40 Commits

Author SHA1 Message Date
agray3
042e95d92f Vectorize load instructions in dmmv f16 CUDA kernel (llama/9816)
* Vectorize load instructions in dmmv f16 CUDA kernel

Replaces scalar with vector load instructions, which substantially
improves performance on NVIDIA HBM GPUs, e.g. gives a 1.27X overall
speedup for Meta-Llama-3-8B-Instruct-F16 BS1 inference evaluation on
H100 SXM 80GB HBM3. On GDDR GPUs, there is a slight (1.01X) speedup.

* addressed comment

* Update ggml/src/ggml-cuda/dmmv.cu

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

---------

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-11-01 10:19:05 +02:00
Johannes Gäßler
936cf3beb7 ggml/ex: calculate accuracy in graph, adapt MNIST (ggml/980) 2024-10-05 15:23:51 +03:00
Johannes Gäßler
bb57ecb85e CUDA: remove bad assert (ggml/972) 2024-10-03 12:22:17 +03:00
Ivan
2fc1d20f9e cuda: add q8_0->f32 cpy operation (llama/9571)
llama: enable K-shift for quantized KV cache
It will fail on unsupported backends or quant types.
2024-09-24 19:45:08 +03:00
R0CKSTAR
13f41af43e musa: enable building fat binaries, enable unified memory, and disable Flash Attention on QY1 (MTT S80) (llama/9526)
* mtgpu: add mp_21 support

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* mtgpu: disable flash attention on qy1 (MTT S80); disable q3_k and mul_mat_batched_cublas

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* mtgpu: enable unified memory

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* mtgpu: map cublasOperation_t to mublasOperation_t (sync code to latest)

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

---------

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
2024-09-24 19:45:08 +03:00
Johannes Gäßler
adf2474b10 CUDA: enable Gemma FA for HIP/Pascal (llama/9581) 2024-09-24 19:45:08 +03:00
Molly Sophia
008816a257 RWKV v6: RWKV_WKV op CUDA implementation (llama/9454)
* ggml: CUDA unary op EXP

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* ggml: rwkv_wkv op CUDA impl

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

---------

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>
2024-09-24 19:45:08 +03:00
agray3
f0a7d65b3d Update CUDA graph on scale change plus clear nodes/params (llama/9550)
* Avoid using saved CUDA graph if scale changes and reset nodes/params on update

Fixes https://github.com/ggerganov/llama.cpp/issues/9451

* clear before resize
2024-09-24 19:45:08 +03:00
Georgi Gerganov
d245d7aec7 ggml : fix builds (llama/0)
ggml-ci
2024-09-24 19:45:08 +03:00
Johannes Gäßler
c0761c95f5 CUDA: fix sum.cu compilation for CUDA < 11.7 (llama/9562) 2024-09-24 19:45:08 +03:00
Johannes Gäßler
a53b69a003 CUDA: fix --split-mode row race condition (llama/9413) 2024-09-24 19:45:08 +03:00
R0CKSTAR
d1c9b47360 musa: remove Clang builtins mapping (llama/9421)
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
2024-09-24 19:45:08 +03:00
Johannes Gäßler
67725ac8f3 CUDA: fix variable name conflict for Windows build (llama/9382) 2024-09-24 19:45:08 +03:00
Georgi Gerganov
26225f1fb0 cuda : fix FA Q src index (1 -> 0) (llama/9374) 2024-09-24 19:45:08 +03:00
Johannes Gäßler
c7515b0995 ggml/examples: add backend support for numerical optimization (ggml/949)
* CUDA eval works

* stochastic gradient descent op

* Adam except decay

* CUDA CROSS_ENTROPY_LOSS_BACK

* CUDA mnist-fc training works

* backend CLI arg

* refactor gguf load

* remove sched from opt_step_adam

* implement l1 regularization (weight decay)

* extra call to add optimizer

* initialize gradients with ggml_graph_reset

* gradient accumulation

* increment iter per eval instead of epoch

* adjust backend interfaces

* fix ggml_graph_reset without backend

* fix ggml graph export/import

* fixup

* rename

* revert ggml_opt changes

* more general CUDA repeat_back

* update documentation, fix CNN

* validation split

* add clarifying comment

* optimize PyTorch training

* adjust buffer size, thread count

* fix 0.0f validation split

* Update examples/mnist/mnist-common.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* fix gradient accumulation

* tensor flag for accumulators -> tensor hash set

* Update include/ggml.h

Co-authored-by: slaren <slarengh@gmail.com>

* Update tests/test-backend-ops.cpp

Co-authored-by: slaren <slarengh@gmail.com>

* Update tests/test-backend-ops.cpp

Co-authored-by: slaren <slarengh@gmail.com>

* fix test prints

* Update src/ggml-backend.c

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* better CUDA support for noncontiguous out_prod

