Rebase b10549 - #217

Merged

gianni-cor merged 537 commits into

tetherto:temp-10549from

gagallo7:qvac-b10549

Sep 7, 2026

This pull request is big! We're only showing the most recent 250 commits

Commits on Aug 26, 2026

ggml-opencl: support antialiased BILINEAR upscale when upsampling

olyasir
authored and
JenySadadia
committed
479e77a
feat: QVAC-21320 Mali GPU projector optimizations — disable FA, warptile, layernorm fusion

iancris
authored and
JenySadadia
committed
c666851
fix: QVAC-21320 engage the Mali FA-disable on temp-9341 (coopmat != efficient FA)

authored and
JenySadadia
committed
e618e47
fix: QVAC-21320 keep clip FA on backends without the efficient-FA query (Metal/CUDA)

authored and
JenySadadia
committed
7ed0aaa
fix: QVAC-21320 tiled NORM dispatch — never exceed maxComputeWorkGroupCount

authored and
JenySadadia
committed
4c367ae
fix: QVAC-21914 downgrade non-coopmat clip FA hard-disable to AUTO (budget-aware)

iancris
authored and
JenySadadia
committed
322b50a
fix: QVAC-21914 bound ggml-opencl submissions (periodic clFlush + FA q-chunking)

iancris
authored and
JenySadadia
committed
399c794
fix: QVAC-21914 review fixes — work-budget flush, memory-clamp rework, tests

iancris
authored and
JenySadadia
committed
2435637
fix: QVAC-21914 make clip_fa_effective_min_kv inline (Windows DLL link)

iancris
authored and
JenySadadia
committed
a7de831
fix: QVAC-21914 address PR review — warn on env clamp, cover n_head guard

iancris
authored and
JenySadadia
committed
b269fe8
fix: QVAC-21914 flush after enqueuing the budget-crossing node

iancris
authored and
JenySadadia
committed
f428081
chore: QVAC-21914 address PR review — flush-cadence wording, nits

iancris
authored and
JenySadadia
committed
e42944e
vulkan: restore the nb00 element stride in the fused norm shader so non-contiguous (permuted) inputs read the right columns

zoq
authored and
JenySadadia
committed
6f0fbfa
cuda: Fix OUT_PROD op support claim

gagallo7
authored and
JenySadadia
committed
16d3267
ggml-opencl: build flash-attn kernels without finite-math

olyasir
authored and
JenySadadia
committed
ff2886c
ggml-opencl: treat null attention mask as bidirectional, not causal

olyasir
authored and
JenySadadia
committed
47823ac
ggml-opencl: add trailing barrier in f32/f16 flash-attn tile loop + guard upscale zero dims

iancris
authored and
gagallo7
committed
c30d480
server tests: detect cache restore via timings, not exact log text
gagallo7
committed
d06cb74
ci: install jinja2 explicitly in the venv so test-jinja-py doesn't depend on the torch pin surviving pip install

zoq
authored and
gagallo7
committed
5831b77
ci: add backend op-coverage manifest guard + SVE variant tripwire
gagallo7
committed
d2c1d56
tests: add no-mask n_q == n_kv FLASH_ATTN_EXT cases (vision-tower shape)

gagallo7
and
claude
committed
9ac8d74
tests: add partial-tile FLASH_ATTN_EXT cases across KV-type kernel variants

gagallo7
and
claude
committed
588330e
vulkan: add GFX1151 q8_0 BK64 matmul path

gianni-cor
and
gagallo7
committed
caafff6
cpu: optimize DeepSeek4 HC post

gianni-cor
authored and
gagallo7
committed
d9cc483
vulkan: add Lightning Indexer

gianni-cor
authored and
gagallo7
committed
937b2e8
vulkan: add DeepSeek4 HC comb

gianni-cor
authored and
gagallo7
committed
07c9b55
vulkan: add DeepSeek4 HC pre

gianni-cor
authored and
gagallo7
committed
581ad23
vulkan: add DeepSeek4 HC post

gianni-cor
authored and
gagallo7
committed
dd72ed5
tests: cover DeepSeek4 fused operations

