Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2027 commits
Select commit Hold shift + click to select a range
63e3e5e
fix: QVAC-21320 tiled NORM dispatch — never exceed maxComputeWorkGrou…
iancris Jul 6, 2026
2528a79
fix: QVAC-21914 downgrade non-coopmat clip FA hard-disable to AUTO (b…
iancris Jul 7, 2026
64a6f45
fix: QVAC-21914 bound ggml-opencl submissions (periodic clFlush + FA …
iancris Jul 7, 2026
f98e2f0
fix: QVAC-21914 review fixes — work-budget flush, memory-clamp rework…
iancris Jul 8, 2026
ae58bd5
fix: QVAC-21914 make clip_fa_effective_min_kv inline (Windows DLL link)
iancris Jul 8, 2026
ab05a0d
fix: QVAC-21914 address PR review — warn on env clamp, cover n_head g…
iancris Jul 8, 2026
cedd217
fix: QVAC-21914 flush after enqueuing the budget-crossing node
iancris Jul 8, 2026
d8c5e0d
chore: QVAC-21914 address PR review — flush-cadence wording, nits
iancris Jul 8, 2026
b8fced9
vulkan: restore the nb00 element stride in the fused norm shader so n…
zoq Jul 13, 2026
b6a2e30
cuda: Fix OUT_PROD op support claim
gagallo7 Jul 13, 2026
82f7d73
ggml-opencl: build flash-attn kernels without finite-math
olyasir Jul 14, 2026
361cac3
ggml-opencl: treat null attention mask as bidirectional, not causal
olyasir Jul 14, 2026
23e9f63
ggml-opencl: add trailing barrier in f32/f16 flash-attn tile loop + g…
iancris Jul 14, 2026
08b49c7
server tests: detect cache restore via timings, not exact log text
gagallo7 Jul 30, 2026
69d5b10
ci: install jinja2 explicitly in the venv so test-jinja-py doesn't de…
zoq Jul 7, 2026
a430532
ci: add backend op-coverage manifest guard + SVE variant tripwire
gagallo7 Jul 30, 2026
6fb34e6
tests: add no-mask n_q == n_kv FLASH_ATTN_EXT cases (vision-tower shape)
gagallo7 Jul 30, 2026
82f59eb
tests: add partial-tile FLASH_ATTN_EXT cases across KV-type kernel va…
gagallo7 Jul 30, 2026
ebbebf4
vulkan: add GFX1151 q8_0 BK64 matmul path
gianni-cor Jul 24, 2026
c9f7c0e
cpu: optimize DeepSeek4 HC post
gianni-cor Jul 24, 2026
fae8843
vulkan: add Lightning Indexer
gianni-cor Jul 24, 2026
2c22367
vulkan: add DeepSeek4 HC comb
gianni-cor Jul 24, 2026
a92f3c9
vulkan: add DeepSeek4 HC pre
gianni-cor Jul 24, 2026
4f911da
vulkan: add DeepSeek4 HC post
gianni-cor Jul 24, 2026
5ec775d
tests: cover DeepSeek4 fused operations
gianni-cor Jul 24, 2026
e53230b
vulkan: add typed K cache support to Lightning Indexer
gianni-cor Jul 24, 2026
df4458f
tests: cover typed Lightning Indexer K caches
gianni-cor Jul 24, 2026
8a897fd
vulkan: support quantized concat
gianni-cor Jul 25, 2026
f05b15a
metal: support quantized concat
gianni-cor Jul 25, 2026
1114bc6
metal: support strided f16 add
gianni-cor Jul 25, 2026
b7caa2a
meta: handle Lightning Indexer split state
gianni-cor Jul 27, 2026
c130feb
ci: extend macOS ARM test timeout
gianni-cor Jul 28, 2026
0aa6003
ci: make nproc portable and verify it in self-hosted deps
gagallo7 Jul 31, 2026
8782913
ci: fail fast when jinja2 is missing from the CI venv
gagallo7 Jul 31, 2026
a68d2c9
Add CHANGELOG.md
makaveli10 Aug 4, 2026
3ec631e
ggml-opt: abort training when the backend graph compute fails instead…
zoq Aug 5, 2026
d4f3516
ggml-metal: default n_cb back to 1 on iOS to fix the A18 GPU hang whe…
zoq Aug 5, 2026
3b0bc0c
ggml-metal: only track memory ranges when the encoder is concurrent, …
zoq Aug 6, 2026
93ae011
ggml-metal: avoid per-iteration whole-block copies in the quantized O…
zoq Aug 6, 2026
0dd5f09
vulkan: drain glslc stdout/stderr concurrently
gagallo7 Aug 7, 2026
81f6a08
QVAC-21550 infra: roll out canonical security baseline (TruffleHog + …
GSServita Jul 7, 2026
3c2d848
QVAC-22747 infra: add weekly schedule to CodeScan caller (#189)
GSServita Jul 30, 2026
4bb0944
QVAC-22740 infra: bump CodeScan caller to qvac-actions 0.3.0 (#195)
GSServita Jul 31, 2026
43d7a33
chore(ci): remove qvac-collabora-merge from CODEOWNERS
sidj-thr Jun 17, 2026
6d4976c
squash! metal: port OUT_PROD, SILU_BACK, SOFT_MAX_BACK, RMS_NORM_BACK…
gagallo7 Aug 12, 2026
6d99a73
tests: relax dsv4 bit-exact check to ulp-based tolerance
zoq Aug 12, 2026
d1d8d4c
squash! hip: allow-list rwkv_wkv_f32<128> VGPRs
gagallo7 Aug 12, 2026
3d89ce0
ggml-webgpu: fix CI errors from #25025 and #25262 (#26566)
