forked from ggml-org/llama.cpp
-
Notifications
You must be signed in to change notification settings - Fork 2
Expand file tree
/
Copy pathserver-chat.cpp
More file actions
630 lines (566 loc) · 25.9 KB
/
server-chat.cpp
File metadata and controls
630 lines (566 loc) · 25.9 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
#include "server-chat.h"
#include "server-common.h"
#include <sstream>
json server_chat_convert_responses_to_chatcmpl(const json & response_body) {
if (!response_body.contains("input")) {
throw std::invalid_argument("'input' is required");
}
if (!json_value(response_body, "previous_response_id", std::string{}).empty()) {
throw std::invalid_argument("llama.cpp does not support 'previous_response_id'.");
}
const json input_value = response_body.at("input");
json chatcmpl_body = response_body;
chatcmpl_body.erase("input");
std::vector<json> chatcmpl_messages;
if (response_body.contains("instructions")) {
chatcmpl_messages.push_back({
{"role", "system"},
{"content", json_value(response_body, "instructions", std::string())},
});
chatcmpl_body.erase("instructions");
}
if (input_value.is_string()) {
// #responses_create-input-text_input
chatcmpl_messages.push_back({
{"role", "user"},
{"content", input_value},
});
} else if (input_value.is_array()) {
// #responses_create-input-input_item_list
static auto exists_and_is_array = [](const json & j, const char * key) -> bool {
return j.contains(key) && j.at(key).is_array();
};
static auto exists_and_is_string = [](const json & j, const char * key) -> bool {
return j.contains(key) && j.at(key).is_string();
};
for (json item : input_value) {
bool merge_prev = !chatcmpl_messages.empty() && chatcmpl_messages.back().value("role", "") == "assistant";
if (exists_and_is_string(item, "content")) {
// #responses_create-input-input_item_list-input_message-content-text_input
// Only "Input message" contains item["content"]::string
// After converting item["content"]::string to item["content"]::array,
// we can treat "Input message" as sum of "Item-Input message" and "Item-Output message"
item["content"] = json::array({
json {
{"text", item.at("content")},
{"type", "input_text"}
}
});
}
if (exists_and_is_array(item, "content") &&
exists_and_is_string(item, "role") &&
(item.at("role") == "user" ||
item.at("role") == "system" ||
item.at("role") == "developer")
) {
// #responses_create-input-input_item_list-item-input_message
std::vector<json> chatcmpl_content;
for (const json & input_item : item.at("content")) {
const std::string type = json_value(input_item, "type", std::string());
if (type == "input_text") {
if (!input_item.contains("text")) {
throw std::invalid_argument("'Input text' requires 'text'");
}
chatcmpl_content.push_back({
{"text", input_item.at("text")},
{"type", "text"},
});
} else if (type == "input_image") {
// While `detail` is marked as required,
// it has default value("auto") and can be omitted.