gianni-cor
authored and
gagallo7
committed
26d28fb
vulkan: add typed K cache support to Lightning Indexer

gianni-cor
authored and
gagallo7
committed
3cc4893
tests: cover typed Lightning Indexer K caches

gianni-cor
authored and
gagallo7
committed
d5038fa
vulkan: support quantized concat

gianni-cor
authored and
gagallo7
committed
e05c1b4
metal: support quantized concat

gianni-cor
authored and
gagallo7
committed
5c74a92
metal: support strided f16 add

gianni-cor
authored and
gagallo7
committed
1a6dd2e
meta: handle Lightning Indexer split state

gianni-cor
authored and
gagallo7
committed
1897ad0
ci: extend macOS ARM test timeout

gianni-cor
authored and
gagallo7
committed
31eff5c
ci: make nproc portable and verify it in self-hosted deps
gagallo7
committed
d14b82a
ci: fail fast when jinja2 is missing from the CI venv
gagallo7
committed
3727a65
Add CHANGELOG.md

makaveli10
authored and
gagallo7
committed
4fcb071
ggml-opt: abort training when the backend graph compute fails instead of accumulating garbage results

zoq
authored and
gagallo7
committed
93d4311
ggml-metal: default n_cb back to 1 on iOS to fix the A18 GPU hang when resuming finetuning from a pause checkpoint

zoq
authored and
gagallo7
committed
eb211b7
ggml-metal: only track memory ranges when the encoder is concurrent, fixing the A18 GPU hang from barriers encoded into serial encoders

zoq
authored and
gagallo7
committed
b2be108
ggml-metal: avoid per-iteration whole-block copies in the quantized OUT_PROD kernels so finetuning backward passes stay under the iOS GPU watchdog

zoq
authored and
gagallo7
committed
c0d8083
vulkan: drain glslc stdout/stderr concurrently
gagallo7
committed
1a9eda5
QVAC-21550 infra: roll out canonical security baseline (TruffleHog + CodeQL) to qvac-fabric-llm.cpp (#177)

GSServita
authored and
gagallo7
committed
dc55c98
QVAC-22747 infra: add weekly schedule to CodeScan caller (#189)

GSServita
authored and
gagallo7
committed
58bc4a7
QVAC-22740 infra: bump CodeScan caller to qvac-actions 0.3.0 (#195)

GSServita
authored and
gagallo7
committed
2fd8cb4
chore(ci): remove qvac-collabora-merge from CODEOWNERS

sidj-thr
authored and
gagallo7
committed
831d508
tests: relax dsv4 bit-exact check to ulp-based tolerance

zoq
authored and
gagallo7
committed
7169098
tests: add more cases and perf mode support for backward GDN

korcan-h
authored and
gagallo7
committed
da9e430
ggml-metal: add getter function for threadExecutionWidth

korcan-h
authored and
gagallo7
committed
def2976
ggml-metal: allow ggml_metal_library_get_pipeline_* fns to derive thread dispatch sizes

korcan-h
authored and
gagallo7
committed
7b1d167
ggml-metal: optimizes gated-delta-net back op

korcan-h
authored and
gagallo7
committed
4707626
ggml-cuda: adds support for backwards gated delta-net

korcan-h
authored and
gagallo7
committed
9ba8061
ggml-cuda: adds sigmoid-back

korcan-h
authored and
gagallo7
committed
c19b609
ggml-cuda: implement geglu backward

makaveli10
authored and
gagallo7
committed
1902e8d
ggml-cuda: implement ssm-conv-back-sx

korcan-h
authored and
gagallo7
committed
51a9a1c
ggml-cuda: implement ssm-conv-back-c

korcan-h
authored and
gagallo7
committed
102bca8
QVAC-23076 feat: add VisionPsy Nano and its Flash preprocessing rule

yingying0906
authored and
gagallo7
committed
8c4e945
QVAC-23075 fix: ask the Adreno transpose predicate about the parent, not the view

yingying0906
authored and
gagallo7
committed
91b7742
QVAC-23075 fix: reconstruct the q8_0 parent before reading a view back on Adreno