yomaytk Aug 11, 2026
6349262
tests: add more cases and perf mode support for backward GDN
korcan-h Aug 10, 2026
ea17f3a
ggml-metal: add getter function for threadExecutionWidth
korcan-h Aug 10, 2026
88e8627
ggml-metal: allow ggml_metal_library_get_pipeline_* fns to derive thr…
korcan-h Aug 10, 2026
f07b377
ggml-metal: optimizes gated-delta-net back op
korcan-h Aug 10, 2026
ad6b938
ggml-cuda: adds support for backwards gated delta-net
korcan-h Aug 13, 2026
88f8a05
ggml-cuda: adds sigmoid-back
korcan-h Aug 17, 2026
9f77126
ggml-cuda: implement geglu backward
makaveli10 Aug 17, 2026
fd20a88
ggml-cuda: implement ssm-conv-back-sx
korcan-h Aug 18, 2026
cdf9eee
ggml-cuda: implement ssm-conv-back-c
korcan-h Aug 18, 2026
9864532
QVAC-23075 feat: add VisionPsy Nano and its Flash preprocessing rule
yingying0906 Aug 10, 2026
fb7a3f6
QVAC-23075 fix: ask the Adreno transpose predicate about the parent, …
yingying0906 Aug 10, 2026
56b5f68
QVAC-23075 fix: reconstruct the q8_0 parent before reading a view bac…
yingying0906 Aug 12, 2026
3199094
QVAC-23075 fix: size the aspect-preserving refine in double, not float
yingying0906 Aug 12, 2026
6cc2d55
QVAC-23075 fix: reject a no-upscale cap below one slice
yingying0906 Aug 12, 2026
1716572
QVAC-23075 fix: keep the NUL terminator out of string_format's result
yingying0906 Aug 12, 2026
42c05a2
QVAC-23075 fix: stop letterboxing the VisionPsy refined image
yingying0906 Aug 12, 2026
28ecc63
QVAC-23075 fix: require the slicing sizes whenever VisionPsy loads, f…
yingying0906 Aug 12, 2026
67086f9
QVAC-23075 test: golden tests for the idefics3 refined size and slice…
yingying0906 Aug 13, 2026
9014aee
QVAC-23075 test: pin the projector alias to general.name
yingying0906 Aug 13, 2026
e90e8f1
QVAC-23075 test: load-time coverage for the preprocessing metadata an…
yingying0906 Aug 13, 2026
f1dfa68
QVAC-23075 test: check the VisionPsy tile structure in tests.sh, not …
yingying0906 Aug 13, 2026
22a16f5
QVAC-23075 fix: reject a zero image_size on idefics3 too, not only Vi…
yingying0906 Aug 13, 2026
ae83d65
QVAC-23075 docs: say that libmtmd consumers must rebuild on every update
yingying0906 Aug 13, 2026
7832231
QVAC-23075 test: assert the VisionPsy prompt structure, not only the …
yingying0906 Aug 13, 2026
3558406
QVAC-23075 fix: format string_format into the result, not a vector
yingying0906 Aug 13, 2026
50d5721
QVAC-23075 test: skip the two tests that need unexported symbols on W…
yingying0906 Aug 13, 2026
19cbde4
squash! ci : functionally test the OpenCL ops on an Intel CPU ICD
yingying0906 Aug 13, 2026
221188e
QVAC-23075 fix: bound the idefics3-style preprocessing metadata at load
yingying0906 Aug 14, 2026
160101b
opencl : fix Adreno MoE repack aliasing
Alok-Ranjan23 Aug 12, 2026
d37bc4e
opencl : keep MoE repack words in GEMM order
Alok-Ranjan23 Aug 12, 2026
12ea236
opencl : add CPU MoE repack diagnostic
Alok-Ranjan23 Aug 14, 2026
54d4c33
opencl : default MoE CPU repack on E031.47
Alok-Ranjan23 Aug 14, 2026
3fe566d
opencl : detect Adreno 830 MoE repack workaround
Alok-Ranjan23 Aug 14, 2026
dae86a8
opencl : disable Q4_K MoE MUL_MAT_ID on Adreno 8xx
Alok-Ranjan23 Aug 14, 2026
f5ec1e7
opencl : scope Q4_K MoE guard to the E031.47 compiler
Alok-Ranjan23 Aug 14, 2026
979fd17
opencl : extend MoE guard to every affected quantization
Alok-Ranjan23 Aug 14, 2026
93470c9
opencl : skip padded MoE output stores
Alok-Ranjan23 Aug 14, 2026
421afcd
opencl : fix Adreno nibble-packing miscompile in trans4_ns repack
Alok-Ranjan23 Aug 17, 2026
4a5ec54
opencl : fix get_scale_min_k4 miscompile in MoE matmul kernels
Alok-Ranjan23 Aug 17, 2026
49341e4
test : cover Adreno MoE repack shapes in test-backend-ops
Alok-Ranjan23 Aug 17, 2026
ed0f5ed
opencl : re-enable MoE MUL_MAT_ID on Adreno 8xx
Alok-Ranjan23 Aug 17, 2026
8e8948f
test : fail Adreno repack cases on a broken weight roundtrip
Alok-Ranjan23 Aug 17, 2026
098f48a
squash! Vulkan: Add MUL_MAT_MAT and MUL_MAT_VEC support for TQ1
gagallo7 Aug 19, 2026
108b204
ggml-cuda: implement gelu backward
makaveli10 Aug 20, 2026
f590e03
ggml-cuda: implement l2_norm backward
makaveli10 Aug 20, 2026
c604302
ggml-vulkan: fix l2_norm_back for non-contiguous inputs
makaveli10 Aug 20, 2026
