* optimise GGML_OP_SUM
* add non-contiguous tests by permuting the input
* change tests to require full contiguity of OP_SUM
* cuda : add check GGML_OP_SUM
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* CUDA set scheduling strategy to spinning for cc121
* Using prop.major and prop.minor, include HIP and MUSA
* Exclude HIP and MUSA
* Remove trailing whitespace
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Remove empty line
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* metal : pad K, V and Mask when needed
* cont : simplify
* cuda : add TODO about KV padding requirement
* metal : add comments
* metal : remove mask padding requirement
* First attempt
* No permute during convert (fixes qk tensors), proper norm application.
* RoPE = NeoX
* Coherence!
* Migrate xielu params from tensors to hyperparameters
* Simple CUDA kernel
* Revert stupid LLM refactorings
* Chat template support
* configchecker / flake8 errors
* Reorder unary.cu
* I do conclude that LLMs are, in fact, stupid.
* Fix after merge
* Final newline
* Make xIELU an UNARY_OP
* Final newline
* Correctly account for parameter shift
* Argh.
* Update ggml/src/ggml-cpu/unary-ops.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Refactor: remove unused methods, inline and factorize softplus, add const modifiers
* Revert CUDA changes, implement xIELU as a separate OP
* Pesky newline
* Add float2half / half2float for F16 inputs/outputs
* CUDA variants, attempt 2
* Actually, attempt 3
* Update ggml/src/ggml-cuda/unary.cu
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Missing convert header
* Proper formula and reference for xIELU in the comments.
* Modify unary-ops.cpp to add the functor-based logic besides the template system to retain optimizations
* Apply suggestions from code review
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Add tensor mappings for Apertus to global list instead
* Fix lazy on scalars
* Update ggml/src/ggml-cuda/unary.cu
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Add comment about the constraints on positive/negative alpha
* Change `softplus` to `ggml_softplus`
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* HIP: Disable ROCWMMA fatt on CDNA when compiled against ROCWMMA 2.0.0
rocwmma 2.0.0 includes a bug in the code fakeing fp16 accumulation on CDNA
* CUDA: Fix volta condition in ggml_cuda_should_use_wmma_fattn
* CUDA: mul_mat_id for mmf for bs <= 64 for f16 and bs <= 32 for f32
This commit adds mul_mat_id support for ncols_dst >= 16. It does this by
packing ncols_dst tiles into the blockDim.y.
My tests on a RTX 3090 show that this is faster than the cuBLAS fallback
for f16 till bs=64, and for f32 till bs=32
* Review: refactor if statement
* CUDA: add a fused top-K MoE kernel
This kernel does the following:
1. softmax over the logits per token [n_experts, n_tokens]
2. argmax reduce over the top-k (n_experts_used) logits
3. write weights + ids to global memory
It is intended as fusion of softmax->top-k->get_rows pipeline for MoE models
* Refactor into ggml_cuda_should_use_topk_moe
* Review: Use better coalescing pattern, use WARP_SIZE, store logits into registers before
* Review: format + micro-optimizations
* Fix bug: fix tie breakers
* Add optional norm + clean-up code
* Use smem for final write
* Add bounds check
* Use better memory pattern for writeback
* implement set_rows with i32 index
* template fix
* test quantized path
warnings--
* Apply suggestions from code review
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* forgotten name change
* deduplicate cuda/sycl and test-fix
* indent++
* vulkan: support set_rows with i32 index type (#16162)
* disable i32 index for webgpu for now
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Jeff Bolz <jbolz@nvidia.com>
* CUDA: Optimize PAD_REFLECT_1D
feat: add more test cases for PAD_REFLECT_1D
* use fast_div to improve performance
* Apply suggestion from JohannesGaessler
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Apply suggestion from JohannesGaessler
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* optimize
* use a concise expression to further speedup the cuda kernel
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* ggml : remove adding extra dim timestep embedding
This commit updates the ggml_timestep_embedding function to no longer
add an extra dimension when the specified dimension is odd.
The motivation for this change is that this introduces an unnecessary
dimension when the dimension is odd, which caused an issue in the
kernels which were not expecting this extra dimension and it resulted in
uninitialized memory for the second to last dimension.
* ggml-cuda : fix padding in timestep embedding kernel
This commit removes the zeroing out of the last dimension now that we
are not adding the extra padding dimension.
* ggml-metal : fix padding in timestep embedding kernel
This commit fixes the zero padding for odd dimensions in
the timestep embedding kernel
* ggml-opencl : fix padding in timestep embedding kernel
This commit fixes the zero padding for odd dimensions in
the timestep embedding kernel.