yingying0906
authored and
gagallo7
committed
631f756
QVAC-23075 fix: size the aspect-preserving refine in double, not float

yingying0906
authored and
gagallo7
committed
760844a
QVAC-23075 fix: reject a no-upscale cap below one slice

yingying0906
authored and
gagallo7
committed
afafe60
QVAC-23075 fix: keep the NUL terminator out of string_format's result

yingying0906
authored and
gagallo7
committed
964119e
QVAC-23075 fix: stop letterboxing the VisionPsy refined image

yingying0906
authored and
gagallo7
committed
9c71651
QVAC-23075 fix: require the slicing sizes whenever VisionPsy loads, flag off included

yingying0906
authored and
gagallo7
committed
3bec113
QVAC-23075 test: golden tests for the idefics3 refined size and slice grid

yingying0906
authored and
gagallo7
committed
1e8614c
QVAC-23075 test: pin the projector alias to general.name

yingying0906
authored and
gagallo7
committed
6ad87c9
QVAC-23075 test: load-time coverage for the preprocessing metadata and the no-upscale override

yingying0906
authored and
gagallo7
committed
dd2df30
QVAC-23075 test: check the VisionPsy tile structure in tests.sh, not just the answer

yingying0906
authored and
gagallo7
committed
34b748a
QVAC-23075 fix: reject a zero image_size on idefics3 too, not only VisionPsy

yingying0906
authored and
gagallo7
committed
f06ec1a
QVAC-23075 docs: say that libmtmd consumers must rebuild on every update

yingying0906
authored and
gagallo7
committed
383a002
QVAC-23075 test: assert the VisionPsy prompt structure, not only the encode count

yingying0906
authored and
gagallo7
committed
b351f10
QVAC-23075 fix: format string_format into the result, not a vector

yingying0906
authored and
gagallo7
committed
8650707
QVAC-23075 test: skip the two tests that need unexported symbols on Windows DLL builds

yingying0906
authored and
gagallo7
committed
b80d05c
QVAC-23075 fix: bound the idefics3-style preprocessing metadata at load

yingying0906
authored and
gagallo7
committed
9bd3df4
opencl : fix Adreno MoE repack aliasing

Alok-Ranjan23
authored and
gagallo7
committed
06f8011
opencl : keep MoE repack words in GEMM order

Alok-Ranjan23
authored and
gagallo7
committed
8623b95
opencl : add CPU MoE repack diagnostic

Alok-Ranjan23
authored and
gagallo7
committed
27a3ed3
opencl : default MoE CPU repack on E031.47

Alok-Ranjan23
authored and
gagallo7
committed
92ffe44
opencl : detect Adreno 830 MoE repack workaround

Alok-Ranjan23
authored and
gagallo7
committed
4cf105d
opencl : disable Q4_K MoE MUL_MAT_ID on Adreno 8xx

Alok-Ranjan23
authored and
gagallo7
committed
eb58138
opencl : scope Q4_K MoE guard to the E031.47 compiler

Alok-Ranjan23
authored and
gagallo7
committed
06bc2a7
opencl : extend MoE guard to every affected quantization

Alok-Ranjan23
authored and
gagallo7
committed
0576793
opencl : skip padded MoE output stores

Alok-Ranjan23
authored and
gagallo7
committed
699911d
opencl : fix Adreno nibble-packing miscompile in trans4_ns repack

Alok-Ranjan23
authored and
gagallo7
committed
c0d6089
opencl : fix get_scale_min_k4 miscompile in MoE matmul kernels

Alok-Ranjan23
authored and
gagallo7
committed
16e71e6
test : cover Adreno MoE repack shapes in test-backend-ops

Alok-Ranjan23
authored and
gagallo7
committed
29c83df
opencl : re-enable MoE MUL_MAT_ID on Adreno 8xx

Alok-Ranjan23
authored and
gagallo7
committed
4a4cf4e
test : fail Adreno repack cases on a broken weight roundtrip

