I've been seeing significantly worse performance for tg with flash attention
enabled vs disabled, and it seems to be related to the submit heuristic.
Change the heuristic to check how many bytes worth of weight matrix are
used and flush every 100MB, and ramp up after the first few submits.
This seems to resolve the issue, and also increases perf for non-FA a bit.
* opencl: more profiling timing
* opencl: generate trace for profiling
* opencl: reduce profiling overhead
* Populate profiling timing info at the end rather than after each
kernel run
* opencl: fix for chrome tracing
* Enable CUDA Graph on CTK < 12.x
`cudaGraphExecUpdate` API was changed on 12.x. For this reason CUDA graph support was disabled on older CUDA toolkit. This change enables CUDA support in CTK version < 12.x by using older API if CTK < 12.x.
* Fix compilation errors with MUSA
* Disable CUDA Graph for MUSA
* cmake: Factor out compiler flag function from ggml
llama.cpps's build requires it, too, and we may want to make use of it
without add_subdirectory(ggml).
* cmake: Enable building against system ggml
This facilitates package maintenance for Linux distributions, where the
libggml library most likely will be shipped as an individual package
upon which a llama.cpp package depends.
When fattn-wmma was ported over to warp64 various bits that also touch fattn-vec where converted to
selectable warp size, however the fattn-vec kernels dont work with 64 wide warps for now, so we need
to avoid launching them with parameters for warp64
refactor mmqv to unify the calculation of nwarps and rows per block between host and device code.
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
This patch nudges the llama.cpp a bit to be supported on PoCL which
doesn't support OpenCL C CL2.0. The issue is solved by querying the
device for the supported OpenCL C versions and using the highest one
available.
This commit updates the compilation of default.metallib to skip the
intermediate .air (Apple Intermediate Representation) file.
The motivation for this change is to simplify the custom command a
little and avoid generating and then removing the .air file.
* ggml_compute_forward_concat() for arbitrary tensor type
* Check that tensors' type match
* ggml-cpu.c: check type of source tensors
* ggml-cpu.c: move tensor type check to ggml_compute_forward_concat()
* ggml.c: check concatenated tensor type
* Remove tensor type check from ggml_compute_forward_concat() in ggml-cpu.c
..., as it was moved to ggml.c.
* metal : refactor im2col parameters into a struct
* metal: Change im2col offset types from int32_t to uint64_t to support larger memory offsets
* metal : refactor sum_rows parameters into a struct
* metal : refactor soft_max parameters into a struct
* metal : refactor diag_mask_inf parameters into a struct
* metal : refactor ssm_conv parameters into a struct
* metal : refactor ssm_scan parameters into a struct
* metal : refactor get_rows parameters into a struct
* metal : refactor group_norm parameters into a struct
* metal : refactor conv_transpose_1d parameters into a struct
* metal : refactor upscale parameters into a struct
* metal : refactor pad parameters into a struct
* metal : refactor pad_reflect_1d parameters into a struct
* metal : refactor arange parameters into a struct
* metal : refactor timestep_embedding parameters into a struct
* metal : refactor argsort parameters into a struct
* metal : refactor leaky_relu parameters into a struct
* metal : refactor pool_2d parameters into a struct
* metal : fix trailing whitespace
---------
Co-authored-by: alexju <alexju@tencent.com>
This commit updates the custom command to build the default.metallib
file to use the correct path to ../ggml-common.h by using the variable
METALLIB_COMMON.
The motivation for this change is that currently when building and
specifying GGML_METAL_EMBED_LIBRARY=OFF the following error is
generated:
```console
[ 11%] Linking CXX shared library ../../bin/libggml.dylib
[ 11%] Built target ggml
make[2]: *** No rule to make target `ggml/src/ggml-metal/ggml-common.h', needed by `bin/default.metallib'. Stop.
make[1]: *** [ggml/src/ggml-metal/CMakeFiles/ggml-metal-lib.dir/all] Error 2
```
With the above change the build could progress but there was a follow
on error about not being able to find the ggml-common.h file in
ggml-metal.metal where is was included as a relative path:
```console
[ 11%] Compiling Metal kernels
/Users/danbev/work/llama.cpp/build/bin/ggml-metal.metal:6:10: error: '../ggml-common.h' file not found, did you mean 'ggml-common.h'?
^~~~~~~~~~~~~~~~~~
"ggml-common.h"
1 error generated.
```
Removing the relative path then allowed the build to complete
successfully.
Fix the following error:
```
ggml-alloc.c:99: not enough space in the buffer
ggml_tallocr_alloc: not enough space in the buffer to allocate blk.17.ffn_down.weight (needed 27525120, available 27521024)
```
which occurs when `ggml_backend_opencl_context::alignment` is larger
than `cl_ptr_base` (hard-coded to `0x1000`).
Also, fix `ggml_backend_opencl_context::alignment` was set to
`CL_DEVICE_MEM_BASE_ADDR_ALIGN` which was treated as bytes but the
value is reported in bits.
* ggml-cpu: Faster IQ1 mul_mat_vec on AVX2 using BMI2 instructions
* cmake: Add GGML_BMI2 build option
* ggml: enable BMI2 on relevant CPU variants
* ggml-cpu: include BMI2 in backend score
* ggml-cpu: register BMI2 in ggml_backend_cpu_get_features
* ggml-cpu: add __BMI2__ define when using MSVC
-- it might happen if ggml is loaded from 2 separate libraries since each one of them will expose the class. This is more of a guard since we want to use only Metal as embedded library and don't care about the other case.
