Skip to content

feat: add MTP draft model file support #118

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

Merged
danielcherubini merged 9 commits into from
Jun 13, 2026
Merged

feat: add MTP draft model file support #118

Show file tree
Hide file tree
Changes from 8 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
4d3b401
docs: add MTP draft model file support implementation plan
danielcherubini Jun 13, 2026
a6c5f54
feat: add QuantKind::Mtp and ModelConfig.mtp_model with DB migration
danielcherubini Jun 13, 2026
917b8aa
feat: add MTP filename support to pull request handler and download s…
danielcherubini Jun 13, 2026
4381ec5
feat: inject --mtp-model flag gated on draft-mtp spec type
danielcherubini Jun 13, 2026
e654674
feat: add MTP support to pull wizards and web types
danielcherubini Jun 13, 2026
cb88a3c
feat: add MTP draft model dropdown to Spec Decoding form
danielcherubini Jun 13, 2026
e34d3f0
docs: mark MTP draft model file support plan as completed
danielcherubini Jun 13, 2026
f4a902f
fix: rename --mtp-model to --spec-draft-model (correct llama.cpp flag)
danielcherubini Jun 13, 2026
a06a1e3
fix: reset mmproj/mtp selections on new repo search
danielcherubini Jun 13, 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
1 change: 1 addition & 0 deletions crates/tama-cli/src/commands/model/create.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,6 +65,7 @@ pub(super) async fn cmd_create(
model: Some(model_id_arg.to_string()),
quant: quant_name.clone(),
mmproj: None,
mtp_model: None,
port: None,
health_check: None,
enabled: true,
Expand Down
6 changes: 6 additions & 0 deletions crates/tama-cli/src/commands/model/pull.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -350,6 +350,7 @@ pub(crate) fn cmd_scan(config: &Config) -> Result<()> {
enabled: true,
selected_quant: None,
selected_mmproj: None,
selected_mtp_model: None,
context_length: None,
gpu_layers: None,
port: None,
Expand Down Expand Up @@ -510,6 +511,7 @@ mod tests {
enabled: true,
selected_quant: None,
selected_mmproj: None,
selected_mtp_model: None,
context_length: None,
gpu_layers: None,
port: None,
Expand Down Expand Up @@ -573,6 +575,7 @@ mod tests {
enabled: true,
selected_quant: None,
selected_mmproj: None,
selected_mtp_model: None,
context_length: None,
gpu_layers: None,
port: None,
Expand Down Expand Up @@ -628,6 +631,7 @@ mod tests {
enabled: true,
selected_quant: None,
selected_mmproj: None,
selected_mtp_model: None,
context_length: None,
gpu_layers: None,
port: None,
Expand Down Expand Up @@ -675,6 +679,7 @@ mod tests {
enabled: true,
selected_quant: None,
selected_mmproj: None,
selected_mtp_model: None,
context_length: None,
gpu_layers: None,
port: None,
Expand Down Expand Up @@ -714,6 +719,7 @@ mod tests {
enabled: true,
selected_quant: None,
selected_mmproj: None,
selected_mtp_model: None,
context_length: None,
gpu_layers: None,
port: None,
Expand Down
1 change: 1 addition & 0 deletions crates/tama-cli/src/handlers/server/add.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -145,6 +145,7 @@ pub async fn cmd_server_add(
model: extracted.model.clone(),
quant: quant_name,
mmproj: None,
mtp_model: None,
port: extracted.port,
health_check: None,
enabled: true,
Expand Down
1 change: 1 addition & 0 deletions crates/tama-cli/tests/tests.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -256,6 +256,7 @@ async fn test_cmd_server_edit_valid_profile_succeeds() {
model: None,
quant: None,
mmproj: None,
mtp_model: None,
port: None,
health_check: None,
enabled: true,
Expand Down
43 changes: 43 additions & 0 deletions crates/tama-core/src/config/resolve/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -303,6 +303,49 @@ impl Config {
}
}