99f7694
Merge pull request #207 from gagallo7/qvac-b10297
gianni-cor Aug 21, 2026
454dcb5
ggml : add vector index C API foundation
dev-nid Jul 27, 2026
5341d6f
ggml : harden vector-index foundation
dev-nid Jul 27, 2026
9fee82d
ggml : fix vector-index shared build coverage
dev-nid Jul 28, 2026
3441005
ggml : harden vector-index snapshot tests
dev-nid Jul 28, 2026
269ee25
ggml : harden vector index persistence and packaging
dev-nid Jul 28, 2026
d0b9bcd
ggml : fix vector-index API and package coverage
dev-nid Jul 28, 2026
57717f1
ggml : fix vector-index packaging and test coverage
dev-nid Jul 28, 2026
4e5e1ac
ggml : renumber vector-index error codes
dev-nid Jul 30, 2026
d7afc1a
ggml: harden vector index foundation
dev-nid Jul 31, 2026
58c6217
ggml : fix vector-index snapshot file handling
dev-nid Jul 31, 2026
4194862
ggml-vector-index : amortize add capacity growth
dev-nid Aug 3, 2026
1028b56
ggml-vector-index : harden snapshot rename durability
dev-nid Aug 4, 2026
8aae819
ggml : reject oversized vector index snapshots on write
dev-nid Aug 6, 2026
6724a32
ggml : add vector index snapshot helper header
dev-nid Aug 6, 2026
23eb318
ggml : harden vector-index durability and packaging
dev-nid Aug 10, 2026
7d075b2
ggml : harden vector-index durability and packaging
dev-nid Aug 10, 2026
09b4cf8
tests : cover failed vector-index snapshot overwrite
dev-nid Aug 11, 2026
c9e950c
Fixing metal run issues in CI by now uses the venv Python for both pi…
dev-nid Aug 11, 2026
29d3136
Revert "Fixing metal run issues in CI by now uses the venv Python for…
dev-nid Aug 11, 2026
9fbf6c6
ci : use venv python for model conversion
dev-nid Aug 11, 2026
d357109
ci : revert venv python conversion
dev-nid Aug 11, 2026
e6dadc0
ggml : add q4 q8 vector index search
dev-nid Jul 29, 2026
09c4537
ggml : restore vector-index utility APIs
dev-nid Jul 29, 2026
67c5d4e
ggml : reject trailing vector-index snapshots
dev-nid Jul 29, 2026
eeedefa
docs : expand vector-index component map
dev-nid Jul 29, 2026
535fcd4
ggml : fix AVX2 vector-index q4 q8 search build
dev-nid Jul 30, 2026
8175363
ggml : fix vector-index q4/q8 search issues
dev-nid Jul 31, 2026
378d2be
ggml : harden vector index persistence
dev-nid Jul 31, 2026
fef1c64
ggml : fix vector-index all_finite linkage
dev-nid Aug 4, 2026
01f334b
ggml : fix vector-index helper linkage
dev-nid Aug 4, 2026
0623f8d
Ensure the standalone vector-index package links libm and includes it in
dev-nid Aug 4, 2026
08e2c6b
ggml : preserve vector index ranking precision
dev-nid Aug 6, 2026
40da56e
ggml : fix vector-index test hook linkage
dev-nid Aug 6, 2026
e9926d3
ggml : fix vector-index PR2a review issues
dev-nid Aug 10, 2026
56d0e10
ggml : add vector-index filter and IVF coverage
dev-nid Jul 29, 2026
83cb86f
tests : strengthen vector-index filter and IVF coverage
dev-nid Jul 31, 2026
f3f9a2f
tests : cover IVF partial nprobe routing
dev-nid Jul 31, 2026
9579dca
ggml : validate IVF state before empty search
dev-nid Aug 6, 2026
79806ee
tests : strengthen vector-index filter and IVF coverage
dev-nid Aug 10, 2026
d509d5f
ggml : add vector-index snapshots and mmap loading
dev-nid Jul 29, 2026
4fcbc0c
ggml : harden vector index persistence
dev-nid Aug 3, 2026
93f1620
ggml : harden vector index persistence
dev-nid Aug 3, 2026
859e0f8
vector-index : fix delta replay CI failures
dev-nid Aug 5, 2026
c976f53
ggml : clarify vector-index mmap and delta contracts
dev-nid Aug 10, 2026
22d20b4
ggml : add vector-index delta log persistence
dev-nid Jul 29, 2026
e36975b
ggml : fix vector-index delta rebase fallout
dev-nid Aug 3, 2026
4040740
ggml : harden vector index delta logging
dev-nid Aug 3, 2026
7d7a984
vector-index: address persistence review issues
dev-nid Aug 3, 2026
36d1662
tests : harden vector-index contention and fault coverage
dev-nid Aug 4, 2026
878a58a
ggml : fix vector-index pkg-config metadata
dev-nid Aug 5, 2026
e741f32
ggml : fix vector-index static pkg-config libs
dev-nid Aug 5, 2026
09afef8
ggml-vector-index : block reuse after unsynced delta append
dev-nid Aug 6, 2026
ae37496
ggml : harden vector-index delta logging
dev-nid Aug 10, 2026
3184bae
ggml : add TurboVec q2 q4 CPU indexes
dev-nid Jul 27, 2026
