tmp

llama/patches/0016-add-ollama-debug.patch

@@ -0,0 +1,33 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Fri, 7 Feb 2025 16:51:55 -0800
+Subject: [PATCH] add ollama debug
+
+---
+ ggml/src/ggml-cpu/ggml-cpu.c | 6 ++++++
+ 1 file changed, 6 insertions(+)
+
+diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
+index b307d554..a23bbe98 100644
+--- a/ggml/src/ggml-cpu/ggml-cpu.c
++++ b/ggml/src/ggml-cpu/ggml-cpu.c
+@@ -13,6 +13,8 @@
+ #include "amx/amx.h"
+ #include "ggml.h"
+ 
++#include "ollama-debug.h"
++
+ #if defined(_MSC_VER) || defined(__MINGW32__)
+ #include <malloc.h> // using malloc.h with MSC/MINGW
+ #elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
+@@ -13645,6 +13647,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
+ 
+         ggml_compute_forward(&params, node);
+ 
++#ifdef OLLAMA_DEBUG
++        ollama_debug(node, false);
++#endif
++
+         if (state->ith == 0 && cplan->abort_callback &&
+                 cplan->abort_callback(cplan->abort_callback_data)) {
+             tp->abort = true;

ml/backend/ggml/ggml.go

@@ -99,15 +99,17 @@ func New(r *os.File) (ml.Backend, error) {
 				}),
 				backend: C.ggml_backend_dev_init(d.d, nil),
 			})
-		case C.GGML_BACKEND_DEVICE_TYPE_GPU:
+
-			slog.Info("gpu", "device", d)
+			C.ggml_backend_cpu_set_n_threads(cpus[len(cpus)-1].backend, C.int(1))
-			gpus = append(gpus, Context{
+		// case C.GGML_BACKEND_DEVICE_TYPE_GPU:
-				ctx: C.ggml_init(C.struct_ggml_init_params{
+		// 	slog.Info("gpu", "device", d)
-					mem_size: C.size_t(int(C.ggml_tensor_overhead()) * (len(meta.Tensors().Items()) + 1 + int(meta.KV().BlockCount())*2)),
+		// 	gpus = append(gpus, Context{
-					no_alloc: true,
+		// 		ctx: C.ggml_init(C.struct_ggml_init_params{
-				}),
+		// 			mem_size: C.size_t(int(C.ggml_tensor_overhead()) * (len(meta.Tensors().Items()) + 1 + int(meta.KV().BlockCount())*2)),
-				backend: C.ggml_backend_dev_init(d.d, nil),
+		// 			no_alloc: true,
-			})
+		// 		}),
+		// 		backend: C.ggml_backend_dev_init(d.d, nil),
+		// 	})
 		}
 	}
 

ml/backend/ggml/ggml/.rsync-filter

@@ -1,5 +1,6 @@
 protect *.go
 protect *-embed.*
+protect ollama-debug.*
 include include/
 include src/
 include src/CMakeLists.txt

ml/backend/ggml/ggml/include/ollama-debug.h

@@ -0,0 +1,11 @@
+#include "ggml.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void ollama_debug(const struct ggml_tensor *tensor, bool verbose);
+
+#ifdef __cplusplus
+}
+#endif

ml/backend/ggml/ggml/src/debug.go

@@ -0,0 +1,6 @@
+//go:build debug
+
+package ggml
+
+// #cgo CPPFLAGS: -DOLLAMA_DEBUG
+import "C"

ml/backend/ggml/ggml/src/ggml-cpu/ggml-cpu.c

@@ -13,6 +13,8 @@
 #include "amx/amx.h"
 #include "ggml.h"
 
+#include "ollama-debug.h"
+
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <malloc.h> // using malloc.h with MSC/MINGW
 #elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
@@ -13645,6 +13647,8 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
 
         ggml_compute_forward(&params, node);
 
