cui-llama.rn 1.0.2 → 1.0.3

package/cpp/ggml-quants.h

@@ -1,132 +1,132 @@
-#pragma once
-
-#define LM_GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
-#include "ggml.h"
-
-// GGML internal header
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// 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_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_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);
-
-// 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_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_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);
-
-// 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_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);
-
-#ifdef __cplusplus
-}
-#endif
+#pragma once
+
+#define LM_GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// 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_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_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);
+
+// 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_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_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);
+
+// 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_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);
+
+#ifdef __cplusplus
+}
+#endif

package/cpp/ggml.c

@@ -3341,7 +3341,7 @@ bool lm_ggml_are_same_stride(const struct lm_ggml_tensor * t0, const struct lm_g
 }
 
 // check if t1 can be represented as a repeatition of t0
-static inline bool lm_ggml_can_repeat(const struct lm_ggml_tensor * t0, const struct lm_ggml_tensor * t1) {
+bool lm_ggml_can_repeat(const struct lm_ggml_tensor * t0, const struct lm_ggml_tensor * t1) {
     static_assert(LM_GGML_MAX_DIMS == 4, "LM_GGML_MAX_DIMS is not 4 - update this function");
 
     return lm_ggml_is_empty(t0) ? lm_ggml_is_empty(t1) :
@@ -13699,6 +13699,7 @@ static void lm_ggml_compute_forward_soft_max(
     }
 }
 
+
 // lm_ggml_compute_forward_soft_max_back
 
 static void lm_ggml_compute_forward_soft_max_back_f32(
@@ -19018,7 +19019,7 @@ void lm_ggml_graph_export(const struct lm_ggml_cgraph * cgraph, const char * fna
         FILE * fout = lm_ggml_fopen(fname, "wb");
 
         if (!fout) {
-            fprintf(stderr, "%s: failed to open %s\n", __func__, fname);
+            fprintf(stderr, "%s: failed to open %s: %s\n", __func__, fname, strerror(errno));
             return;
         }
 
@@ -19155,7 +19156,7 @@ struct lm_ggml_cgraph * lm_ggml_graph_import(const char * fname, struct lm_ggml_
     {
         FILE * fin = lm_ggml_fopen(fname, "rb");
         if (!fin) {
-            fprintf(stderr, "%s: failed to open %s\n", __func__, fname);
+            fprintf(stderr, "%s: failed to open %s: %s\n", __func__, fname, strerror(errno));
             return result;
         }
 
@@ -19478,7 +19479,7 @@ void lm_ggml_graph_dump_dot(const struct lm_ggml_cgraph * gb, const struct lm_gg
 
     fprintf(fp, "digraph G {\n");
     fprintf(fp, "  newrank = true;\n");
-    fprintf(fp, "  rankdir = LR;\n");
+    fprintf(fp, "  rankdir = TB;\n");
 
     for (int i = 0; i < gb->n_nodes; i++) {
         struct lm_ggml_tensor * node = gb->nodes[i];
@@ -19540,7 +19541,7 @@ void lm_ggml_graph_dump_dot(const struct lm_ggml_cgraph * gb, const struct lm_gg
         }
 
         fprintf(fp, "CONST %d [%" PRId64 ", %" PRId64 "]", i, node->ne[0], node->ne[1]);
-        if (lm_ggml_nelements(node) < 5) {
+        if (lm_ggml_nelements(node) < 5 && node->data != NULL) {
             fprintf(fp, " | (");
             for (int j = 0; j < lm_ggml_nelements(node); j++) {
                 if (node->type == LM_GGML_TYPE_I8 || node->type == LM_GGML_TYPE_I16 || node->type == LM_GGML_TYPE_I32) {
@@ -20829,6 +20830,7 @@ struct lm_gguf_context * lm_gguf_init_empty(void) {
 struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gguf_init_params params) {
     FILE * file = lm_ggml_fopen(fname, "rb");
     if (!file) {
+        fprintf(stderr, "%s: failed to open '%s': '%s'\n", __func__, fname, strerror(errno));
         return NULL;
     }
 
@@ -21013,7 +21015,7 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
             lm_gguf_tensor_info_sanitize(info);
 
             // make sure there is no duplicated tensor names
-            for (uint64_t j = 0; j < i; ++j) {
+            for (uint64_t j = 0; j < i && ok; ++j) {
                 if (strcmp(info->name.data, ctx->infos[j].name.data) == 0) {
                     fprintf(stderr, "%s: duplicated tensor name %s\n", __func__, info->name.data);
                     ok = false;
@@ -21995,6 +21997,14 @@ int lm_ggml_cpu_has_rpc(void) {
 #endif
 }
 
+int lm_ggml_cpu_has_cann(void) {
+#if defined(LM_GGML_USE_CANN)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int lm_ggml_cpu_has_gpublas(void) {
     return lm_ggml_cpu_has_cuda() || lm_ggml_cpu_has_vulkan() || lm_ggml_cpu_has_kompute() || lm_ggml_cpu_has_sycl();
 }