if (!input_item.contains("image_url")) {
throw std::invalid_argument("'image_url' is required");
}
chatcmpl_content.push_back({
{"image_url", json {
{"url", input_item.at("image_url")}
}},
{"type", "image_url"},
});
} else if (type == "input_file") {
throw std::invalid_argument("'input_file' is not supported by llamacpp at this moment");
} else {
throw std::invalid_argument("'type' must be one of 'input_text', 'input_image', or 'input_file'");
}
}
if (item.contains("type")) {
item.erase("type");
}
if (item.contains("status")) {
item.erase("status");
}
item["content"] = chatcmpl_content;
chatcmpl_messages.push_back(item);
} else if (exists_and_is_string(item, "role") &&
item.at("role") == "assistant" &&
exists_and_is_string(item, "type") &&
item.at("type") == "message"
) {
// #responses_create-input-input_item_list-item-output_message
auto chatcmpl_content = json::array();
// Handle both string content and array content
if (item.contains("content") && item.at("content").is_string()) {
// String content - convert to text content part
chatcmpl_content.push_back({
{"text", item.at("content")},
{"type", "text"},
});
} else if (exists_and_is_array(item, "content")) {
// Array content - process each item
for (const auto & output_text : item.at("content")) {
const std::string type = json_value(output_text, "type", std::string());
if (type == "output_text" || type == "input_text") {
// Accept both output_text and input_text (string content gets converted to input_text)
if (!exists_and_is_string(output_text, "text")) {
throw std::invalid_argument("'Output text' requires 'text'");
}
chatcmpl_content.push_back({
{"text", output_text.at("text")},
{"type", "text"},
});
} else if (type == "refusal") {
if (!exists_and_is_string(output_text, "refusal")) {
throw std::invalid_argument("'Refusal' requires 'refusal'");
}
chatcmpl_content.push_back({
{"refusal", output_text.at("refusal")},
{"type", "refusal"},
});
} else {
throw std::invalid_argument("'type' must be one of 'output_text' or 'refusal'");
}
}
}
if (merge_prev) {
auto & prev_msg = chatcmpl_messages.back();
if (!exists_and_is_array(prev_msg, "content")) {
prev_msg["content"] = json::array();
}
auto & prev_content = prev_msg["content"];
prev_content.insert(prev_content.end(), chatcmpl_content.begin(), chatcmpl_content.end());
} else {
item.erase("status");
item.erase("type");
item["content"] = chatcmpl_content;
chatcmpl_messages.push_back(item);
}
} else if (exists_and_is_string(item, "arguments") &&
exists_and_is_string(item, "call_id") &&
exists_and_is_string(item, "name") &&
exists_and_is_string(item, "type") &&
item.at("type") == "function_call"
) {
// #responses_create-input-input_item_list-item-function_tool_call
json tool_call = {
{"function", json {
{"arguments", item.at("arguments")},
{"name", item.at("name")},
}},
{"id", item.at("call_id")},
{"type", "function"},
};
if (merge_prev) {
auto & prev_msg = chatcmpl_messages.back();
if (!exists_and_is_array(prev_msg, "tool_calls")) {
prev_msg["tool_calls"] = json::array();
}
prev_msg["tool_calls"].push_back(tool_call);
} else {
chatcmpl_messages.push_back(json {
{"role", "assistant"},
{"tool_calls", json::array({tool_call})}
});
}
} else if (exists_and_is_string(item, "call_id") &&
(exists_and_is_string(item, "output") || exists_and_is_array(item, "output")) &&
exists_and_is_string(item, "type") &&
item.at("type") == "function_call_output"
) {
// #responses_create-input-input_item_list-item-function_tool_call_output
if (item.at("output").is_string()) {
chatcmpl_messages.push_back(json {
{"content", item.at("output")},
{"role", "tool"},
{"tool_call_id", item.at("call_id")},
});
} else {
json chatcmpl_outputs = item.at("output");
for (json & chatcmpl_output : chatcmpl_outputs) {
if (!chatcmpl_output.contains("type") || chatcmpl_output.at("type") != "input_text") {
throw std::invalid_argument("Output of tool call should be 'Input text'");
}
chatcmpl_output["type"] = "text";
}
chatcmpl_messages.push_back(json {
{"content", chatcmpl_outputs},
