* vulkan: Use larger workgroups for mul_mat_vec when M is small
Also use subgroup instructions for (part of) the reduction when supported.
Without this, the more expensive reductions would eat into the benefits of
the larger workgroups.
* update heuristic for amd/intel
Co-authored-by: 0cc4m <picard12@live.de>
---------
Co-authored-by: 0cc4m <picard12@live.de>
- Launch an appropriate number of invocations (next larger power of two).
32 invocations is common and the barrier is much cheaper there.
- Specialize for "needs bounds checking" vs not.
- Make the code less branchy and [[unroll]] the loops. In the final code,
I see no branches inside the main loop (only predicated stores) when
needs_bounds_check is false.
- Always sort ascending, then apply the ascending vs descending option when
doing the final stores to memory.
- Copy the values into shared memory, makes them slightly cheaper to access.
* vulkan: fuse adds
Fuse adds that have the same shape, which are common in MoE models.
It will currently fuse up to 6 adds, because we assume no more than
8 descriptors per dispatch. But this could be changed.
* check runtimeDescriptorArray feature
* disable multi_add for Intel due to likely driver bug
* vulkan: Add missing bounds checking to scalar/coopmat1 mul_mat_id
* vulkan: Support mul_mat_id with f32 accumulators, but they are not hooked up
- There's no explicit way to request f32 precision for mul_mat_id, but there
probably should be, and this gets the code in place for that.
- A couple fixes to check_results.
- Remove casts to fp16 in coopmat1 FA shader (found by inspection).
* add F16/F16 fa support
* fix kernel init
* use mad instead of fma
* use inline function
* mark FA with sinks as unsupported for now
* add pragma unroll to loops
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>
* vulkan: perf_logger improvements
- Account for batch dimension in flops calculation.
- Fix how "_VEC" is detected for mat_mul_id.
- Fix "n" dimension for mat_mul_id (in case of broadcasting).
- Include a->type in name.
* use <=mul_mat_vec_max_cols rather than ==1
* 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
* ggml-rpc: chunk send()/recv() to avoid EINVAL for very large tensors over RPC (macOS & others). Fixes#15055
* ggml-rpc: rename RPC_IO_CHUNK->MAX_CHUNK_SIZE, use std::min() for cap, switch to GGML_LOG_ERROR, handle 0-length send/recv
* rpc: drop n==0 special case in send_data(); retry in loop per review
* rpc: remove trailing whitespace in send_data()
---------
Co-authored-by: Shinnosuke Takagi <nosuke@nosukenoMacBook-Pro.local>
* 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
* feat(cann): add optional support for ACL Graph execution
This commit adds support for executing ggml computational graphs using
Huawei's ACL graph mode via the USE_CANN_GRAPH flag. The support can be
enabled at compile time using the CMake option:
-DUSE_CANN_GRAPH=ON
By default, ACL graph execution is **disabled**, and the fallback path
uses node-by-node execution.
Key additions:
- CMake option to toggle graph mode
- Graph capture and execution logic using
- Tensor property matching to determine whether graph update is required
- Safe fallback and logging if the environment variable LLAMA_SET_ROWS
is unset or invalid
This prepares the backend for performance improvements in repetitive graph
execution scenarios on Ascend devices.
Signed-off-by: noemotiovon <757486878@qq.com>
* Fix review comments
Signed-off-by: noemotiovon <757486878@qq.com>
* remane USE_CANN_GRAPH to USE_ACL_GRAPH
Signed-off-by: noemotiovon <757486878@qq.com>
* fix typo
Signed-off-by: noemotiovon <757486878@qq.com>
---------
Signed-off-by: noemotiovon <757486878@qq.com>
* Add paramater buffer pool, batching of submissions, refactor command building/submission
* Add header for linux builds
* Free staged parameter buffers at once
* Format with clang-format
* Fix thread-safe implementation
* Use device implicit synchronization
* Update workflow to use custom release
* Remove testing branch workflow
* Disable set_rows until it's implemented
* Fix potential issue around empty queue submission
* Try synchronous submission
* Try waiting on all futures explicitly
* Add debug
* Add more debug messages
* Work on getting ssh access for debugging
* Debug on failure
* Disable other tests
* Remove extra if
* Try more locking
* maybe passes?
* test
* Some cleanups
* Restore build file
* Remove extra testing branch ci
* cmake: Add GGML_BACKEND_DIR option
This can be used by distributions to specify where to look for backends
when ggml is built with GGML_BACKEND_DL=ON.
* Fix phrasing
* Add parameter buffer pool, batching of submissions, refactor command building/submission
* Add header for linux builds
* Free staged parameter buffers at once
* Format with clang-format
* Fix thread-safe implementation
* Use device implicit synchronization
* Update workflow to use custom release
* Remove testing branch workflow
- Increase tile size for k-quants, to match non-k-quants
- Choose more carefully between large and medium tiles, considering how it
interacts with split_k
- Allow larger/non-power of two split_k, and make the splits a multiple of 256
- Use split_k==3 to when >1/2 and <=2/3 of the SMs would hae been used
* vulkan: optimizations for direct convolution
- Empirically choose a better tile size. Reducing BS_K/BS_NPQ helps fill
the GPU. The new size should be amenable to using coopmat, too.
- Fix shmem bank conflicts. 16B padding should work with coopmat.
- Some explicit loop unrolling.
- Skip math/stores work for parts of the tile that are OOB.
- Apply fastdiv opt.
- Disable shuffles for NV.
