KVFlash: bounded KV residency (lookahead sparse attention) for dflash#373

Open

davide221 wants to merge 22 commits into

mainfrom

proto/kv-pager

Commits on Jun 12, 2026

feat(common): KVFlash pager core + chunk-relevance scorer interface

davide221
and
wozcode
committed
feat(qwen35): wire KVFlash into the daemon (--kvflash / --kvflash-tau)

davide221
and
wozcode
committed
test: KVFlash verification suite (test_kvflash)

davide221
and
wozcode
committed
docs: optimizations/kvflash (README, RESULTS, DESIGN)

davide221
and
wozcode
committed
feat(kvflash): port bounded KV residency to qwen35moe, laguna, gemma4

davide221
and
wozcode
committed
fix(kvflash): address cubic review findings on PR #373

davide221
and
wozcode
committed
refactor(kvflash): consolidate per-backend duplication into common helpers

davide221
and
wozcode
committed
docs(kvflash): hub README card + hero + Q8_0 footnote on the 256K rows

davide221
and
wozcode
committed
docs(kvflash): state the KV quant in the table headers

davide221
and
wozcode
committed
docs: KVFlash flags in the main README server-flags reference

davide221
and
wozcode
committed
docs: keep the main-README KVFlash intro model-agnostic

davide221
and
wozcode
committed
feat(kvflash): pooled chunked prefill, --kvflash auto, drafter scorer without compression

davide221
and
wozcode
committed
feat(kvflash): drafter-scored residency is the default policy

davide221
and
wozcode
committed
feat(kvflash): cross-tokenizer drafter scoring for laguna/gemma4 + --kvflash-policy

davide221
and
wozcode
committed
feat(kvflash): VRAM-aware auto pool sizing

davide221
and
wozcode
committed
fix(kvflash): pre-ship audit — cubic round 2 + doc refresh

davide221
and
wozcode
committed
ci: give the ROCm GPU job its own concurrency group

davide221
and
wozcode
committed
ci: fail the ROCm job fast with a KFD diagnosis instead of hanging

davide221
and
wozcode
committed
feat(kvflash): --ddtree runs on the pool (gate removed)

davide221
and
wozcode
committed
ci: ROCm probe survives a D-state hang

davide221
and
wozcode
committed
feat(kvflash): gemma4 spec decode on the pool + gemma draft-loader regression fix

davide221
and
wozcode
committed

Commits on Jun 13, 2026

fix(draft): convert_dflash_to_gguf reads arch from config.json (was 27B-hardcoded)

davide221
and
wozcode
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

KVFlash: bounded KV residency (lookahead sparse attention) for dflash#373

KVFlash: bounded KV residency (lookahead sparse attention) for dflash#373
davide221 wants to merge 22 commits into
mainfrom
proto/kv-pager

Commits on Jun 12, 2026

feat(common): KVFlash pager core + chunk-relevance scorer interface

feat(qwen35): wire KVFlash into the daemon (--kvflash / --kvflash-tau)

test: KVFlash verification suite (test_kvflash)

docs: optimizations/kvflash (README, RESULTS, DESIGN)

feat(kvflash): port bounded KV residency to qwen35moe, laguna, gemma4

fix(kvflash): address cubic review findings on PR #373

refactor(kvflash): consolidate per-backend duplication into common helpers

docs(kvflash): hub README card + hero + Q8_0 footnote on the 256K rows

docs(kvflash): state the KV quant in the table headers

docs: KVFlash flags in the main README server-flags reference

docs: keep the main-README KVFlash intro model-agnostic

feat(kvflash): pooled chunked prefill, --kvflash auto, drafter scorer without compression

feat(kvflash): drafter-scored residency is the default policy

feat(kvflash): cross-tokenizer drafter scoring for laguna/gemma4 + --kvflash-policy

feat(kvflash): VRAM-aware auto pool sizing

fix(kvflash): pre-ship audit — cubic round 2 + doc refresh

ci: give the ROCm GPU job its own concurrency group

ci: fail the ROCm job fast with a KFD diagnosis instead of hanging

feat(kvflash): --ddtree runs on the pool (gate removed)

ci: ROCm probe survives a D-state hang

feat(kvflash): gemma4 spec decode on the pool + gemma draft-loader regression fix

Commits on Jun 13, 2026

fix(draft): convert_dflash_to_gguf reads arch from config.json (was 27B-hardcoded)