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

Merged

ngxson merged 65 commits into

ggml-org:masterfrom

unslothai:qwen4exp/qwen3.8-flash-next

Aug 27, 2026

Commits on Aug 26, 2026

gguf: add qwen4exp (Qwen3.8-Flash-Next) arch and converter
danielhanchen
committed
6e5b8b7
llama: load qwen4exp (Qwen3.8-Flash-Next) hparams and tensors
danielhanchen
committed
331fa7e
qwen4exp: shorten comments
danielhanchen
committed
058b7a1
llama: qwen4exp text graph with hyper-connections, GDN and MoE
danielhanchen
committed
c9ad86f
llama: qwen4exp PLE n-gram hash embedding
danielhanchen
committed
ad4fa3f
llama: carry the qwen4exp PLE conv state across ubatches
danielhanchen
committed
c89e67b
llama: fix the qwen4exp PLE conv state and unblock test-llama-archs
danielhanchen
committed
ddf0980
llama: optional indexer key cache in llama_memory_hybrid
danielhanchen
committed
ff24f38
llama: QSA sparse attention for qwen4exp
danielhanchen
committed
c88c916
llama: give the qwen4exp indexer cache the attention cache's slots
danielhanchen
committed
1b59910
tests: record what the qwen4exp arch-test skip actually observes
danielhanchen
committed
dd89627
tests: cover qwen4exp in test-llama-archs
danielhanchen
committed
40e9041
convert: stream the qwen4exp PLE table instead of concatenating it
danielhanchen
committed
53c0f62
quantize: fall back to F16 for 32-block types with an odd ncols
danielhanchen
committed
570c26d
quantize: let --tensor-type name per_layer_token_embd
danielhanchen
committed
4597a86
quantize: size the output buffer exactly instead of nelements * 4
danielhanchen
committed
25aa77a
qwen4exp: hash the image placeholder for multimodal batches
danielhanchen
committed
eb95d12
qwen4exp: support a non-unified KV cache in QSA

danielhanchen
authored and
danielhanchen
committed
bea3b12
llama: keep the qwen4exp top-k attention mask arch-local
danielhanchen
committed
7fee670
llama: hold the qwen4exp indexer cache in a new llama_memory_hybrid_idx
danielhanchen
committed
035e227

Commits on Aug 27, 2026

llama: save and restore the qwen4exp indexer KV cache
danielhanchen
committed
cfbdc0a
llama: make the qwen4exp PLE n-gram history per context and serialise it
danielhanchen
committed
d22d2be
qwen4exp: tidy comments and simplify image token read

danielhanchen
authored and
danielhanchen
committed
d4a943f
qwen4exp: support a quantized KV cache in the QSA attention path

danielhanchen
authored and
danielhanchen
committed
0ac4b18
llama: give qwen4exp a large-graph node budget

danielhanchen
authored and
danielhanchen
committed
c52ed2a
qwen4exp: drop an unused variable that breaks -Werror builds

danielhanchen
authored and
danielhanchen
committed
6a69a0c
quantize: dequantize and quantize large tensors in row bands

danielhanchen
authored and
danielhanchen
committed
ef9fa1b
llama: segment the qwen4exp fused QKV for tensor split
danielhanchen
committed
5674c73
llama: fix the qwen4exp PLE history seq_rm(-1) iterator invalidation and the fatal-warning build
danielhanchen
committed
24ea62d
qwen4exp: include llama-impl.h explicitly for llama_mul_mat_hadamard

danielhanchen
authored and
danielhanchen
committed
0b19188
convert: fix the qwen4exp lint and type-check failures
danielhanchen
committed
213df58
llama: give the qwen4exp indexer cache its own tensor names

danielhanchen
authored and
danielhanchen
committed
fbe1773
qwen4exp: double the Q split granularity for tensor parallelism

danielhanchen
authored and
danielhanchen
committed
6c5afc8
qwen4exp: keep the indexer cache in step across server slots
danielhanchen
committed
8b3ed0a
kv-cache: clear the cache once when restoring a whole context
danielhanchen
committed
f43070a
kv-cache: check the mirrored slot layout on a whole-context restore too
danielhanchen
committed
d1fe507
qwen4exp: give the PLE conv history its own mirrored recurrent row
danielhanchen
committed
bf53960
Merge remote-tracking branch 'origin/master' into qwen4exp/qwen3.8-flash-next
ngxson
committed
b5b5fc3
no more ple_hist (use master version)
ngxson
committed
5eeb97d
llama: give the qwen4exp full memory context its indexer cache

danielhanchen
and
ServeurpersoCom
committed
1caf6c5
qwen4exp: shrink the PLE hparams storage
danielhanchen
committed
c325977
llama: opt-in random-access mmap advice for host-resident gather tables

danielhanchen
authored and
danielhanchen
committed
95da4ba
llama: narrow the random-access mmap advice to the gather table

danielhanchen
authored and
danielhanchen
committed
5832557
llama: fold the random-access prefetch into its own feature flag
danielhanchen
committed
b8bdf73
Merge commit 'refs/pull/27742/head' of github.com:ggerganov/llama.cpp into qwen4exp/qwen3.8-flash-next
ngxson
committed
69a1269
FACP (Fewer Acronym Classes Please)
CISC
committed
e5227ab
qwen4exp: bias the QSA selection per block, not per cell
danielhanchen
committed
ef68766
clean up code comments
ngxson
committed
5b34ce2
Merge commit 'refs/pull/27742/head' of github.com:ggerganov/llama.cpp into qwen4exp/qwen3.8-flash-next
ngxson
committed
56672bf
clean up new comments
ngxson
committed
2439147
revert LLAMA_MMAP_RANDOM
ngxson
committed
41b4b70
Merge remote-tracking branch 'origin/master' into qwen4exp/qwen3.8-flash-next
ngxson
committed
539c35e
nits
ngxson
committed
af1ffaf
replace some changes with #27795
ngxson
committed
250b614
improve the m-rope image for get_prev_tokens
ngxson
committed
fbc5135
LazyChunkedTensor
ngxson
committed
a510c82
Merge remote-tracking branch 'origin/master' into qwen4exp/qwen3.8-flash-next
ngxson
committed
3d7baae
fix lint
ngxson
committed
072f519
add some validations
ngxson
committed
d7ea031
reduce input nodes
ngxson
committed
e361303
trim output tokens
ngxson
committed
dfb5c13
nits
ngxson
committed
2014107
some more sanity checks
ngxson
committed
61b62d8
fix llm_graph_input_ple reuse
ngxson
committed
d807f04
exclude from webgpu test
ngxson
committed
eaf9376