Add native MTP generation metadata to layer packages

crates/mesh-llm-config/src/model.rs

@@ -424,6 +424,8 @@ pub struct SkippyConfig {
 #[derive(Clone, Debug, Default, Deserialize, Serialize, PartialEq)]
 #[serde(deny_unknown_fields)]
 pub struct SpeculativeConfig {
+    #[serde(default)]
+    pub strategy: Option<String>,
     #[serde(default)]
     pub mode: Option<String>,
     #[serde(default)]

crates/mesh-llm-config/src/model/built_in_schema.rs

@@ -580,6 +580,7 @@ fn skippy_settings(prefix: &str) -> Vec<ConfigSettingSchema> {
 
 fn speculative_settings(prefix: &str) -> Vec<ConfigSettingSchema> {
     vec![
+        basic_setting(&format!("{prefix}.strategy"), ConfigValueSchema::String),
         basic_setting(&format!("{prefix}.mode"), ConfigValueSchema::String),
         basic_setting(
             &format!("{prefix}.draft_model_path"),

crates/mesh-llm-config/src/validate.rs

@@ -727,6 +727,11 @@ fn validate_skippy(config: &SkippyConfig, base_path: &str) -> DiagnosticResult {
 }
 
 fn validate_speculative(config: &SpeculativeConfig, base_path: &str) -> DiagnosticResult {
+    validate_optional_enum(
+        config.strategy.as_deref(),
+        &["auto", "disabled", "native-mtp-n1"],
+        &format!("{base_path}.strategy"),
+    )?;
     validate_optional_enum(
         config.mode.as_deref(),
         &["auto", "disabled", "draft", "ngram"],
@@ -1544,6 +1549,29 @@ gpu_id = "metal:0"
         );
     }
 
+    #[test]
+    fn speculative_strategy_rejects_unknown_values() {
+        let config: MeshConfig = toml::from_str(
+            r#"
+[defaults.speculative]
+strategy = "mystery-oracle"
+"#,
+        )
+        .expect("config should parse before validation");
+
+        let diagnostics = validate_config_diagnostics(&config);
+        assert_eq!(diagnostics.len(), 1);
+        assert_eq!(
+            diagnostics[0].path.as_ref().map(ConfigPath::render),
+            Some("defaults.speculative.strategy".to_string())
+        );
+        assert!(
+            diagnostics[0]
+                .message
+                .contains("defaults.speculative.strategy must be one of")
+        );
+    }
+
     #[test]
     fn legacy_validation_errors_derive_compatible_string_messages() {
         let config: MeshConfig = toml::from_str(

crates/mesh-llm-host-runtime/src/inference/skippy/certification.rs

@@ -624,6 +624,7 @@ mod tests {
             source_model_bytes: Some(42),
             layer_count: 2,
             activation_width: 4096,
+            generation: None,
             projector_path: None,
             layers: vec![StagePackageLayerInfo {
                 layer_index: 0,

crates/mesh-llm-host-runtime/src/inference/skippy/deployment.rs

@@ -24,6 +24,7 @@ pub(crate) struct StageDeploymentContext<'a> {
     pub(crate) kv_cache: KvCachePolicy,
     pub(crate) flash_attn_type: FlashAttentionType,
     pub(crate) projector_path: Option<String>,
+    pub(crate) native_mtp_enabled: bool,
 }
 
 pub(crate) fn remote_stage_load_request(
@@ -57,6 +58,7 @@ pub(crate) fn remote_stage_load_request(
         cache_type_k: context.kv_cache.cache_type_k().to_string(),
         cache_type_v: context.kv_cache.cache_type_v().to_string(),
         flash_attn_type: context.flash_attn_type,
+        native_mtp_enabled: context.native_mtp_enabled,
         shutdown_generation: 1,
         coordinator_term: 0,
         coordinator_id: None,
@@ -105,6 +107,7 @@ pub(crate) fn stage0_config(
         filter_tensors_on_load: true,
         selected_device,
         kv_cache: None,
+        native_mtp_enabled: context.native_mtp_enabled,
         load_mode: LoadMode::LayerPackage,
         bind_addr: "127.0.0.1:0".to_string(),
         upstream: None,
@@ -200,6 +203,7 @@ mod tests {
             source_model_bytes: Some(100),
             layer_count: 4,
             activation_width: 1024,
+            generation: None,
             projector_path: Some("/tmp/package/projectors/mmproj.gguf".to_string()),
             layers: vec![StagePackageLayerInfo {
                 layer_index: 0,
@@ -230,6 +234,7 @@ mod tests {
             kv_cache: KvCachePolicy::for_model_size(0),
             flash_attn_type: FlashAttentionType::Auto,
             projector_path: Some("/models/mmproj.gguf".to_string()),
+            native_mtp_enabled: false,
         };
         let request = remote_stage_load_request(
             &context,
@@ -254,6 +259,7 @@ mod tests {
         );
         assert_eq!((request.layer_start, request.layer_end), (4, 8));
         assert!(request.projector_path.is_none());
+        assert!(!request.native_mtp_enabled);
 