* Add include files for std::min/max and std::toupper/tolower
* win32: move _USE_MATH_DEFINES before includes to ensure M_PI is defined
* Use GGML_RESTRICT instead of "restrict" keyword everywhere, and use "__restrict" in MSVC plain C mode
* win32: only use __restrict in MSVC if C11/C17 support is not enabled
---------
Co-authored-by: Marcus Groeber <Marcus.Groeber@cerence.com>
Adds GGML_HIP_ROCWMMA_FATTN and rocwmma header check
Adds rocWMMA support to fattn-wmma-f16
---
Signed-off-by: Carl Klemm <carl@uvos.xyz>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
Co-authored-by: Ben Jackson <ben@ben.com>
* Support fp16 unary operations in the CUDA backend
* cpu: increase fp16 support for unary operators in the CPU backend
* cuda: increase fp16 support for unary operators in the CUDA backend
* Add test cases for fp16 unary operators
* metal: update supports_op for unary operators that don't support fp16, to prevent test-backend-ops from failing
* metal: fix PR comments for unary op support after fp16 unary tests
* Support float16-to-float16 add/sub/mul/div operations in the CUDA backend
* Add fp16 support for add/sub/mul/div on the CPU backend
* Add test cases for fp16 add/sub/mul/div
* Upgrade init_tensor API to return a ggml_status
To prepare for an 'abort-free' ggml
(ggml not to abort on OOMs but return a OOM status),
as agreeed with Diego in the ggml repo,
upgrade the init_tensor() and view_init() APIs
to return a ggml_status.
* misc fixes
---------
Co-authored-by: slaren <slarengh@gmail.com>
* vulkan: implement specialized MMV kernels for IQ2 quantizations
* vulkan: add MMV kernels for IQ3 quants
* vulkan: Increase MMV batch size and unroll IQ LUT setup
* vulkan: fix init_iq_shmem for WG sizes larger than tables
* vulkan: common batch size for all I-quants
* Added SVE Support for Q2_K Quantized Models
* Use 4-space indentation in the switch cases
* removed comments lines
* Remove the loop Retain the curly bracess for better understanding of code
* Remove the comment like added for q3_k_q8_k kernel
---------
Co-authored-by: vithulep <p.m.vithule1517@gmail.com>
* Fix dependencies between ggml and backends
ggml backends link only to ggml-base and ggml links to all backends.
* Fix installation of ggml backends
Set up GNUInstallDirs before setting the installation directory of ggml backends
* opt performance by reorder for Intel GPU
* detect hw type and save opt feature, and print opt feature
* correct name
* support optimize graph once when compute graph, record the opt status in tensor->extra, make CI passed
* add env variable GGML_SYCL_DISABLE_OPT for debug
* use syclex::architecture replace the custom hw define, update the guide for GGML_SYCL_DISABLE_OPT
* add performance data
* mv getrows functions to separeted files
* fix global variables
---------
Co-authored-by: arthw <14088817+arthw@users.noreply.github.com>
* MUSA: support ARM64 and enable __dp4a .etc
* fix cross entropy loss op for musa
* update
* add cc info log for musa
* add comment for the MUSA .cc calculation block
---------
Co-authored-by: Bodhi Hu <huaishun.hu@mthreads.com>
* ggml-cpu: Add CPU backend support for KleidiAI library
* Add environmental variable GGML_KLEIDIAI_SME
* Add support for multithread LHS conversion
* Switch kernel selection order to dotprod and i8mm
* updates for review comments
* More updates for review comments
* Reorganize and rename KleidiAI files
* Move ggml-cpu-traits.h to source file
* Update cmake for SME build and add alignment for SME
* Remove append GGML_USE_CPU_KLEIDIAI to the GGML_CDEF_PUBLIC list
* vulkan: initial support for IQ1_S and IQ1_M quantizations
* vulkan: define MMV kernels for IQ1 quantizations
* devops: increase timeout of Vulkan tests again
* vulkan: simplify ifdef for init_iq_shmem
* musa: Update MUSA SDK version to rc3.1.1
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
* musa: Remove workaround in PR #10042
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
---------
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
* ggml-cpu : add chunking support to mul_mat_id
* allocate chunk counter in wdata
parallelize src1 quantization by column to allows parallelization even when there is only one row
* disable for arm
* cleanup
* better way to disable for arm
* fix uninitialized counter when using 1 thread only
* revert test-backend-ops changes
* Bug fix for clamp_f32
When using tensors larger than 1d clamp operation does not work due to the restriction of returning if ith is not 0.
* Bug fix for clamp_f32
* Bug fix for clamp_f32
After the barrier in last iteration is executed, still the loop termination
condition will be executed. However main thread can destroy the cgraph object
and its nodes already, then another thread will access it, but the thing is already gone.
Also trouble can happen when n_nodes == 0 or abort is called, but I'm not sure if the
prior situation is possible.
Last syncronization should be done after the loop to ensure the cgraph/cplan won't be
accessed after the main thread exits from the function.
* ggml : optimize convert f32<->f16 for loongarch_asx
* ggml : optimize loongarch_asx extend i16,i8,u8 to i32,i16
* ggml : Fix warnings when run cpu CI locally on LoongArch
Add bounds checking in `rpc_server::copy_tensor` to prevent out-of-bounds writes
+ Check if `(uint8_t *)dst->data + ggml_nbytes(src)` remains within the destination buffer’s allocated region.
This makes git as a dependency optional, and is useful in the case where
ggml is built not from git, but from a tarball, or a distribution source
package.
This conditional also affects GGML_BUILD_COMMIT. Nothing seems to be
using it, though, so there doesn't seem much value factor it out, or
even require it.