Alok-Ranjan23
authored and
gagallo7
committed
13a37f7
ggml-cuda: implement gelu backward

makaveli10
authored and
gagallo7
committed
575921b
ggml-cuda: implement l2_norm backward

makaveli10
authored and
gagallo7
committed
09b4788
ggml-vulkan: fix l2_norm_back for non-contiguous inputs

makaveli10
authored and
gagallo7
committed
325e283
fixup! QVAC-23075 test: golden tests for the idefics3 refined size and slice grid
gagallo7
committed
1450539
ggml : add vector index C API foundation

dev-nid
authored and
gagallo7
committed
697967f
ggml : harden vector-index foundation

dev-nid
authored and
gagallo7
committed
fa2d97d
ggml : fix vector-index shared build coverage

dev-nid
authored and
gagallo7
committed
757dd5f
ggml : harden vector-index snapshot tests

dev-nid
authored and
gagallo7
committed
0818fab
ggml : harden vector index persistence and packaging

dev-nid
authored and
gagallo7
committed
e483482
ggml : fix vector-index API and package coverage

dev-nid
authored and
gagallo7
committed
b0e3ca5
ggml : fix vector-index packaging and test coverage

dev-nid
authored and
gagallo7
committed
32873c3
ggml : renumber vector-index error codes

dev-nid
authored and
gagallo7
committed
2ae44c4
ggml: harden vector index foundation

dev-nid
authored and
gagallo7
committed
df0d244
ggml : fix vector-index snapshot file handling

dev-nid
authored and
gagallo7
committed
c340b39
ggml-vector-index : amortize add capacity growth

dev-nid
authored and
gagallo7
committed
5221b52
ggml-vector-index : harden snapshot rename durability

dev-nid
authored and
gagallo7
committed
e0974be
ggml : reject oversized vector index snapshots on write

dev-nid
authored and
gagallo7
committed
1f0921a
ggml : add vector index snapshot helper header

dev-nid
authored and
gagallo7
committed
1885b9e
ggml : harden vector-index durability and packaging

dev-nid
authored and
gagallo7
committed
7fde694
ggml : harden vector-index durability and packaging

dev-nid
authored and
gagallo7
committed
6d48eb2
tests : cover failed vector-index snapshot overwrite

dev-nid
authored and
gagallo7
committed
ae64979
Fixing metal run issues in CI by now uses the venv Python for both pip installs and the jinja2 import check.

dev-nid
authored and
gagallo7
committed
6a5d923
Revert "Fixing metal run issues in CI by now uses the venv Python for both pip installs and the jinja2 import check."

dev-nid
authored and
gagallo7
committed
9038218
ci : use venv python for model conversion

dev-nid
authored and
gagallo7
committed
647836d
ci : revert venv python conversion

dev-nid
authored and
gagallo7
committed
b8c4318
ggml : add q4 q8 vector index search

dev-nid
authored and
gagallo7
committed
aa8efb9
ggml : restore vector-index utility APIs

dev-nid
authored and
gagallo7
committed
cc9a41d
ggml : reject trailing vector-index snapshots

dev-nid
authored and
gagallo7
committed
891b3f0
docs : expand vector-index component map

dev-nid
authored and
gagallo7
committed
6ebd2c2
ggml : fix AVX2 vector-index q4 q8 search build

dev-nid
authored and
gagallo7
committed
41672b8
ggml : fix vector-index q4/q8 search issues

dev-nid
authored and
gagallo7
committed
fcdea9d
ggml : harden vector index persistence

dev-nid
authored and
gagallo7
committed
f9471ef
ggml : fix vector-index all_finite linkage

dev-nid
authored and
gagallo7
committed
d27b7ca
ggml : fix vector-index helper linkage

dev-nid
authored and
gagallo7
committed
a0ceac0
Ensure the standalone vector-index package links libm and includes it in

dev-nid
authored and
gagallo7
committed
1eda6d1
ggml : preserve vector index ranking precision

dev-nid
authored and
gagallo7
committed
1f94984
ggml : fix vector-index test hook linkage

dev-nid
authored and
gagallo7
committed
9795a2a
ggml : fix vector-index PR2a review issues