1785e5e
tests : relax TurboVec x86 parity checks
dev-nid Jul 27, 2026
73795e0
tests : update delta append stale-writer expectations
dev-nid Jul 27, 2026
85ff377
ggml : catch up stale delta writers
dev-nid Jul 28, 2026
a080c53
tests : add TurboVec mutation and v3 corruption coverage
dev-nid Jul 28, 2026
055d236
ggml : document and test turbovec numeric parity
dev-nid Jul 30, 2026
2c1821d
tests : fix rebased vector-index test helpers
dev-nid Jul 31, 2026
47dafab
ggml : fix vector-index rebase follow-ups
dev-nid Aug 3, 2026
b64efc9
tests : fix vector-index rebase test fallout
dev-nid Aug 3, 2026
4c1341b
harden TurboVec replay and calibration handling
dev-nid Aug 3, 2026
3e6dc05
ggml : fix vector index persistence and TurboVec compatibility
dev-nid Aug 3, 2026
0759124
tests : restore vector-index delta replay coverage
dev-nid Aug 4, 2026
596131c
ggml : fix q8 q4 quantization rounding mode
dev-nid Aug 5, 2026
d1417dd
ggml : optimize TurboVec query search
dev-nid Aug 5, 2026
610b70f
ggml : make TurboVec parity deterministic
dev-nid Aug 6, 2026
2513e09
tests : keep vector-index production test public
dev-nid Aug 6, 2026
5f35c2a
tests : avoid writer starvation in vector-index concurrency case
dev-nid Aug 7, 2026
f4cfa48
tests : fix rebased vector-index helpers
dev-nid Aug 10, 2026
21602b7
tests : preserve vector-index persistence contracts
dev-nid Aug 10, 2026
d4a2a40
ggml : fix rebased vector-index delta replay
dev-nid Aug 10, 2026
cdbd68b
ggml : fix vector-index test hook declaration
dev-nid Aug 10, 2026
3fc662d
tests : strengthen vector-index delta and mmap coverage
dev-nid Aug 11, 2026
2b6f222
Merge pull request #211 from dev-nid/turbovec-cpu-pr1-core
gianni-cor Aug 24, 2026
6cb6f78
DSV4: sm tensor
am17an Jul 31, 2026
8df0ede
set coarser granularity for head splits
am17an Aug 3, 2026
4ec1ad8
rpc : allow `-sm tensor` on RDMA enabled devices
am17an Aug 4, 2026
b60682a
fix split
am17an Aug 5, 2026
62f63a5
no spec decoding
amangupta-tether Aug 5, 2026
0ca2df4
implement vulkan tensor parallel
amangupta-tether Aug 10, 2026
06ec528
cleanup: remove minor negotiation
amangupta-tether Aug 12, 2026
c7c1238
cleanup: add rpc sync
amangupta-tether Aug 12, 2026
11017ea
cleanup: add timeout
amangupta-tether Aug 12, 2026
7505107
address review comments
amangupta-tether Aug 12, 2026
4fccf82
address further review comments
amangupta-tether Aug 13, 2026
8e63961
enable fusion with the RPC backend
amangupta-tether Aug 14, 2026
5b58565
add checked overflow, add all-reduce op ids
am17an Aug 17, 2026
597bb8d
extend bounds check for 1-d helpers too
am17an Aug 17, 2026
57c6199
add wire bf16 opt-out
am17an Aug 18, 2026
a1c7d47
align with upstream
amangupta-tether Aug 24, 2026
b790b68
address review comments
amangupta-tether Aug 25, 2026
658f70c
fix webgpu tests
amangupta-tether Aug 25, 2026
897814c
port fix from 210, fix zero size tensor
amangupta-tether Aug 25, 2026
c057875
RPC: Pipeline parallelism
amangupta-tether Aug 6, 2026
623e5fa
backend: support multi-buffer tensor resolution
amangupta-tether Aug 10, 2026
c166c6e
tests: add RPC backend integration coverage
amangupta-tether Aug 10, 2026
e0f8fdc
fix windows test
am17an Aug 19, 2026
a8b2484
add double signal guard
am17an Aug 19, 2026
f8be8b2
address review comments
amangupta-tether Aug 25, 2026
5cc1f0d
address remaining review comments
amangupta-tether Aug 25, 2026
a0a7da5
RPC: add shared comm for drafter
am17an Aug 20, 2026
85e7804
vulkan improvements
am17an Aug 21, 2026
60f4a19
add rpc init test
amangupta-tether Aug 24, 2026
29db2fb
Merge pull request #210 from tetherto/mtp_drafter
amangupta-tether Aug 25, 2026
3e101bf
metal : fuse MoE topk softmax/argsort/get_rows
amangupta-tether Aug 26, 2026
a76c5a4
metal : fuse decode gate+up SWIGLU
amangupta-tether Aug 26, 2026
8f5adff
metal : fuse MUL_MAT_ID with routing-weight MUL
amangupta-tether Aug 26, 2026
d98f855
metal : fuse SSM_CONV+SILU and use 32-wide decode conv
amangupta-tether Aug 26, 2026
8d45565
metal : cut Q4_K matvec rows per simdgroup
amangupta-tether Aug 26, 2026
fb8a4ea
metal : vectorize gated_delta_net inner loop
amangupta-tether Aug 26, 2026
ab572e2
metal : fuse UNARY+MUL and Q4_0 gate+up SWIGLU
amangupta-tether Aug 26, 2026