{"role", "tool"},
{"tool_call_id", item.at("call_id")},
});
}
} else if (exists_and_is_array(item, "summary") &&
exists_and_is_string(item, "type") &&
item.at("type") == "reasoning") {
// #responses_create-input-input_item_list-item-reasoning
if (!exists_and_is_array(item, "content")) {
throw std::invalid_argument("item['content'] is not an array");
}
if (item.at("content").empty()) {
throw std::invalid_argument("item['content'] is empty");
}
if (!exists_and_is_string(item.at("content")[0], "text")) {
throw std::invalid_argument("item['content']['text'] is not a string");
}
if (merge_prev) {
auto & prev_msg = chatcmpl_messages.back();
prev_msg["reasoning_content"] = item.at("content")[0].at("text");
} else {
chatcmpl_messages.push_back(json {
{"role", "assistant"},
{"content", json::array()},
{"reasoning_content", item.at("content")[0].at("text")},
});
}
} else {
throw std::invalid_argument("Cannot determine type of 'item'");
}
}
} else {
throw std::invalid_argument("'input' must be a string or array of objects");
}
chatcmpl_body["messages"] = chatcmpl_messages;
if (response_body.contains("tools")) {
if (!response_body.at("tools").is_array()) {
throw std::invalid_argument("'tools' must be an array of objects");
}
std::vector<json> chatcmpl_tools;
for (json resp_tool : response_body.at("tools")) {
json chatcmpl_tool;
if (json_value(resp_tool, "type", std::string()) != "function") {
throw std::invalid_argument("'type' of tool must be 'function'");
}
resp_tool.erase("type");
chatcmpl_tool["type"] = "function";
if (!resp_tool.contains("strict")) {
resp_tool["strict"] = true;
}
chatcmpl_tool["function"] = resp_tool;
chatcmpl_tools.push_back(chatcmpl_tool);
}
chatcmpl_body.erase("tools");
chatcmpl_body["tools"] = chatcmpl_tools;
}
if (response_body.contains("max_output_tokens")) {
chatcmpl_body.erase("max_output_tokens");
chatcmpl_body["max_tokens"] = response_body["max_output_tokens"];
}
return chatcmpl_body;
}
// Edits the cch section of an "x-anthropic-billing-header" system prompt.
// Does nothing to any other prompt.
//
// This is a claude message with a "cch=ef01a" attribute that breaks prefix caching.
// The cch stamp is a whitebox end-to-end integrity hint. It's not meaningful as a
// system prompt data, particularly to llama.cpp, but its presence means the prefix
// cache will not get past it: It changes on each request.
//
// Reference: https://github.com/ggml-org/llama.cpp/pull/21793
// Example header:
// ```
// x-anthropic-billing-header: cc_version=2.1.101.e51; cc_entrypoint=cli; cch=a5145;You are Claude Code, Anthropic's official CLI for Claude.
// ^^^^^
// ```
static void normalize_anthropic_billing_header(std::string & system_text) {
if (system_text.rfind("x-anthropic-billing-header:", 0) != 0) {
return;
}
const size_t header_prefix_length = strlen("x-anthropic-billing-header:");
const size_t cch_length = 5;
const size_t index_cch = system_text.find("cch=", header_prefix_length);
if (index_cch == std::string::npos) {
return;
}
const size_t index_replace = index_cch + 4;
if (index_replace + cch_length < system_text.length() && system_text[index_replace + cch_length] == ';') {
for (size_t i = 0; i < cch_length; ++i) {
system_text[index_replace + i] = 'f';
}
} else {
LOG_ERR("anthropic string not as expected: %s", system_text.c_str());
}
}
json server_chat_convert_anthropic_to_oai(const json & body) {
json oai_body;
// Convert system prompt
json oai_messages = json::array();
auto system_param = json_value(body, "system", json());
if (!system_param.is_null()) {
std::string system_content;
if (system_param.is_string()) {
system_content = system_param.get<std::string>();
normalize_anthropic_billing_header(system_content);
} else if (system_param.is_array()) {
for (const auto & block : system_param) {
if (json_value(block, "type", std::string()) == "text") {
auto system_text = json_value(block, "text", std::string());
normalize_anthropic_billing_header(system_text);
system_content += system_text;
}
}
}
oai_messages.push_back({
