Spaces:
Running
Running
File size: 16,762 Bytes
5f923cd | 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 | // Copyright 2025 The ODML Authors.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may
// may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "runtime/conversation/internal_callback_util.h"
#include <algorithm>
#include <cstddef>
#include <string>
#include <utility>
#include <vector>
#include "absl/functional/any_invocable.h" // from @com_google_absl
#include "absl/status/status.h" // from @com_google_absl
#include "absl/status/statusor.h" // from @com_google_absl
#include "absl/strings/string_view.h" // from @com_google_absl
#include "nlohmann/json_fwd.hpp" // from @nlohmann_json
#include "runtime/conversation/channel_util.h"
#include "runtime/conversation/io_types.h"
#include "runtime/conversation/model_data_processor/config_registry.h"
#include "runtime/conversation/model_data_processor/model_data_processor.h"
#include "runtime/engine/io_types.h"
namespace litert::lm {
namespace {
// Returns the number of overlapping characters between the suffix of string
// `a` and the prefix of string `b`.
size_t SuffixPrefixOverlap(absl::string_view a, absl::string_view b) {
if (a.empty() || b.empty()) {
return 0;
}
size_t max_overlap = std::min(a.length(), b.length());
for (size_t len = max_overlap; len > 0; --len) {
if (a.substr(a.length() - len) == b.substr(0, len)) {
return len;
}
}
return 0;
};
// Sends standard text and tool calls to the user callback. The text and/or tool
// calls are first wrapped in a Message object via
// `model_data_processor.ToMessage` before being sent.
void SendMessage(
absl::AnyInvocable<void(absl::StatusOr<Message>)>& user_callback,
absl::string_view text, const ModelDataProcessor& model_data_processor,
DataProcessorArguments processor_args) {
if (text.empty()) {
return;
}
auto message = model_data_processor.ToMessage(
Responses(TaskState::kProcessing, {std::string(text)}), processor_args);
if (!message.ok()) {
user_callback(message.status());
return;
}
user_callback(std::move(message.value()));
}
// Sends streamed text associated with a specific channel. It wraps the text in
// a JsonMessage under the provided `target_channel_name` with a role of
// "assistant" and bypasses `model_data_processor.ToMessage` formatting.
void SendMessageToChannel(
absl::AnyInvocable<void(absl::StatusOr<Message>)>& user_callback,
absl::string_view text, absl::string_view channel_name) {
if (text.empty()) {
return;
}
JsonMessage json_msg;
json_msg["role"] = "assistant";
json_msg["channels"] = nlohmann::ordered_json::object();
json_msg["channels"][std::string(channel_name)] = std::string(text);
user_callback(Message(json_msg));
}
// Sends remaining un-flushed text at the end of generation and then invokes
// `complete_message_callback` on the final message.
void SendCompleteMessage(
absl::AnyInvocable<void(absl::StatusOr<Message>)>& user_callback,
absl::string_view accumulated_response_text,
const ModelDataProcessor& model_data_processor,
DataProcessorArguments processor_args, int cursor,
absl::AnyInvocable<void(Message)>& complete_message_callback,
const std::string& active_channel_name,
const std::vector<Channel>& channels) {
// Send remaining un-flushed text at the end of generation.
if (cursor < accumulated_response_text.size()) {
if (!active_channel_name.empty()) {
SendMessageToChannel(user_callback,
accumulated_response_text.substr(cursor),
active_channel_name);
} else {
SendMessage(user_callback, accumulated_response_text.substr(cursor),
model_data_processor, processor_args);
}
}
// Wrap the accumulated response text in a `Responses` object.
Responses responses(TaskState::kProcessing,
{std::string(accumulated_response_text)});
// Extract channel content from the responses. Modifies responses in place.
auto extracted_channels = ExtractChannelContent(channels, responses);
if (!extracted_channels.ok()) {
user_callback(extracted_channels.status());
return;
}
auto complete_message =
model_data_processor.ToMessage(responses, processor_args);
if (!complete_message.ok()) {
user_callback(complete_message.status());
return;
}
InsertChannelContentIntoMessage(*extracted_channels, *complete_message);
if (complete_message_callback) {
complete_message_callback(*complete_message);
}
user_callback(Message(JsonMessage()));
}
// Returns the complete list of channels the parser should search for, including
// any tool call code blocks (treated as a special channel with no target
// message field) and custom channels passed in by the user config.