* ggml-sycl : fix padding in timestep embedding kernel
This commit fixes the zero padding for odd dimensions in
the timestep embedding kernel.
* ggml-vulkan : fix padding in timestep embedding kernel
This commit fixes the zero padding for odd dimensions in
the timestep embedding kernel.
* ggml-cpu : fix padding in timestep embedding function
This commit removes the zeroing out of the last dimension now that we
are not adding the extra padding dimension.
* fix im2col_3d to respect non-contiguous inputs (views)
The CUDA 3D im2col kernel computed source addresses assuming compact layout (products of dims), ignoring nb[] strides.
This patch switches im2col_3d source indexing to use true strides derived from src1->nb[] (in elements), mirroring the approach used in the 2D CUDA im2col path. Destination indexing is unchanged.
* use ggml_element_size() for src strides
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* ggml-backend : add GGML_BACKEND_DEVICE_TYPE_IGPU device type
ggml-backend : add device id to device props
llama : only use iGPU devices if there are no GPU devices
llama : do not use multiple devices from different backends with the same device id
* Add fastdiv and fastmodulo to k_bin_bcast kernel
* Address review comments
* `prod_` instead of `prod` suffix
* Add test case for `k_bin_bcast_unravel` in CUDA backend
* CUDA: Add mul_mat_id support the mmf
Add support for mul_mat_id for bs < 16
* Review: use warp_size, fix should_use_mmf condition
* Launch one block per expert, stride along n_expert_used
* templatize mul_mat_id
* Pad shmem to 16 bytes, add helper function mul_mat_f_switch_ids
* Reduce compile times by dividing mmf into f16, bf16 and f32 variants
* Divide mmf by ncols_dst
* Add missing files
* Fix MUSA/HIP builds
* vulkan: sort graph to allow more parallel execution
Add a backend proc to allow the backend to modify the graph. The
vulkan implementation looks at which nodes depend on each other
and greedily reorders them to group together nodes that don't
depend on each other. It only reorders the nodes, doesn't change
the contents of any of them.
With #15489, this reduces the number of synchronizations needed.
* call optimize_graph per-split
* cuda : fix supports_op condition for get_rows when src1->ne2 > 1
ggml-ci
* ggml : add comment about ggml_get_rows
ggml-ci
* cuda : add FIXME [no ci]
* cuda : update support condition
ggml-ci
* ggml: allow casting between f32 and i32
* fix cuda
* add vulkan
* fix CPU non-cont
* add non-cont test case
* add note
* extend test number range
* correct note
* add cont version for vulkan
* Add fastdiv, use it in modulo and use modulo in rms_norm_f32
Fastdiv is much faster way to do integer division, which was identified
as bottleneck in rms_norm_f32
* Support more `block_size` values in `rms_norm_f32`
This makes us more flexible in selecting the optimal threads w.r.t
paralellizing across a col vs. launch-overheads of threads and mio
throttles
* Update ggml/src/ggml-cuda/common.cuh
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Replace modulo with fastmodulo in `rms_norm_f32`
* Use `BinPackArguments=true` for formating function calls
Will file a separate PR to adjust .clang-format file
* Update ggml/src/ggml-cuda/common.cuh
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Use uint3 for both `fastdiv` and `fastmodulo`
The compiler seems to reliably optimize away the unused .z component in
the fastdiv use-case, see https://godbolt.org/z/rx8KPrKr3
* More constrained type declarations
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Rename fastdiv and fastmodulo variables to shared variable name
As suggest by JohannesGaessler, this increases clarity of the intended
use
* Pack fastdiv/fastmodulo constants into uint2/uint3 objects
By packing constants to be used together into a struct, we are less
likely to make errors.
* Rename function parameter of fastmodulo
`modulo_consts` is more fitting/descriptive
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* CUDA: fix build error from ambiguous __half conversions in conv2d
Building conv2d with half precision failed because `__half` defines
multiple implicit conversion operators (to float, int, short, etc.),
causing ambiguous overload resolution when multiplying with float.
Introduce a templated `to_float` helper that explicitly converts
`__half` via `__half2float`, while passing through float unchanged.
Use this helper in conv2d accumulation to ensure unambiguous and
correct promotion to float.
Fixes some build errors with half-precision kernels on CUDA.
ggml-ci
* CUDA: Replace custom to_float helper with unified ggml_cuda_cast and add half‑>float conversion
* CUDA: Add missing convert.cuh header
* CUDA: remove unnecessary extension in ggml_cuda_cast
* CUDA: Address review comment, remove second type template argument
Prior to this change, we faced undefined cublasLt references when
attempting to compile 'llama-cli' with GGML_STATIC=ON on Linux.
