207 Commits

Author	SHA1	Message	Date
Radoslav Gerganov	00b7317e63	vulkan : do not use tensor->extra (#9407) ... * vulkan : do not use tensor->extra This patch allows using the Vulkan backend with the RPC backend as tensor->extra is no longer used. Ref: #8536 * Adapt GGML_VULKAN_CHECK_RESULTS to extra removal (#2) --------- Co-authored-by: 0cc4m <picard12@live.de>	2024-10-02 13:49:16 +03:00
Johannes Gäßler	e98c1c188e	test: fix OPT_STEP_ADAMW for test-backend-ops (ggml/974)	2024-10-01 16:07:40 +03:00
Salvatore Mesoraca	cb00020504	vulkan : mul_mat: fix UB with small warps (ggml/952) ... When the device's warp size is less than 16, it is possible for loadstride_a (mul_mm.comp:114) and loadstride_b (mul_mm.comp:115) to be set to 0. Because they are calculated as: the workgroup size, multiplied by LOAD_VEC_* (which can be 1) and divided by 16. And the workgroup size is set to be the same as the warp/subgroup size. The loadstride_* variables are used as increments in the loops that populate the buffers used for the multiplication. When they are 0 they cause an infinite loop. But infinite loops without side-effects are UB and the values of loadstride_* are known at compile time. So, the compiler quietly optimizes all the loops away. As a consequence, the buffers are not populated and the multiplication result is just a matrix with all elements set to 0. We prevent the UB by making sure that the workgroup size will never be less than 16, even if our device has a smaller warp size (e.g. 8). Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com>	2024-10-01 16:07:39 +03:00
Borislav Stanimirov	6c5322481a	ggml : fix ggml_cast (ggml/973)	2024-10-01 16:07:39 +03:00
Johannes Gäßler	7254cdf7e8	ggml: fix gradient allocation logic (ggml/966) ... * ggml: fix gradient allocation logic * gradient allocation in ggml_build_backward_expand * fixup * fix test-backend-ops grad * suggestions by slaren * fix test1.c * fix legacy opt API * fix test-grad0 * remove keep arg	2024-10-01 16:07:38 +03:00
Georgi Gerganov	cad341d889	metal : reduce command encoding overhead (#9698) ... * metal : reduce command encoding overhead ggml-ci * metal : add comments	2024-10-01 16:00:25 +03:00
Georgi Gerganov	c919d5db39	ggml : define missing HWCAP flags (#9684) ... ggml-ci Co-authored-by: Willy Tarreau <w@1wt.eu>	2024-09-29 21:18:23 +03:00
Johannes Gäßler	aaa4099925	CUDA: remove bad assert (ggml/972)	2024-09-29 21:15:37 +03:00
Jeff Bolz	641002fba8	vulkan : multithread pipeline creation (ggml/963)	2024-09-29 21:15:37 +03:00
Jeff Bolz	0de8b203f1	vulkan : fix build for GGML_VULKAN_RUN_TESTS, add TFLOPS to log (ggml/961)	2024-09-29 21:15:37 +03:00
Salvatore Mesoraca	544f409b4b	vulkan : argsort barriers must be under uniform control flow (ggml/951) ... a return before a barrier (that happens only in some threads in a workgroup) leads to UB. While the old code actually works on some devices, it fails on some others (i.e. "smaller" GPUs). BTW, I think it would be better to set specialization constants when the graph is built, in that way the local workgroup could be sized appropriately. But it would take a lot of work. Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com>	2024-09-29 21:15:37 +03:00
Georgi Gerganov	6084bfb261	ggml : fix GGML_MAX_N_THREADS + improve formatting (ggml/969)	2024-09-29 21:15:35 +03:00
Dan Johansson	6a0f779484	ggml : add run-time detection of neon, i8mm and sve (#9331) ... * ggml: Added run-time detection of neon, i8mm and sve Adds run-time detection of the Arm instructions set features neon, i8mm and sve for Linux and Apple build targets. * ggml: Extend feature detection to include non aarch64 Arm arch * ggml: Move definition of ggml_arm_arch_features to the global data section	2024-09-28 15:06:16 +03:00
Markus Tavenrath	89f9944981	Enable use to the rebar feature to upload buffers to the device. (#9251)	2024-09-28 12:05:05 +02:00
R0CKSTAR	7691654c68	mtgpu: enable VMM (#9597) ... Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>	2024-09-26 03:27:40 +02:00
