hugo/DS4Server
1
0
Fork 0
Code Issues 7 Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Integrate DS4 execution parity in Rust

Browse Source
This commit is contained in:
Georg Bauer
2026-07-26 17:58:05 +02:00
parent c9f0c3661c
commit 4420b81117
20 changed files with 11643 additions and 358 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
src/app/view/preferences.rs
View File

@@ -3,6 +3,19 @@ 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
.model
.supports_dspark()
.then_some(Message::PreferenceLegacyMtpChanged);
let legacy_mtp = hint(
checkbox(
"Enable legacy MTP for this model",
self.preference_draft.legacy_mtp_enabled,
)
.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
.model
@@ -14,7 +27,7 @@ impl App {
self.preference_draft.dspark_enabled,
)
.on_toggle_maybe(dspark_toggle),
"Speculative decoding with the managed DSpark draft artifact: a small model proposes tokens that the main model verifies in one pass. Usually a large speedup, and it cannot be combined with SSD streaming.",
"Speculative decoding with the managed DSpark draft artifact: a small model proposes tokens that the main model verifies in one pass. Usually a large speedup; the target model may also stream routed experts from SSD.",
);
let glm_mtp_toggle: Option<fn(bool) -> Message> = (self.preference_draft.model
== ModelChoice::Glm52)
@@ -311,6 +324,7 @@ 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",
@@ -326,7 +340,7 @@ impl App {
"Lets the draft model only propose, never decide: every token is sampled by the full model. Gives up some of the speedup in exchange for output identical to non-speculative decoding.",
),
text(if self.preference_draft.model.supports_dspark() {
"DSpark uses the managed support artifact; entering a threshold or enabling strict mode also enables DSpark."
"Legacy MTP and DSpark use separate managed support artifacts; entering a DSpark threshold or enabling strict mode selects DSpark."
} else if self.preference_draft.model == ModelChoice::Glm52 {
"GLM MTP is integrated; DSpark is unavailable for this model."
} else {
@@ -339,9 +353,10 @@ impl App {
|engine| {
let settings = engine.speculative;
format!(
"Engine: MTP draft {} • margin {} • GLM MTP {} • timing {} • DSpark {} • confidence {}{} • target-only {}",
"Engine: MTP draft {} • margin {} • legacy MTP {} • GLM MTP {} • timing {} • DSpark {} • confidence {}{} • target-only {}",
settings.mtp_draft_tokens,
settings.mtp_margin,
if self.preference_draft.legacy_mtp_enabled { "on" } else { "off" },
if settings.glm_mtp { "on" } else { "off" },
if settings.glm_mtp_timing { "on" } else { "off" },
if settings.dspark { "on" } else { "off" },
@@ -357,7 +372,7 @@ impl App {
hint(
checkbox("Enable SSD-backed model streaming", self.preference_draft.ssd_streaming)
.on_toggle(Message::PreferenceSsdChanged),
"Leaves the routed expert weights on disk and pages them in as they are needed, so a model larger than this machine's memory still runs. Every cache miss waits for the SSD, and DSpark cannot run alongside it.",
"Leaves the routed expert weights on disk and pages them in as they are needed, so a model larger than this machine's memory still runs. Every cache miss waits for the SSD; speculative support weights remain resident while target experts stream.",
),
hint(
checkbox("Skip automatic expert preload", self.preference_draft.ssd_streaming_cold)
@@ -381,7 +396,7 @@ 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. SSD streaming and DSpark are mutually exclusive.")
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.")
.size(12),
text(engine.as_ref().map_or_else(
|| "Effective SSD settings will appear after valid values are entered."