dev-nid
authored and
gagallo7
committed
b5b1daf
ggml : add vector-index filter and IVF coverage

dev-nid
authored and
gagallo7
committed
b85efab
tests : strengthen vector-index filter and IVF coverage

dev-nid
authored and
gagallo7
committed
ae215a3
tests : cover IVF partial nprobe routing

dev-nid
authored and
gagallo7
committed
bddb52d
ggml : validate IVF state before empty search

dev-nid
authored and
gagallo7
committed
3067e1d
tests : strengthen vector-index filter and IVF coverage

dev-nid
authored and
gagallo7
committed
fec189e
ggml : add vector-index snapshots and mmap loading

dev-nid
authored and
gagallo7
committed
6fc1d30
ggml : harden vector index persistence

dev-nid
authored and
gagallo7
committed
1870a98
ggml : harden vector index persistence

dev-nid
authored and
gagallo7
committed
dc0fdc4
vector-index : fix delta replay CI failures

dev-nid
authored and
gagallo7
committed
e236802
ggml : clarify vector-index mmap and delta contracts

dev-nid
authored and
gagallo7
committed
ad282f2
ggml : add vector-index delta log persistence

dev-nid
authored and
gagallo7
committed
d804f11
ggml : fix vector-index delta rebase fallout

dev-nid
authored and
gagallo7
committed
2f3c03b
ggml : harden vector index delta logging

dev-nid
authored and
gagallo7
committed
dbdc626
vector-index: address persistence review issues

dev-nid
authored and
gagallo7
committed
75889b3
tests : harden vector-index contention and fault coverage

dev-nid
authored and
gagallo7
committed
c373497
ggml : fix vector-index pkg-config metadata

dev-nid
authored and
gagallo7
committed
ee81940
ggml : fix vector-index static pkg-config libs

dev-nid
authored and
gagallo7
committed
72b58c5
ggml-vector-index : block reuse after unsynced delta append

dev-nid
authored and
gagallo7
committed
5d28e81
ggml : harden vector-index delta logging

dev-nid
authored and
gagallo7
committed
fd63b3d
ggml : add TurboVec q2 q4 CPU indexes

dev-nid
authored and
gagallo7
committed
1704b69
tests : relax TurboVec x86 parity checks

dev-nid
authored and
gagallo7
committed
b78be31
tests : update delta append stale-writer expectations

dev-nid
authored and
gagallo7
committed
e08be34
ggml : catch up stale delta writers

dev-nid
authored and
gagallo7
committed
a308dc7
tests : add TurboVec mutation and v3 corruption coverage

dev-nid
authored and
gagallo7
committed
0e73c35
ggml : document and test turbovec numeric parity

dev-nid
authored and
gagallo7
committed
eaa4be3
tests : fix rebased vector-index test helpers

dev-nid
authored and
gagallo7
committed
2396c80
ggml : fix vector-index rebase follow-ups

dev-nid
authored and
gagallo7
committed
0281f83
tests : fix vector-index rebase test fallout

dev-nid
authored and
gagallo7
committed
c8fe648
harden TurboVec replay and calibration handling

dev-nid
authored and
gagallo7
committed
a5961b5
ggml : fix vector index persistence and TurboVec compatibility

dev-nid
authored and
gagallo7
committed
d0e8bb0
tests : restore vector-index delta replay coverage

dev-nid
authored and
gagallo7
committed
6f9ffb0
ggml : fix q8 q4 quantization rounding mode

dev-nid
authored and
gagallo7
committed
bf2d561
ggml : optimize TurboVec query search

dev-nid
authored and
gagallo7
committed
50e83ff
ggml : make TurboVec parity deterministic

dev-nid
authored and
gagallo7
committed
3f88c52
tests : keep vector-index production test public

dev-nid
authored and
gagallo7
committed
7d5a333
tests : avoid writer starvation in vector-index concurrency case

dev-nid
authored and
gagallo7
committed
1e3b9ea
tests : fix rebased vector-index helpers

dev-nid
authored and
gagallo7
committed
b05589d
tests : preserve vector-index persistence contracts