2ceeb6c
cleanup
amangupta-tether Aug 26, 2026
93430e4
Optimize for dense model
amangupta-tether Aug 27, 2026
2e431fe
model: add Qwen3.8-Flash-Next (qwen4exp) (#27742)
danielhanchen Aug 27, 2026
576af58
llama-context: exclude ACCEL backends from the pipeline-parallel gate
dev-nid Aug 28, 2026
e56b889
llama: model_loader: add TENSOR_READ_LAZY (#27794)
ngxson Aug 27, 2026
3669bfd
address review comments
amangupta-tether Aug 28, 2026
ab483bd
Merge pull request #220 from tetherto/fix/pipeline-parallel-ignore-ac…
gianni-cor Aug 28, 2026
3dd70c8
metal: address remaining review notes on gdn fusion
amangupta-tether Aug 28, 2026
7c32b6c
Merge pull request #216 from tetherto/metal-opt
gianni-cor Aug 28, 2026
fa53740
llama: add token ID tracking to KV cell (#27762)
ngxson Aug 26, 2026
80e06b8
QVAC-23796 infra: point the self-hosted CI jobs at the qvac fleet
covert-oddity Aug 28, 2026
973836d
rpc: connect to multiple endpoints concurrently, not serially (#221)
dev-nid Aug 28, 2026
b79b1c6
Merge pull request #218 from tetherto/qwen4-next
gianni-cor Aug 28, 2026
44db907
Merge pull request #225 from tetherto/infra/QVAC-23796-temp-10297
gianni-cor Aug 28, 2026
c7fdd86
CUDA: fix build warning for non blackwell builds
amangupta-tether Aug 28, 2026
2951370
fit: budget against real memory availability
simon-iribarren Aug 25, 2026
3a8d167
fit: repair the shared-memory fold per review
simon-iribarren Aug 26, 2026
5c629bf
fit: extract the decision arithmetic into pure, tested functions
simon-iribarren Aug 26, 2026
c3ce47f
test: gate the Metal expectation on GGML_METAL, not on the host arch
simon-iribarren Aug 26, 2026
5854ddd
fit: split the shared pool budget across shares_host devices
simon-iribarren Aug 27, 2026
2324760
Merge remote-tracking branch 'origin/master' into merge-temp-10297-in…
gianni-cor Aug 28, 2026
a756518
deps: update dependencies flagged by Socket
gianni-cor Aug 28, 2026
e9e7f5b
Potential fix for pull request finding 'CodeQL / Workflow does not co…
gianni-cor Aug 28, 2026
ccff56f
deps: pin transformers to compatible 5.x
gianni-cor Aug 28, 2026
aed05e1
conversion: remove unsupported qwen4exp decorators
gianni-cor Aug 28, 2026
68b89b6
tests: relax OpenVINO skinny matmul tolerance
gianni-cor Aug 28, 2026
2f21f39
style: fix editorconfig violations
gianni-cor Aug 28, 2026
a0d4d93
ui: escape backslashes in nerdamer latex names
gianni-cor Aug 28, 2026
73b5f30
ci: limit self-hosted workflow token permissions
gianni-cor Aug 28, 2026
94357c6
Merge pull request #214 from tetherto/fix/fit-host-memory-budget
gianni-cor Aug 28, 2026
82ffc97
ui: escape nerdamer latex names in one pass
gianni-cor Aug 28, 2026
5cb0282
cuda: guard NVFP4 helper on Blackwell support
gianni-cor Aug 28, 2026
aa90461
ci: keep gguf split test binaries stable
gianni-cor Aug 28, 2026
a6a2ef2
Merge pull request #226 from tetherto/nvfp4-fix-block
gianni-cor Aug 28, 2026
a3ce387
context : disable non-fused GDN and LID ops
ggerganov Aug 28, 2026
1e9bcd3
model: qwen4exp: reduce number of graph splits
ngxson Aug 28, 2026
787a93e
kv-cells: stop the sequence scan once all sequences are seen
ServeurpersoCom Aug 30, 2026
65c7c96
vulkan: top_k radix select for k >= 1024 for Qwen 3.8 Flash Next (#28…
0cc4m Aug 31, 2026
9d181fe
Merge pull request #229 from amangupta-tether/qwen4-next-fixes
amangupta-tether Aug 31, 2026
cd91714
Merge remote-tracking branch 'origin/temp-10297' into merge-temp-1029…
gianni-cor Sep 1, 2026
f5f2151
ui: avoid constants barrel in settings registry
gianni-cor Sep 1, 2026
4250a5e
Merge branch 'master' into merge-temp-10297-into-master
gianni-cor Sep 4, 2026
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
The table of contents is too big for display.
Diff view
Diff view
  •  
  •  
  •  
66 changes: 54 additions & 12 deletions .github/workflows/build-apple.yml
Original file line number Diff line number Diff line change
Expand Up @@ -8,21 +8,35 @@ on:
paths: [
'.github/workflows/build-apple.yml',
'**/CMakeLists.txt',
'**/.cmake',
'**/*.cmake',
'**/*.cmake.in',
'**/*.pc.in',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp',
'**/*.swift',
'**/*.m',
'**/*.metal'
'**/*.metal',
'scripts/vector-index-package-smoke.sh'
]

