mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 30 Packages Projects Releases Wiki Activity

spec : parallel drafting support (#22838)

Browse Source 
* spec : refactor

* spec : drop support for incompatible vocabs

* spec : update common_speculative_init()

* cont : pass seq_id

* cont : dedup ctx_seq_rm_type

* server : sketch the ctx_dft decode loop

* server : draft prompt cache and checkpoints

* server : improve ctx names

* server, spec : transition to unified spec context

* cont : sync main and drft contexts

* cont : async drft eval when possible

* cont : handle non-ckpt models

* cont : pass correct n_past for drafting

* cont : process images throught the draft context

* spec : handle draft running out of context

* server : fix mtmd draft processing

* server : fix URL for draft model

* server : add comment

* server : clean-up + dry

* speculative-simple : update

* spec : fix n_past type

* server : fix slot ctx_drft ptr

* tools : update readme

* naming : improve consistency

* spec : refactor for multi-sequence speculative context

* cont : prepare params

* cont : prepare params

* spec : support parallel drafts

* server : support parallel drafting

* llama : reuse device buffers when possible

* server, spec : clean-up

* cont : clean-up

* cont : minor

* spec : reset `drafting` flag at the end

* spec : introduce `common_speculative_process()`

* spec : allow for multiple spec types (chain of speculators)

* replace old type field of type common_speculative_type in the
  common_params_speculative struct with a vector to allow multiple
  types to be specified

* introduce common_get_enabled_speculative_impls(const std::vector<enum common_speculative_type>)
  to figure out which implementations the user has enabled

* introduce common_speculative_type_from_names(const std::vector<std::string> & names)
  to parse the already user provided spec types

* all speculators run sequentially, best one wins (we verify its drafted tokens)

* maximize expected accepted tokens for current round by calculating the
  product between the probability of accepting current token (n_acc_tokens / n_gen_drafts)
  and the draft's length

---------

Co-authored-by: Petros Sideris <petros.sideris@nokia.com>
This commit is contained in:
Georgi Gerganov
2026-05-11 19:09:43 +03:00
committed by GitHub
parent 928b486b0c
commit 68e7ea3eab
14 changed files with 1286 additions and 1071 deletions
Download Patch File Download Diff File Expand all files Collapse all files
common/arg.cpp
+5 -36
View File
@@ -622,10 +622,6 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
for (auto & seq_breaker : params.sampling.dry_sequence_breakers) {
string_process_escapes(seq_breaker);
}
for (auto & pair : params.speculative.draft.replacements) {
string_process_escapes(pair.first);
string_process_escapes(pair.second);
}
}
if (!params.kv_overrides.empty()) {
@@ -3518,13 +3514,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.speculative.draft.p_min = std::stof(value);
}
).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_P_MIN"));
add_opt(common_arg(
{"--spec-draft-ctx-size", "-cd", "--ctx-size-draft"}, "N",
string_format("size of the prompt context for the draft model (default: %d, 0 = loaded from model)", params.speculative.draft.n_ctx),
[](common_params & params, int value) {
params.speculative.draft.n_ctx = value;
}
).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_CTX_SIZE"));
add_opt(common_arg(
{"--spec-draft-device", "-devd", "--device-draft"}, "<dev1,dev2,..>",
"comma-separated list of devices to use for offloading the draft model (none = don't offload)\n"
@@ -3561,32 +3550,12 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_MODEL"));
add_opt(common_arg(
{"--spec-draft-replace", "--spec-replace"}, "TARGET", "DRAFT",
"translate the string in TARGET into DRAFT if the draft model and main model are not compatible",
[](common_params & params, const std::string & tgt, const std::string & dft) {
params.speculative.draft.replacements.push_back({ tgt, dft });
}
).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
add_opt(common_arg(
{"--spec-type"}, "[none|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]",
{"--spec-type"}, common_speculative_all_types_str(),
string_format("type of speculative decoding to use when no draft model is provided (default: %s)\n",
common_speculative_type_to_str(params.speculative.type).c_str()),
common_speculative_type_name_str(params.speculative.types).c_str()),
[](common_params & params, const std::string & value) {
if (value == "none") {
params.speculative.type = COMMON_SPECULATIVE_TYPE_NONE;
} else if (value == "ngram-cache") {
params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_CACHE;
} else if (value == "ngram-simple") {
params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE;
} else if (value == "ngram-map-k") {
params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K;
} else if (value == "ngram-map-k4v") {
params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V;
} else if (value == "ngram-mod") {
params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MOD;
} else {
throw std::invalid_argument("unknown speculative decoding type without draft model");
}
const auto enabled_types = string_split<std::string>(value, ',');
params.speculative.types = common_speculative_types_from_names(enabled_types);
}
).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_TYPE"));
add_opt(common_arg(
@@ -4075,7 +4044,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--spec-default"},
string_format("enable default speculative decoding config"),
[](common_params & params) {
params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MOD;
params.speculative.types = { COMMON_SPECULATIVE_TYPE_NGRAM_MOD };
params.speculative.ngram_mod.n_match = 24;
params.speculative.ngram_mod.n_min = 48;
params.speculative.ngram_mod.n_max = 64;