{"role", "system"},
{"content", system_content}
});
}
// Convert messages
if (!body.contains("messages")) {
throw std::runtime_error("'messages' is required");
}
const json & messages = body.at("messages");
if (messages.is_array()) {
for (const auto & msg : messages) {
std::string role = json_value(msg, "role", std::string());
if (!msg.contains("content")) {
if (role == "assistant") {
continue;
}
oai_messages.push_back(msg);
continue;
}
const json & content = msg.at("content");
if (content.is_string()) {
oai_messages.push_back(msg);
continue;
}
if (!content.is_array()) {
oai_messages.push_back(msg);
continue;
}
json tool_calls = json::array();
json converted_content = json::array();
json tool_results = json::array();
std::string reasoning_content;
bool has_tool_calls = false;
for (const auto & block : content) {
std::string type = json_value(block, "type", std::string());
if (type == "text") {
converted_content.push_back(block);
} else if (type == "thinking") {
reasoning_content += json_value(block, "thinking", std::string());
} else if (type == "image") {
json source = json_value(block, "source", json::object());
std::string source_type = json_value(source, "type", std::string());
if (source_type == "base64") {
std::string media_type = json_value(source, "media_type", std::string("image/jpeg"));
std::string data = json_value(source, "data", std::string());
std::ostringstream ss;
ss << "data:" << media_type << ";base64," << data;
converted_content.push_back({
{"type", "image_url"},
{"image_url", {
{"url", ss.str()}
}}
});
} else if (source_type == "url") {
std::string url = json_value(source, "url", std::string());
converted_content.push_back({
{"type", "image_url"},
{"image_url", {
{"url", url}
}}
});
}
} else if (type == "tool_use") {
tool_calls.push_back({
{"id", json_value(block, "id", std::string())},
{"type", "function"},
{"function", {
{"name", json_value(block, "name", std::string())},
{"arguments", json_value(block, "input", json::object()).dump()}
}}
});
has_tool_calls = true;
} else if (type == "tool_result") {
std::string tool_use_id = json_value(block, "tool_use_id", std::string());
auto result_content = json_value(block, "content", json());
std::string result_text;
if (result_content.is_string()) {
result_text = result_content.get<std::string>();
} else if (result_content.is_array()) {
for (const auto & c : result_content) {
if (json_value(c, "type", std::string()) == "text") {
result_text += json_value(c, "text", std::string());
}
}
}
tool_results.push_back({
{"role", "tool"},
{"tool_call_id", tool_use_id},
{"content", result_text}
});
}
}
if (!converted_content.empty() || has_tool_calls || !reasoning_content.empty()) {
json new_msg = {{"role", role}};
if (!converted_content.empty()) {
new_msg["content"] = converted_content;
} else if (has_tool_calls || !reasoning_content.empty()) {
new_msg["content"] = "";
}
if (!tool_calls.empty()) {
new_msg["tool_calls"] = tool_calls;
}
if (!reasoning_content.empty()) {
new_msg["reasoning_content"] = reasoning_content;
}
oai_messages.push_back(new_msg);
}
for (const auto & tool_msg : tool_results) {
oai_messages.push_back(tool_msg);
}
}
}
oai_body["messages"] = oai_messages;
// Convert tools
if (body.contains("tools")) {
const json & tools = body.at("tools");
if (tools.is_array()) {
json oai_tools = json::array();
for (const auto & tool : tools) {
oai_tools.push_back({
{"type", "function"},
{"function", {
{"name", json_value(tool, "name", std::string())},
{"description", json_value(tool, "description", std::string())},
{"parameters", tool.contains("input_schema") ? tool.at("input_schema") : json::object()}
}}
});
}
oai_body["tools"] = oai_tools;
}
}
// Convert tool_choice
if (body.contains("tool_choice")) {
const json & tc = body.at("tool_choice");
if (tc.is_object()) {
std::string type = json_value(tc, "type", std::string());
if (type == "auto") {
oai_body["tool_choice"] = "auto";
} else if (type == "any" || type == "tool") {
oai_body["tool_choice"] = "required";
}
}
