* CANN: fix RoPE cache issue on multi-device
RoPE cache only needs to be computed once per token.
However, in multi-device scenarios, not every device starts
computation from layer 0, which may lead to unallocated memory
issues and precision errors.
This commit records the first layer of each device to avoid
the above issues.
* CANN: Optimize first-layer detection method
* CANN: Remove trailing whitespace
* CANN: Only cache the data that can be determined as unchanged through the parameters.
* CANN: Update function comment
Exposes ggml_backend_sched_split_graph() to allow splitting the graph without allocating compute buffers and uses it to split the graph for the automatic Flash Attention check.
* vulkan: mul_mat_id coopmat2 optimizations
Add a path for when the tile fits in BN/2, similar to what we have for mul_mat.
Only call fetch_scales/store_scales once per QUANT_K block, and once at the
beginning in case start_k is not aligned.
* Also add a path for BN/4 - worth a couple more percent
This commit removes the portability_enumeration_ext variable from the
ggml_vk_instance_portability_enumeration_ext_available function as it
is initialized to false but never modified, making it redundant.
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.
* CANN(flash-attn): refactor mask handling and improve performance
1. Refactored the mask computation in Flash Attention, unified the logic without separating prefill and decode.
2. Optimized performance in non-alibi scenarios by reducing one repeat operation.
3. Updated operator management to explicitly mark unsupported cases on 310P devices and when dim is not divisible by 16.
Signed-off-by: noemotiovon <757486878@qq.com>
* [CANN]: fix review
Signed-off-by: noemotiovon <757486878@qq.com>
* [CANN]: Optimization FA BNSD to BSND
Signed-off-by: noemotiovon <757486878@qq.com>
---------
Signed-off-by: noemotiovon <757486878@qq.com>
The original implementation unconditionally returned true for this operation, leading to a failure when the tensor's first dimension (ne[0]) was not a multiple of WARP_SIZE. This caused an GGML_ASSERT(ncols % WARP_SIZE == 0) failure in ggml-sycl/norm.cpp.
This change updates the ggml_backend_sycl_device_supports_op check to correctly return true for GGML_OP_RMS_NORM only when the first dimension of the tensor is a multiple of WARP_SIZE, ensuring the operation can be performed without error.
This patch improves GEMM for FP32 Data Type on PowerPC
Implements GEMM on large blocks with configurable block size mc, nc, kc
(default: 256, 256, 256).
Packing Function optimized to access blocks as per memory layout.
GEMM Optimized to work on larger blocks.
Isolated Packing from GEMM Operations for better MMA utilization.
Verified functionality and correctness uing llama-cli and stand alone
test case (performs matmul and compares final mattrix C result with base).
Minor code refactoring changes:
Replace macro with inline function
Code Indent made consistent with 4 spaces
Performance Testing:
Observed 50% ~ 70% improvement in Prompt Processing Speed mesured using
llama-bench with Meta-Llama3-8B FP32 Model. Similar gains observed with
Mistral-7b-Instruct-v0.3 Model.
model Size Params Backend Threads Test Patch Base
llama 8B all F32 29.92 GiB 8.03 B CPU 20 pp512 98.58 60.3
llama 8B all F32 29.92 GiB 8.03 B CPU 20 pp1024 95.88 57.36
llama 8B all F32 29.92 GiB 8.03 B CPU 20 pp2048 85.46 53.26
llama 8B all F32 29.92 GiB 8.03 B CPU 20 pp4096 68.66 45.78
llama 8B all F32 29.92 GiB 8.03 B CPU 20 pp6144 57.35 40.44
25 ~ 30% improvement in llama-batched-bench with Metla-Llama3-8B in
Prompt Processing Speed for large prompts (256, 512, 1024, 2048, 4096)tokens with various batch
sizes ( 1, 2, 4, 8, 16)
Signed-off-by: Shalini Salomi Bodapati <Shalini.Salomi.Bodapati@ibm.com>
* metal : optmize FA vec for large heads and sequences
* metal : adjust small-batch mul mv kernels
ggml-ci
* batched-bench : fix total speed computation
ggml-ci
* cont : add comments
ggml-ci
* metal : mul_mm_id remove hdst
* metal : remove mul_mm_id hsrc1
* metal : mul_mm_id simplify + add test
* metal : opt mul_mm_id map0
* metal : optimize mul_mm_id id gathering
* metal : mul/div opt
* metal : optimize mul_mm_id_map0
ggml-ci
* 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: use subgroup function for mul_mat_id shader even without coopmat
* vulkan: fix compile warnings
* vulkan: properly check for subgroup size control and require full subgroups for subgroup mul_mat_id
* vulkan: disable subgroup mul_mat_id on devices with subgroups < 16
The scalar FA shader already handled multiples of 8. The coopmat1 FA
shader assumed 16x16x16 and the shared memory allocations need the HSK
dimensions padded to a multiple of 16. NVIDIA's coopmat2 implementation
requires multiples of 16 for N and K, and needs the matrix dimensions
padded and loads clamped.