// Inject --spec-draft-model from model card, only if:
// 1. mtp_model is set
// 2. The referenced quant has kind = Mtp
// 3. draft-mtp is in spec_decoding.spec_types (user enabled it)
// No backend gate — mirrors --mmproj; silently ignored by non-llama.cpp
// backends if they don't recognise the flag.
if let (Some(ref model_id), Some(ref mtp_name)) = (&server.model, &server.mtp_model) {
let has_draft_mtp = server
.spec_decoding
.spec_types
.iter()
.any(|t| t == "draft-mtp");
if has_draft_mtp {
if let Some(mtp_entry) = server.quants.get(mtp_name.as_str()) {
if mtp_entry.kind == crate::config::QuantKind::Mtp {
let models_dir = self.models_dir()?;
let mtp_path = repo_path(&models_dir, model_id).join(&mtp_entry.file);
let already_has_draft = grouped.iter().any(|e| {
matches!(crate::config::flag_name(e), Some("--spec-draft-model"))
});
if !already_has_draft {
let path_str = mtp_path.to_string_lossy();
let quoted = crate::config::quote_value(&path_str);
grouped.push(format!("--spec-draft-model {}", quoted));
}
} else {
tracing::warn!(
"mtp_model '{}' for model '{}' has kind={:?}, expected Mtp",
mtp_name,
model_id,
mtp_entry.kind
);
}
} else {
tracing::warn!(
"mtp_model '{}' not found in ModelConfig for model '{}'",
mtp_name,
model_id
);
}
}
}
Comment on lines +306 to +347

@coderabbitai coderabbitai Bot Jun 13, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

⚠️ Potential issue | 🟠 Major | ⚡ Quick win

Guard --spec-draft-model injection by backend capability.

This block injects a llama.cpp-specific flag before backend gating. On non-llama backends, it can still resolve paths via self.models_dir()? and fail arg construction (or pass unsupported flags), which is a startup-risk path.

Suggested fix 
-        // Inject --spec-draft-model from model card, only if:
+        let is_llama_cpp_backend = backend_is_llama_cpp(&server.backend);
+
+        // Inject --spec-draft-model from model card, only if:
         // 1. mtp_model is set
         // 2. The referenced quant has kind = Mtp
         // 3. draft-mtp is in spec_decoding.spec_types (user enabled it)
-        // No backend gate — mirrors --mmproj; silently ignored by non-llama.cpp
-        // backends if they don't recognise the flag.
-        if let (Some(ref model_id), Some(ref mtp_name)) = (&server.model, &server.mtp_model) {
+        // 4. backend is llama.cpp-compatible
+        if is_llama_cpp_backend
+            && let (Some(ref model_id), Some(ref mtp_name)) = (&server.model, &server.mtp_model)
+        {
             let has_draft_mtp = server
                 .spec_decoding
                 .spec_types
@@
-        let is_llama_cpp_backend = backend_is_llama_cpp(&server.backend);
+        // already computed above
🤖 Prompt for AI Agents 
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@crates/tama-core/src/config/resolve/mod.rs` around lines 306 - 347, The code
injects the llama.cpp-only "--spec-draft-model" flag without checking backend
capability; change the block that uses server.model and server.mtp_model so it
first checks the backend supports this flag (e.g. call a capability check like
self.backend.supports_spec_draft_model() or add a method
supports_spec_draft_model() and use it) and only then call self.models_dir(),
repo_path(...) and push the formatted "--spec-draft-model" into grouped; ensure
the existing checks (mtp kind == QuantKind::Mtp, existence in server.quants, and
duplicate flag detection via crate::config::flag_name) remain unchanged but are
nested under the new backend-capability guard so non-llama backends never
compute paths or inject the flag.


// Inject -c (context length) only if not already present.
let ctx = ctx_override.or(server.context_length).or_else(|| {
server
Expand Down
Loading
Loading