char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Martin Changggml-et: Initial ET backend (#24179)082b326fc

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
2.0 KiB54 linesraw
1#pragma once
2
3#include "ggml-cpu.h"
4#include "ggml-et-common.h"
5#include "ggml-impl.h"
6
7// Configuration for CPU comparison
8struct ggml_et_cpu_compare_config {
9    bool   enabled;           // Whether to enable CPU comparison
10    bool   use_cpu_result;    // Whether to replace ET result with CPU result
11    bool   log_differences;   // Whether to log detailed element differences
12    float  tolerance;         // Relative tolerance for comparison (default: 1e-5f)
13    size_t max_log_elements;  // Maximum number of elements to log (default: 10)
14};
15
16// Default configuration
17static const ggml_et_cpu_compare_config ggml_et_cpu_compare_default_config = {
18    /* .enabled = */ false,
19    /* .use_cpu_result = */ false,
20    /* .log_differences = */ true,
21    /* .tolerance = */ 1e-5f,
22    /* .max_log_elements = */ 10
23};
24
25// CPU comparison context for a single operation
26struct ggml_et_cpu_compare_ctx {
27    ggml_backend_t cpu_backend;
28    ggml_context * ggml_ctx;
29    ggml_tensor *  cpu_src0;
30    ggml_tensor *  cpu_src1;
31    ggml_tensor *  cpu_src2;
32    ggml_tensor *  cpu_dst;
33    ggml_cgraph *  cpu_graph;
34    void *         cpu_src0_data;
35    void *         cpu_src1_data;
36    void *         cpu_src2_data;
37    void *         cpu_dst_data;
38    void *         et_dst_data;
39    size_t         src0_size;
40    size_t         src1_size;
41    size_t         src2_size;
42    size_t         dst_size;
43};
44
45// Phase 1: Initialize CPU comparison context and copy source buffers (call before ET kernel)
46bool ggml_et_cpu_compare_init_pre(ggml_et_cpu_compare_ctx * ctx, const ggml_tensor * node, ggml_op op);
47
48// Phase 2: Execute CPU computation and compare with ET result (call after ET kernel)
49bool ggml_et_cpu_compare_compute_and_check(ggml_et_cpu_compare_ctx *          ctx,
50                                           const ggml_tensor *                node,
51                                           const ggml_et_cpu_compare_config * config);
52
53// Free CPU comparison context resources
54void ggml_et_cpu_compare_free(ggml_et_cpu_compare_ctx * ctx);