131
src/app/view/stats.rs
View File

@@ -198,6 +198,55 @@ impl App {
.spacing(7)
.into(),
);
let ssd_activity = stats_panel(
"SSD STREAMING ACTIVITY · LAST 24 SECONDS",
column![
mini_chart(
&self.metrics_history,
|point| point.ssd_requests_per_second,
Color::from_rgb8(240, 180, 70),
),
row![
text("Selected loads")
.size(12)
.color(Color::from_rgb8(240, 180, 70)),
Space::with_width(Length::Fill),
text(format!("{:.1}/s", latest.ssd_requests_per_second))
.size(12)
.color(muted_text()),
],
mini_chart(
&self.metrics_history,
|point| point.ssd_bytes_per_second,
Color::from_rgb8(67, 194, 203),
),
row![
text("Requested expert data")
.size(12)
.color(Color::from_rgb8(67, 194, 203)),
Space::with_width(Length::Fill),
text(format_rate(latest.ssd_bytes_per_second))
.size(12)
.color(muted_text()),
],
mini_chart(
&self.metrics_history,
|point| point.ssd_wait_ms_per_second,
Color::from_rgb8(220, 80, 86),
),
row![
text("Inference wait")
.size(12)
.color(Color::from_rgb8(220, 80, 86)),
Space::with_width(Length::Fill),
text(format!("{:.0} ms/s", latest.ssd_wait_ms_per_second))
.size(12)
.color(muted_text()),
],
]
.spacing(7)
.into(),
);
let model = stats_panel(
"MODEL CORE",
@@ -249,6 +298,86 @@ impl App {
.spacing(9)
.into(),
);
let acceptance = if stats.drafted_tokens == 0 {
0.0
} else {
stats.accepted_draft_tokens as f64 / stats.drafted_tokens as f64
};
let target_passes = stats
.speculative_cycles
.saturating_add(stats.verifier_passes);
let effective_speedup = if target_passes == 0 {
1.0
} else {
stats
.speculative_cycles
.saturating_add(stats.accepted_draft_tokens) as f64
/ target_passes as f64
};
let speculative = stats_panel(
"SPECULATIVE DECODING",
column![
metric_row(
"Mode",
match stats.speculative_mode {
1 => "Legacy MTP",
2 => "DSpark",
_ => "Off",
},
),
metric_row("Cycles", format_count(stats.speculative_cycles)),
metric_row("Drafted", format_count(stats.drafted_tokens)),
metric_row("Accepted", format_count(stats.accepted_draft_tokens)),
metric_row("Acceptance", format!("{:.1}%", acceptance * 100.0)),
metric_row(
"Target verifier passes",
format_count(stats.verifier_passes)
),
metric_row("Verifier wall time", format_milliseconds(stats.verifier_ms)),
metric_row(
"Effective target-pass speedup",
format!("{effective_speedup:.2}×")
),
]
.spacing(9)
.into(),
);
let ssd = stats_panel(
"SSD EXPERT STREAMING",
column![
metric_row("State", if stats.ssd_enabled { "Enabled" } else { "Off" }),
metric_row("Resident weights", format_bytes(stats.ssd_resident_bytes)),
metric_row("Expert cache", format_bytes(stats.ssd_cache_bytes)),
metric_row(
"Cache capacity",
format!("{} experts", stats.ssd_cache_experts)
),
metric_row(
"Preloaded",
format!("{} experts", stats.ssd_preloaded_experts)
),
metric_row(
"Selected-load requests",
format_count(stats.ssd_selected_requests)
),
metric_row(
"Requested expert bytes",
format_bytes(stats.ssd_requested_bytes)
),
metric_row("Selected-load wait", format_milliseconds(stats.ssd_wait_ms)),
metric_row(
"Average load wait",
format_milliseconds(
stats
.ssd_wait_ms
.checked_div(stats.ssd_selected_requests)
.unwrap_or(0)
)
),
]
.spacing(9)
.into(),
);
let cache = stats_panel(
"KV CACHE",
column![
@@ -352,10 +481,12 @@ impl App {
heading,
headline,
throughput,
ssd_activity,
kv_io,
requests,
disc,
row![model, runtime].spacing(10),
row![speculative, ssd].spacing(10),
row![cache, server].spacing(10),
text("Counters are published by the runtime with relaxed atomics and sampled by the UI every 200 ms.")
.size(11)