DeepSeek V4 #24162

Open

am17an wants to merge 37 commits into

ggml-org:masterfrom

Commits on Jun 20, 2026

convert: add dsv4 conversion
am17an
committed
add basic setup
am17an
committed
add llm_graph_input_dsv4
am17an
committed
add save-load state
am17an
committed
add sinkhorn eps - correction by @fairydreaming
am17an
committed
add rope fix
am17an
committed
cleanup dead code
am17an
committed
fix bugs
am17an
committed
support pro model: added by @fairydreaming
am17an
committed
remove redundant V cache
am17an
committed
Chat template

pwilkin
authored and
am17an
committed
remove debugging leftovers

sszymczy
authored and
am17an
committed
Add mechanism for inlining templates based on architecture

pwilkin
authored and
am17an
committed
s/deepseek-v4-flash/deepseek4/g
am17an
committed
s/deepseek-v4-flash/deepseek4/g continued

sszymczy
authored and
am17an
committed
enable graph reuse
am17an
committed
enable FA
am17an
committed
fix test llama archs
am17an
committed
rename
am17an
committed
compatibility with antirez ds4 GGUFs

sszymczy
authored and
am17an
committed
simplified set_gguf_parameters() by calling super class method, replaced moe.score_func with expert_gating_func.

sszymczy
authored and
am17an
committed
reserve worst-case kv-cache
am17an
committed
revert max split inputs
am17an
committed
address review comments
am17an
committed
add padding to enable FA
am17an
committed
pad only the final value of plan.n_kv to 256

sszymczy
authored and
am17an
committed
remove built-in cpp chat template

ngxson
authored and
am17an
committed
cont: remove cpp built-in template

ngxson
authored and
am17an
committed
rm outdated test

ngxson
authored and
am17an
committed
(cont) remove outdated test

ngxson
authored and
am17an
committed
replace ggml_view_3d() with ggml_reshape_3d()

authored and
am17an
committed
only support n_seq=1 for now
am17an
committed
remove unused var
am17an
committed
cont: remove unused var
am17an
committed

Commits on Jun 26, 2026

use scale bias
am17an
committed

Commits on Jun 27, 2026