cui-llama.rn 1.2.6 → 1.3.3

package/cpp/ggml-quants.h

@@ -11,136 +11,89 @@
 extern "C" {
 #endif
 
+// NOTE: these functions are defined as LM_GGML_API because they used by the CPU backend
+
 // Quantization
-void quantize_row_q4_0_ref(const float * LM_GGML_RESTRICT x, block_q4_0 * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1_ref(const float * LM_GGML_RESTRICT x, block_q4_1 * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0_ref(const float * LM_GGML_RESTRICT x, block_q5_0 * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1_ref(const float * LM_GGML_RESTRICT x, block_q5_1 * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0_ref(const float * LM_GGML_RESTRICT x, block_q8_0 * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1_ref(const float * LM_GGML_RESTRICT x, block_q8_1 * LM_GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q2_K_ref(const float * LM_GGML_RESTRICT x, block_q2_K * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K_ref(const float * LM_GGML_RESTRICT x, block_q3_K * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K_ref(const float * LM_GGML_RESTRICT x, block_q4_K * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K_ref(const float * LM_GGML_RESTRICT x, block_q5_K * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K_ref(const float * LM_GGML_RESTRICT x, block_q6_K * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K_ref(const float * LM_GGML_RESTRICT x, block_q8_K * LM_GGML_RESTRICT y, int64_t k);
-
-void quantize_row_tq1_0_ref(const float * LM_GGML_RESTRICT x, block_tq1_0 * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_tq2_0_ref(const float * LM_GGML_RESTRICT x, block_tq2_0 * LM_GGML_RESTRICT y, int64_t k);
-
-void quantize_row_iq3_xxs_ref(const float * LM_GGML_RESTRICT x, block_iq3_xxs * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_nl_ref (const float * LM_GGML_RESTRICT x, block_iq4_nl  * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs_ref (const float * LM_GGML_RESTRICT x, block_iq4_xs  * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_iq3_s_ref  (const float * LM_GGML_RESTRICT x, block_iq3_s   * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_iq2_s_ref  (const float * LM_GGML_RESTRICT x, block_iq2_s   * LM_GGML_RESTRICT y, int64_t k);
-
-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_iq3_xxs(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);
-void quantize_row_iq3_s  (const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
-void quantize_row_iq2_s  (const float * LM_GGML_RESTRICT x, void * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_q4_0_ref(const float * LM_GGML_RESTRICT x, block_q4_0 * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_q4_1_ref(const float * LM_GGML_RESTRICT x, block_q4_1 * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_q5_0_ref(const float * LM_GGML_RESTRICT x, block_q5_0 * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_q5_1_ref(const float * LM_GGML_RESTRICT x, block_q5_1 * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_q8_0_ref(const float * LM_GGML_RESTRICT x, block_q8_0 * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_q8_1_ref(const float * LM_GGML_RESTRICT x, block_q8_1 * LM_GGML_RESTRICT y, int64_t k);
+
+LM_GGML_API void quantize_row_q2_K_ref(const float * LM_GGML_RESTRICT x, block_q2_K * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_q3_K_ref(const float * LM_GGML_RESTRICT x, block_q3_K * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_q4_K_ref(const float * LM_GGML_RESTRICT x, block_q4_K * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_q5_K_ref(const float * LM_GGML_RESTRICT x, block_q5_K * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_q6_K_ref(const float * LM_GGML_RESTRICT x, block_q6_K * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_q8_K_ref(const float * LM_GGML_RESTRICT x, block_q8_K * LM_GGML_RESTRICT y, int64_t k);
+
+LM_GGML_API void quantize_row_tq1_0_ref(const float * LM_GGML_RESTRICT x, block_tq1_0 * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_tq2_0_ref(const float * LM_GGML_RESTRICT x, block_tq2_0 * LM_GGML_RESTRICT y, int64_t k);
+
+LM_GGML_API void quantize_row_iq3_xxs_ref(const float * LM_GGML_RESTRICT x, block_iq3_xxs * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_iq4_nl_ref (const float * LM_GGML_RESTRICT x, block_iq4_nl  * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_iq4_xs_ref (const float * LM_GGML_RESTRICT x, block_iq4_xs  * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_iq3_s_ref  (const float * LM_GGML_RESTRICT x, block_iq3_s   * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void quantize_row_iq2_s_ref  (const float * LM_GGML_RESTRICT x, block_iq2_s   * LM_GGML_RESTRICT y, int64_t k);
 