We add linking with CUDA::cublasLt_static when CUDA version is greater
than 10.1.
* CUDA: optimize get_int_from_table_16
* CUDA: use v_perm_b32 to replace byte_perm on AMD GPUs
* revise documentation
---------
Co-authored-by: xix <xiapc@outlook.com>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* vulkan : support ggml_mean
* vulkan : support sum, sum_rows and mean with non-contiguous tensors
* vulkan : fix subbuffer size not accounting for misalign offset
* tests : add backend-op tests for non-contiguous sum_rows
* cuda : require contiguous src for SUM_ROWS, MEAN support
* sycl : require contiguous src for SUM, SUM_ROWS, ARGSORT support
* require ggml_contiguous_rows in supports_op and expect nb00=1 in the shader
* Add Pad Reflect 1D CUDA support
* Update ggml/src/ggml-cuda/pad_reflect_1d.cu
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* musa: fix build warnings
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
* fix warning: comparison of integers of different signs: 'const int' and 'unsigned int' [-Wsign-compare]
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
---------
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
add expicit conversion operator to support older versions of rocm
Switch over to hip_bf16 from legacy hip_bfloat16
Simplify RDNA3 define
Reduce swap over of new hipblas api to rocm 6.5 as this version is used for rocm 7.0 previews
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* examples/finetune -opt SGD (stochastic gradient descent) memory opt
add unit tested GGML_OPT_OPTIMIZER_SGD to ggml - avoids allocating
m, v tensors.
support finetune.cpp arg -opt SGD (or sgd). (default adamw as before)
llama 3.2-1b-F32 result: observed 11gb gpu ram (41 sec/epoch)
when using SGD instead of 19gb (55 sec/epoch) using adamw.
(wikipedia 100 lines finetune)
(
using the same GPU memory, adamw can only do before OOM 512
batch/context, reaching:
train: [███████▉] data=0000140/0000140 loss=0.02575±0.00099 acc=99.52±0.03% t=00:00:47 ETA=00:00:00
val: [███████▉] data=0000008/0000008 loss=4.76565±0.28810 acc=41.46±0.77% t=00:00:00 ETA=00:00:00
SGD is superior, though it converges slower, with max before OOM 1728
batch/context (esp see the better validation perf):
train: [███████▉] data=0000039/0000039 loss=0.00371±0.00010 acc=99.96±0.01% t=00:00:41 ETA=00:00:00
val: [███████▉] data=0000003/0000003 loss=5.11406±0.76034 acc=48.01±0.69% t=00:00:01 ETA=00:00:00
)
note: when finetuning long enough (or w/ enough -lr),
validation accuracy *eventually* drops ('catastrophic forgetting')
-lr-half (halflife) option useful for SGD to avoid oscillation or
super slow underdamped learning (makes setting -lr more forgiving).
terminal -lr for now is set by lr-halvings i.e. if you want at most
1/8 the inital -lr you set -lr-halvings 3.
note: objective loss not directly comparable between adamw, sgd? -
check perplexity or accuracy or consider relative improvements
for convergence
new finetune args -wd 1e-9 to enable weight decay in sgd or adamw,
and max -epochs N (default 2 as before)
cache (1 - wd*alpha) in 'adamw' opt struct -
no noticeable perf benefit, disabled (still done
for new SGD though)
since opt. memory is pre-allocated, the ggml_opt_get_optimizer_params
would probably be able to change between SGD and AdamW with each epoch
but would need to use adamw for the first (unconfirmed - no cmdline arg
to set such a policy yet)
test-opt checks adamw as before and now sgd (except for a few disabled
tests for sgd only; probably just needs logging values and adding
alternate reference values); tolerance on the 'regression'
test is broader for sgd (so we don't need many more epochs)
* Vulkan: Implement GGML_OP_OPT_STEP_SGD
* tests: Fix OPT_STEP_SGD test-backend-ops
* SGD op param store weight-decay and not 1-alpha*wd
* minor + cosmetic changes
* fix vulkan sgd
* try CI fix
---------
Co-authored-by: 0cc4m <picard12@live.de>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Factor out `reduce_rows_f32` from common.cuh
This increases iteration cycle speed by not having to recompile
every kernel all the time
* Hide memory-latency by loop unrolling in reduce_rows_f32
* Further optimizations to `reduce_rows_f32`
1. Increase threadblock size to better hide latency of memory requests.
As a consequence of bigger threadblocks, do 2-step summation, using
shared memory to communicate results between invocations
2. Use sum_temp array to reduce waits on sum
3. Adjust num_unroll to reflext bigger threadblock
4. Improve default block_dims, increase support for more block_dims
* Add perf tests for `reduce_rows_f32` kernel
* Add heuristic to toggle 128/512 threads based on sm count
Break even point was the minimum of the following multiples.