* CUDA: use mma PTX instructions for FlashAttention
* __shfl_sync workaround for movmatrix
* add __shfl_sync to HIP
Co-authored-by: Diego Devesa <slarengh@gmail.com>
* vulkan: initial support for IQ3_S
* vulkan: initial support for IQ3_XXS
* vulkan: initial support for IQ2_XXS
* vulkan: initial support for IQ2_XS
* vulkan: optimize Q3_K by removing branches
* vulkan: implement dequantize variants for coopmat2
* vulkan: initial support for IQ2_S
* vulkan: vertically realign code
* port failing dequant callbacks from mul_mm
* Fix array length mismatches
* vulkan: avoid using workgroup size before it is referenced
* tests: increase timeout for Vulkan llvmpipe backend
---------
Co-authored-by: Jeff Bolz <jbolz@nvidia.com>
* Add option to not print stack on abort
Add option/envvar to disable stack printing on abort.
Also link some unittests with Threads to fix link errors on
ubuntu/g++11.
* Update ggml/src/ggml.c
---------
Co-authored-by: Diego Devesa <slarengh@gmail.com>
loops with bounds not known at compile time can not be unrolled.
when ncols_template == 0, the bounds of the loop are not constexpr, thus llvm cant unroll the loops here.
This disables the workaround on rocblas fixed versions (>=4.0.0) to eliminate the runtime cost and unnecessary VRAM allocation of loading all tensile objects.
Implemented ggml_sycl_op_soft_max() F16 src1(mask) support for which a pragma deprecation warning was added during #5021.
To do this, had to decouple it from ggml_sycl_op_flatten which always considered src1 to be of fp32 type(many OP functions are dependent on it).
* SYCL: SOFTMAX F16 mask support and other fixes
* test-backend-ops: Add F16 mask test cases
This fixes segmentation fault error when running tests when no metal
devices are available (for example, when not linked with Core Graphics
framework or otherwise).
With robustbufferaccess disabled, this shader was showing OOB stores. There
is a bounds check in the code, but the workgrouop dimensions were reversed vs
CUDA and it was running the wrong number of threads. So fix the workgroup
dimensions and disable robustness for this pipeline.
mul mat and flash attention shaders were loading f32 types directly into
A/B matrices, which happens to work but is technically invalid usage.
For FA, we can load it as an Accumulator matrix and convert and this
is not in the inner loop and is cheap enough. For mul mat, it's more
efficient to do this conversion in a separate pass and have the input(s)
be f16.
coopmat2 requires SPIR-V 1.6 (related using to LocalSizeId). LocalSizeId
requires maintenance4 be enabled, and SPIR-V 1.6 requires Vulkan 1.3.
* Implement host pool for matrix_info
Creating a new memory pool on the host to store memory location for
matrix_info needed to launch gemm_batch from oneMKL/oneMath.
Removing complex support in gemm_batch since it is not used in llama.cpp
* Remove unnecessary headers and cast
* Reorder member variable to avoid warning on initialization
* Formatting
* Remove unused variable
* Address PR review feedback - remove warning
---------
Signed-off-by: nscipione <nicolo.scipione@codeplay.com>
Add code similar to mul_mm_cm2 to force alignment of strides, to avoid
a performance regression.
Add noncontiguous FA tests in test-backend-ops.
Fixes#11268.
* vulkan: support copy from f32 to q4_0/q4_1/q5_0/q5_1/q8_0/iq4_nl
Shaders are based on cpy.cu.
* vulkan: support copy from q4_0/q4_1/q5_0/q5_1/q8_0/iq4_nl to f32
* ggml: copy q->f32 assumes some contiguity in the destination
* Add SVE support for q4_K_q8_K
* Update ggml/src/ggml-cpu/ggml-cpu-quants.c
change to use K_SCALE_SIZE
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* q6_k scale caching
* 16 bit unpack
* q4_k test (slow)
* revert it
* q3_k
* q2_k
* little stuff
* try precalculating products of a and q2_k scales
* Revert "try precalculating products of a and q2_k scales"
This reverts commit 65110b81f23f66331a50c6e889a7c1ab9470a86b.
* unpack should be u16, add vim swap to gitignore (about time)
* better q4_k scales
* q5_k
* better q6_k with separate paths for all threads and partial threads in use, plus some more optimizations
* q2_k better dequant
* q3_k optimizations
* q3_k use hmask simd from cpu avx version
* make the caches happy
* q3_k separate out calculation
* q2_k separate out
* little stuff
* use calc_superblock everywhere
* q2_k optimize scale calculation
* more barriers
* fix: ggml: fix vulkan-shaders-gen build
The vulkan-shaders-gen target was not being built correctly
in case of cross-compilation.
Other outputs need to be built for the cross compile target,
but vulkan-shaders-gen needs to be built for the host.
* refactor: ggml: Improve vulkan-shaders-gen toolchain setup
- Add GGML_SHADERS_GEN_TOOLCHAIN CMake option.
- Auto-detect host toolchain if not set.
* refactor: ggml: Improve vulkan-shaders-gen toolchain setup
Use configure_file to generate host_toolchain.cmake from template
* fix: ggml: Fix compile error
Fix compile error not finding vulkan-shaders-gen
* fix: vulkan-shaders-gen build and path handling
Fix build issues with vulkan-shaders-gen:
- Add target dependency for correct build order
- Use CMAKE_HOST_SYSTEM_NAME for executable suffix
- Fix MSVC output directory in host toolchain
- Normalize path handling for cross-compilation
* fix: improve host compiler detection in vulkan shader build
Improve host compiler detection for vulkan shader generation:
- Add NO_CMAKE_FIND_ROOT_PATH to all compiler searches
- Consolidate compiler detection logic
- Fix Windows-specific MSVC detection
- Ensure correct compiler search in cross-compilation
* refactor: Simplify CMake function for detecting host compiler
Simplified the CMake function to improve the process of detecting the host compiler.