 // Dequantization
-void dequantize_row_q4_0(const block_q4_0 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_q4_1(const block_q4_1 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_0(const block_q5_0 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_1(const block_q5_1 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_q8_0(const block_q8_0 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-//void dequantize_row_q8_1(const block_q8_1 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-
-void dequantize_row_q2_K(const block_q2_K * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_q3_K(const block_q3_K * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_q4_K(const block_q4_K * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_K(const block_q5_K * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_q6_K(const block_q6_K * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_q8_K(const block_q8_K * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-
-void dequantize_row_tq1_0(const block_tq1_0 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_tq2_0(const block_tq2_0 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-
-void dequantize_row_iq2_xxs(const block_iq2_xxs * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq2_xs (const block_iq2_xs  * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq2_s  (const block_iq2_s   * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq3_xxs(const block_iq3_xxs * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq1_s  (const block_iq1_s   * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq1_m  (const block_iq1_m   * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq4_nl (const block_iq4_nl  * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq4_xs (const block_iq4_xs  * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq3_s  (const block_iq3_s   * LM_GGML_RESTRICT x, float * 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);
+LM_GGML_API void dequantize_row_q4_0(const block_q4_0 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_q4_1(const block_q4_1 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_q5_0(const block_q5_0 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_q5_1(const block_q5_1 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_q8_0(const block_q8_0 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+//LM_GGML_API void dequantize_row_q8_1(const block_q8_1 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+
+LM_GGML_API void dequantize_row_q2_K(const block_q2_K * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_q3_K(const block_q3_K * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_q4_K(const block_q4_K * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_q5_K(const block_q5_K * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_q6_K(const block_q6_K * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_q8_K(const block_q8_K * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+
+LM_GGML_API void dequantize_row_tq1_0(const block_tq1_0 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_tq2_0(const block_tq2_0 * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+
+LM_GGML_API void dequantize_row_iq2_xxs(const block_iq2_xxs * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_iq2_xs (const block_iq2_xs  * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_iq2_s  (const block_iq2_s   * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_iq3_xxs(const block_iq3_xxs * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_iq1_s  (const block_iq1_s   * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_iq1_m  (const block_iq1_m   * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_iq4_nl (const block_iq4_nl  * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_iq4_xs (const block_iq4_xs  * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
+LM_GGML_API void dequantize_row_iq3_s  (const block_iq3_s   * LM_GGML_RESTRICT x, float * LM_GGML_RESTRICT y, int64_t k);
 
 // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
-size_t quantize_iq2_xxs(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq2_xs (const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq2_s  (const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq3_xxs(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq1_s  (const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq1_m  (const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq4_nl (const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq4_xs (const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq3_s  (const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-size_t quantize_tq1_0(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_tq2_0(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-size_t quantize_q2_K(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q3_K(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_K(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_K(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q6_K(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_0(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_1(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_0(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_1(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q8_0(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-void iq2xs_init_impl(enum lm_ggml_type type);
-void iq2xs_free_impl(enum lm_ggml_type type);
-void iq3xs_init_impl(int grid_size);
-void iq3xs_free_impl(int grid_size);
+LM_GGML_API size_t quantize_iq2_xxs(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_iq2_xs (const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_iq2_s  (const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_iq3_xxs(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_iq1_s  (const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_iq1_m  (const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_iq4_nl (const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_iq4_xs (const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_iq3_s  (const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+LM_GGML_API size_t quantize_tq1_0(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_tq2_0(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+LM_GGML_API size_t quantize_q2_K(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_q3_K(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_q4_K(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_q5_K(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_q6_K(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_q4_0(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_q4_1(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_q5_0(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_q5_1(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+LM_GGML_API size_t quantize_q8_0(const float * LM_GGML_RESTRICT src, void * LM_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+LM_GGML_API void lm_iq2xs_init_impl(enum lm_ggml_type type);
+LM_GGML_API void lm_iq2xs_free_impl(enum lm_ggml_type type);
+LM_GGML_API void lm_iq3xs_init_impl(int grid_size);
+LM_GGML_API void lm_iq3xs_free_impl(int grid_size);
 
 #ifdef __cplusplus
 }

package/cpp/ggml-threading.cpp

@@ -0,0 +1,12 @@
+#include "ggml-threading.h"
+#include <mutex>
+
+std::mutex lm_ggml_critical_section_mutex;
+
+void lm_ggml_critical_section_start() {
+    lm_ggml_critical_section_mutex.lock();
+}
+
+void lm_ggml_critical_section_end(void) {
+    lm_ggml_critical_section_mutex.unlock();
+}

package/cpp/ggml-threading.h

@@ -0,0 +1,12 @@
+#pragma once
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void lm_ggml_critical_section_start(void);
+void lm_ggml_critical_section_end(void);
+
+#ifdef __cplusplus
+}
+#endif