Skip to content

feat: update speculative decoding flags for llama.cpp v1.5+ #956

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
HelloItMeMort wants to merge 2 commits into
base: main
Choose a base branch
from
+51 −12
Open

feat: update speculative decoding flags for llama.cpp v1.5+ #956

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9ed8e5e
feat: upgrade deprecated speculative decoding runtime flags for llama…
Jun 5, 2026
60e199a
feat: restore legacy backward-compatible speculative decoding flags
Jun 5, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions pkg/inference/backends/llamacpp/llamacpp.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -181,10 +181,10 @@ func (l *llamaCpp) Run(ctx context.Context, socket, model string, _ string, mode
if draftPath != "" {
args = append(args, "--model-draft", draftPath)
if config.Speculative.NumTokens > 0 {
args = append(args, "--draft-max", strconv.Itoa(config.Speculative.NumTokens))
args = append(args, "--spec-draft-n-max", strconv.Itoa(config.Speculative.NumTokens))
}
if config.Speculative.MinAcceptanceRate > 0 {
args = append(args, "--draft-p-min", strconv.FormatFloat(config.Speculative.MinAcceptanceRate, 'f', 2, 64))
args = append(args, "--spec-draft-p-min", strconv.FormatFloat(config.Speculative.MinAcceptanceRate, 'f', 2, 64))
}
}
}
Expand Down
48 changes: 40 additions & 8 deletions pkg/inference/runtime_flags_allowlist.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -114,14 +114,46 @@
"--spm-infill": true,

// Speculative decoding (safe flags only - no file paths)
"--draft": true, "--draft-n": true, "--draft-max": true,
"--draft-min": true, "--draft-n-min": true,
"--draft-p-min": true,
"-cd": true, "--ctx-size-draft": true,
"-devd": true, "--device-draft": true,
"-ngld": true, "--gpu-layers-draft": true, "--n-gpu-layers-draft": true,
"-td": true, "--threads-draft": true,
"-tbd": true, "--threads-batch-draft": true,
// Flags prefixed with --spec-draft-* are the canonical llama.cpp v1.5+ names.
// Short aliases (e.g. -td) and legacy unprefixed variants (e.g. --threads-draft)
// are kept for backward compatibility with older llama.cpp versions.
"--spec-draft-threads": true, "-td": true, "--threads-draft": true,
"--spec-draft-threads-batch": true, "-tbd": true, "--threads-batch-draft": true,
"--spec-draft-cpu-mask": true, "-Cd": true, "--cpu-mask-draft": true,
"--spec-draft-cpu-range": true, "-Crd": true, "--cpu-range-draft": true,
"--spec-draft-cpu-strict": true, "--cpu-strict-draft": true,
"--spec-draft-prio": true, "--prio-draft": true,
"--spec-draft-poll": true, "--poll-draft": true,
"--spec-draft-cpu-mask-batch": true, "-Cbd": true, "--cpu-mask-batch-draft": true,
"--spec-draft-cpu-strict-batch": true, "--cpu-strict-batch-draft": true,
"--spec-draft-prio-batch": true, "--prio-batch-draft": true,
"--spec-draft-poll-batch": true, "--poll-batch-draft": true,
"--spec-draft-override-tensor": true, "-otd": true, "--override-tensor-draft": true,
"--spec-draft-cpu-moe": true, "-cmoed": true, "--cpu-moe-draft": true,
"--spec-draft-n-cpu-moe": true, "--spec-draft-ncmoe": true, "-ncmoed": true, "--n-cpu-moe-draft": true,
"--spec-draft-n-max": true, "--draft-n-max": true, "--draft-max": true,
"--spec-draft-n-min": true, "--draft-n-min": true, "--draft-min": true,
"--spec-draft-p-split": true, "--draft-p-split": true,
"--spec-draft-p-min": true, "--draft-p-min": true,
"--spec-draft-backend-sampling": true, "--no-spec-draft-backend-sampling": true,
"--spec-draft-device": true, "-devd": true, "--device-draft": true,
"--spec-draft-ngl": true, "-ngld": true, "--gpu-layers-draft": true, "--n-gpu-layers-draft": true,
"--spec-type": true,

Check failure on line 141 in pkg/inference/runtime_flags_allowlist.go

View workflow job for this annotation

GitHub Actions / lint (windows) 

File is not properly formatted (gci)
"--spec-ngram-mod-n-min": true,
"--spec-ngram-mod-n-max": true,
"--spec-ngram-mod-n-match": true,
"--spec-ngram-simple-size-n": true,
"--spec-ngram-simple-size-m": true,
"--spec-ngram-simple-min-hits": true,
"--spec-ngram-map-k-size-n": true,
"--spec-ngram-map-k-size-m": true,
"--spec-ngram-map-k-min-hits": true,
"--spec-ngram-map-k4v-size-n": true,
"--spec-ngram-map-k4v-size-m": true,
"--spec-ngram-map-k4v-min-hits": true,
"--spec-draft-type-k": true, "-ctkd": true, "--cache-type-k-draft": true,
"--spec-draft-type-v": true, "-ctvd": true, "--cache-type-v-draft": true,
"-cd": true, "--ctx-size-draft": true,

// LoRA (safe flags only - no file paths)
"--lora-init-without-apply": true,
Expand Down
11 changes: 9 additions & 2 deletions pkg/inference/runtime_flags_allowlist_test.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -162,9 +162,16 @@ func TestLlamaCppAllowedFlags_Categories(t *testing.T) {
"--metrics", "--no-metrics", "--jinja", "--no-jinja",
},
"speculative": {
"--draft", "--draft-max", "--draft-min",
"--spec-draft-n-max", "--draft-n-max", "--draft-max", "--spec-draft-n-min", "--draft-n-min", "--draft-min",
"--spec-draft-p-min", "--draft-p-min",
"--spec-draft-p-split", "--draft-p-split",
"--spec-draft-threads", "-td", "--threads-draft",
"--spec-draft-ngl", "-ngld", "--gpu-layers-draft", "--n-gpu-layers-draft",
"--spec-draft-device", "-devd", "--device-draft",
"--spec-type",
"-cd", "--ctx-size-draft",
"-ngld", "--gpu-layers-draft",
"--spec-draft-type-k", "-ctkd", "--cache-type-k-draft",
"--spec-draft-type-v", "-ctvd", "--cache-type-v-draft",
},
}

Expand Down
Loading