* fix: Remove unnecessary Vulkan library linkage in CMakeLists.txt
Since `vulkan-shader-gen.cpp` only requires the `glslc` executable
and not the Vulkan headers or libraries, CMakeLists.txt needs to
be corrected.
(See: ecc93d0558)
* refactor: Rename host_toolchain.cmake.in
- Rename host_toolchain.cmake.in to cmake/host-toolchain.cmake.in
* refactor: GGML_VULKAN_SHADERS_GEN_TOOLCHAIN
Rename the macro GGML_SHADERS_GEN_TOOLCHAIN to GGML_VULKAN_SHADERS_GEN_TOOLCHAIN
* Refactor: Moves cuda graph executable update step to separate function.
* Refactor: Moves cuda graph update check to separate function.
* Refactor: Moves cuda graph maintenance (update or adjusting copy parameters) to separate function for improved readability.
* Fix: Adds missing reference to maintain_cuda_graph() definition.
* Refactor: Improves structure and abstractions by moving CUDA graph evaluation and capture to its own function.
* Refactor: Moves node graph checks and copy ops into individual function for improved readability.
* Refactor: Removes code permanently excluded from compilation to increase readability.
* Style: Adds missing newline
* Style: Consolidates several neighboring '#ifdef USE_CUDA_GRAPH' into a single one
* Refactor: Makes 'cuda_graph_update_required' a local variable
* remove double lines between functions
---------
Co-authored-by: slaren <slarengh@gmail.com>
Build fails when using HIP and GGML_BACKEND_DL:
```
/usr/bin/ld: ../ggml/src/libggml.so: undefined reference to `ggml_backend_cuda_reg'
collect2: error: ld returned 1 exit status
```
This patch fixes this.
* SYCL: refactor ggml_sycl_compute_forward
* SYCL: add back GGML_USED(dst) to ggml_sycl_cpy
* SYCL: add function name to noop debug
* SYCL: Some device info print refactoring and add details of XMX availability
This change upstreams llamafile's cpu matrix
multiplication kernels for ppc64le using MMA
builtins for quantised int8 datatype.
This change results in 10% - 70% improvement
in total speed(ie all tokens/total time), across
various batch sizes.
The patch is tested with Meta-Lllama-3-8B,
Mistral-7B, Llama-2-7B-chat-hf models on a
IBM POWER10 machine.
Signed-off-by: Amrita H S <amritahs@linux.vnet.ibm.com>
* Disable GL_KHR_cooperative_matrix Vulkan extension if not available.
* Perform Vulkan extensions checks in a more sensible order
* Remove unnecessary #ifdef directive
* GGUF: C++ refactor, backend support, misc fixes
remove ggml_tensor.backend
update CODEOWNERS [no ci]
remove gguf_get_data from API
revise GGUF API data types
* SYCL: Use get_multi_ptr instead of deprecated get_pointer in wkv6
* Revert "SYCL: Use get_multi_ptr instead of deprecated get_pointer in wkv6"
This reverts commit f62dc45f31.
* Reland: Use get_multi_ptr instead of deprecated get_pointer in wkv6
This commit attempts to improve the log message for the inputs of the
splits in the sched_print_assignments function.
The motivation for this change is that currently even if there are no
inputs a colon is displayed at the end of the line, which can make it a
little confusing when reading the output as it could be interpreted as
the line below are inputs when they are in fact nodes. With this change
the colon will only be printed if there actually are inputs.
In common/common.cpp:
* Convert usage of stat() function call to check if file exists to standard library function std::filesystem::exists (error unable to match to correct function signature)
* Additional conditions to check if PATH_MAX is already defined in WIN32 environment (warning it is already defined in MSYS2)
In examples/run/run.cpp:
* Add io.h header inclusion (error cannot find function _get_osfhandle)
* Change initialisers for OVERLAPPED to empty struct (warning about uninitialised members)
* Add initialiser for hFile (warning it may be uninitialised)
* Add cast for curl_off_t percentage value to long int in generate_progress_prefix function (warning that curl_off_t is long long int)
In ggml/src/ggml-opencl/ggml-opencl.cpp:
* Initialise certain declared cl_mem variables to nullptr for greater safety (warning about B_d variable possibly used unassigned)
Make the mul_mat_vec shaders support N>1 (as a spec constant, NUM_COLS) where
the batch_strides are overloaded to hold the row strides. Put the loads from the
B matrix in the innermost loop because it should cache better.
Share some code for reducing the result values to memory in mul_mat_vec_base.
* tests: Add im2col perf tests
* vulkan: optimize im2col, more elements per thread
* vulkan: increase small tile size for NV_coopmat2
* vulkan: change im2col to 512 elements per workgroup
Warning types fixed (observed under MSYS2 GCC 14.2.0):
* format '%ld' expects argument of type 'long int', but argument has type 'size_t'
* llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp:81:46: warning: missing initializer for member '_STARTUPINFOA::lpDesktop' [-Wmissing-field-initializers] (emitted for all struct field except first)
* more perfo with llamafile tinyblas on x86_64.
- add bf16 suport
- change dispache strategie (thanks:
https://github.com/ikawrakow/ik_llama.cpp/pull/71 )
- reduce memory bandwidth
simple tinyblas dispache and more cache freindly
* tinyblas dynamic dispaching
* sgemm: add M blocs.
* - git 2.47 use short id of len 9.
- show-progress is not part of GNU Wget2
* remove not stable test
Change the code to do 16b loads when possible and extract the appropriate
component late, so the code is effectively decoding a pair of elements and
then selecting one. This can allow more commoning to happen in the compiler
when neighboring elements are loaded.