}
// Convert stop_sequences to stop
if (body.contains("stop_sequences")) {
oai_body["stop"] = body.at("stop_sequences");
}
// Handle max_tokens (required in Anthropic, but we're permissive)
if (body.contains("max_tokens")) {
oai_body["max_tokens"] = body.at("max_tokens");
} else {
oai_body["max_tokens"] = 4096;
}
// Pass through common params
for (const auto & key : {"temperature", "top_p", "top_k", "stream", "chat_template_kwargs"}) {
if (body.contains(key)) {
oai_body[key] = body.at(key);
}
}
// Handle Anthropic-specific thinking param
if (body.contains("thinking")) {
json thinking = json_value(body, "thinking", json::object());
std::string thinking_type = json_value(thinking, "type", std::string());
if (thinking_type == "enabled") {
int budget_tokens = json_value(thinking, "budget_tokens", 10000);
oai_body["thinking_budget_tokens"] = budget_tokens;
}
}
// Handle Anthropic-specific metadata param
if (body.contains("metadata")) {
json metadata = json_value(body, "metadata", json::object());
std::string user_id = json_value(metadata, "user_id", std::string());
if (!user_id.empty()) {
oai_body["__metadata_user_id"] = user_id;
}
}
return oai_body;
}
json server_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff) {
json delta = json::object();
if (!diff.reasoning_content_delta.empty()) {
delta["reasoning_content"] = diff.reasoning_content_delta;
}
if (!diff.content_delta.empty()) {
delta["content"] = diff.content_delta;
}
if (diff.tool_call_index != std::string::npos) {
json tool_call;
tool_call["index"] = diff.tool_call_index;
if (!diff.tool_call_delta.id.empty()) {
tool_call["id"] = diff.tool_call_delta.id;
tool_call["type"] = "function";
}
if (!diff.tool_call_delta.name.empty() || !diff.tool_call_delta.arguments.empty()) {
json function = json::object();
if (!diff.tool_call_delta.name.empty()) {
function["name"] = diff.tool_call_delta.name;
}
if (!diff.tool_call_delta.arguments.empty()) {
function["arguments"] = diff.tool_call_delta.arguments;
}
tool_call["function"] = function;
}
delta["tool_calls"] = json::array({ tool_call });
}
return delta;
}
json convert_transcriptions_to_chatcmpl(
const json & inp_body,
const common_chat_templates * tmpls,
const std::map<std::string, raw_buffer> & in_files,
std::vector<raw_buffer> & out_files) {
// TODO @ngxson : this function may need to be improved in the future
// handle input files
out_files.clear();
auto it = in_files.find("file");
if (it != in_files.end()) {
out_files.push_back(it->second);
} else {
throw std::invalid_argument("No input file found for transcription");
}
// handle input data
std::string prompt = json_value(inp_body, "prompt", std::string());
std::string language = json_value(inp_body, "language", std::string());
std::string response_format = json_value(inp_body, "response_format", std::string("json"));
if (response_format != "json") {
throw std::invalid_argument("Only 'json' response_format is supported for transcription");
}
const common_chat_prompt_preset preset = common_chat_get_asr_prompt(tmpls);
if (prompt.empty()) {
prompt = preset.user;
}
if (!language.empty()) {
prompt += string_format(" (language: %s)", language.c_str());
}
prompt += get_media_marker();
json messages = json::array();
if (!preset.system.empty()) {
messages.push_back({{"role", "system"}, {"content", preset.system}});
}
messages.push_back({{"role", "user"}, {"content", prompt}});
json chatcmpl_body = inp_body; // copy all fields
chatcmpl_body["messages"] = messages;
// because input from form-data, everything is string, we need to correct the types here
std::string stream = json_value(inp_body, "stream", std::string("false"));
chatcmpl_body["stream"] = stream == "true";
if (inp_body.contains("max_tokens")) {
std::string inp = inp_body["max_tokens"].get<std::string>();
chatcmpl_body["max_tokens"] = std::stoul(inp);
}
if (inp_body.contains("temperature")) {
std::string inp = inp_body["temperature"].get<std::string>();
chatcmpl_body["temperature"] = std::stof(inp);
}
return chatcmpl_body;
}