llama.cpp
model: add Qwen3.8-Flash-Next (qwen4exp)
#27742
Merged
Overview
Commits
65
Changes
View On GitHub

model: add Qwen3.8-Flash-Next (qwen4exp) #27742

ngxson merged 65 commits into ggml-org:master from unslothai:qwen4exp/qwen3.8-flash-next
danielhanchen
danielhanchen gguf: add qwen4exp (Qwen3.8-Flash-Next) arch and converter
6e5b8b7e
danielhanchen llama: load qwen4exp (Qwen3.8-Flash-Next) hparams and tensors
331fa7e5
danielhanchen qwen4exp: shorten comments
058b7a1b
danielhanchen llama: qwen4exp text graph with hyper-connections, GDN and MoE
c9ad86f6
danielhanchen llama: qwen4exp PLE n-gram hash embedding
ad4fa3fa
danielhanchen llama: carry the qwen4exp PLE conv state across ubatches
c89e67b0
danielhanchen llama: fix the qwen4exp PLE conv state and unblock test-llama-archs
ddf0980e
danielhanchen llama: optional indexer key cache in llama_memory_hybrid
ff24f387
danielhanchen llama: QSA sparse attention for qwen4exp
c88c9166
danielhanchen llama: give the qwen4exp indexer cache the attention cache's slots
1b599101
danielhanchen tests: record what the qwen4exp arch-test skip actually observes
dd896276
danielhanchen tests: cover qwen4exp in test-llama-archs
40e90413
danielhanchen convert: stream the qwen4exp PLE table instead of concatenating it
53c0f624
danielhanchen quantize: fall back to F16 for 32-block types with an odd ncols
570c26d3
danielhanchen quantize: let --tensor-type name per_layer_token_embd
4597a86f
danielhanchen quantize: size the output buffer exactly instead of nelements * 4
25aa77a4
danielhanchen qwen4exp: hash the image placeholder for multimodal batches
eb95d125
qwen4exp: support a non-unified KV cache in QSA
bea3b12d
github-actions github-actions added model
github-actions github-actions added testing
github-actions github-actions added conversion
ggml-gh-bot
ggerganov ggerganov marked this pull request as ready for review 75 days ago
ggerganov ggerganov requested a review from JohannesGaessler JohannesGaessler 75 days ago
ggerganov ggerganov requested a review from CISC CISC 75 days ago
ggerganov ggerganov requested a review from ggerganov ggerganov 75 days ago
ggerganov ggerganov assigned ggerganov ggerganov 75 days ago
danielhanchen
Green-Sky
danielhanchen
ggerganov
ggerganov commented on 2026-08-26
danielhanchen
JJJYmmm
am17an
Rose22
ggerganov
danielhanchen
Dampfinchen
Green-Sky
PmNz8
Dampfinchen
Green-Sky
am17an
lingyezhixing
PmNz8
miifanboy
danielhanchen llama: keep the qwen4exp top-k attention mask arch-local
7fee670a
danielhanchen llama: hold the qwen4exp indexer cache in a new llama_memory_hybrid_idx
035e2273
danielhanchen
ngxson
ngxson commented on 2026-08-26
ServeurpersoCom
jamespa16
ServeurpersoCom
ServeurpersoCom
ggml-org ggml-org deleted a comment from Biggles10-claude on 2026-08-26
ngxson
phhusson
jamespa16
danielhanchen
ServeurpersoCom
coder543
imezx
kyuz0
HDANILO
phhusson
ServeurpersoCom
FlorinAndrei
coder543
lingyezhixing
phhusson
anonimousA
ServeurpersoCom
mikki8
coder543
ServeurpersoCom
phhusson
HDANILO
ServeurpersoCom
coder543
ServeurpersoCom
nazeshinjite
ngxson
ngxson
ngxson commented on 2026-08-26
twentyonemore1
jggc
sergey-automation
phhusson
jessegit
tomasslivka
julek-wolfssl
CesarPetrescu
frau
hansihe
ewired
jamfor999
jkyamog
benklop
Tonoken3
benklop
rumgewieselt
danielhanchen
Rose22
GsonZhao
odellus
tomasslivka
danielhanchen
linuxid10t
danielhanchen llama: save and restore the qwen4exp indexer KV cache
cfbdc0a5
danielhanchen llama: make the qwen4exp PLE n-gram history per context and serialise it
d22d2be2
qwen4exp: tidy comments and simplify image token read
d4a943f9
qwen4exp: support a quantized KV cache in the QSA attention path
0ac4b180
llama: give qwen4exp a large-graph node budget
c52ed2a0
qwen4exp: drop an unused variable that breaks -Werror builds
6a69a0c1
quantize: dequantize and quantize large tensors in row bands
ef9fa1ba
