1_q8_1(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_q8_0_q8_0(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-
-void lm_ggml_vec_dot_q2_K_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_q3_K_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_q4_K_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_q5_K_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_q6_K_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-
-void lm_ggml_vec_dot_tq1_0_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_tq2_0_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-
-void lm_ggml_vec_dot_iq2_xxs_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_iq2_xs_q8_K (int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_iq2_s_q8_K  (int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_iq3_xxs_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_iq1_s_q8_K  (int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_iq1_m_q8_K  (int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_iq4_nl_q8_0 (int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_iq4_xs_q8_K (int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-void lm_ggml_vec_dot_iq3_s_q8_K  (int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
-
-#ifdef __cplusplus
-}
-#endif
+#pragma once
+
+#define LM_GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML CPU internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_row_q4_0(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_1(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q2_K(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+void quantize_row_q3_K(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+
+void quantize_row_tq1_0(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+void quantize_row_tq2_0(const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+
+void quantize_row_iq4_nl (const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs (const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+
+// Dot product
+void lm_ggml_vec_dot_q4_0_q8_0(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_q4_1_q8_1(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_q5_0_q8_0(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_q5_1_q8_1(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_q8_0_q8_0(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+
+void lm_ggml_vec_dot_q2_K_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_q3_K_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_q4_K_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_q5_K_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_q6_K_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+
+void lm_ggml_vec_dot_tq1_0_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_tq2_0_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+
+void lm_ggml_vec_dot_iq2_xxs_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_iq2_xs_q8_K (int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_iq2_s_q8_K  (int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_iq3_xxs_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_iq1_s_q8_K  (int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_iq1_m_q8_K  (int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_iq4_nl_q8_0 (int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_iq4_xs_q8_K (int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+void lm_ggml_vec_dot_iq3_s_q8_K  (int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc);
+
+#ifdef __cplusplus
+}
+#endif

package/cpp/ggml-cpu-traits.cpp

@@ -0,0 +1,36 @@
+#include "ggml-cpu-traits.h"
+
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+
+namespace ggml::cpu {
+tensor_traits::~tensor_traits() {}
+
+extra_buffer_type::~extra_buffer_type() {}
+}  // namespace ggml::cpu
+
+bool lm_ggml_cpu_extra_compute_forward(struct lm_ggml_compute_params * params, struct lm_ggml_tensor * op) {
+    for (auto extra : lm_ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra && extra->context) {
+            auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
+            auto tensor_traits = buf_extra->get_tensor_traits(op);
+            if (tensor_traits && tensor_traits->compute_forward(params, op)) {
+                return true;
+            }
+        }
+    }
+    return false;
+}
+
+bool lm_ggml_cpu_extra_work_size(int n_threads, const struct lm_ggml_tensor * op, size_t * size) {
+    for (auto extra : lm_ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra && extra->context) {
+            auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
+            auto tensor_traits = buf_extra->get_tensor_traits(op);
+            if (tensor_traits && tensor_traits->work_size(n_threads, op, *size)) {
+                return true;
+            }
+        }
+    }
+    return false;
+}

package/cpp/ggml-cpu-traits.h

@@ -0,0 +1,38 @@
+#pragma once
+#include "ggml-backend-impl.h"
+#include "ggml-cpu-impl.h"
+#include "ggml.h"
+
+#ifdef __cplusplus
+#    include <vector>
+extern "C" {
+#endif
+
+// return true if op part of extra "accelerator"
+bool lm_ggml_cpu_extra_compute_forward(struct lm_ggml_compute_params * params, struct lm_ggml_tensor * op);
+bool lm_ggml_cpu_extra_work_size(int n_threads, const struct lm_ggml_tensor * op, size_t * size);
+
+#ifdef __cplusplus
+}
+
+namespace ggml::cpu {
+// register in tensor->extra
+class tensor_traits {
+  public:
+    virtual ~tensor_traits();
+    virtual bool work_size(int n_threads, const struct lm_ggml_tensor * op, size_t & size)        = 0;
+    virtual bool compute_forward(struct lm_ggml_compute_params * params, struct lm_ggml_tensor * op) = 0;
+};
+
+class extra_buffer_type {
+  public:
+    virtual ~extra_buffer_type();
+    virtual bool            supports_op(lm_ggml_backend_dev_t dev, const struct lm_ggml_tensor * op) = 0;
+    virtual tensor_traits * get_tensor_traits(const struct lm_ggml_tensor * op)                   = 0;
+};
+}  // namespace ggml::cpu
+
+// implemented in ggml-cpu.cpp.
+std::vector<lm_ggml_backend_buffer_type_t> & lm_ggml_backend_cpu_get_extra_buffers_type();
+
+#endif