* add comment

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: slaren <slarengh@gmail.com>
2024-09-24 19:45:08 +03:00
slaren
709a22b92d cuda : fix defrag with quantized KV (llama/9319) 2024-09-24 19:45:08 +03:00
Johannes Gäßler
5d6dc19f04 tests: add gradient tests for all backends (ggml/932)
* tests: add gradient checking to test-backend-ops

* remove old comment

* reorder includes

* adjust SIN/COS parameters

* add documentation, use supports_op if possible
2024-09-24 19:45:08 +03:00
Johannes Gäßler
24d8534bd8 CPU/CUDA: Gemma 2 FlashAttention support (llama/8542)
* CPU/CUDA: Gemma 2 FlashAttention support

* apply logit_softcap to scale in kernel

* disable logit softcapping tests on Metal

* remove metal check
2024-08-28 13:22:20 +03:00
Daniel Bevenius
60098d6204 ggml : move rope type enum to ggml.h (llama/8949)
* ggml : move rope type enum to ggml.h

This commit moves the `llama_rope_type` enum from `llama.h` to
`ggml.h` and changes its name to `ggml_rope_type`.

The motivation for this change is to address the TODO in `llama.h` and
use the enum in ggml.

Note: This commit does not change the `mode` parameter to be of type
`enum ggml_rope_type`. The name `mode` and its usage suggest that it
might be more generic and possibly used as a bit field for multiple
flags. Further investigation/discussion may be needed to determine
if `mode` should be restricted to RoPE types.

* squash! ggml : move rope type enum to ggml.h

This commit removes GGML_ROPE_TYPE_NONE and GGML_ROPE_TYPE_GLM from
ggml.h, and back the llama_rope_type enum.

I've kept the assert for GGML_ROPE_TYPE_GLM as I'm not sure if it is
safe to remove it yet.

* squash! ggml : move rope type enum to ggml.h

This commit removes the enum ggml_rope_type from ggml.h and replaces it
with a define (GGML_ROPE_TYPE_NEOX). This define is used in the code to
check if the mode is set to GPT-NeoX. Also the enum llama_rope_type has
been updated to reflect this change.

* squash! ggml : move rope type enum to ggml.h

This commit contains a suggestion enable the GGML_ROPE_TYPE_NEOX
macro/define to be passed to the shader compiler.

* squash! ggml : move rope type enum to ggml.h

This commit fixes the editorconfig-checker warnings.

* squash! ggml : move rope type enum to ggml.h

Update comment for ggml_rope function.

* Revert "squash! ggml : move rope type enum to ggml.h"

This reverts commit 6261222bd0dc0efd51f0fb0435ad3f16a5b52fd6.

* squash! ggml : move rope type enum to ggml.h

Add GGML_ROPE_TYPE_NEOX to rope_common.comp.

* remove extra line

---------

Co-authored-by: slaren <slarengh@gmail.com>
2024-08-28 13:22:20 +03:00
Johannes Gäßler
8954769aa2 feat: ref. cross entropy, add CUDA, fix grad test (ggml/929) 2024-08-28 13:22:20 +03:00
Radoslav Gerganov
b6c05ce82f yolo : add backend support (ggml/924)
* yolo : add backend support

* metal : add sub and sqrt kernels

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-08-21 11:07:13 +03:00
Ronsor
3643120690 feat: add new sin and cos operators (ggml/919)
* ggml : add sin/cos operators

* ggml-cuda : add sin/cos operators

* ggml : add corresponding tests for sin/cos

* ggml : add backward computation for sin/cos operators

* ggml-vulkan : add sin/cos operators

* ggml-vulkan : add sin/cos shader source

* metal : add sin, cos

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-08-21 11:07:13 +03:00
Molly Sophia
4160b930f1 ggml : add epsilon as a parameter for group_norm (llama/8818)
Signed-off-by: Molly Sophia <mollysophia379@gmail.com>
2024-08-08 22:48:46 +03:00
slaren
5218ea21b8 cuda : fix dmmv cols requirement to 2*GGML_CUDA_DMMV_X (llama/8800)
* cuda : fix dmmv cols requirement to 2*GGML_CUDA_DMMV_X

* update asserts

* only use dmmv for supported types

* add test
2024-08-08 22:48:46 +03:00
R0CKSTAR
49ac8872b4 cuda : organize vendor-specific headers into vendors directory (llama/8746)
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
2024-08-08 22:48:46 +03:00
R0CKSTAR
e471adcfa5 feat: Support Moore Threads GPU (llama/8383)
* Update doc for MUSA

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* Add GGML_MUSA in Makefile

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* Add GGML_MUSA in CMake

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* CUDA => MUSA

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* MUSA adds support for __vsubss4