* Migrate to tensor->buffer for checking backend buffer type: 1
* SYCL: common.cpp try to migrate away from tensor->backend
* SYCL: fix assertions and add proper comments
* SYCL: remove extra space
* SYCL: Add back static to ggml_backend_buffer_is_sycl_split function
* SYCL: Add pragma directive to suppress warning spam
* SYCL: Integrate debug logs with GGML_LOG and other fixes
* Revert "SYCL: Integrate debug logs with GGML_LOG and other fixes"
This reverts commit 2607b7de0f.
Let's keep the current SYCL specific logging mechanism for now
* SYCL: Use GGML_SYCL_DEBUG after reverting
* SYCL: reg_get_proc_address func, update to the current func signature
* SYCL: Refactor SYCL buffer checks in ggml_sycl_cpy_tensor_2d
This commit removes the return statement from ggml_gallocr_allocate_node
function.
The motivation behind this change is to make the code more readable and
consistent.
* ggml : add check for grad_accs
This commit adds a check for grad_accs in ggml_graph_get_grad and
ggml_graph_get_grad_acc functions. This is necessary to avoid segfaults
when grad_accs is not initialized.
The motivation for this change is that I find it nice to be able to
print out a computation graph using ggml_graph_print but this function
segfaults when grad_accs is not initialized:
```console
(gdb) p g1
$2 = (ggml_cgraph *) 0x7ffff66004b0
(gdb) p *g1
$3 = {size = 2048, n_nodes = 1, n_leafs = 2, nodes = 0x7ffff6600500,
grads = 0x0, grad_accs = 0x0, leafs = 0x7ffff6604500,
visited_hash_set = {size = 4099, used = 0x7ffff6610518,
keys = 0x7ffff6608500}, order = GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT}
(gdb) p ggml_graph_print(g1)
=== GRAPH ===
n_nodes = 1
Program received signal SIGSEGV, Segmentation fault.
0x0000555555579775 in ggml_graph_get_grad
(cgraph=0x7ffff66004b0,node=0x7ffff6600340)
at /ggml/ggml/src/ggml.c:5990
5990 return igrad != GGML_HASHSET_FULL &&
ggml_bitset_get(cgraph->visited_hash_set.used, igrad) ?
cgraph->grads[igrad] : NULL;
```
* squash! ggml : add check for grad_accs
Fix the check in ggml_graph_get_grad. The check was incorrectly using
cgraph->grad_accs instead of cgraph->grads.
* ensure mul mat shaders work on systems with subgroup size less than 32
more fixes
add test
* only s_warptile_mmq needs to be run with 32 threads or more
* [cl][adreno] Add Adreno GPU support
Add new OpenCL backend to support Adreno GPUs
---------
Co-authored-by: Skyler Szot <quic_sszot@quicinc.com>
Co-authored-by: Shangqing Gu <quic_shawngu@quicinc.com>
Co-authored-by: Alexander Angus <quic_aangus@quicinc.com>
Co-authored-by: Hongqiang Wang <quic_wangh@quicinc.com>
Co-authored-by: Max Krasnyansky <quic_maxk@quicinc.com>
* [cl][ci] Add workflow for CL
* [cl][adreno] Fix memory leak for non SMALL_ALLOC path
* opencl: integrate backend dyn.load interface and fix compiler and format warnings
* opencl: remove small-alloc support and fix build errors for non-opencl platforms
* opencl: fixed merge conflict (MUSA added twice in cmake)
* opencl-ci: use RUNNER_TEMP instead of github.workspace
* opencl: fix embed tool invocation with python3
* opencl: CI workflow fixes
* opencl: Clean up small-alloc in CMake files
* opencl: cleanup ggml-opencl2 header file
* opencl: use ulong for offsets and strides in ADD kernel
* opencl: use cl_ulong for all offsets
* opencl: use cl_ulong for sizes and strides
* opencl: use `GGML_LOG_xxx` instead of `fprintf(stderr, ...)`
* opencl: rename backend `opencl2` -> `opencl`
* opencl: rename kernel files `ggml-opencl2` -> `ggml-opencl`
* opencl: make OpenCL required, remove redundant lib and inc directories
* `ggml-base`, `..` and `.` are added by `ggml_add_backend_library`
* opencl: rename backend - funcs, structs, etc `opencl2` -> `opencl`
* opencl: remove copyright marker since main license already covers
* opencl: replace some more OPENCL2 leftovers
* opencl: remove limits on `tensor_extra`
* opencl: use pools for `tensor_extra`
* opencl: fix compiler warnings with GCC and Clang
Still getting the warning about clCreateCmdQueue being obsolete.
Will fix that separately.
* opencl: fail gracefully if opencl devices are not available
Also for unsupported GPUs.
* opencl: fix MSVC builds (string length error)
* opencl: check for various requirements, allow deprecated API
* opencl: update log message for unsupported GPUs
---------
Co-authored-by: Skyler Szot <quic_sszot@quicinc.com>
Co-authored-by: Shangqing Gu <quic_shawngu@quicinc.com>
Co-authored-by: Alexander Angus <quic_aangus@quicinc.com>
Co-authored-by: Hongqiang Wang <quic_wangh@quicinc.com>
Co-authored-by: Max Krasnyansky <quic_maxk@quicinc.com>
* Fix crash caused by ggml_backend_load_all when launching on AndroidActivity.
Details:
Calling ggml_backend_load_all during initialization in the AndroidActivity project leads to a crash with the error:
terminating with uncaught exception of type std::__ndk1::__fs::filesystem::filesystem_error: filesystem error: in directory_iterator::directory_iterator(...): Permission denied [./].
This issue occurs because AndroidActivity restricts file access due to sandboxing.