pull_request:
types: [opened, synchronize, reopened]
paths: [
'.github/workflows/build-apple.yml',
'ggml/src/ggml-metal/**'
'**/CMakeLists.txt',
'**/*.cmake',
'**/*.cmake.in',
'**/*.pc.in',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp',
'**/*.swift',
'**/*.m',
'**/*.metal',
'scripts/vector-index-package-smoke.sh'
]

concurrency:
Expand All @@ -36,9 +50,6 @@ env:
LLAMA_ARG_LOG_PREFIX: 1
LLAMA_ARG_LOG_TIMESTAMPS: 1

permissions:
contents: read

jobs:
macos-latest-arm64:
runs-on: macos-latest
Expand All @@ -63,10 +74,11 @@ jobs:
-DCMAKE_BUILD_RPATH="@loader_path" \
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_BUILD_BORINGSSL=ON \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=OFF \
-DGGML_METAL_SHADER_DEBUG=ON \
-DGGML_RPC=ON
-DGGML_VECTOR_INDEX=ON \
-DGGML_RPC=ON \
-DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
leaks -atExit -- ./build/bin/test-thread-safety -hf ggml-org/gemma-3-270m-qat-GGUF -ngl 99 -p "$(printf 'hello %.0s' {1..128})" -n 16 -c 512 -ub 32 -np 2 -t 2 -lv 1

Expand Down Expand Up @@ -102,6 +114,7 @@ jobs:
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_BUILD_BORINGSSL=ON \
-DGGML_METAL=OFF \
-DGGML_VECTOR_INDEX=ON \
-DGGML_RPC=ON \
-DGGML_BLAS=OFF \
-DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
Expand All @@ -113,6 +126,39 @@ jobs:
cd build
ctest -L main --verbose --timeout 900

- name: Vector-index static pkg-config smoke
id: vector_index_static_pkg_config
run: |
STATIC_PREFIX="$RUNNER_TEMP/vector-index-static-install"
cmake -S ggml -B build-vector-index-package-static \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_INSTALL_PREFIX="$STATIC_PREFIX" \
-DBUILD_SHARED_LIBS=OFF \
-DGGML_BUILD_TESTS=OFF \
-DGGML_BUILD_EXAMPLES=OFF \
-DGGML_VECTOR_INDEX=ON
cmake --build build-vector-index-package-static -j $(sysctl -n hw.logicalcpu)
cmake --install build-vector-index-package-static
PREFIX="$STATIC_PREFIX" \
CONSUMER_DIR=vector-index-pkg-config-static \
RUN_CMAKE=0 \
PKG_CONFIG_STATIC=1 \
bash scripts/vector-index-package-smoke.sh
cat > ggml-pkg-config-static.c <<'EOF'
#include "ggml.h"

int main(void) {
ggml_time_init();
return ggml_time_us() < 0;
}
EOF
PKG_CONFIG_PATH="$STATIC_PREFIX/lib/pkgconfig" \
cc ggml-pkg-config-static.c \
$(PKG_CONFIG_PATH="$STATIC_PREFIX/lib/pkgconfig" \
pkg-config --cflags --libs --static ggml) \
-o ggml-pkg-config-static
./ggml-pkg-config-static