std::vector<Channel> GetChannels(const ModelDataProcessor& model_data_processor,
const std::vector<Channel>& custom_channels) {
std::vector<Channel> channels;
// Add the tool call channel if the code fence start is not empty.
if (!model_data_processor.CodeFenceStart().empty()) {
channels.push_back({"", std::string(model_data_processor.CodeFenceStart()),
std::string(model_data_processor.CodeFenceEnd())});
}
// Add the custom channels.
for (const auto& channel : custom_channels) {
if (!channel.start.empty()) {
channels.push_back({channel.channel_name, channel.start, channel.end});
}
}
return channels;
}
// Searches the provided text string starting at `cursor` for the earliest
// matching channel start delimiter out of the possible channels provided.
//
// Returns a pointer to the matching channel and mutates `best_start_pos` to
// store its index in the accumulated response text.
//
// Returns nullptr if no channel start delimiter is found.
const Channel* FindNextChannelStart(
const std::vector<Channel>& possible_channels, absl::string_view text,
size_t cursor, size_t& best_start_pos) {
best_start_pos = std::string::npos;
const Channel* best_match = nullptr;
for (const auto& channel : possible_channels) {
size_t start_pos = text.find(channel.start, cursor);
if (start_pos != std::string::npos) {
if (best_start_pos == std::string::npos || start_pos < best_start_pos) {
best_start_pos = start_pos;
best_match = &channel;
}
}
}
return best_match;
}
// Checks if the end of the un-parsed string might potentially be the first part
// of any channel start delimiter. Returns the maximum character substring
// overlap length between the end of the response string and the start of any
// active channels.
size_t FindMaxOverlap(const std::vector<Channel>& channels,
absl::string_view text) {
size_t max_overlap = 0;
for (const auto& channel : channels) {
size_t overlap = SuffixPrefixOverlap(text, channel.start);
if (overlap > max_overlap) {
max_overlap = overlap;
}
}
return max_overlap;
}
// Streams out channel tokens. Only streams text that it safely validates could
// not possibly be a partial overlap of the active channel end delimiter.
void StreamActiveChannel(
absl::AnyInvocable<void(absl::StatusOr<Message>)>& user_callback,
absl::string_view accumulated_response_text, size_t search_start,
size_t& cursor, absl::string_view active_channel_end,
const std::string& active_channel_name) {
// Stream channel content except for potential partial matches of
// the end delimiter.
size_t overlap = SuffixPrefixOverlap(
accumulated_response_text.substr(search_start), active_channel_end);
size_t safe_end = accumulated_response_text.size() - overlap;
if (safe_end > cursor) {
SendMessageToChannel(
user_callback,
accumulated_response_text.substr(cursor, safe_end - cursor),
active_channel_name);
cursor = safe_end;
}
}
} // namespace
// Creates an internal callback that parses the model's raw text responses.
//
// This parser supports "channels" defined by start and end delimiters.
// 1. Tool Calls: These act as a special channel (defined by
// CodeFenceStart/End).
// Because their content needs to be parsed as a whole object (e.g. JSON),
// tool calls do not stream. The parser buffers the text and waits for the
// end delimiter before using `model_data_processor.ToMessage` to emit them.
// This is represented internally by an empty `active_message_field`.
// `custom_channels`. If a custom channel specifies a
// `channel_name`, its text is streamed out immediately as new JSON messages
// with that specific field as it arrives instead of being buffered.
absl::AnyInvocable<void(absl::StatusOr<Responses>)> CreateInternalCallback(
const ModelDataProcessor& model_data_processor,
const DataProcessorArguments processor_args,
const std::vector<Channel>& custom_channels,
absl::AnyInvocable<void(absl::StatusOr<Message>)> user_callback,
absl::AnyInvocable<void()> cancel_callback,
absl::AnyInvocable<void(Message)> complete_message_callback) {
return [&model_data_processor, processor_args, custom_channels,
user_callback = std::move(user_callback),
cancel_callback = std::move(cancel_callback),
complete_message_callback = std::move(complete_message_callback),
accumulated_response_text = std::string(), cursor = size_t(0),
channels = GetChannels(model_data_processor, custom_channels),
inside_channel = false, active_channel_end = std::string(),
active_channel_start_pos = size_t(0),
active_channel_start_size = size_t(0),
active_channel_name =
std::string()](absl::StatusOr<Responses> responses) mutable {
if (!responses.ok()) {
// If the error is due to cancellation, then we should trigger the cancel
// callback for removing the last message from the history.