Reproduction:
In the example folder, the LlamaAndroid project can reproduce the crash by calling ggml_backend_load_all first in Java_android_llama_cpp_LLamaAndroid_backend_1init.
* Update ggml/src/ggml-backend-reg.cpp
---------
Co-authored-by: Diego Devesa <slarengh@gmail.com>
* double the number of rows per workgroup
* Update ggml-vulkan.cpp
* Vulkan: Add VK_EXT_subgroup_size_control support to ensure full subgroups for coopmats
* only increase the number of rows for amd and subgroup size 64
* fix missing NUM_ROWS for mul_mat_vec_iq4_nl_f16_f32, untested
* use subgroup min and max to check for gcn (requires https://github.com/ggerganov/llama.cpp/pull/10721)
* manual merge ggml-vulkan.cpp
* set min and max subgroup size in any case
* Also double the number of rows for Intel GPUs
* Try to reduce some unused and typecast warnings
* Reduce compiler warnings step 2
* add a newline at the end of the file
* Initialize nreduce as size_t
* [SYCL] Remove pragma directives from mmq.cpp
* SYCL: mmq add condition to prevent blocks_per_tile_x_row variable from becoming 0
* SYCL softmax: Initialize nreduce as size_t
* ggml-sycl.cpp: fix some trailing whitespaces
* SYCL: remove the unused variables instead of commenting it out
* SYCL poo2d kernel: set NAN for invalid pooling op
* SYCL gemm.hpp: remove pragma directives
* SYCL gemm.hpp: use const cast to properly support dnnl::memory
* SYCL: wkv6 remove a comment
* SYCL: clean comments step 2
* SYCL: clean comments and variables step 3
* SYCL: Use GGML_UNUSED for unused variables
* SYCL: remove extra empty lines and a comment
* Remove TODO
* cleanup spaces
* add a stdout for unsupported op
* use sycl printf over fprintf
* remove prints for CI
* SYCL ggml-sycl: pool2D use sycl::nan and remove if-else block
---------
Co-authored-by: Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>
* faster uncontiguous concat
* Use a lambda to avoid code duplication
Co-authored-by: Diego Devesa <slarengh@gmail.com>
* Update ggml/src/ggml-cuda/concat.cu
* add constexpr and static assert
---------
Co-authored-by: Diego Devesa <slarengh@gmail.com>
* Vulkan: Add VK_EXT_subgroup_size_control support to ensure full subgroups for coopmats
* Fix subgroup size control extension support check
Add accf32 and accf16 checks for coopmats
* Also disable coopmats on amdvlk
* feat: load all backends from a user-provided search path
* fix: Windows search path
* refactor: rename `ggml_backend_load_all_in_search_path` to `ggml_backend_load_all_from_path`
* refactor: rename `search_path` to `dir_path`
* fix: change `NULL` to `nullptr`
Co-authored-by: Diego Devesa <slarengh@gmail.com>
* fix: change `NULL` to `nullptr`
---------
Co-authored-by: Diego Devesa <slarengh@gmail.com>
Vulkan doesn't mandate a specific rounding mode, but the shader_float_controls
feature allows rounding mode to be requested if the implementation supports it.
* Renames NVIDIA GPU-architecture flags to avoid name clashes with WinAPI. (e.g. CC_PASCAL, GPU architecture or WinAPI pascal compiler flag?)
* Reverts erroneous rename in SYCL-code.
* Renames GGML_CUDA_MIN_CC_DP4A to GGML_CUDA_CC_DP4A.
* Renames the rest of the compute capability macros for consistency.
There are some bugs in the 1.3.296 SDK, so disable this. It isn't strictly
necessary anyway.
Add missing dependency on vulkan-shaders-gen, so shaders get recompiled when it
changes.
Fix coopmat support reporting when glslc doesn't support NV_coopmat2.
* rename ggml-cpu-aarch64.c to .cpp
* reformat extra cpu backend.
- clean Q4_0_N_M and IQ4_0_N_M
- remove from "file" tensor type
- allow only with dynamic repack
- extract cpu extra bufts and convert to C++
- hbm
- "aarch64"
- more generic use of extra buffer
- generalise extra_supports_op
- new API for "cpu-accel":
- amx
- aarch64
* clang-format
* Clean Q4_0_N_M ref
Enable restrict on C++
* add op GGML_OP_MUL_MAT_ID for Q4_0_N_M with runtime repack
* added/corrected control on tensor size for Q4 repacking.
* Update ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* add debug logs on repacks.
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Vulkan: Implement VK_KHR_cooperative_matrix support in the matrix matrix multiplication shader
* Improve performance with better q4_k and q5_k dequant and store unrolling
* Add Vulkan MUL_MAT and MUL_MAT_ID accumulator precision selection
* Rework mulmat shader selection and compilation logic, avoid compiling shaders that won't get used by device
* Vulkan: Implement accumulator switch for specific mul mat mat shaders
* Vulkan: Unroll more loops for more mul mat mat performance
* Vulkan: Add VK_AMD_shader_core_properties2 support to read Compute Unit count for split_k logic
* Disable coopmat support on AMD proprietary driver
* Remove redundant checks
* Add environment variable GGML_VK_DISABLE_COOPMAT to disable VK_KHR_cooperative_matrix support
* Fix rebase typo
* Fix coopmat2 MUL_MAT_ID pipeline selection
* metal : Extend how Llama.cpp locates metal resources (#10675)
* It searches the resource file in the directory where the current
binary is located as well.
* Resolves symbolic links.
Rationale:
When we plug this dependency into a Bazel build and run it in the
context of Bazel (e.g. testing):
* the execution directory is often very different from where the files
are located and no direct control over this (Bazel sandboxing),
* the Bazel sandbox often use symbolic links to make files available.