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* Fix CI build failure

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

---------

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
2024-08-08 22:48:46 +03:00
slaren
dd916a2852 ggml : reduce hash table reset cost (llama/8698)
* ggml : reduce hash table reset cost

* fix unreachable code warnings after GGML_ASSERT(false)

* GGML_ASSERT(false) -> GGML_ABORT("fatal error")

* GGML_ABORT use format string
2024-08-08 22:48:46 +03:00
Jeroen Mostert
86506b0c5c Allow all RDNA2 archs to use sdot4 intrinsic (llama/8629)
The check gating the use of `__builtin_amdgc_sdot4` specifically checks for gfx1030. This causes a severe perf regression for anything gfx103? that's not gfx1030 and not using `HSA_OVERRIDE_GFX_VERSION` (if you've built ROCm to support it). We already have a generic RDNA2 define, let's use it.
2024-08-08 22:48:46 +03:00
Johannes Gäßler
8c4f30497a CUDA: MMQ code deduplication + iquant support (llama/8495)
* CUDA: MMQ code deduplication + iquant support

* 1 less parallel job for CI build
2024-08-08 22:48:46 +03:00
Daniel Bevenius
84493d7f3e cuda : suppress 'noreturn' warn in no_device_code (llama/8414)
* cuda : suppress 'noreturn' warn in no_device_code

This commit adds a while(true) loop to the no_device_code function in
common.cuh. This is done to suppress the warning:

```console
/src/ggml-cuda/template-instances/../common.cuh:346:1: warning:
function declared 'noreturn' should not return [-Winvalid-noreturn]
  346 | }
      | ^
```

The motivation for this is to reduce the number of warnings when
compilng with GGML_HIPBLAS=ON.

Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com>

* squash! cuda : suppress 'noreturn' warn in no_device_code

Update __trap macro instead of using a while loop to suppress the
warning.

Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com>

---------

Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com>
2024-08-08 22:48:46 +03:00
Johannes Gäßler
15d71189e9 CUDA: optimize and refactor MMQ (llama/8416)
* CUDA: optimize and refactor MMQ

* explicit q8_1 memory layouts, add documentation
2024-08-08 22:48:46 +03:00
Johannes Gäßler
5dc636a65a CUDA: MMQ support for iq4_nl, iq4_xs (llama/8278) 2024-07-08 14:53:55 +03:00
Daniele
73703a144f CUDA: revert part of the RDNA1 optimizations (llama/8309)
The change on the launch_bounds was causing a small performance drop in perplexity of 25 t/s
2024-07-08 14:53:55 +03:00
Johannes Gäßler
e89fdceec2 CUDA: fix MMQ stream-k rounding if ne00 % 128 != 0 (llama/8311) 2024-07-08 14:53:55 +03:00
Daniele
95e90823d9 Define and optimize RDNA1 (llama/8085) 2024-07-08 14:53:55 +03:00
Clint Herron
c2c60dc9ba Removes multiple newlines at the end of files that is breaking the editorconfig step of CI. (llama/8258) 2024-07-08 14:53:55 +03:00
Johannes Gäßler
e4bc83ab47 CUDA: refactor and optimize IQ MMVQ (llama/8215)
* CUDA: refactor and optimize IQ MMVQ

* uint -> uint32_t

* __dp4a -> ggml_cuda_dp4a

* remove MIN_CC_DP4A checks

* change default

* try CI fix
2024-07-08 14:53:55 +03:00
Johannes Gäßler
bf88c94da9 CUDA: fix MMQ stream-k for --split-mode row (llama/8167) 2024-07-08 14:53:55 +03:00
John Balis
3eea171cab feat: cuda implementation for ggml_conv_transpose_1d (ggml/854)
* conv transpose 1d passing test for 1d input and kernel

* working for different input and output channel counts, added test for variable stride

* initial draft appears to work with stride other than 1

* working with all old and new conv1d  tests

* added a test for large tensors

* removed use cuda hardcoding

* restored test-conv-transpose.c

* removed unused arugments, and fixed bug where test failure would cause subsequent tests to fail

* fixed accumulator bug

* added test to test-backend-ops

* fixed mistake

* addressed review

* fixed includes

* removed blank lines

* style and warning fixes

* return failure when test fails

* fix supports_op

---------

Co-authored-by: slaren <slarengh@gmail.com>
2024-07-08 14:53:55 +03:00
Georgi Gerganov
e30c679928
whisper : reorganize source code + improve CMake (#2256)
* scripts : update sync [no ci]

* files : reorganize [no ci]

* sync : llama.cpp

* cmake : link math library

* cmake : build normal ggml library

* files : move headers to include

* objc : fix path to ggml-metal.h

* ci : fix WHISPER_CUDA -> GGML_CUDA

* scripts : sync LICENSE [no ci]
2024-06-26 19:34:09 +03:00