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

Support DeepSeek V4 Flash 0731

Browse Source
This commit is contained in:
Georg Bauer
2026-08-29 20:28:50 +02:00
parent ad855b321e
commit f1c177b754
23 changed files with 10510 additions and 2324 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
src/app/preferences.rs
View File

@@ -45,6 +45,7 @@ pub(super) struct PreferenceDraft {
pub(super) glm_mtp_timing: bool,
pub(super) dspark_confidence_threshold: String,
pub(super) dspark_strict: bool,
pub(super) dspark_exact_sampling: bool,
pub(super) ssd_streaming: bool,
pub(super) ssd_streaming_cold: bool,
pub(super) ssd_cache: String,
@@ -111,6 +112,7 @@ impl PreferenceDraft {
glm_mtp_timing: speculative.glm_mtp_timing,
dspark_confidence_threshold: optional_string(speculative.dspark_confidence_threshold),
dspark_strict: speculative.dspark_strict,
dspark_exact_sampling: speculative.dspark_exact_sampling,
ssd_streaming: runtime.ssd.enabled,
ssd_streaming_cold: runtime.ssd.cold,
ssd_cache: optional_string(runtime.ssd.cache),
@@ -226,6 +228,7 @@ impl PreferenceDraft {
&self.dspark_confidence_threshold,
)?,
dspark_strict: self.dspark_strict,
dspark_exact_sampling: self.dspark_exact_sampling,
})
}
@@ -267,6 +270,7 @@ impl PreferenceDraft {
self.glm_mtp_timing = speculative.glm_mtp_timing;
self.dspark_confidence_threshold = optional_string(speculative.dspark_confidence_threshold);
self.dspark_strict = speculative.dspark_strict;
self.dspark_exact_sampling = speculative.dspark_exact_sampling;
self.ssd_streaming = ssd.enabled;
self.ssd_streaming_cold = ssd.cold;
self.ssd_cache = optional_string(ssd.cache);
@@ -693,12 +697,16 @@ impl App {
self.preference_error = None;
}
Message::PreferenceLegacyMtpChanged(enabled) => {
self.preference_draft.legacy_mtp_enabled =
self.preference_draft.acceleration_model.supports_dspark() && enabled;
self.preference_draft.legacy_mtp_enabled = self
.preference_draft
.acceleration_model
.supports_legacy_mtp()
&& enabled;
if self.preference_draft.legacy_mtp_enabled {
self.preference_draft.dspark_enabled = false;
self.preference_draft.dspark_confidence_threshold.clear();
self.preference_draft.dspark_strict = false;
self.preference_draft.dspark_exact_sampling = false;
}
self.preference_error = None;
}
@@ -708,6 +716,7 @@ impl App {
if !self.preference_draft.dspark_enabled {
self.preference_draft.dspark_confidence_threshold.clear();
self.preference_draft.dspark_strict = false;
self.preference_draft.dspark_exact_sampling = false;
} else {
self.preference_draft.legacy_mtp_enabled = false;
}
@@ -903,6 +912,15 @@ impl App {
}
self.preference_error = None;
}
Message::PreferenceDsparkExactSamplingChanged(value) => {
self.preference_draft.dspark_exact_sampling =
self.preference_draft.acceleration_model.supports_dspark() && value;
if self.preference_draft.dspark_exact_sampling {
self.preference_draft.dspark_enabled = true;
self.preference_draft.legacy_mtp_enabled = false;
}
self.preference_error = None;
}
Message::PreferenceSsdChanged(value) => {
self.preference_draft.ssd_streaming = value;
self.preference_error = None;
@@ -1011,14 +1029,14 @@ mod tests {
assert_eq!(draft.context_tokens, "32768");
draft.context_tokens = "456".into();
draft
.select_generation(ModelChoice::DeepSeekV4Flash, ReasoningMode::High)
.select_generation(ModelChoice::DeepSeekV4Flash0731, ReasoningMode::High)
.unwrap();
assert_eq!(draft.context_tokens, "123");
draft.select_acceleration(ModelChoice::Glm52).unwrap();
assert!(!draft.ssd_streaming);
draft
.select_acceleration(ModelChoice::DeepSeekV4Flash)
.select_acceleration(ModelChoice::DeepSeekV4Flash0731)
.unwrap();
assert!(draft.ssd_streaming);
}

20
src/app/view/preferences.rs
View File

@@ -6,7 +6,7 @@ impl App {
let legacy_mtp_toggle: Option<fn(bool) -> Message> = self
.preference_draft
.acceleration_model
.supports_dspark()
.supports_legacy_mtp()
.then_some(Message::PreferenceLegacyMtpChanged);
let legacy_mtp = hint(
toggle(self.preference_draft.legacy_mtp_enabled)
@@ -36,6 +36,11 @@ impl App {
.acceleration_model
.supports_dspark()
.then_some(Message::PreferenceDsparkStrictChanged);
let dspark_exact_toggle: Option<fn(bool) -> Message> = self
.preference_draft
.acceleration_model
.supports_dspark()
.then_some(Message::PreferenceDsparkExactSamplingChanged);
let effective = self
.preference_draft
.effective_for(
@@ -74,7 +79,7 @@ impl App {
text_input("Automatic", &self.preference_draft.directional_steering_ffn);
let mut steering_attn = text_input("0", &self.preference_draft.directional_steering_attn);
let mut dspark_confidence = text_input(
"0.9 (DS4 default)",
"0.6 (DS4 default)",
&self.preference_draft.dspark_confidence_threshold,
);
if self.preference_draft.model != ModelChoice::Glm52 {
@@ -509,7 +514,7 @@ impl App {
dspark,
preference_input_row(
"DSpark confidence threshold",
"How sure the draft model must be, from 0 to 1, before its token is handed to the verifier. Lower forwards more guesses for more speed and more rejected work; blank uses DS4's 0.9.",
"How sure the draft model must be, from 0 to 1, before its token is handed to the verifier. Lower forwards more guesses for more speed and more rejected work; blank uses DS4's 0.6, or 0.8 for exact sampling.",
dspark_confidence,
),
hint(
@@ -518,6 +523,12 @@ impl App {
.on_toggle_maybe(dspark_strict_toggle),
"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.",
),
hint(
toggle(self.preference_draft.dspark_exact_sampling)
.label("Use exact DSpark sampling")
.on_toggle_maybe(dspark_exact_toggle),
"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."
} else if self.preference_draft.acceleration_model == ModelChoice::Glm52 {
@@ -532,7 +543,7 @@ impl App {
|engine| {
let settings = engine.speculative;
format!(
"Engine: MTP draft {} • margin {} • legacy MTP {} • GLM MTP {} • timing {} • DSpark {} • confidence {}{} • target-only {}",
"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" },
@@ -542,6 +553,7 @@ impl App {
settings.dspark_confidence_threshold,
if settings.dspark_confidence_threshold_set { " explicit" } else { " default" },
if settings.dspark_strict { "on" } else { "off" },
if settings.dspark_exact_sampling { "on" } else { "off" },
)
},
))