whisper.rn 0.5.0-rc.9 → 0.5.0

package/cpp/ggml-cpu/ops.h

@@ -29,6 +29,7 @@ extern "C" {
 
 void wsp_ggml_compute_forward_dup(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_add(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
+void wsp_ggml_compute_forward_add_id(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_add1(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_acc(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_sum(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
@@ -82,13 +83,7 @@ void wsp_ggml_compute_forward_arange(const struct wsp_ggml_compute_params * para
 void wsp_ggml_compute_forward_timestep_embedding(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_argsort(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_leaky_relu(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
-void wsp_ggml_compute_forward_flash_attn_ext(
-    const struct wsp_ggml_compute_params * params,
-    const struct wsp_ggml_tensor * q,
-    const struct wsp_ggml_tensor * k,
-    const struct wsp_ggml_tensor * v,
-    const struct wsp_ggml_tensor * mask,
-    struct wsp_ggml_tensor * dst);
+void wsp_ggml_compute_forward_flash_attn_ext(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_flash_attn_back(
         const struct wsp_ggml_compute_params * params,
         const bool masked,
@@ -112,7 +107,7 @@ void wsp_ggml_compute_forward_cross_entropy_loss(const struct wsp_ggml_compute_p
 void wsp_ggml_compute_forward_cross_entropy_loss_back(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_opt_step_adamw(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 void wsp_ggml_compute_forward_mul_mat(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
-
+void wsp_ggml_compute_forward_opt_step_sgd(const struct wsp_ggml_compute_params * params, struct wsp_ggml_tensor * dst);
 #ifdef __cplusplus
 }
 #endif

package/cpp/ggml-cpu/quants.c

@@ -46,6 +46,10 @@ void wsp_quantize_row_q8_1_generic(const float * WSP_GGML_RESTRICT x, void * WSP
     wsp_quantize_row_q8_1_ref(x, y, k);
 }
 
+void wsp_quantize_row_mxfp4(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k) {
+    wsp_quantize_row_mxfp4_ref(x, y, k);
+}
+
 //
 // 2-6 bit quantization in super-blocks
 //
@@ -181,6 +185,37 @@ void wsp_ggml_vec_dot_q4_1_q8_1_generic(int n, float * WSP_GGML_RESTRICT s, size
     *s = sumf;
 }
 
+void wsp_ggml_vec_dot_mxfp4_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_MXFP4 == 0);
+    static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same");
+
+    const block_mxfp4 * WSP_GGML_RESTRICT x = vx;
+    const block_q8_0 * WSP_GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_MXFP4;
+
+    int ib = 0;
+    float sumf = 0;
+
+    for (; ib < nb; ++ib) {
+        const float d = WSP_GGML_CPU_FP16_TO_FP32(y[ib].d)*WSP_GGML_E8M0_TO_FP32_HALF(x[ib].e);
+
+        int sumi1 = 0;
+        int sumi2 = 0;
+        for (int j = 0; j < QK_MXFP4/2; ++j) {
+            sumi1 += y[ib].qs[j +          0] * kvalues_mxfp4[x[ib].qs[j] & 0xf];
+            sumi2 += y[ib].qs[j + QK_MXFP4/2] * kvalues_mxfp4[x[ib].qs[j] >>  4];
+        }
+        sumf += d * (sumi1 + sumi2);
+    }
+    *s = sumf;
+}
+
 void wsp_ggml_vec_dot_q5_0_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;

package/cpp/ggml-cpu/quants.h

@@ -19,6 +19,8 @@ void wsp_quantize_row_q5_1(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RE
 void wsp_quantize_row_q8_0(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
 void wsp_quantize_row_q8_1(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
 
+void wsp_quantize_row_mxfp4(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
+
 void wsp_quantize_row_q2_K(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
 void wsp_quantize_row_q3_K(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
 void wsp_quantize_row_q4_K(const float * WSP_GGML_RESTRICT x, void * WSP_GGML_RESTRICT y, int64_t k);
@@ -39,6 +41,8 @@ void wsp_ggml_vec_dot_q5_0_q8_0(int n, float * WSP_GGML_RESTRICT s, size_t bs, c
 void wsp_ggml_vec_dot_q5_1_q8_1(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
 void wsp_ggml_vec_dot_q8_0_q8_0(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
 
+void wsp_ggml_vec_dot_mxfp4_q8_0(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+
 void wsp_ggml_vec_dot_q2_K_q8_K(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
 void wsp_ggml_vec_dot_q3_K_q8_K(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
 void wsp_ggml_vec_dot_q4_K_q8_K(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
@@ -67,8 +71,12 @@ void wsp_ggml_vec_dot_q4_1_q8_1_generic(int n, float * WSP_GGML_RESTRICT s, size
 void wsp_ggml_vec_dot_q5_0_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
 void wsp_ggml_vec_dot_q5_1_q8_1_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
 void wsp_ggml_vec_dot_q8_0_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+
+void wsp_ggml_vec_dot_mxfp4_q8_0_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+
 void wsp_ggml_vec_dot_tq1_0_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
 void wsp_ggml_vec_dot_tq2_0_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
+
 void wsp_ggml_vec_dot_q2_K_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
 void wsp_ggml_vec_dot_q3_K_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);
 void wsp_ggml_vec_dot_q4_K_q8_K_generic(int n, float * WSP_GGML_RESTRICT s, size_t bs, const void * WSP_GGML_RESTRICT vx, size_t bx, const void * WSP_GGML_RESTRICT vy, size_t by, int nrc);