macos-latest-ios-xcode:
runs-on: macos-latest

Expand All @@ -130,7 +176,6 @@ jobs:
run: |
sysctl -a
cmake -B build -G Xcode \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_OPENSSL=OFF \
-DLLAMA_BUILD_APP=OFF \
Expand Down Expand Up @@ -181,7 +226,6 @@ jobs:
run: |
sysctl -a
cmake -B build -G Xcode \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_BUILD_COMMON=OFF \
-DLLAMA_BUILD_APP=OFF \
Expand Down Expand Up @@ -215,7 +259,6 @@ jobs:
run: |
sysctl -a
cmake -B build -G Xcode \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_BUILD_COMMON=OFF \
-DLLAMA_BUILD_APP=OFF \
Expand Down Expand Up @@ -260,7 +303,6 @@ jobs:
run: |
sysctl -a
cmake -B build -G Xcode \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_OPENSSL=OFF \
-DLLAMA_BUILD_APP=OFF \
Expand Down
3 changes: 0 additions & 3 deletions .github/workflows/build-cache.yml
Original file line number Diff line number Diff line change
Expand Up @@ -9,9 +9,6 @@ concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true

permissions:
contents: read

jobs:
ubuntu-24-vulkan-cache:
runs-on: ubuntu-24.04
Expand Down
3 changes: 0 additions & 3 deletions .github/workflows/build-cann.yml
Original file line number Diff line number Diff line change
Expand Up @@ -33,9 +33,6 @@ env:
LLAMA_ARG_LOG_PREFIX: 1
LLAMA_ARG_LOG_TIMESTAMPS: 1

permissions:
contents: read

jobs:
openEuler-latest-cann:
defaults:
Expand Down
23 changes: 23 additions & 0 deletions .github/workflows/build-cmake-pkg.yml
Original file line number Diff line number Diff line change
Expand Up @@ -21,10 +21,33 @@ jobs:
-DLLAMA_BUILD_TOOLS=OFF \
-DLLAMA_BUILD_EXAMPLES=OFF \
-DLLAMA_BUILD_APP=OFF \
-DGGML_VECTOR_INDEX=ON \
-DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release
cmake --install build --prefix "$PREFIX" --config Release

PREFIX="$PREFIX" \
CONSUMER_DIR=vector-index-cmake-pkg \
RUN_PKG_CONFIG=0 \
bash scripts/vector-index-package-smoke.sh

GGML_PREFIX="$(pwd)"/inst-ggml
cmake -S ggml -B build-ggml \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_INSTALL_PREFIX="$GGML_PREFIX" \
-DCMAKE_INSTALL_LIBDIR=lib/x86_64-linux-gnu \
-DBUILD_SHARED_LIBS=OFF \
-DGGML_BUILD_TESTS=OFF \
-DGGML_BUILD_EXAMPLES=OFF \
-DGGML_VECTOR_INDEX=ON
cmake --build build-ggml --config Release
cmake --install build-ggml --config Release
PREFIX="$GGML_PREFIX" \
CONSUMER_DIR=vector-index-pkg-config \
LIBDIR=lib/x86_64-linux-gnu \
PKG_CONFIG_STATIC=1 \
bash scripts/vector-index-package-smoke.sh

export LLAMA_CONFIG="$PREFIX"/share/llama/llama-config.cmake
tclsh <<'EOF'
set build(commit) [string trim [exec git rev-parse --short HEAD]]
Expand Down
132 changes: 113 additions & 19 deletions .github/workflows/build-cpu.yml
Original file line number Diff line number Diff line change
Expand Up @@ -7,32 +7,36 @@ on:
- master
paths: [
'.github/workflows/build-cpu.yml',
# Disable. There are some modifications in the fork
# so that ggml dynamic builds work with vcpkg.
# '.github/workflows/build-cmake-pkg.yml',
'.github/workflows/build-cmake-pkg.yml',
'.github/workflows/build-linux-cross.yml',
'scripts/vector-index-package-smoke.sh',
'**/CMakeLists.txt',
'**/.cmake',
'**/*.cmake',
'**/*.cmake.in',
'**/*.pc.in',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp',
'**/*.inc',
]

pull_request:
types: [opened, synchronize, reopened]
paths: [
'.github/workflows/build-cpu.yml',
# Disable. There are some modifications in the fork
# so that ggml dynamic builds work with vcpkg.
# '.github/workflows/build-cmake-pkg.yml',
'.github/workflows/build-cmake-pkg.yml',
'.github/workflows/build-linux-cross.yml',
'scripts/vector-index-package-smoke.sh',
'**/CMakeLists.txt',
'**/.cmake',
'**/*.cmake',
'**/*.cmake.in',
'**/*.pc.in',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp'
'**/*.cpp',
'**/*.inc'
]

concurrency:
Expand All @@ -46,13 +50,7 @@ env:
LLAMA_ARG_LOG_PREFIX: 1
LLAMA_ARG_LOG_TIMESTAMPS: 1

permissions:
contents: read

jobs:
# build-cmake-pkg:
# uses: ./.github/workflows/build-cmake-pkg.yml

ubuntu:
strategy:
matrix:
Expand Down Expand Up @@ -111,6 +109,7 @@ jobs:
run: |
cmake -B build \
-DLLAMA_FATAL_WARNINGS=ON \
-DGGML_VECTOR_INDEX=ON \
-DGGML_RPC=ON
time cmake --build build --config Release -j $(nproc)

