qwen35: DSpark speculative decoding (Qwen3.8-27B drafters) - #625

Merged

davide221 merged 37 commits into

mainfrom

Aug 24, 2026

Commits on Aug 18, 2026

qwen35: DSpark speculative decoding support
davide221
committed
cc75a80
qwen35: per-step verify length for DSpark confidence gate
davide221
committed
b57738b
ggml: fused DeltaNet decode kernels for HIP
davide221
committed
86eba54
ggml: 64x64 MMQ tiles for dense verify widths on RDNA
davide221
committed
cd2fc08
qwen35: stacked projections and fused DeltaNet decode graph
davide221
committed
fa345d6
qwen35: adaptive speculation policy and chain-path profiling
davide221
committed
e9e8cf4
ggml: FA vec kernel splits short KV spans across two blocks
davide221
committed
f1281da
qwen35: adaptive policy probe step reacts fast
davide221
committed
cb08540
qwen35: adaptive policy uses the measured spec/plain step-time ratio
davide221
committed
9df0a0a

Commits on Aug 19, 2026

qwen35: DFlash 2 drafter support (dynamic convs + candidate selector)
davide221
committed
ebc69e6
ggml: skip the pathological mmq_x=32 small tile on RDNA
davide221
committed
324eb8b
Merge origin/main into qwen38-dspark
davide221
committed
480e60a

Commits on Aug 20, 2026

ggml: tree-mode support for the grouped-cols GDN kernel
davide221
committed
8f2d4fb
qwen35: DFlash2 selector-scored DDTree candidates
davide221
committed
4465965
qwen35: fix and gate the chunked delta-net prefill path
davide221
committed
2ff7afa
ggml: binary exponentiation for the fp64 RoPE angle
davide221
committed
9a32fda
ggml: dual-tile MMQ dispatch on RDNA4
davide221
committed
a9f8296
ggml: non-temporal weight loads in the IQ4_XS decode GEMV
davide221
committed
1e0f49f

Commits on Aug 21, 2026

Commits on Aug 24, 2026

qwen35: narrow the draft block at long context, and fix review findings
mrciffa
committed
3d077c4
qwen35: lower the long-context draft-narrowing threshold to 8192
mrciffa
committed
7630f73
qwen35: make the tuned decode path the default, so no env vars are needed
mrciffa
committed
d461bad
readme: make Qwen3.8-27B the headline target
mrciffa
committed
d3e738d
qwen35: cap the verify width, not the draft, at long context
mrciffa
committed
5c61ee0
docs: correct comments that still described the old defaults
mrciffa
committed
85e7276
test: expect disk cache version 2
mrciffa
committed
910a208
qwen35: warn when a prompt outgrows --fa-window (#653)

davide221
and
mrciffa
authored
10a94ab
fix(qwen35) span the whole KV pool for slot-mapped verify attention (#652)

davide221
and
mrciffa
authored
d23951d
qwen35: fix review findings from the pre-merge audit
davide221
committed
15a242a
readme: recommend the Unsloth UD-IQ4_XS target and fix the drafter quickstart
davide221
committed
f28d85c