Gemma4: Full DFlash Integration (Speculative Decode + BSA Prefill + Prefix Cache)#232

Merged

davide221 merged 19 commits into

Luce-Org:mainfrom

howard0su:gemma4

May 22, 2026

Commits on May 21, 2026

gemma4: fix loader + graph for actual GGUF format

howard0su
and
Copilot
committed
gemma4: implement DFlashTarget for speculative decode (G4)

howard0su
and
Copilot
committed
gemma4: fix attention scale, tokenizer decode, and server integration

howard0su
and
Copilot
committed
gemma4: implement real park/unpark for VRAM management

howard0su
and
Copilot
committed
gemma4: implement G5 SWA ring-buffer, G6 fa_window, G3 compress

howard0su
and
Copilot
committed
gemma4: wire DFlash speculative decode into Gemma4 backend

howard0su
and
Copilot
committed
prefix_cache: add Gemma family detection for chat markers

howard0su
and
Copilot
committed
gemma4: fix DFlash spec-decode acceptance rate

howard0su
and
Copilot
committed
gemma4 dflash: fix SWA causal masking and rope_theta

howard0su
and
Copilot
committed
draft: use F16 mask directly, remove unnecessary F32 cast

howard0su
and
Copilot
committed
draft: rename draft_dflash_graph.cpp → draft_graph.cpp to match header

howard0su
and
Copilot
committed
draft: remove DFLASH27B_ROPE_THETA constant, read from GGUF only

howard0su
and
Copilot
committed
gemma4 spec-decode: replace snapshot/replay with KV truncation

howard0su
and
Copilot
committed
gemma4: add BSA sparse-FA prefill path + unified flash_prefill_forward dispatch

howard0su
and
Copilot
committed
gemma4: full feature mirror resync after prefix cache restore

howard0su
and
Copilot
committed
gemma4: check graph_compute return in prefill_bsa per-layer embed

howard0su
and
Copilot
committed
gemma4: fix namespace dflash27b -> dflash::common after rebase

howard0su
and
Copilot
committed

Commits on May 22, 2026

gemma4: fix MoE GELU contig + loader tensor name mismatches

authored and
davide221
committed
Merge branch 'main' into gemma4
davide221
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Gemma4: Full DFlash Integration (Speculative Decode + BSA Prefill + Prefix Cache)#232

Gemma4: Full DFlash Integration (Speculative Decode + BSA Prefill + Prefix Cache)#232
davide221 merged 19 commits into
Luce-Org:mainfrom
howard0su:gemma4

Commits on May 21, 2026

gemma4: fix loader + graph for actual GGUF format

gemma4: implement DFlashTarget for speculative decode (G4)

gemma4: fix attention scale, tokenizer decode, and server integration

gemma4: implement real park/unpark for VRAM management

gemma4: implement G5 SWA ring-buffer, G6 fa_window, G3 compress

gemma4: wire DFlash speculative decode into Gemma4 backend

prefix_cache: add Gemma family detection for chat markers

gemma4: fix DFlash spec-decode acceptance rate

gemma4 dflash: fix SWA causal masking and rope_theta

draft: use F16 mask directly, remove unnecessary F32 cast

draft: rename draft_dflash_graph.cpp → draft_graph.cpp to match header

draft: remove DFLASH27B_ROPE_THETA constant, read from GGUF only

gemma4 spec-decode: replace snapshot/replay with KV truncation

gemma4: add BSA sparse-FA prefill path + unified flash_prefill_forward dispatch

gemma4: full feature mirror resync after prefix cache restore

gemma4: check graph_compute return in prefill_bsa per-layer embed

gemma4: fix namespace dflash27b -> dflash::common after rebase

Commits on May 22, 2026

gemma4: fix MoE GELU contig + loader tensor name mismatches

Merge branch 'main' into gemma4