hugo/DS4Server
1
0
Fork 0
Code Issues 2 Pull Requests Actions Packages Projects Releases Wiki Activity

Remove deprecated Flash preview and legacy MTP

Browse Source
This commit is contained in:
Georg Bauer
2026-08-31 08:04:20 +02:00
parent 7f88e23884
commit 0fa15bb68b
24 changed files with 246 additions and 1200 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
src/app/view/preferences.rs
View File

@@ -3,17 +3,6 @@ use iced::widget::column;
impl App {
pub(super) fn preferences_panel(&self) -> Element<'_, Message> {
let legacy_mtp_toggle: Option<fn(bool) -> Message> = self
.preference_draft
.acceleration_model
.supports_legacy_mtp()
.then_some(Message::PreferenceLegacyMtpChanged);
let legacy_mtp = hint(
toggle(self.preference_draft.legacy_mtp_enabled)
.label("Enable legacy MTP for this model")
.on_toggle_maybe(legacy_mtp_toggle),
"Uses the managed one-stage MTP support GGUF. The target model verifies every drafted token; it is mutually exclusive with DSpark.",
);
let dspark_toggle: Option<fn(bool) -> Message> = self
.preference_draft
.acceleration_model
@@ -131,7 +120,7 @@ impl App {
|engine| engine.artifacts.model.display().to_string(),
),
engine
.and_then(|engine| engine.artifacts.mtp.as_ref())
.and_then(|engine| engine.artifacts.support.as_ref())
.map_or_else(String::new, |path| format!(
" • support: {}",
path.display()
@@ -602,18 +591,6 @@ impl App {
.spacing(12)
.align_y(Alignment::Center),
text("SPECULATIVE DECODING").size(11).color(muted_text()),
preference_input_row(
"MTP draft tokens",
"How many tokens the multi-token-prediction head guesses ahead for the main model to check in a single pass. More drafting pays off on predictable text and is wasted work on surprising text; the engine caps it at 16.",
text_input("1", &self.preference_draft.mtp_draft_tokens)
.on_input(Message::PreferenceMtpDraftChanged),
),
preference_input_row(
"MTP verifier margin",
"How much more likely the main model must find a drafted token before accepting it. A high margin accepts few drafts and stays close to plain decoding; a low one accepts more and rolls back more often.",
text_input("3", &self.preference_draft.mtp_margin)
.on_input(Message::PreferenceMtpMarginChanged),
),
hint(
toggle(self.preference_draft.glm_mtp)
.label("Enable integrated GLM MTP")
@@ -626,7 +603,6 @@ impl App {
.on_toggle_maybe(glm_mtp_timing_toggle),
"Records per-stage timings of the speculative path to the log, to show where the acceleration actually goes. A diagnostic aid that costs a little throughput.",
),
legacy_mtp,
dspark,
preference_input_row(
"DSpark confidence threshold",
@@ -646,11 +622,11 @@ impl App {
"For non-zero temperatures, applies DS4's exact acceptance and corrected rejection sampling. Off uses the faster opportunistic mode: sample a boundary token, then accept DSpark tokens only while they match the target's greedy path.",
),
text(if self.preference_draft.acceleration_model.supports_dspark() {
"Legacy MTP and DSpark use separate managed support artifacts; entering a DSpark threshold or enabling strict mode selects DSpark."
"DeepSeek V4 Flash 0731 uses its managed DSpark support artifact."
} else if self.preference_draft.acceleration_model == ModelChoice::Glm52 {
"GLM MTP is integrated; DSpark is unavailable for this model."
} else {
"No managed MTP support artifact is available for this model."
"No speculative-decoding support is available for this model."
})
.size(12),
text(acceleration_engine.map_or_else(
@@ -659,10 +635,7 @@ impl App {
|engine| {
let settings = engine.speculative;
format!(
"Engine: MTP draft {} • margin {} • legacy MTP {} • GLM MTP {} • timing {} • DSpark {} • confidence {}{} • target-only {} • exact sampling {}",
settings.mtp_draft_tokens,
settings.mtp_margin,
if self.preference_draft.legacy_mtp_enabled { "on" } else { "off" },
"Engine: GLM MTP {} • timing {} • DSpark {} • confidence {}{} • target-only {} • exact sampling {}",
if settings.glm_mtp { "on" } else { "off" },
if settings.glm_mtp_timing { "on" } else { "off" },
if settings.dspark { "on" } else { "off" },
@@ -705,7 +678,7 @@ impl App {
text_input("Automatic", &self.preference_draft.ssd_preload_experts)
.on_input(Message::PreferenceSsdPreloadChanged),
),
text("A blank full-layer value is automatic; an explicit 0 disables fully resident GLM layers. Flash legacy MTP and DSpark support weights remain resident when target experts stream.")
text("A blank full-layer value is automatic; an explicit 0 disables fully resident GLM layers. DSpark support weights remain resident when target experts stream.")
.size(12),
text(acceleration_engine.map_or_else(
|| "Effective SSD settings will appear after valid values are entered."