With this patch, we can have the resource file added to the target,
can build and run tests in the context of Bazel.
* Update ggml/src/ggml-metal/ggml-metal.m
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update ggml/src/ggml-metal/ggml-metal.m
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* ggml_pad_reflect_1d defined in header
* implemented on CPU
* called the forward pass
* impl Metal kernel
* added Metal kernel
* added OP_PAD_REFLECT_1D in test-backend-ops.cpp
* add test-pad-reflect-1d test case
* test case support multiple backend
* [SYCL] Move to Compile Time backend selection on oneMKL Interface for NVIDIA backend
Move to compile time selection to backend to avoid latency at run time.
Add it to all mkl gemm calls and only for NVIDIA backend.
Signed-off-by: nscipione <nicolo.scipione@codeplay.com>
* Formatting
* Address PR comments to increase readibility
---------
Signed-off-by: nscipione <nicolo.scipione@codeplay.com>
* kqmax_new_j in every thread within warp is same after operate at line 199,this reduce can be omit
* same problem in vec32
---------
Co-authored-by: ZhaoXiaoYu <zhao.xiaoyu@zte.com.cn>
* subgroup 64 version with subgroup add. 15% faster
scalable version
tested for subgroup sizes 16-128
* check for subgroup multiple of 16 and greater than 16
* subgroup sizes are always a power of 2 (https://github.com/KhronosGroup/GLSL/issues/45)
* force 16 sequential threads per block
* make 16 subgroup size a constant
This is an incremental improvement over #9118 to get work to the GPU a bit
sooner. The first part is to start with a smaller number of nodes before
the first submit, and ramp it up to the current 100 nodes/submit. The
second part is to reduce the dryrun overhead for all the nodes that just
need to request descriptor space.
With these changes I get around 1-2% speedup on RTX 4070 combined with my
old Haswell-era CPU.
* CANN: Fix the bug build fail on Ascend310P under two cases:
1) Manual specify SOC_TYPE
2) Under some unusual compile environment
* Update the cann backend News content: Support F16 and F32 data type model for Ascend 310P NPU.
* fix CANN compile fail bug: the assert in ascend kernel function doesn't supportted on some CANN version
There have been reports of failure to compile on systems with <= 32KB
of shared memory (e.g. #10037). This change makes the large tile size
fall back to a smaller size if necessary, and makes mul_mat_id fall
back to CPU if there's only 16KB of shared memory.
* improve inferencing performance for ascend npu.
Co-authored-by: Frank Mai <thxCode@thxcode0824@gmail.com>
* some modification after review
* some modifications after review
* restore some modifications
* restore some modifications
---------
Co-authored-by: shanshan shen <shanshanshen333@gmail.com>
Co-authored-by: Frank Mai <thxCode@thxcode0824@gmail.com>
The vulkan-shaders-gen was not parsing the --no-clean argument correctly.
Because the previous code was parsing the arguments which have a value only
and the --no-clean argument does not have a value, it was not being parsed
correctly. This commit can now correctly parse arguments that don't have values.
* llama : accept a list of devices to use to offload a model
* accept `--dev none` to completely disable offloading
* fix dev list with dl backends
* rename env parameter to LLAMA_ARG_DEVICE for consistency
* CANN Support Ascend310P to accelerate F32 and F16 Model
* Add compile option soc type macro ASCEND_310P to ggml-cann lib
* Remove unused code
* Remove the ascend soc_type hard code compile option in CMakelist.txt
* vulkan: Use pipeline_robustness to disable robustness in mul_mat_vec.
Add some early returns for nonexistent rows in mul_mat_vec shaders. These
can only be hit when dispatching a 2D grid of workgroups. Fix the logic
for the 2D grid of workgroups to round up.
Enable the pipeline robustness extension if it's available, and use it to
disable robustness for these pipelines. The instructions to do the bounds
checking contend for the same ALU resources as the bit twiddling dequant
instructions.
* vulkan: Add GLSL structure aliases for quant types to allow larger loads
In Vulkan it's not possible to cast pointer types, so instead you have to
declare an aliased binding for the memory with a different type. This
commit adds aliases for the quant formats using 16b ints, and in a few
places where the struct size is a multiple of 4 also using 32b ints.
Currently only q4_k's aliases are used, but others will be used in
subsequent commits.
* vulkan: use larger loads in q5_k and q6_k shaders.
Similar to the optimization I did in q4_k recently, this vectorizes some loads
and reduces the number of bit twiddling instructions.
* vulkan: use larger K step per iteration in mul_mat_vec.
Add vec4 dequantization functions, and use them to do K=8 per iteration in
mul_mat_vec. This uses 16b loads for the quant values and 128b loads for B
which helps reduce the load on the memory system.
The K_PER_ITER==2 logic is still there, just for F16/F32, and really only
because they support unaligned sizes.
Tweak the num_iters/unrolling logic to be simpler and catch a couple missed
unrolling opportunities.
* Add option to set the SYCL architecture for all targets
* Convert GGML_SYCL_HIP_TARGET to the more generic GGML_SYCL_ARCH option
* Document that setting GGML_SYCL_ARCH can improve the performance
* vulkan: Optimize soft_max
Large soft_max could already saturate memory, but small/medium sizes were
pretty slow. The bulk of the gains for them comes from using a smaller
workgroup size, and making the workgroup size match the subgroup size also
makes the barriers much cheaper.
Cache some values in locals to avoid refetching/recomputing. And stamp
out a few "template instantiations" so smaller cases will fully unroll.
Add a missing early return for OOB rows. This happens when there are more
than 512 rows and the dispatch is 512 x H.