dev-nid
authored and
gagallo7
committed
edb586b
ggml : fix rebased vector-index delta replay

dev-nid
authored and
gagallo7
committed
09effb0
ggml : fix vector-index test hook declaration

dev-nid
authored and
gagallo7
committed
3d6e0e6
tests : strengthen vector-index delta and mmap coverage

dev-nid
authored and
gagallo7
committed
d1ff063
DSV4: sm tensor

am17an
authored and
gagallo7
committed
5cc888a
set coarser granularity for head splits

am17an
authored and
gagallo7
committed
6c2bcba
rpc : allow `-sm tensor` on RDMA enabled devices

am17an
authored and
gagallo7
committed
5696c96
fix split

am17an
authored and
gagallo7
committed
5f05ff6
no spec decoding

amangupta-tether
authored and
gagallo7
committed
92ff6b6
implement vulkan tensor parallel

amangupta-tether
authored and
gagallo7
committed
ea67b8c
cleanup: remove minor negotiation

amangupta-tether
authored and
gagallo7
committed
c952567
cleanup: add rpc sync

amangupta-tether
authored and
gagallo7
committed
49d5ada
cleanup: add timeout

amangupta-tether
authored and
gagallo7
committed
0a9b061
address review comments

amangupta-tether
authored and
gagallo7
committed
b948971
address further review comments

amangupta-tether
authored and
gagallo7
committed
32395af
enable fusion with the RPC backend

amangupta-tether
authored and
gagallo7
committed
c3927a6
add checked overflow, add all-reduce op ids

am17an
authored and
gagallo7
committed
340e438
extend bounds check for 1-d helpers too

am17an
authored and
gagallo7
committed
ce1ee4b
add wire bf16 opt-out

am17an
authored and
gagallo7
committed
3e0b455
align with upstream

amangupta-tether
authored and
gagallo7
committed
e4072dc
address review comments

amangupta-tether
authored and
gagallo7
committed
8c9b94f
fix webgpu tests

amangupta-tether
authored and
gagallo7
committed
08ea46e
port fix from 210, fix zero size tensor

amangupta-tether
authored and
gagallo7
committed
be2cc81
RPC: Pipeline parallelism

amangupta-tether
authored and
gagallo7
committed
0263046
backend: support multi-buffer tensor resolution

amangupta-tether
authored and
gagallo7
committed
e26e69e
tests: add RPC backend integration coverage

amangupta-tether
authored and
gagallo7
committed
fb997fc
fix windows test

am17an
authored and
gagallo7
committed
1ecd31b
add double signal guard

am17an
authored and
gagallo7
committed
6cc2e7b
address review comments

amangupta-tether
authored and
gagallo7
committed
a97cd71
address remaining review comments

amangupta-tether
authored and
gagallo7
committed
1aa5f65
RPC: add shared comm for drafter

am17an
authored and
gagallo7
committed
5657f23
vulkan improvements

am17an
authored and
gagallo7
committed
9aa389e
add rpc init test

amangupta-tether
authored and
gagallo7
committed
6eeb22e

Commits on Aug 27, 2026

Commits on Aug 28, 2026

metal : fuse MoE topk softmax/argsort/get_rows

amangupta-tether
authored and
gagallo7
committed
691f9f8
metal : fuse decode gate+up SWIGLU

amangupta-tether
authored and
gagallo7
committed
ee3ab0b
metal : fuse MUL_MAT_ID with routing-weight MUL

amangupta-tether
authored and
gagallo7
committed
de950c1
metal : fuse SSM_CONV+SILU and use 32-wide decode conv

amangupta-tether
authored and
gagallo7
committed
1923f61
metal : cut Q4_K matvec rows per simdgroup

amangupta-tether
authored and
gagallo7
committed
ed9af1d
metal : vectorize gated_delta_net inner loop

amangupta-tether
authored and
gagallo7
committed
cf38343
metal : fuse UNARY+MUL and Q4_0 gate+up SWIGLU

amangupta-tether
authored and
gagallo7
committed
fc6f1dd
cleanup

amangupta-tether
authored and
gagallo7
committed
ad5ae74
Optimize for dense model

