Sync master with upstream release b6293#219

Merged

Minh141120 merged 29 commits intodevfrom

update-dev-from-master-2025-08-27-00-11

Aug 27, 2025

Commits on Aug 27, 2025

CANN: ROPE cache sin/cos repeat (#15501)

noemotiovon
authored and
Minh141120
committed
convert : support interns1-mini (#15412)

RunningLeon
authored and
Minh141120
committed
metal : add FA kernels for HS=40 (#15559)

ggerganov
authored and
Minh141120
committed
convert : update Ernie 4.5 dense architecture name (#15555)

ownia
authored and
Minh141120
committed
batched-bench : fix unified KV cache handling + pp timing (#15562)

ggerganov
authored and
Minh141120
committed
model-conversion : add model card template for embeddings [no ci] (#15557)

danbev
authored and
Minh141120
committed
model-conversion : set pooling type to none in logits.cpp (#15564)

danbev
authored and
Minh141120
committed
CUDA: MoE helper in device code, better tile sizes (#15525)

JohannesGaessler
authored and
Minh141120
committed
metal: fix regression when no metal devices are present (#15531)

booxter
authored and
Minh141120
committed
tests: Generate unique input values for count_equal (#15487)

jeffbolznv
authored and
Minh141120
committed
vulkan: fix min subgroup 16 condition for mmid subgroup optimization (#15565)

0cc4m
authored and
Minh141120
committed
opencl: fix support ops condition for `rms_norm` (#15560)

lhez
authored and
Minh141120
committed
CUDA: Accelerate MXFP4 table lookup using `__byte_perm` (#15451)

authored and
Minh141120
committed
vulkan: Remove splitting for mul_mat_id (#15568)

jeffbolznv
authored and
Minh141120
committed
Add a warning for special devices (#15563)

authored and
Minh141120
committed
metal : remove contiguous assertion for src0 in IM2COL (#15577)

CISC
authored and
Minh141120
committed
gguf-py : remove erroneous FFN_GATE entry (#15583)

CISC
authored and
Minh141120
committed
model : support MiniCPM-V 4.5 (#15575)

tc-mb
authored and
Minh141120
committed
metal : improve `MUL_MAT_ID` (#15541)

ggerganov
authored and
Minh141120
committed
context : print graph stats for memory-less contexts (#15586)

ggerganov
authored and
Minh141120
committed
mtmd : support Kimi VL model (#15458)

ngxson
authored and
Minh141120
committed
metal : optimize FA vec for large sequences and BS <= 8 (#15566)

ggerganov
authored and
Minh141120
committed
CUDA: return -1 for nonexistent compiled arch (#15587)

JohannesGaessler
authored and
Minh141120
committed
model-conversion : add qat-q4 quantization targets (#15588)

danbev
authored and
Minh141120
committed
graph : fix assert in memory-less build_attn (#15590)

ggerganov
authored and
Minh141120
committed
llamafile: PowerPC Sgemm Optimization (#15558)

shalinib-ibm
authored and
Minh141120
committed
tests: add performance test for mul mat id (#15543)

netrunnereve
authored and
Minh141120
committed
mtmd : fix mtmd ios build (#15579)

fidoriel
authored and
Minh141120
committed
SYCL: fix rms_norm_mul_add for tensor dim not a multiple of sg_size (#15592)

qnixsynapse
authored and
Minh141120
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Sync master with upstream release b6293#219

Sync master with upstream release b6293#219
Minh141120 merged 29 commits intodevfrom
update-dev-from-master-2025-08-27-00-11

Commits on Aug 27, 2025

CANN: ROPE cache sin/cos repeat (#15501)

convert : support interns1-mini (#15412)

metal : add FA kernels for HS=40 (#15559)

convert : update Ernie 4.5 dense architecture name (#15555)

batched-bench : fix unified KV cache handling + pp timing (#15562)

model-conversion : add model card template for embeddings [no ci] (#15557)

model-conversion : set pooling type to none in logits.cpp (#15564)

CUDA: MoE helper in device code, better tile sizes (#15525)

metal: fix regression when no metal devices are present (#15531)

tests: Generate unique input values for count_equal (#15487)

vulkan: fix min subgroup 16 condition for mmid subgroup optimization (#15565)

opencl: fix support ops condition for `rms_norm` (#15560)

CUDA: Accelerate MXFP4 table lookup using `__byte_perm` (#15451)

vulkan: Remove splitting for mul_mat_id (#15568)

Add a warning for special devices (#15563)

metal : remove contiguous assertion for src0 in IM2COL (#15577)

gguf-py : remove erroneous FFN_GATE entry (#15583)

model : support MiniCPM-V 4.5 (#15575)

metal : improve `MUL_MAT_ID` (#15541)

context : print graph stats for memory-less contexts (#15586)

mtmd : support Kimi VL model (#15458)

metal : optimize FA vec for large sequences and BS <= 8 (#15566)

CUDA: return -1 for nonexistent compiled arch (#15587)

model-conversion : add qat-q4 quantization targets (#15588)

graph : fix assert in memory-less build_attn (#15590)

llamafile: PowerPC Sgemm Optimization (#15558)

tests: add performance test for mul mat id (#15543)

mtmd : fix mtmd ios build (#15579)

SYCL: fix rms_norm_mul_add for tensor dim not a multiple of sg_size (#15592)

Sync master with upstream release b6293#219Minh141120 merged 29 commits intodevjanhq/llama.cpp:devfrom update-dev-from-master-2025-08-27-00-11janhq/llama.cpp:update-dev-from-master-2025-08-27-00-11Copy head branch name to clipboard

Commits

Commits on Aug 27, 2025

Sync master with upstream release b6293#219
Minh141120 merged 29 commits intodevfrom
update-dev-from-master-2025-08-27-00-11