* vulkan: Further soft_max optimizations
Restore the workgroup size of 512 case, use it for >1024.
Use unrollable loops for more iteration counts.
* metal : add kernel arg structs (wip)
* metal : fattn args
ggml-ci
* metal : cont + avoid potential int overflow [no ci]
* metal : mul mat struct (wip)
* cont : mul mat vec
* cont : pass by reference
* cont : args is first argument
* cont : use char ptr
* cont : shmem style
* cont : thread counters style
* cont : mul mm id
ggml-ci
* cont : int safety + register optimizations
ggml-ci
* metal : GGML_OP_CONCAT
ggml-ci
* metal : GGML_OP_ADD, GGML_OP_SUB, GGML_OP_MUL, GGML_OP_DIV
* metal : GGML_OP_REPEAT
* metal : GGML_OP_CPY
* metal : GGML_OP_RMS_NORM
* metal : GGML_OP_NORM
* metal : add TODOs for rest of ops
* ggml : add ggml-metal-impl.h
ggml-ci
Compute two result elements per workgroup (for Q{4,5}_{0,1}). This reuses
the B loads across the rows and also reuses some addressing calculations.
This required manually partially unrolling the loop, since the compiler
is less willing to unroll outer loops.
Add bounds-checking on the last iteration of the loop. I think this was at
least partly broken before.
Optimize the Q4_K shader to vectorize most loads and reduce the number of
bit twiddling instructions.
* use 128 bit loads (i've tried 256->128 to death and its slower)
* double accumulator
* avx bf16 vec dot
* +3% q4_0 inference
* +7% tg +5% pp compared to master
* slower f16c version, kep for reference
* 256b version, also slow. i tried :)
* revert f16
* faster with madd
* split to functions
* Q8_0 and IQ4_NL, 5-7% faster
* fix potential overflow (performance reduced)
* 16 bit add for q4_0 only
* merge
* sycl: Use syclcompat::dp4a
* Using the syclcompat version allow the compiler to optimize the
operation with native function
* Update news section
* Update CI Windows oneAPI version to 2025.0
* Reword doc
* Call syclcompat::dp4a inside dpct::dp4a
This reverts commit 90cb61d692.
Reuse the index calculations across all of src0/src1/dst. Add a shader
variant for when src0/src1 are the same dimensions and additional modulus
for src1 aren't needed. Div/mod are slow, so add "fast" div/mod that
have a fast path when the calculation isn't needed or can be done more
cheaply.
* Fixes broken build for the SYCL CUDA backend caused by non-explicit gemm call in outprod (merged in with RWKV6 in
Optimize RWKV6 Operator Naming and Implement Multi-core CPU/ SYCL Acceleration #10133)
* Marks permuted MUL_MAT as unsupported to be able to run test-backend-ops
* Fixes asserts in norm to fix debug builds.
* tests: Fix memory bandwidth calculation for perf tests
Add a flops calculation for flash attention.
Add one GGML_OP_CPY perf test.
* vulkan: Optimize contiguous copies
Add a variant of the copy shader for when the tensors are contiguous. Avoid
the complex addressing calculations, and do four elements per invocation
to hide some other overhead.
Apply similar changes to the scale shader, since scale is always contiguous.
Add a "progress bar" for shader compiles.
Fixes#9582
Spawning too many concurrent copies of glslc leads to "Failed to create pipes"
errors on Linux. This change applies the same throttling we use for
multithreaded pipeline creation.
Jeff Bolz
lhez
R0CKSTAR
Łukasz Ślusarczyk
Prajwal B Mehendarkar
fj-y-saito
0cc4m
Molly Sophia
Gaurav Garg
Guus Waals
Daniele
Christian Kastner
Akarshan Biswas
aubreyli
fairydreaming
Chenguang Li
Alberto Cabrera Pérez
uvos
jklincn
BB-fat
Eve
Henry Linjamäki
Jason C.H
Daniel Bevenius
vmobilis
Rémy O
xiaofei
Johannes Gäßler
Henry Linjamäki
simon886212
Plamen Minev
mgroeber9110
David Huang
cmdr2
Diego Devesa
midnight
petterreinholdtsen
cmdr2
ag2s20150909
Erik Scholz
William Tambellini
Prashant Vithule
hipudding
Daniele
Vladimir Vuksanovic
Judd
Gian-Carlo Pascutto
Neo Zhang Jianyu
Akarshan Biswas
Aaron Teo
PureJourney
Bodhi
Charles Xu
Hale Chan
Adrian Kretz
Georgi Gerganov
Jinyang He
Jeffrey Morgan
Xuan-Son Nguyen
bandoti
Richard
Weizhao Ouyang
Sheldon Robinson
Maxim Evtush
Danny Milosavljevic
Wagner Bruna
Karol Kontny
Patrick Peng
junchao-zhao
Charles Duffy
fxzjshm
Jhen-Jie Hong
Olivier Chafik
uvos
Rémy Oudompheng
issixx
Nikita Sarychev
someone13574
Haus1
Ihar Hrachyshka
Bernhard M. Wiedemann
amd-dwang
Radoslav Gerganov
Nicolò Scipione
Junil Kim
Andreas Kieslinger
hydai
amritahs-ibm
Mathieu Baudier
matt23654
Gilad S.
Srihari-mcw
Peter
Djip007
yuri@FreeBSD
Adrien Gallouët
gn64
Zhiyuan Li
HimariO
谢乃闻
a3sh
stduhpf
Robert Ormandi
PAB
Frankie Robertson
mahorozte
Adrien Gallouët
Shupei Fan
Chenguang Li
Sergio López
leo-pony
Shanshan Shen
haopeng
Romain Biessy
FirstTimeEZ
Dan Johansson
thewh1teagle