| GPU Model | Nrow SM Count Multiple |
| ----------- | ----------- |
| RTX 4000 SFF ADA | 2.0x |
| RTX 6000 ADA | 2.5x |
| RTX PRO 6000 Blackwell Max-Q | 3.04x |
| RTX PRO 4500 Blackwell | 3.15x |
* Ensure perf gains also for small ncols and large nrows
Alternative to this, one could have also made the number of unrollings
template-able, but that would require compiling the kernel multiple
times, increasing binary size unnecessarily
* Modify perf and unit-tests
* Apply auto-formatting by clang
* Fix CI build failure
See https://github.com/ggml-org/llama.cpp/actions/runs/16798370266/job/47573716079?pr=15132#step:7:486
Building with VS generator worked though.
* Remove sm_count property from `ggml_backend_cuda_context`
Requested by @JohannesGaessler, and should fix remaining CI issues as a
side-effect
* Add CUB-based implementation for GGML_OP_MEAN
Currently this branch is only executed for nrows==1
* Add heuristics to execute CUB branch only when it brings perf
Heuristics were determined on the following HW:
* RTX 4000 SFF ADA
* RTX 6000 ADA
* RTX PRO 6000 Blackwell Max-Q
* RTX PRO 4500 Blackwell
* Add unit-test for CUB-based mean
Tests should run with CUDA Graphs enabled per default on NVGPUs
* Rename `USE_CUB` to `GGML_CUDA_USE_CUB`
Suggested by @JohannesGaessler
* Unindent Preprocessor directives
See
https://github.com/ggml-org/llama.cpp/pull/15132#discussion_r2269213506
* musa: fix failures in test-backend-ops for mul_mat_id op
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
* Address review comments
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
---------
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
* cuda: refactored ssm_scan to use CUB
* fixed compilation error when when not using CUB
* assign L to constant and use size_t instead of int
* deduplicated functions
* change min blocks per mp to 1
* Use cub load and store warp transpose
* suppress clang warning
This is useful for testing for regressions on GCN with CDNA hardware.
With GGML_HIP_MMQ_MFMA=Off and GGML_CUDA_FORCE_MMQ=On we can conveniently test the GCN code path on CDNA. As CDNA is just GCN renamed with MFMA added and limited use ACC registers, this provides a good alternative for regression testing when GCN hardware is not available.
llvm with the amdgcn target dose not support unrolling loops with conditional break statements, when those statements can not be resolved at compile time. Similar to other places in GGML lets simply ignore this warning.
This commit adds support for MFMA instructions to MMQ. CDNA1/GFX908 CDNA2/GFX90a and CDNA3/GFX942 are supported by the MFMA-enabled code path added by this commit. The code path and stream-k is only enabled on CDNA3 for now as it fails to outperform blas in all cases on the other devices.
Blas is currently only consistently outperformed on CDNA3 due to issues in the amd-provided blas libraries.
This commit also improves the awareness of MMQ towards different warp sizes and as a side effect improves the performance of all quant formats besides q4_0 and q4_1, which regress slightly, on GCN gpus.
* musa: apply mublas API changes
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
* musa: update musa version to 4.2.0
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
* musa: restore MUSA graph settings in CMakeLists.txt
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
* musa: disable mudnnMemcpyAsync by default
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
* musa: switch back to non-mudnn images
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
* minor changes
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
* musa: restore rc in docker image tag
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
---------
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
Sam/Samuel
Julius Tischbein
Aman Gupta
Johannes Gäßler
Anav Prasad
Diego Devesa
ai-fonsi
Georgi Gerganov
Piotr Wilkin (ilintar)
R0CKSTAR
uvos
Sigbjørn Skjæret
Gregor Jasny
Jeff Bolz
Bowen Han
Daniel Bevenius
Jake Karnes
Oliver Simons
Xuan-Son Nguyen
leejet
Akarshan Biswas
mnehete32
compilade
matiaslin
Yoshi_likes_e4
Qeeweew
Acly
Yavor Ivanov
Jonathan Graehl
David Zhao
AN Long
deepsek