* Three tiles sizes for CONV_2D, and a heuristic to choose
* reallow collectives for pre-Turing
* make SHMEM_PAD a spec constant
* fixes for intel perf - no shmem padding, placeholder shader core count
* shader variants with/without unrolling
* 0cc4m's fixes for AMD perf
Co-authored-by: 0cc4m <picard12@live.de>
---------
Co-authored-by: 0cc4m <picard12@live.de>
* Initial Q2_K Block Interleaving Implementation
* Addressed review comments and clean up of the code
* Post rebase fixes
* Initial CI/CD fixes
* Update declarations in arch-fallback.h
* Changes for GEMV Q2_K in arch-fallback.h
* Enable repacking only on AVX-512 machines
* Update comments in repack.cpp
* Address q2k comments
---------
Co-authored-by: Manogna-Sree <elisetti.manognasree@multicorewareinc.com>
The pipeline member can be cast to VkPipeline.
This is a VkPipeline_T* on 64 bit but a uint64_t on 32 bit.
Cf. VK_DEFINE_NON_DISPATCHABLE_HANDLE documentation.
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.
* remove redundant code in riscv
* remove redundant code in arm
* remove redundant code in loongarch
* remove redundant code in ppc
* remove redundant code in s390
* remove redundant code in wasm
* remove redundant code in x86
* remove fallback headers
* fix x86 ggml_vec_dot_q8_0_q8_0
* SYCL: Add set_rows support for quantized types
This commit adds support for GGML_OP_SET_ROWS operation for various
quantized tensor types (Q8_0, Q5_1, Q5_0, Q4_1, Q4_0, IQ4_NL) and BF16
type in the SYCL backend.
The quantization/dequantization copy kernels were moved from cpy.cpp
to cpy.hpp to make them available for set_rows.cpp.
This addresses part of the TODOs mentioned in the code.
* Use get_global_linear_id() instead
ggml-ci
* Fix formatting
ggml-ci
* Use const for ne11 and size_t variables in set_rows_sycl_q
ggml-ci
* Increase block size for q kernel to 256
ggml-ci
* Cleanup imports
* Add float.h to cpy.hpp
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.
* feat: Add s_off as a parameter in the args struct
This may not be necessary, but it more closely mirrors the CUDA kernel
Branch: GraniteFourPerf
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* perf: Parallelize mamba2 SSM_SCAN metal kernel over d_state
This is a first attempt at optimizing the metal kernel. The changes here
are:
- Launch the kernel with a thread group of size d_state
- Use simd groups and shared memory to do the summation for the y
computation
When tested with G4 tiny preview, this shows roughly a 3x speedup on
prefill and 15% speedup on decode.
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix: Update logic to correctly do the multi-layer parallel sum
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix: Correctly size the shared memory bufer and assert expected size relationships
Branch: GraniteFourPerf
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* refactor: Compute block offsets once rather than once per token
Branch: GraniteFourPerf
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Use local variable for state recursion
Branch: GraniteFourPerf
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Use a secondary simd_sum instead of a for loop
Branch: GraniteFourPerf
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Add assertion and comment about relationship between simd size and num simd groups
Branch: GraniteFourPerf
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Parallelize of d_state for mamba-1
Branch: GraniteFourPerf
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Parallel sum in SSM_CONV
Branch: GraniteFourPerf
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* Revert "feat: Parallel sum in SSM_CONV"
After discussion with @compilade, the size of the parallelism here is
not worth the cost in complexity or overhead of the parallel for.
https://github.com/ggml-org/llama.cpp/pull/14743#discussion_r2223395357
This reverts commit 16bc059660.
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* refactor: Simplify shared memory sizing
Branch: GraniteFourPerf
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
Co-Authored-By: Georgi Gerganov <ggerganov@gmail.com>
---------
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Neither "g" nor "x" are valid portPos specifiers per the official
[graphviz documents](https://graphviz.org/docs/attr-types/portPos/):
> If a compass point is used, it must have the form "n","ne","e","se","s","sw","w","nw","c","_".
I tested locally for it to fall back to default portPos specifier if an
invalid portPos is specified. As a consequence, we can remove associated
code.
* 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>
* CMake config: Create target only once
Fix error on repeated find_package(ggml).
For simplicity, check only for the top-level ggml::ggml.
* CMake config: Add CUDA link libs
* CMake config: Add OpenCL link libs
* CMake config: Use canonical find_dependency
Use set and append to control link lib variables.
Apply more $<LINK_ONLY...>.
* CMake config: Wire OpenMP dependency
This commit removes the inclusion of `<cstdlib>`.
The motivation for this change is that this source file does not seem to
use any functions from this header and the comment about `qsort` is a
little misleading/confusing.
Jeff Bolz
Dong Won Kim
rmatif
lhez
Georgi Gerganov
Aaron Teo
Johannes Gäßler
uvos
Jason Ni
Sigbjørn Skjæret
Jonathan Graehl
Judd
Oliver Simons
Tak-RS
Romain Biessy
Aman Gupta
Chenguang Li
R0CKSTAR
hipudding
Charles Xu
David Zhao
compilade
AN Long
Christian Kastner
Reese Levine
Diego Devesa
leejet
lhez
Srihari-mcw
diannao
Ruben Ortlam
Kai Pastor
xctan
Akarshan Biswas
Alberto Cabrera Pérez
Erik Scholz
deepsek
Gabe Goodhart
Chris Rohlf
Daniel Bevenius
Donghyeon Jeong