RPC_CMD_SET_TENSOR always returns an empty response and we send this 4
times per token. We can improve TG speed if we don't wait for this empty
response.
The performance impact of this change depends on the network latency.
This commit adds a check to makes sure that the target exists before
trying to add compile options to ignore warnings when using MSVC.
The motivation for this is currently the build is broken depending on
the cmake options provided. With this fix it should be possible to build
even if the targets are not actually available.
Refs: https://github.com/ggml-org/whisper.cpp/pull/3090#issuecomment-2842760104
* whisper: suppress Windows compiler warnings
This commit disables compiler warnings on window using MSVC.
The motivation for these changes is that some compilers generate
warnings for these conversion, for example Windows MSVC, and
there are quite a few of them. This makes it a little difficult to
spot new warnings that may be introduced and also can be difficult
for users/embedders of ggml where these warnings are hard to separate
from their own warnings.
* squash! whisper: suppress Windows compiler warnings
Move ggml related warnings into ggml. This commit also fixes the
indentation and adds a missing whitespace to the if statement.
* tune matmul for gcn
* this one is more power efficient
* Update ggml/src/ggml-vulkan/ggml-vulkan.cpp
Co-authored-by: 0cc4m <picard12@live.de>
* disable this tune for the proprietary driver
---------
Co-authored-by: 0cc4m <picard12@live.de>
Add RPC_CMD_HELLO for getting the version of the protocol implemend by
the server. Follow the semantic versioning rules at https://semver.org
Hopefully this bring better user experience when we make breaking
changes at the protocol level and avoid issues like #12465
* graph : make mla compatible with FA
* metal : add exp FA kernels for DeepSeek models
ggml-ci
* llama : minor naming updates
ggml-ci
* ggml : disable FA for DS head sizes
* tests : add FA tests for MLA shapes
ggml-ci
Submit operators using asynchronous threads to improve performance.
Use the environment variable GGML_CANN_ASYNC_MODE to control whether
asynchronous submission is enabled. It is disabled by default.
Testing shows a 10%–20% performance improvement in scenarios with
small parameter sizes, especially in quantized models.
The grouped query attention optmization doesn't require a power of two ratio,
the only thing relying on it was the modulo operation written as bitwise &.
split_k need not depend on gqa_ratio - enable it any time there's only one
workgroup in the X dimension. The shader gets the split index from the x coord,
and multiple workgroups in the X dimension (pre-split) indicates a larger
FA operation that wouldn't need splitting.
Replace compile-time `GGML_HIP_UMA` with environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY`. This unifies the usage on NVIDIA and AMD GPUs, and allows a single binary to be shared between integrated and dedicated GPUs.
Multiple optional memory pools are provided for CANN, including VMM,
priority queue-based, and traditional memory pools.
1.When the memory pool is available and GGML_CANN_DISABLE_VMM_POOL
is not defined, the VMM pool is selected by default.
2.Otherwise, if GGML_CANN_ENABLE_BUF_PRIO_POOL is defined,
the priority queue-based memory pool is used.
3.If neither condition is met, the default memory pool is used.
The current usage of the SYCL-Graph extension checks for
the `sycl_ext_oneapi_graph` device aspect. However, it is also
possible to support `sycl_ext_oneapi_limied_graph` devices that
don't support update
* SYCL: Add fp16 support to some elementwise OP kernels
* remove comment
ggml-ci
* Use static_cast directly
* remove not needed cast from tanh
* Use static cast and remove unneeded castings
* Adjust device_support_op for unary OPs
* Use cast_data and typed_data struct to deduplicate casting code
* [CANN] Support ELU and CONV_TRANSPOSE_1D
* [CANN]Modification review comments
* [CANN]Modification review comments
* [CANN]name adjustment
* [CANN]remove lambda used in template
* [CANN]Use std::func instead of template
* [CANN]Modify the code according to the review comments
---------
Signed-off-by: noemotiovon <noemotiovon@gmail.com>
q4_k and q5_k had a lot of redundant global loads where the same 16B of
scale information is repeatedly loaded and decoded during each loop iteration.
This change restructures the loops to more explicitly iterate over whole
blocks in the outer loop (with unrolled inner loop) and to copy/decode the
scale data into shared memory once at the start of each outer loop. The copy
is pipelined so the scale load from global memory is relatively cheap.
This improves q4_k/q5_k model prompt processing performance by around 5-7%.
I briefly tried applying this to q6_k and q4_0, and it didn't help for q6_k
and hurt for q4_0.
The big "else" path in mul_mm_cm2.comp that had all the clamped/unclamped
variants isn't used as often as it originally was (e.g. due to the padded_N
change), so I trimmed it down to offset some of the new complexity of the
semi-manual loop unrolling.
