get_rows & dequantize function implementation for repacked weights of type q4_K (q4_Kx8)

ggml/src/ggml-cpu/repack.cpp

@@ -1180,6 +1180,11 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
 
                     size += sizeof_mmid_row_mapping*ne02*(ne12 + 1);
 
+                    return true;
+                }
+            case GGML_OP_GET_ROWS:
+                {
+                    size = 0;  // GET_ROWS (standard and repacked) doesn't need a work buffer
                     return true;
                 }
             default:
@@ -1197,6 +1202,9 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             case GGML_OP_MUL_MAT_ID:
                 forward_mul_mat_id(params, op);
                 return true;
+            case GGML_OP_GET_ROWS:
+                forward_get_rows(params, op);
+                return true;
             default:
                 // GGML_ABORT("fatal error");
                 break;
@@ -1405,6 +1413,145 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
 #undef MMID_MATRIX_ROW
     }
 
+    void forward_get_rows(const ggml_compute_params * params,
+                          ggml_tensor * dst) {
+        const ggml_tensor * src0 = dst->src[0];
+
+        switch (src0->type) {
+            case GGML_TYPE_Q4_K:
+                ggml_compute_forward_get_rows_q4_Kx8(params, dst);
+                break;
+            default:
+                GGML_ABORT("fatal error");
+                break;
+        }
+    }
+
+    static void ggml_compute_forward_get_rows_q4_Kx8(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+        const ggml_tensor * src0 = dst->src[0];
+        const ggml_tensor * src1 = dst->src[1];
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+        const int64_t nc = ne00;
+        const int64_t nr = ggml_nelements(src1);
+
+        assert(ne0 == nc);
+        assert(ne02 == ne11);
+        assert(nb00 == ggml_type_size(src0->type));
+        assert(ggml_nrows(dst) == nr);
+
+        const int ith = params->ith;
+        const int nth = params->nth;
+
+        // rows per thread
+        const int dr = (nr + nth - 1) / nth;
+
+        // row range for this thread
+        const int ir0 = dr * ith;
+        const int ir1 = MIN(ir0 + dr, nr);
+
+        constexpr int nrows_interleaved = 8;
+        const size_t sizeof_one_repacked_block = sizeof(block_q4_Kx8);
+
+        const int num_repacked_blocks_per_row_width = nc / QK_K;
+
+        const size_t stride_between_actual_row_groups = num_repacked_blocks_per_row_width * sizeof_one_repacked_block;
+
+        for (int64_t i = ir0; i < ir1; ++i) {
+            const int64_t i12 = i / (ne11 * ne10);
+            const int64_t i11 = (i - i12 * ne11 * ne10) / ne10;
+            const int64_t i10 = (i - i12 * ne11 * ne10 - i11 * ne10);
+            const int64_t i01 = *(int32_t *)((char *)src1->data + i10 * nb10 + i11 * nb11 + i12 * nb12);  // original logical row
+
+            GGML_ASSERT(i01 >= 0 && i01 < ne01);
+
+            const int row_group_idx = i01 / nrows_interleaved;
+            const int row_idx_in_group = i01 % nrows_interleaved;
+
+            const char * base_ptr_for_higher_dims_in_src0 = (const char *)src0->data + i11 * nb02 + i12 * nb03;
+
+            // Pointer to the first block_q4_Kx8 of the identified row_group_idx
+            const block_q4_Kx8 * p_first_repacked_block_of_group_x8 = (const block_q4_Kx8 *)(base_ptr_for_higher_dims_in_src0 + row_group_idx * stride_between_actual_row_groups);
+
+            dequantize_row_q4_Kx8(
+                p_first_repacked_block_of_group_x8,
+                (float *)((char *)dst->data + i10 * nb1 + i11 * nb2 + i12 * nb3), nc, row_idx_in_group);
+        }
+    }
+
+    /**
+     * Dequantizes a single logical row from the repacked q4_Kx8 data format.
+     *
+     * @param p_repacked_blocks Pointer to the start of the 'block_q4_Kx8' structures for the entire row.
+     * @param y                 Output buffer for the dequantized float values.
+     * @param k                 Total number of elements (columns) in the logical row.
+     * @param row_idx_in_group  The index (0-7) of the logical row to extract from the interleaved data.
+     */
+
+    static void dequantize_row_q4_Kx8(
+        const void * GGML_RESTRICT p_repacked_blocks,
+        float * GGML_RESTRICT y,
+        int64_t k,
+        int row_idx_in_group) {
+
+        assert(k % QK_K == 0);
+        assert(row_idx_in_group >= 0 && row_idx_in_group < 8);
+
+        const int nb = k / QK_K;
+        const block_q4_Kx8 * blocks = (const block_q4_Kx8 *)p_repacked_blocks;
+
+        for (int i = 0; i < nb; i++) {
+            const block_q4_Kx8 * current_block = &blocks[i];
+
+            const float d_super_block = GGML_FP16_TO_FP32(current_block->d[row_idx_in_group]);
+            const float dmin_super_block = GGML_FP16_TO_FP32(current_block->dmin[row_idx_in_group]);
+
+            const uint8_t * ptr_qs_base = current_block->qs;
+            const uint8_t * ptr_repacked_scales = (const uint8_t *)current_block->scales;
+            int is = 0, chunk_group_start_idx = 0;
+            for (int j = 0; j < QK_K; j += 64) {
+
+                uint8_t sc1, m1_val, sc2, m2_val;
+                const uint8_t *scales_repacked_data;
+
+                scales_repacked_data = &ptr_repacked_scales[(is + 0) * 12];
+                get_scale_min_k4(row_idx_in_group, scales_repacked_data, &sc1, &m1_val);
+
+                scales_repacked_data = &ptr_repacked_scales[(is + 1) * 12];
+                get_scale_min_k4(row_idx_in_group, scales_repacked_data, &sc2, &m2_val);
+
+                const float d1 = d_super_block * sc1;
+                const float m1 = dmin_super_block * m1_val;
+                const float d2 = d_super_block * sc2;
+                const float m2 = dmin_super_block * m2_val;
+
+                for (int idx = 0; idx < 4; idx++) {
+                    const uint8_t * ptr_qs_chunk = ptr_qs_base + ((chunk_group_start_idx + idx) * 64) + row_idx_in_group * 8;
+                    for (int l = 0; l < 8; ++l) *y++ = d1 * (ptr_qs_chunk[l] & 0xF) - m1; // 16 elements of quants
+                }
+
+                for (int idx = 0; idx < 4; idx++) {
+                    const uint8_t * ptr_qs_chunk = ptr_qs_base + ((chunk_group_start_idx + idx) * 64) + row_idx_in_group * 8;
+                    for (int l = 0; l < 8; ++l) *y++ = d2 * (ptr_qs_chunk[l] >> 4) - m2; // 16 elements of quants
+                }
+                is += 2;
+                chunk_group_start_idx += 4;
+            }
+        }
+    }
+
+    static inline void get_scale_min_k4(int j, const uint8_t *GGML_RESTRICT s, uint8_t *GGML_RESTRICT d, uint8_t *GGML_RESTRICT m) {
+        if (j < 4) {
+            *d = s[j] & 63;
+            *m = s[j + 4] & 63;
+        } else {
+            *d = (s[j + 4] & 0xF) | ((s[j - 4] >> 6) << 4);
+            *m = (s[j + 4] >> 4) | ((s[j - 0] >> 6) << 4);
+        }
+    }
+
     int repack(struct ggml_tensor * t, const void * data, size_t data_size) override {
         GGML_LOG_DEBUG("%s: repack tensor %s with %s_%dx%d\n", __func__, t->name, ggml_type_name(t->type),
                        (int) NB_COLS, (int) INTER_SIZE);
@@ -1538,12 +1685,23 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
             //if (op->src[1]->type == GGML_TYPE_Q8_0) {
             //    return true;
             //}
+        } else if (op->op == GGML_OP_GET_ROWS
+            && op->src[0]->buffer
+            && (ggml_n_dims(op->src[0]) == 2)
+            && op->src[0]->buffer->buft == ggml_backend_cpu_repack_buffer_type()
+            && ggml_repack_get_optimal_repack_type(op->src[0])) {
+            if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
+                return false;
+            }
+            if (op->src[0]->type == GGML_TYPE_Q4_K) {
+                return true;
+            }
         }
         return false;
     }
 
     ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
-        if (op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_MUL_MAT_ID) {
+        if (op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_MUL_MAT_ID || op->op == GGML_OP_GET_ROWS) {
             if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_repack_buffer_type()) {
                 return (ggml::cpu::tensor_traits *) op->src[0]->extra;
             }

src/whisper.cpp

@@ -1437,24 +1437,25 @@ static bool weight_buft_supported(const whisper_hparams & hparams, ggml_tensor *
         // GPU and default CPU backend support all operators
         op_supported = true;
     } else {
-        switch (op) {
-            // The current extra_buffer_type implementations only support GGML_OP_MUL_MAT
-            case GGML_OP_MUL_MAT: {
-                ggml_init_params params = {
-                    /*.mem_size   =*/ 2 * ggml_tensor_overhead(),
-                    /*.mem_buffer =*/ nullptr,
-                    /*.no_alloc   =*/ true,
-                };
+        ggml_init_params params = {
+            /*.mem_size   =*/ 2 * ggml_tensor_overhead(),
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
 
-                ggml_context_ptr ctx_ptr { ggml_init(params) };
-                if (!ctx_ptr) {
-                    throw std::runtime_error("failed to create ggml context");
-                }
-                ggml_context * ctx = ctx_ptr.get();
+        ggml_context_ptr ctx_ptr { ggml_init(params) };
+        if (!ctx_ptr) {
+            throw std::runtime_error("failed to create ggml context");
+        }
+        ggml_context * ctx = ctx_ptr.get();
 
-                ggml_tensor * op_tensor = nullptr;
+        ggml_tensor * op_tensor = nullptr;
+
+        int64_t n_ctx = hparams.n_audio_ctx;
 
-                int64_t n_ctx = hparams.n_audio_ctx;
+        switch (op) {
+            // The current extra_buffer_type implementations only support GGML_OP_MUL_MAT & GGML_OP_GET_ROWS (repacked - q4_K)
+            case GGML_OP_MUL_MAT: {
                 ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], n_ctx, w->ne[2], w->ne[3]);
                 op_tensor = ggml_mul_mat(ctx, w, b);
 
@@ -1466,6 +1467,18 @@ static bool weight_buft_supported(const whisper_hparams & hparams, ggml_tensor *
                 w->buffer = nullptr;
                 break;
             }
+            case GGML_OP_GET_ROWS: {
+                ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_ctx);
+                op_tensor = ggml_get_rows(ctx, w, b);
+
+                // create a temporary dummy buffer for the weight so that supports_op can check the buffer type
+                GGML_ASSERT(w->buffer == nullptr);
+                w->buffer = ggml_backend_buft_alloc_buffer(buft, 0);
+                op_supported = ggml_backend_dev_supports_op(dev, op_tensor);
+                ggml_backend_buffer_free(w->buffer);
+                w->buffer = nullptr;
+                break;
+            }
             default: {
                 op_supported = false;
                 break;