+        ollama_debug(node, true);
+
         if (state->ith == 0 && cplan->abort_callback &&
                 cplan->abort_callback(cplan->abort_callback_data)) {
             tp->abort = true;

ml/backend/ggml/ggml/src/ollama-debug.c

@@ -0,0 +1,110 @@
+#include <string.h>
+
+#include "ollama-debug.h"
+
+static int mul(int64_t *dims, int ndims) {
+    int result = 1;
+    for (int i = 0; i < ndims; i++) {
+        result *= dims[i];
+    }
+
+    return result;
+}
+
+static void repeat(char c, int n) {
+    for (int i = 0; i < n; i++) {
+        fprintf(stderr, "%c", c);
+    }
+}
+
+static void print_tensor(const void *tensor, void (*cb)(const void *, int),
+                         int shape,
+                         int64_t *dims, int ndims, int stride,
+                         int nitems, int pad) {
+    fprintf(stderr, "[");
+    for (int i = 0; i < dims[0]; i++) {
+        if (i >= nitems && i < dims[0] - nitems) {
+            fprintf(stderr, "... (%lld more), ", dims[0] - 2 * nitems);
+            int skip = dims[0] - 2 * nitems;
+            if (ndims > 1) {
+                stride += mul(dims + 1, ndims - 1) * skip;
+                repeat('\n', ndims - 1);
+                repeat(' ', shape - ndims + 1 + pad);
+            }
+            i += skip - 1;
+        } else if (ndims > 1) {
+            print_tensor(tensor, cb, shape, dims + 1, ndims - 1, stride,
+                         nitems, pad);
+            stride += mul(dims + 1, ndims - 1);
+            if (i < dims[0] - 1) {
+                fprintf(stderr, ", ");
+                repeat('\n', ndims - 1);
+                repeat(' ', shape - ndims + 1 + pad);
+            }
+        } else {
+            cb(tensor, stride + i);
+            if (i < dims[0] - 1) {
+                fprintf(stderr, ", ");
+            }
+        }
+    }
+    fprintf(stderr, "]");
+}
+
+static void print_tensor_f16(const void *tensor, int i) {
+    fprintf(stderr, "%f", ggml_fp16_to_fp32(((const ggml_fp16_t *)tensor)[i]));
+}
+
+static void print_tensor_f32(const void *tensor, int i) {
+    fprintf(stderr, "%f", ((const float *)tensor)[i]);
+}
+
+static void print_tensor_i32(const void *tensor, int i) {
+    fprintf(stderr, "%d", ((const int32_t *)tensor)[i]);
+}
+
+static void ollama_debug_tensor(const struct ggml_tensor *tensor, bool verbose, const char *prefix, int indent) {
+    fprintf(stderr, "%s%s %s (%s): [%lld %lld %lld %lld]\n", prefix, tensor->name,
+            ggml_op_name(tensor->op), ggml_type_name(tensor->type), tensor->ne[0],
+            tensor->ne[1], tensor->ne[2], tensor->ne[3]);
+
+    if (!verbose) {
+        return;
+    }
+
+    for (int i = 0; i < indent; i++) {
+        fprintf(stderr, " ");
+    }
+
+    switch (tensor->type) {
+    case GGML_TYPE_F16:
+        print_tensor(ggml_get_data(tensor), print_tensor_f16, ggml_n_dims(tensor),
+                     (int64_t *)tensor->ne, ggml_n_dims(tensor), 0, 3, indent);
+        break;
+    case GGML_TYPE_F32:
+        print_tensor(ggml_get_data(tensor), print_tensor_f32, ggml_n_dims(tensor),
+                     (int64_t *)tensor->ne, ggml_n_dims(tensor), 0, 3, indent);
+        break;
+    case GGML_TYPE_I32:
+        print_tensor(ggml_get_data(tensor), print_tensor_i32, ggml_n_dims(tensor),
+                     (int64_t *)tensor->ne, ggml_n_dims(tensor), 0, 3, indent);
+        break;
+    default:
+        fprintf(stderr, "<unsupported type>\n");
+        return;
+    }
+
+    fprintf(stderr, "\n");
+}
+
+void ollama_debug(const struct ggml_tensor *tensor, bool verbose) {
+    ollama_debug_tensor(tensor, verbose, ">>> ", 4);
+
+    if (tensor->src[0] != NULL) {
+        ollama_debug_tensor(tensor->src[0], verbose, " ?? ", 4);
+    }
+
+    if (tensor->src[1] != NULL) {
+        ollama_debug_tensor(tensor->src[1], verbose, " ?? ", 4);
+    }
+}