* ggml : FA supports F32 V
* graph : cast KV to F16 when the KV cache is not used
ggml-ci
* server : add test that exercises embeddings with FA enabled
ggml-ci
nem1 must be a multiple of GGML_KQ_MASK_PAD, and GGML_KQ_MASK_PAD is a multiple
of the number of rows in the matrix. The KV dim is a multiple of the number of
columns for the aligned shader.
There seems to be a bubble waking up from waitForFences, which costs a few
percent performance and also increased variance in performance. This change
inserts an "almost_ready" fence when the graph is about 80% complete and we
waitForFences for the almost_ready fence and then spin (with _mm_pauses) waiting
for the final fence to be signaled.
* Prefer vector flash decoding kernel for Gemma models
Vector flash decoding kernel was not being picked for models with head dimension 256. Gemma models are in this category.
Removing this limit improves e2e performance by upto 12% in gen phase throughput for Gemm models.
* Update ggml/src/ggml-cuda/fattn.cu
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* CUDA: Simplify and improve CUDA graphs through use of indirect copy pointers
Previously there was complexity in the CUDA graphs implementation due
frequently changing parameters to copy kernels associated with K and V
cache pointers. This patch simplifies by using indirection to avoid
such parameters frequently changing, avoiding the need for frequent
graph updates.
Fixes#12152
* Addressed comments
* fix HIP builds
* properly sync to stream
* removed ggml_cuda_cpy_fn_ptrs
* move stream sync before free
* guard to only use indirection with graphs
* style fixes
* check for errors
---------
Co-authored-by: slaren <slarengh@gmail.com>
When using group query attention, we have one workgroup per KV batch and this
can be very few workgroups (e.g. just 8 in some models). Enable split_k to
spread the work across SMs. This helps a lot when the KV cache is large.
When adjacent batches of Q share the same batches of K/V, batch them into
the same workgroup. For example, when:
dst(128,32,1,1) = FA(q(128,1,32,1), k(128,16640,8,1), v(128,16640,8,1))
previously we would run 32 workgroups computing 1 result each, now we will
run 8 workgroups computing 4 results each.
This doesn't directly translate to better performance (at least when you have
>=32 SMs), but in a subsequent change I'll enable split_k which will scale much
better with 4x fewer workgroups.
* add bf16 support
* use convert_from_bf16_cuda instead of convert_unary_cuda for f32
* revert 7ec5085
* move functionality into convert_unary with constexpr
* cpu: refactor SIMD mappings and vectorized op functions into separate files
* Fix warning for ggml_float to float
* Fix warnings
* cpu: move all the operations (except mul_mat) to a separate c++ file
* fix whitespace
* Update ggml/src/ggml-cpu/vec.h
Co-authored-by: Diego Devesa <slarengh@gmail.com>
* Fix PR comments - use GGML_UNUSED, use cassert in ops.cpp
* Reverse the order of import for ops.h and vec.h, to match what was present in ggml-cpu.c previously
---------
Co-authored-by: Diego Devesa <slarengh@gmail.com>
* Rename oneMKL Interface to oneMath
* Use oneMath for Intel vendor
* Rename occurences to mkl
* clang-format
* Silence verbose warnings
* Set oneMath HIP_TARGETS
* Fix silence warnings
* Remove step to build oneMath from build instructions
* Use fixed oneMath version
* Remove INTEL_CPU
* Fold CMake oneDNN conditions
* Use Intel oneMKL for Intel devices
* Improve CMake message
* Link against MKL::MKL_SYCL::BLAS only
* Move oneMath documentation to Nvidia and AMD sections
If users already set CMAKE_C_COMPILER_LAUNCHER globally, setting it in
cmake again will lead to conflict and compile fail.
Signed-off-by: Jay <BusyJay@users.noreply.github.com>
* ggml : FA with different K, V head sizes (CPU)
ggml-ci
* metal : add FA with HS=192
* metal : extend FA to support different K and V head sizes
ggml-ci
* metal : add FA vector kernels for heads K 192 and V 128
ggml-ci
* ggml : restrict op on other backends to equal head sizes
ggml-ci
* metal : optimize FA-vec kernel
ggml-ci
* metal : FA remove mq registers
* metal : improve MoE mul_mat_id condition
ggml-ci
* metal : fix comments + remove unnecessary addition
ggml-ci
* metal : avoid too much shared memory usage with mul_mat_id
ggml-ci
* vulkan: fix coopmat shader generation when cross-compiling
Previously the status of coopmat{,2} support isn't passed to the
vulkan-shaders-gen project building on the host, which leads to build
failure because of the cross-compiling code expecting coopmat{,2}
shaders that didn't get generated.
