feat(speculation): add generic adaptive activation - #626

Draft

Graffioh wants to merge 43 commits into

Luce-Org:mainfrom

Graffioh:codex/qwen38-dspark-adaptive

Commits on Aug 18, 2026

feat(qwen35): complete concurrent speculative serving
Graffioh
committed
8902901
bench(qwen36): verify concurrent feature matrix
Graffioh
committed
3ad0796
fix(qwen36): harden feature smoke paths
Graffioh
committed
631a9e1
feat(qwen36): tune concurrent serving for Strix Halo
Graffioh
committed
5f7c5df
perf(qwen35): adapt packed prefill budgeting
Graffioh
committed
ea5b737
fix(qwen35): deduplicate prefill policy helper
Graffioh
committed
e3cee32
qwen35: DSpark speculative decoding support
davide221
committed
cc75a80
Merge PR #625 DSpark support onto concurrent serving
Graffioh
committed
2b83366
test(feature-gate): repair stacked baseline assertions
Graffioh
committed
49a3d61
feat(server): add per-request decode policy
Graffioh
committed
2e9be33
feat(draft): export calibrated pre-norm hidden state
Graffioh
committed
37423f2
feat(qwen35): add concurrent DSpark chain execution
Graffioh
committed
2fdfa64
feat(concurrency): add adaptive speculation gate
Graffioh
committed
e13ee33
feat(qwen35): wire adaptive DSpark speculation
Graffioh
committed
df0195a
feat(draft): batch concurrent DSpark drafting
Graffioh
committed
4bf139d
feat(harness): add Qwen3.8 DSpark oracle matrix
Graffioh
committed
ab61d32

Commits on Aug 19, 2026

feat(concurrency): calibrate adaptive speculation confidence
Graffioh
committed
0fdc5f9
fix(concurrency): activate calibration after 32 rounds
Graffioh
committed
9afca0b
refactor(concurrency): make adaptive gate confidence-only
Graffioh
committed
50d4d51
bench(qwen38): codify PR 625 baseline
Graffioh
committed
167682e
qwen35: per-step verify length for DSpark confidence gate

davide221
authored and
Graffioh
committed
c0fbe36
ggml: fused DeltaNet decode kernels for HIP

davide221
authored and
Graffioh
committed
248b501
ggml: 64x64 MMQ tiles for dense verify widths on RDNA

davide221
authored and
Graffioh
committed
7a1c77e
qwen35: stacked projections and fused DeltaNet decode graph

davide221
authored and
Graffioh
committed
d50f928
qwen35: adaptive speculation policy and chain-path profiling

davide221
authored and
Graffioh
committed
bcd0742
ggml: FA vec kernel splits short KV spans across two blocks

davide221
authored and
Graffioh
committed
56b2e05
qwen35: adaptive policy probe step reacts fast

davide221
authored and
Graffioh
committed
6136339
qwen35: adaptive policy uses the measured spec/plain step-time ratio

davide221
authored and
Graffioh
committed
b0957a7
qwen35: DFlash 2 drafter support (dynamic convs + candidate selector)

davide221
authored and
Graffioh
committed
fb63013
ggml: skip the pathological mmq_x=32 small tile on RDNA

davide221
authored and
Graffioh
committed
547f5c2
bench(qwen38): pin PR 625 q8 draft baseline
Graffioh
committed
3821738
bench(concurrency): profile DSpark adaptive rounds
Graffioh
committed
b9c0405
feat(concurrency): make DSpark activation request-sticky
Graffioh
committed
7354617
fix(concurrency): enforce one-shot DSpark activation
Graffioh
committed
fccb39a
refactor(concurrency): isolate adaptive request scoring
Graffioh
committed
1626eeb
feat(concurrency): add adaptive DFlash2 activation
Graffioh
committed
79ac1b4
refactor(speculation): extract generic activation primitives
Graffioh
committed
460c2dc
feat(speculation): add speculator interface and DFlash2 adapter
Graffioh
committed
a34aca2
refactor(speculation): remove concurrent DSpark drafting
Graffioh
committed
6f1486a
chore(speculation): consolidate activation harness and telemetry
Graffioh
committed
375543e

Commits on Aug 20, 2026

perf(qwen35): commit mixed speculation directly
Graffioh
committed
07a37a1
bench(concurrency): validate direct adaptive refill
Graffioh
committed
7e876ef
fix: replan adaptive speculation by cohort
Graffioh
committed
8eeb4e4