         let stage0 = stage0_config(
             &context,
@@ -280,5 +286,6 @@ mod tests {
             stage0.projector_path.as_deref(),
             Some("/models/mmproj.gguf")
         );
+        assert!(!stage0.native_mtp_enabled);
     }
 }

crates/mesh-llm-host-runtime/src/inference/skippy/family_policy.rs

@@ -445,6 +445,7 @@ mod tests {
             filter_tensors_on_load: false,
             selected_device: None,
             kv_cache: None,
+            native_mtp_enabled: true,
             load_mode: LoadMode::RuntimeSlice,
             bind_addr: "127.0.0.1:0".to_string(),
             upstream: None,

crates/mesh-llm-host-runtime/src/inference/skippy/materialization.rs

@@ -11,6 +11,7 @@ use hf_hub::progress::{DownloadEvent, Progress, ProgressEvent, ProgressHandler};
 use serde::{Deserialize, Serialize};
 use sha2::{Digest, Sha256};
 use skippy_protocol::{LoadMode, StageConfig};
+use skippy_runtime::package::PackageGenerationInfo;
 use skippy_runtime::package::{
     self, LayerPackageInfo, PackageIntegrityOptions, PackageStageRequest,
 };
@@ -92,6 +93,7 @@ pub struct StagePackageInfo {
     pub source_model_bytes: Option<u64>,
     pub layer_count: u32,
     pub activation_width: u32,
+    pub generation: Option<PackageGenerationInfo>,
     pub projector_path: Option<String>,
     pub layers: Vec<StagePackageLayerInfo>,
 }
@@ -1337,6 +1339,7 @@ fn stage_package_info(package_ref: &str, info: LayerPackageInfo) -> Result<Stage
         source_model_bytes: info.source_model_bytes,
         layer_count: info.layer_count,
         activation_width,
+        generation: info.generation,
         projector_path: info
             .projectors
             .first()
@@ -1576,6 +1579,7 @@ mod tests {
             cache_type_k: "f16".to_string(),
             cache_type_v: "f16".to_string(),
             flash_attn_type: FlashAttentionType::Auto,
+            native_mtp_enabled: true,
             shutdown_generation: 1,
             coordinator_term: 0,
             coordinator_id: None,
@@ -2085,6 +2089,7 @@ mod tests {
             cache_type_k: "f16".to_string(),
             cache_type_v: "f16".to_string(),
             flash_attn_type: skippy_protocol::FlashAttentionType::Auto,
+            native_mtp_enabled: true,
             shutdown_generation: 0,
             coordinator_term: 0,
             coordinator_id: None,

crates/mesh-llm-host-runtime/src/inference/skippy/mod.rs

@@ -155,6 +155,7 @@ pub(crate) struct SkippyModelLoadOptions {
     pub(crate) projector_path: Option<PathBuf>,
     pub(crate) telemetry: SkippyTelemetryOptions,
     pub(crate) openai_guardrails: Option<OpenAiGuardrailsConfig>,
+    pub(crate) native_mtp_enabled: bool,
 }
 
 #[derive(Clone, Debug)]
@@ -238,6 +239,7 @@ impl SkippyModelLoadOptions {
             projector_path: None,
             telemetry: SkippyTelemetryOptions::off(),
             openai_guardrails: Some(OpenAiGuardrailsConfig::disabled_for_skippy()),
+            native_mtp_enabled: true,
         }
     }
 
