PR #24162 DeepSeek V4 - SemanticDiff

github-actions added model

github-actions added python

github-actions added ggml

github-actions added script

github-actions added testing

fairydreaming commented on 2026-06-07

ngxson commented on 2026-06-11

am17an force pushed from 415c6d07 to e5865927 66 days ago

am17an changed the title ~~[WIP] DeepSeek V4~~ DeepSeek V4 63 days ago

am17an marked this pull request as ready for review 63 days ago

am17an requested a review from

JohannesGaessler 63 days ago

am17an requested a review from

ggerganov 63 days ago

am17an requested a review from

CISC 63 days ago

am17an requested a review 63 days ago

ggerganov commented on 2026-06-16

ggerganov assigned

ggerganov 62 days ago

CISC commented on 2026-06-16

fairydreaming commented on 2026-06-16

ggerganov commented on 2026-06-17

ggerganov approved these changes on 2026-06-17

ggerganov commented on 2026-06-17

ggerganov commented on 2026-06-18

am17an requested a review 58 days ago

github-actions added documentation

github-actions added Vulkan

github-actions added examples

github-actions added devops

github-actions added server

github-actions added SYCL

github-actions added Apple Metal

github-actions added CUDA

github-actions added jinja parser

github-actions added WebGPU

github-actions added server/ui

am17an force pushed from dd03bebd to 25c8398e 58 days ago

am17an removed review request 58 days ago

am17an commented on 2026-06-20

dfriehs commented on 2026-06-20

github-actions added conversion

GEN-BIT commented on 2026-06-27

am17an commented on 2026-06-27

ggerganov commented on 2026-06-28

am17an force pushed from fd0955f2 to e16065fd 50 days ago

ggerganov commented on 2026-06-28

convert: add dsv4 conversion

aa0d5f41

add basic setup

d70705d5

add llm_graph_input_dsv4

d871b49a

add save-load state

3735a82a

add sinkhorn eps - correction by @fairydreaming

36333ad5

add rope fix

5927082f

cleanup dead code

b2d36ac9

fix bugs

67bfd0fa

support pro model: added by @fairydreaming

10e7f361

remove redundant V cache

7ab3378c

Chat template

b7631f01

remove debugging leftovers

737b5a4b

Add mechanism for inlining templates based on architecture

0fa3bb08

s/deepseek-v4-flash/deepseek4/g

91920e1a

s/deepseek-v4-flash/deepseek4/g continued

f07c4a36

enable graph reuse

ea11d094

enable FA

8db46cc9

fix test llama archs

76fd0083

rename

517849c7

compatibility with antirez ds4 GGUFs

5ca52d52

simplified set_gguf_parameters() by calling super class method, repla…

8ecac3a9

reserve worst-case kv-cache

21ac24da

revert max split inputs

374d51d8

address review comments

41838689

add padding to enable FA

8971d843

pad only the final value of plan.n_kv to 256

494591cb

remove built-in cpp chat template

bb632f45

cont: remove cpp built-in template

3b93912d

rm outdated test

0d33ad0e

replace ggml_view_3d() with ggml_reshape_3d()

f9f4fb30

only support n_seq=1 for now

4b67acd0

remove unused var

10214b66

cont: remove unused var

c2a43966

use scale bias

c7d65399

use correct ptr for can_reuse

b16391ce

remove gen-chat-inline-templates.py

411c2bcd

simplify graph reuse

5764dc41

cont: cleanup

63359743

remove unused inputs

175ac851

enable partial checkpointing

9885fa76

add correct shape for kq_mask + set llama_model_n_swa to 0 for dsv4

55fdfa82

precompute source_idx + add comment about dummy write

8eb9c860

support multi-seq

45056404

remove restored_trim_pos

344965da

use split_equal when possible

0b809789

am17an force pushed from 4dd09f3c to 0b809789 50 days ago

am17an requested a review from

CISC 49 days ago

ggerganov approved these changes on 2026-06-29

fix indent

ce2b545e

CISC commented on 2026-06-29

address review comments

b8983f9f

CISC commented on 2026-06-29

CISC approved these changes on 2026-06-29

use LLM_KV

fc8f11dd

am17an removed documentation

am17an removed testing

am17an removed Vulkan

am17an removed SYCL

am17an removed Apple Metal

am17an removed CUDA

am17an removed WebGPU

am17an removed server/ui

am17an removed devops

CISC approved these changes on 2026-06-29

am17an removed examples

github-actions added testing

CISC commented on 2026-06-29

fix ci

f180ae8b

am17an force pushed from 62eab9de to f180ae8b 49 days ago

CISC approved these changes on 2026-06-29

CISC removed script

CISC removed python

CISC removed server

CISC removed ggml

CISC removed jinja parser

am17an merged 8c146a83 into master 49 days ago

am17an deleted the dsv4 branch 49 days ago

fairydreaming commented on 2026-06-30

llama.cpp
DeepSeek V4
#24162

Merged

DeepSeek V4 #24162

llama.cpp DeepSeek V4 #24162 Merged

DeepSeek V4 #24162

llama.cpp
DeepSeek V4
#24162

Merged