Fix this by passing the coopmat{,2} support status to vulkan-shaders
subproject.
Signed-off-by: Icenowy Zheng <uwu@icenowy.me>
* Only call coop-mat shaders once
* Fix whitespace
---------
Signed-off-by: Icenowy Zheng <uwu@icenowy.me>
Co-authored-by: bandoti <141645996+bandoti@users.noreply.github.com>
This patch enables usage of MMA when one of the
dimensions of the matrix(ie either M or N) is 1. This
is useful in case of token generation where N < 2.
The concept of 'GEMV Forwarding' is used where when one
of the matrix has a single row/column, the elements are
broadcasted, instead of using packing routine to prepack
the matrix elements.
This change results in 5% - 15% improvement in total
speed(ie all tokens/total time), across various batch
sizes. This is in comparision with the corresponding
dot product implementation.
The patch is tested with FP32 models of Meta-Lllama-3-8B,
Mistral-7B, Llama-2-7B-chat-hf on a IBM POWER10 machine.
Signed-off-by: Amrita H S <amritahs@linux.vnet.ibm.com>
* rpc : send hash when tensor data is above some fixed threshold
ref #10095
* rpc : put cache under $HOME/.cache/llama.cpp
* try to fix win32 build
* another try to fix win32 build
* remove llama as dependency
This change upstreams llamafile's cpu matrix
multiplication kernels for ppc64le ISA using MMA
builtins. This patch handles matrix multiplication
between quantised datatypes, block_q4_0 and
block_q8_0.
This change results in 5% - 50% 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>
The OOB calculation could be wrong if the last iteration was during one of
the unrolled loops. Adjust the unrolling counts to avoid this. Add a couple
new backend tests that hit this failure on NVIDIA GPUs.
* tests: add mul_mat perf/functional tests for p021/nc vulkan shaders
* vulkan: Optimize mul_mat_vec p021 and nc shaders.
These shaders are used in attention calculations, and when the KV cache grows
large they start to dominate the run time. For the nc shader (which is called
with large 'k' dimension), use unrolling and vector loads. For the p021 shader
(which is called with large 'm' and small 'k' dimensions), take advantage of
grouped query attention to reuse loads from the A matrix for the whole group,
and reduce the number of workgroups (too much overhead from tiny dispatches).
Using subgroupAdd in the p021 shader also helps, use that conditionally.
* [SYCL] Fix build on Windows when ccache enabled (llama/9954)
* take effect only on windows and force it to icl
---------
Co-authored-by: Romain Biessy <romain.biessy@codeplay.com>
* Add block interleaving support for Q4_K quantization
* Remove whitespaces and fix CI/CD issues
* Update pointer of bsums from int16_t to const int16_t
* Add vector version of quantize_q8_K_4x8 function
* Update code formatting based on review comments
- Find out active blocks per SM using cudaOccupancyMaxActiveBlocksPerMultiprocessor API. Use this value to determine the optimal parallel_blocks value.
- Prefer vector flash attention kernels over MMA kernel for BS=1
Fixes Issue: #12182
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* ci: add visionOS build workflow
Add a new GitHub Actions workflow for building on visionOS with CMake and Xcode.
* ggml: Define _DARWIN_C_SOURCE for visionOS to fix missing u_xxx typedefs
* ci: remove define hacks for u_xxx system types
---------
Co-authored-by: Giovanni Petrantoni <7008900+sinkingsugar@users.noreply.github.com>
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.
SXX
Neo Zhang Jianyu
Radoslav Gerganov
Diego Devesa
Daniel Bevenius
Georgi Gerganov
lhez
Johannes Gäßler
Eve
Akarshan Biswas
Jeff Bolz
Jeffrey Morgan
Alan Gray
hipudding
kimminsu
Chenguang Li
David Huang
Srihari-mcw
Xinpeng Dou
Acly
Ewan Crawford
Aaron Teo
cmdr2
Prajwal B Mehendarkar
Piotr Kubaj
Sigbjørn Skjæret
zhouwg
R0CKSTAR
0cc4m
Nicolò Scipione
Ronny Brendel
Gaurav Garg
a3sh
bandoti
Junil Kim
Wagner Bruna
Romain Biessy
Jay
Xuan-Son Nguyen
Sandro Hanea
Icenowy Zheng
amritahs-ibm
Slobodan Josic
Dan Johansson
stduhpf
蕭澧邦
Svetlozar Georgiev
Guus Waals
Łukasz Ślusarczyk
fj-y-saito
Molly Sophia
Daniele
Christian Kastner
aubreyli
fairydreaming
Alberto Cabrera Pérez