if (cancel_callback && absl::IsCancelled(responses.status())) {
cancel_callback();
}
user_callback(responses.status());
return;
}
// If there are no more new responses, it means the model has finished
// generating content, trigger the complete message callback and return an
// OK status to indicate the inference is done.
if (responses->GetTaskState() == TaskState::kDone ||
responses->GetTaskState() == TaskState::kMaxNumTokensReached) {
SendCompleteMessage(user_callback, accumulated_response_text,
model_data_processor, processor_args, cursor,
complete_message_callback,
inside_channel ? active_channel_name : "", channels);
cursor = accumulated_response_text.size();
return;
}
// Else, add the new response text to the accumulated text and process the
// response text.(Which sends to the user callback accordingly.)
if (responses->GetTaskState() == TaskState::kProcessing) {
// If there are no new responses, it is just a state update and we can
// return early.
if (responses->GetTexts().empty()) {
return;
}
// Append the new response text to the accumulated text.
accumulated_response_text += responses->GetTexts()[0];
// Loop through the accumulated response text and send to the user
// callback accordingly.
while (cursor < accumulated_response_text.size()) {
if (!inside_channel) {
size_t channel_start_pos;
const Channel* next_channel = FindNextChannelStart(
channels, accumulated_response_text, cursor, channel_start_pos);
if (next_channel != nullptr) {
// A channel start delimiter was found.
// The text from the cursor up to the channel start is normal text
// and can be sent to the user callback.
SendMessage(user_callback,
absl::string_view(accumulated_response_text)
.substr(cursor, channel_start_pos - cursor),
model_data_processor, processor_args);
// Move cursor up to channel start.
cursor = channel_start_pos;
inside_channel = true;
active_channel_end = next_channel->end;
active_channel_start_pos = channel_start_pos;
active_channel_start_size = next_channel->start.size();
active_channel_name = next_channel->channel_name;
// For custom channels, move the cursor past the start delimiter so
// that it is not included in the resulting streamed content.
// For tool calls (empty message field), we leave the cursor alone
// to buffer the complete block including delimiters.
if (!active_channel_name.empty()) {
cursor += active_channel_start_size;
}
} else {
// A channel start delimiter was not found. We still need to check
// if there's a partial match of any channel start at the very end
// of the string.
size_t max_overlap = FindMaxOverlap(
channels,
absl::string_view(accumulated_response_text).substr(cursor));
if (max_overlap > 0) {
// There's a partial match of a channel at the end of the
// string.
size_t possible_start_pos =
accumulated_response_text.size() - max_overlap;
// Call the callback with text up to the potential start of the
// channel.
SendMessage(user_callback,
accumulated_response_text.substr(
cursor, possible_start_pos - cursor),
model_data_processor, processor_args);
// Move cursor up to potential start of channel.
cursor = possible_start_pos;
// Break for the next token.
break;
} else {
// Remaining string is text.
SendMessage(user_callback,
accumulated_response_text.substr(cursor),
model_data_processor, processor_args);
cursor = accumulated_response_text.size();
}
}
}
if (inside_channel) {
// Look for channel end.
size_t search_start =
std::max(static_cast<size_t>(cursor),
active_channel_start_pos + active_channel_start_size);
size_t end_pos =
accumulated_response_text.find(active_channel_end, search_start);
if (end_pos != std::string::npos) {
// A channel end delimiter was found.
if (!active_channel_name.empty()) {
// Flush the active stream channel.
SendMessageToChannel(user_callback,
absl::string_view(accumulated_response_text)
.substr(cursor, end_pos - cursor),
active_channel_name);
} else {
// Treat as tool call: include everything up to and including the
// end delimiter.
SendMessage(
user_callback,
accumulated_response_text.substr(
cursor, end_pos + active_channel_end.size() - cursor),
model_data_processor, processor_args);
}
// Move cursor past the end of the channel block.
cursor = end_pos + active_channel_end.size();
inside_channel = false;
} else {
// We're inside a channel or tool call but the end has not been
// found.
if (!active_channel_name.empty()) {
// If we're inside a channel, stream the text, but stop before
// any potential partial match of the channel's end delimiter.
StreamActiveChannel(user_callback, accumulated_response_text,
search_start, cursor, active_channel_end,
active_channel_name);
}
// Break for the next token.
break;
}
}
}
}
};
}
} // namespace litert::lm
|