amangupta-tether
authored and
gagallo7
committed
2db25da
model: add Qwen3.8-Flash-Next (qwen4exp) (#27742)

authored and
gagallo7
committed
f0d0320
llama-context: exclude ACCEL backends from the pipeline-parallel gate

dev-nid
authored and
gagallo7
committed
7fa667c
llama: model_loader: add TENSOR_READ_LAZY (#27794)

ngxson
authored and
gagallo7
committed
0dcf151
address review comments

amangupta-tether
authored and
gagallo7
committed
6b739be
metal: address remaining review notes on gdn fusion

amangupta-tether
authored and
gagallo7
committed
0ccfc8b
llama: add token ID tracking to KV cell (#27762)

ngxson
authored and
gagallo7
committed
88c3e6d
QVAC-23796 infra: point the self-hosted CI jobs at the qvac fleet

covert-oddity
authored and
gagallo7
committed
c126dd7
rpc: connect to multiple endpoints concurrently, not serially (#221)

dev-nid
authored and
gagallo7
committed
26bcabf
fit: budget against real memory availability

simon-iribarren
authored and
gagallo7
committed
d648cf2
fit: repair the shared-memory fold per review

simon-iribarren
authored and
gagallo7
committed
3c5f1ee
fit: extract the decision arithmetic into pure, tested functions

simon-iribarren
authored and
gagallo7
committed
d015012
test: gate the Metal expectation on GGML_METAL, not on the host arch

simon-iribarren
authored and
gagallo7
committed
5534cec
fit: split the shared pool budget across shares_host devices

simon-iribarren
authored and
gagallo7
committed
5ce28c2
cuda: suppress GCC 11 -Wstringop-overflow false positive in host code

gagallo7
and
claude
committed
a2d972f

Commits on Aug 30, 2026

llama: override LLAMA_LOAD_MODE_AUTO to NONE for memory-buffer loads
zoq
committed
a29f422

Commits on Aug 31, 2026

context : disable non-fused GDN and LID ops

ggerganov
authored and
gagallo7
committed
62ae601
model: qwen4exp: reduce number of graph splits

ngxson
authored and
gagallo7
committed
f65eb76
kv-cells: stop the sequence scan once all sequences are seen

ServeurpersoCom
authored and
gagallo7
committed
51082b3
vulkan: top_k radix select for k >= 1024 for Qwen 3.8 Flash Next (#28032)

0cc4m
authored and
gagallo7
committed
9aab30d

Commits on Sep 1, 2026

ggml-cuda: implement mat-mul-id back A
korcan-h
committed
7b08331
ggml-cuda: implement mat-mul-id back B
korcan-h
committed
3677dd4

Commits on Sep 2, 2026

Commits on Sep 3, 2026

ggml-vulkan: bounds-check the ALIGNED tile loads in mul_mm
korcan-h
committed
6a8e94d

Commits on Sep 4, 2026

ggml-cuda: add cross-entropy loss masked forward/backward
makaveli10
committed
d4ce813
ggml-cuda: add count-equal masked
makaveli10
committed
002ed24
ggml-vulkan: read the upstream gradient on the device in cross_entropy_loss_masked_back
makaveli10
committed
271b1b6
common: only enable thinking when the SFT data carries reasoning
makaveli10
committed
976b9b9
ggml-vulkan: fix get_rows_back for batched (ne2/ne3) inputs
makaveli10
committed
117f93c
vulkan: optimize bounds-check in mul_mm ALIGNED tile loads

gagallo7
and
korcan-h
committed
6239750
vulkan: add GFX1151 q8_0 BK64 matmul path

gagallo7
authored and
korcan-h
committed
bba74e8
QVAC-23773: configurable MoE cache for hot experts (#232)

authored and
gagallo7
committed
8e9b76b
ggml-cuda: support batched and strided get_rows_back for the MoE router backward
zoq
committed
b584747
ggml: support quantized expert weights in mul_mat_id_back_b on CUDA and Vulkan
zoq
committed
b7341b1

Commits on Sep 6, 2026