@@ -444,6 +446,7 @@ impl SkippyModelHandle {
             speculative_window: embedded_args.speculative_window,
             adaptive_speculative_window: embedded_args.adaptive_speculative_window,
             draft_n_gpu_layers: embedded_args.draft_n_gpu_layers,
+            native_mtp_enabled: embedded_args.native_mtp_enabled,
             activation_width: embedded_args.activation_width,
             wire_dtype: embedded_args.wire_dtype,
             reply_credit_limit: embedded_args.reply_credit_limit,
@@ -539,6 +542,7 @@ impl SkippyModelHandle {
             speculative_window: embedded_args.speculative_window,
             adaptive_speculative_window: embedded_args.adaptive_speculative_window,
             draft_n_gpu_layers: embedded_args.draft_n_gpu_layers,
+            native_mtp_enabled: embedded_args.native_mtp_enabled,
             activation_width: embedded_args.activation_width,
             wire_dtype: embedded_args.wire_dtype,
             reply_credit_limit: embedded_args.reply_credit_limit,
@@ -710,6 +714,7 @@ impl SkippyModelHandle {
             speculative_window: embedded_args.speculative_window,
             adaptive_speculative_window: embedded_args.adaptive_speculative_window,
             draft_n_gpu_layers: embedded_args.draft_n_gpu_layers,
+            native_mtp_enabled: embedded_args.native_mtp_enabled,
             activation_width: embedded_args.activation_width,
             wire_dtype: embedded_args.wire_dtype,
             reply_credit_limit: embedded_args.reply_credit_limit,
@@ -832,6 +837,7 @@ impl SkippyModelHandle {
             speculative_window: embedded_args.speculative_window,
             adaptive_speculative_window: embedded_args.adaptive_speculative_window,
             draft_n_gpu_layers: embedded_args.draft_n_gpu_layers,
+            native_mtp_enabled: embedded_args.native_mtp_enabled,
             activation_width: embedded_args.activation_width,
             wire_dtype: embedded_args.wire_dtype,
             reply_credit_limit: embedded_args.reply_credit_limit,
@@ -1054,6 +1060,7 @@ pub(crate) fn single_stage_config(options: &SkippyModelLoadOptions) -> Result<St
         filter_tensors_on_load: false,
         selected_device: options.selected_device.clone().map(Into::into),
         kv_cache: None,
+        native_mtp_enabled: options.native_mtp_enabled,
         load_mode: LoadMode::RuntimeSlice,
         bind_addr: "127.0.0.1:0".to_string(),
         upstream: None,
@@ -1236,6 +1243,7 @@ mod tests {
             layer_count,
             activation_width: 4096,
             tensor_count: 100,
+            generation: None,
         }
     }
 

crates/mesh-llm-host-runtime/src/inference/skippy/package.rs

@@ -7,6 +7,7 @@ use std::{
 use anyhow::{Context, Result};
 use serde::Serialize;
 use sha2::{Digest, Sha256};
+use skippy_runtime::package::PackageGenerationInfo;
 
 #[derive(Clone, Debug, Eq, PartialEq)]
 pub struct SkippyPackageIdentity {
@@ -19,6 +20,7 @@ pub struct SkippyPackageIdentity {
     pub layer_count: u32,
     pub activation_width: u32,
     pub tensor_count: u64,
+    pub generation: Option<PackageGenerationInfo>,
 }
 
 #[derive(Clone, Debug, Eq, PartialEq, Serialize)]
@@ -109,6 +111,7 @@ pub fn synthetic_direct_gguf_package(
         layer_count: compact.layer_count,
         activation_width: compact.embedding_size,
         tensor_count,
+        generation: None,
     })
 }
 
@@ -328,6 +331,7 @@ pub fn identity_from_layer_package(package_ref: &str) -> Result<SkippyPackageIde
         layer_count: info.layer_count,
         activation_width,
         tensor_count: info.layers.iter().map(|l| l.tensor_count as u64).sum(),
+        generation: info.generation,
     })
 }
 

crates/mesh-llm-host-runtime/src/inference/skippy/resolver.rs

@@ -5,6 +5,12 @@ mod support;
 mod translation;
 mod types;
 
+#[cfg(test)]
+mod test_support;
+
+#[cfg(test)]
+mod native_mtp_tests;
+
 #[cfg(test)]
 mod tests;