xdCloudy
danielhanchen llama: segment the qwen4exp fused QKV for tensor split
5674c73a
danielhanchen llama: fix the qwen4exp PLE history seq_rm(-1) iterator invalidation …
24ea62df
qwen4exp: include llama-impl.h explicitly for llama_mul_mat_hadamard
0b19188e
odellus
danielhanchen convert: fix the qwen4exp lint and type-check failures
213df585
pixel4696
ggml-org ggml-org deleted a comment from Gwonk1 on 2026-08-27
ryant00000
llama: give the qwen4exp indexer cache its own tensor names
fbe17732
qwen4exp: double the Q split granularity for tensor parallelism
6c5afc86
danielhanchen
ggml-org ggml-org deleted a comment from adamlawi on 2026-08-27
ggerganov
ServeurpersoCom
DrSchnitzel57
danielhanchen qwen4exp: keep the indexer cache in step across server slots
8b3ed0a4
danielhanchen kv-cache: clear the cache once when restoring a whole context
f43070aa
danielhanchen kv-cache: check the mirrored slot layout on a whole-context restore too
d1fe507f
danielhanchen qwen4exp: give the PLE conv history its own mirrored recurrent row
bf539602
coolthor
chmryan
mattepiu
hyungraekim-bagelcode
ggml-org ggml-org locked as too heated and limited conversation to collaborators 74 days ago
ngxson
ngxson Merge remote-tracking branch 'origin/master' into qwen4exp/qwen3.8-fl…
b5b5fc30
ngxson no more ple_hist (use master version)
5eeb97d1
danielhanchen llama: give the qwen4exp full memory context its indexer cache
1caf6c57
danielhanchen qwen4exp: shrink the PLE hparams storage
c3259772
llama: opt-in random-access mmap advice for host-resident gather tables
95da4ba8
llama: narrow the random-access mmap advice to the gather table
58325573
danielhanchen llama: fold the random-access prefetch into its own feature flag
b8bdf73b
ngxson
ngxson Merge commit 'refs/pull/27742/head' of github.com:ggerganov/llama.cpp…
69a1269f
CISC
ngxson
danielhanchen
danielhanchen
danielhanchen
CISC
danielhanchen
danielhanchen
ngxson
danielhanchen
ngxson
CISC FACP (Fewer Acronym Classes Please)
e5227ab2
ngxson
danielhanchen
danielhanchen qwen4exp: bias the QSA selection per block, not per cell
ef687669
ngxson
ngxson clean up code comments
5b34ce29
ngxson Merge commit 'refs/pull/27742/head' of github.com:ggerganov/llama.cpp…
56672bf6
ngxson clean up new comments
24391470
ngxson revert LLAMA_MMAP_RANDOM
41b4b707
ngxson Merge remote-tracking branch 'origin/master' into qwen4exp/qwen3.8-fl…
539c35e9
ngxson nits
af1ffaf3
ngxson
ngxson replace some changes with #27795
250b6144
ngxson
CISC
ngxson
danielhanchen
ggerganov
ngxson
ngxson improve the m-rope image for get_prev_tokens
fbc5135a
ServeurpersoCom
ngxson LazyChunkedTensor
a510c82e
ngxson Merge remote-tracking branch 'origin/master' into qwen4exp/qwen3.8-fl…
3d7baaeb
ngxson
CISC
ServeurpersoCom
ngxson fix lint
072f519f
ggml-org ggml-org unlocked this conversation 73 days ago
solarkyle
ggerganov
mjungnickel18
ServeurpersoCom
ggerganov
ggerganov
ggerganov commented on 2026-08-27
solarkyle
mjungnickel18
ngxson add some validations
d7ea031a
ngxson
ngxson commented on 2026-08-27
ggerganov
ngxson
kyunghyunr
ngxson reduce input nodes
e3613033
ggerganov
ngxson trim output tokens
dfb5c139
ngxson nits
2014107a
ngxson some more sanity checks
61b62d84
sirfyyn
Kononnable
ggerganov
ngxson fix llm_graph_input_ple reuse
d807f04e
kyunghyunr
ggerganov
ggerganov approved these changes on 2026-08-27
ServeurpersoCom
ServeurpersoCom approved these changes on 2026-08-27
ServeurpersoCom
ngxson
coeniebeyers
ngxson exclude from webgpu test
eaf93765
ngxson ngxson merged 6c84c7d5 into master 73 days ago
erazortt
ServeurpersoCom
erazortt
SamuelBerezhinskiy
NightMustEnd
GlidLov
drew913s
sjvrensburg
lviobio
anthonykim3197
MCFBMCFB
fulvian

Login to write a write a comment.

Login via GitHub

Reviewers
Assignees
Labels
model testing conversion
Milestone
No milestone
Terms Of Use Privacy Policy Refund Policy Impressum