Expand All @@ -120,6 +119,101 @@ jobs:
cd build
ctest -L main --verbose --timeout 900

- name: Static vector-index smoke
if: ${{ matrix.build == 'x64' || matrix.build == 'arm64' }}
id: vector_index_static
run: |
cmake -B build-vector-index-static \
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_BUILD_TESTS=ON \
-DGGML_VECTOR_INDEX=ON \
-DBUILD_SHARED_LIBS=OFF
cmake --build build-vector-index-static --target test-vector-index test-vector-index-hooks test-vector-index-faults -j $(nproc)
ctest --test-dir build-vector-index-static -R '^test-vector-index(-hooks|-faults)?$' --output-on-failure

- name: Vector-index package smoke
if: ${{ matrix.build == 'x64' || matrix.build == 'arm64' }}
id: vector_index_package
run: |
PREFIX="$RUNNER_TEMP/vector-index-install"
cmake --install build --prefix "$PREFIX" --config Release

mkdir -p vector-index-consumer
cat > vector-index-consumer/CMakeLists.txt <<'EOF'
cmake_minimum_required(VERSION 3.16)
project(vector_index_consumer C CXX)
find_package(ggml CONFIG REQUIRED)
add_executable(vector_index_consumer main.c)
target_link_libraries(vector_index_consumer PRIVATE ggml::vector-index)
EOF
cat > vector-index-consumer/main.c <<'EOF'
#include "ggml-vector-index.h"

#include <stdint.h>
#include <stdio.h>

int main(void) {
ggml_vec_index_t * idx = ggml_vec_index_create(2, 32);
if (idx == 0) {
return 1;
}
const float vector[2] = { 1.0f, 0.0f };
const uint64_t id = 7;
float score = 0.0f;
uint64_t out_id = 0;
if (ggml_vec_index_add(idx, vector, 1, &id) != GGML_VEC_INDEX_OK ||
ggml_vec_index_search(idx, vector, 1, 1, &score, &out_id) != GGML_VEC_INDEX_OK ||
out_id != id) {
ggml_vec_index_free(idx);
return 2;
}
const char * snapshot_path = "vector-index-consumer.tvim";
if (ggml_vec_index_write(idx, snapshot_path) != GGML_VEC_INDEX_OK) {
ggml_vec_index_free(idx);
return 3;
}
ggml_vec_index_free(idx);
idx = ggml_vec_index_load(snapshot_path);
remove(snapshot_path);
if (idx == 0 || ggml_vec_index_contains(idx, id) != 1) {
ggml_vec_index_free(idx);
return 4;
}
ggml_vec_index_free(idx);
return 0;
}
EOF

cmake -S vector-index-consumer -B vector-index-consumer/build \
-DCMAKE_PREFIX_PATH="$PREFIX/share"
cmake --build vector-index-consumer/build
vector-index-consumer/build/vector_index_consumer

export PKG_CONFIG_PATH="$PREFIX/lib/pkgconfig"
cc vector-index-consumer/main.c \
$(pkg-config --cflags --libs ggml-vector-index) \
-o vector-index-pkg-config
LD_LIBRARY_PATH="$PREFIX/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}" \
./vector-index-pkg-config

STATIC_PREFIX="$RUNNER_TEMP/vector-index-static-install"
STATIC_LIBDIR="lib/$(gcc -print-multiarch)"
cmake -S ggml -B build-vector-index-package-static \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_INSTALL_PREFIX="$STATIC_PREFIX" \
-DCMAKE_INSTALL_LIBDIR="$STATIC_LIBDIR" \
-DBUILD_SHARED_LIBS=OFF \
-DGGML_BUILD_TESTS=OFF \
-DGGML_BUILD_EXAMPLES=OFF \
-DGGML_VECTOR_INDEX=ON
cmake --build build-vector-index-package-static -j $(nproc)
cmake --install build-vector-index-package-static
PREFIX="$STATIC_PREFIX" \
CONSUMER_DIR=vector-index-consumer-static \
LIBDIR="$STATIC_LIBDIR" \
PKG_CONFIG_STATIC=1 \
bash scripts/vector-index-package-smoke.sh

- name: Test llama2c conversion
id: llama2c_test
run: |
Expand All @@ -137,23 +231,23 @@ jobs:
env:
OPENBLAS_VERSION: 0.3.23
SDE_VERSION: 9.33.0-2024-01-07
VULKAN_VERSION: 1.4.313.2
VULKAN_VERSION: 1.4.357.0

strategy:
matrix:
include:
- build: 'x64-cpu-static'
arch: 'x64'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VECTOR_INDEX=ON -DBUILD_SHARED_LIBS=OFF'
- build: 'x64-openblas'
arch: 'x64'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VECTOR_INDEX=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
- build: 'x64-vulkan'
arch: 'x64'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
- build: 'arm64'
arch: 'arm64'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_VECTOR_INDEX=ON'

steps:
- name: Clone
Expand Down
Loading
Loading