Store the FA pipelines in a map, indexed by the pipeline state.
* vulkan: optimize rms_norm, and allow the work to spread across multiple SMs
There are really two parts to this change:
(1) Some optimizations similar to what we have in soft_max, to unroll with
different numbers of iterations.
(2) A fusion optimization where we detect add followed by rms_norm, and make
the add shader atomically accumulate the values^2 into memory. Then the
rms_norm shader can just load that sum. This allows the rms_norm to be
parallelized across multiple workgroups, it just becomes a simple per-element
multiply.
The fusion optimization is currently only applied when the rms_norm is on a
single vector. This previously always ran on a single SM. It could apply more
broadly, but when there are other dimensions the work can already spread across
SMs, and there would be some complexity to tracking multiple atomic sums.
* Change add+rms_norm optimization to write out an array of partial sums
rather than using atomic add, to make it deterministic. The rms_norm
shader fetches a subgroup's worth in parallel and uses subgroupAdd to
add them up.
* complete rebase against fused adds - multi_add shader can also compute partial sums
* fix validation errors
* disable add_rms_fusion for Intel due to possible driver bug
* resolve against #15489, sync after clearing partial sums
Track a list of nodes that need synchronization, and only sync if the new node
depends on them (or overwrites them). This allows some overlap which can
improve performance, and centralizes a big chunk of the synchronization logic.
The remaining synchronization logic involves writes to memory other than the
nodes, e.g. for dequantization or split_k. Each of these allocations has a bool
indicating whether they were in use and need to be synced. This should be
checked before they are written to, and set to true after they are done being
consumed.
* 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
- Spread the work across the whole workgroup. Using more threads seems to
far outweigh the synchronization overhead.
- Specialize the code for when the division is by a power of two.
* Begin work on set_rows
* Work on set rows
* Add error buffers for reporting unsupported SET_ROWS indices
* Remove extra comments
* Work on templating for different types in shaders
* Work on shader type generation
* Working q4_0 mul_mat and some templating for different types
* Add q4_0_f16 matmul and fix device init
* Add matmul support for basic quantization types
* Add q2_k and q3_k quantization
* Add rest of k-quants
* Get firt i-quant working
* Closer to supporting all i-quants
* Support rest of i-quants
* Cleanup code
* Fix python formatting
* debug
* Bugfix for memset
* Add padding to end of buffers on creation
* Simplify bit-shifting
* Update usage of StringView
* 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>
* vulkan: Reuse conversion results in prealloc_y
Cache the pipeline and tensor that were most recently used to fill prealloc_y,
and skip the conversion if the current pipeline/tensor match.
* don't use shared pointer for prealloc_y_last_pipeline_used
* 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>
* 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>
hipudding
Georgi Gerganov
Diego Devesa
Jeff Bolz
Daniel Bevenius
Charles Xu
Johannes Gäßler
Chenguang Li
Aman Gupta
mnehete32
Aaron Teo
compilade
matiaslin
uvos
xctan
rmatif
Akarshan Biswas
shalinib-ibm
Sigbjørn Skjæret
Yoshi_likes_e4
Qeeweew
lhez
Ruben Ortlam
Ihar Hrachyshka
Acly
Reese Levine
Yavor Ivanov
Dong Won Kim
Xuan-Son Nguyen
R0CKSTAR
SHUAI YANG
Marvin Gießing
rmatif
Jason Ni
Jonathan Graehl
Judd
Oliver Simons
Tak-RS
Romain Biessy