Charles Xu	1e43630218	ggml : remove assert for AArch64 GEMV and GEMM Q4 kernels (#9217) ... * ggml : remove assert for AArch64 GEMV and GEMM Q4 kernels * added fallback mechanism when the offline re-quantized model is not optimized for the underlying target. * fix for build errors * remove prints from the low-level code * Rebase to the latest upstream	2024-09-25 16:12:20 +03:00
Dou Xinpeng	904837e0cb	cann: fix crash when llama-bench is running on multiple cann devices (#9627)	2024-09-25 11:30:38 +08:00
Eric Zhang	70392f1f81	ggml : add AVX512DQ requirement for AVX512 builds (#9622)	2024-09-24 11:03:21 +03:00
Georgi Gerganov	c038931615	examples : adapt to ggml.h changes (ggml/0) ... ggml-ci	2024-09-24 11:00:52 +03:00
Georgi Gerganov	cea1486ecf	log : add CONT level for continuing previous log entry (#9610)	2024-09-24 10:15:35 +03:00
Max Krasnyansky	c087b6f11d	threads: fix msvc build without openmp (#9615) ... We're missing atomic_thread_fence() in MSVC builds when openmp is disabled.	2024-09-23 21:18:48 -07:00
Ivan	116efee0ee	cuda: add q8_0->f32 cpy operation (#9571) ... llama: enable K-shift for quantized KV cache It will fail on unsupported backends or quant types.	2024-09-24 02:14:24 +02:00
Max Krasnyansky	f0c7b5edf8	threads: improve ggml_barrier scaling with large number of threads (#9598) ... Make sure n_barrier and n_barrier_passed do not share the cache line to avoid cache line bouncing. This optimization shows performance improvements even for n_threads <= 8 cases. Resurect TSAN (Thread Sanitizer) check so that we can avoid doing expensive read-modify-write in the normal case and just use thread-fence as originally intended. --- Here is the original description and suggestions from Willy Tarreau : There's currently some false sharing between n_barrier and n_barrier_passed that is amplified in ggml_barrier() by the fact that all threads need to increment n_barrier when entering, while all previous threads continue to read n_barrier_passed, waiting for the last one to release them all. The side effect is that all these readers are slowing down all new threads by making the cache line bounce back and forth between readers and writers. Just placing them in two distinct cache lines is sufficient to boost the performance by 21% on a 80-core ARM server compared to the no-openmp version, and by 3% compared to the openmp version. Note that the variables could have been spread apart in the structure as well, but it doesn't seem that the size of this threadpool struct is critical so here we're simply aligning them. Finally, the same issue was present when leaving the barrier since all threads had to update the n_barrier_passed counter, though only one would add a non-zero value. This alone is responsible for half of the cost due to undesired serialization. It might be possible that using a small array of n_barrier counters could make things even faster on many-core systems, but it would likely complicate the logic needed to detect the last thread. Co-authored-by: Willy Tarreau <w@1wt.eu>	2024-09-23 11:42:43 -07:00
Srihari-mcw	1e7b9299c6	ggml : AVX512 gemm for Q4_0_8_8 (#9532) ... * AVX512 version of ggml_gemm_q4_0_8x8_q8_0 * Remove zero vector parameter passing * Rename functions and rearrange order of macros * Edit commments * style : minor adjustments * Update x to start from 0 --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2024-09-23 17:06:38 +03:00
Georgi Gerganov	bf9c1013ac	metal : use F32 prec for K*Q in vec FA (#9595) ... ggml-ci	2024-09-23 11:27:47 +03:00
Akarshan Biswas	e62e9789cd	Revert "[SYCL] fallback mmvq (#9088)" (#9579) ... This reverts commit 50addec9a5.	2024-09-23 11:28:06 +08:00
R0CKSTAR	c35e586ea5	musa: enable building fat binaries, enable unified memory, and disable Flash Attention on QY1 (MTT S80) (#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>	2024-09-22 16:55:49 +02:00
Molly Sophia	912c331d3d	Fix merge error in #9454 (#9589) ... Signed-off-by: Molly Sophia <mollysophia379@gmail.com>	2024-09-22 15:26:50 +02:00
Johannes Gäßler	a5b57b08ce	CUDA: enable Gemma FA for HIP/Pascal (#9581)	2024-09-22 09:34:52 +02:00
Molly Sophia	2a63caaa69	RWKV v6: RWKV_WKV op CUDA implementation (#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>	2024-09-22 04:29:12 +02:00
slaren	d09770cae7	ggml-alloc : fix list of allocated tensors with GGML_ALLOCATOR_DEBUG (#9573)	2024-09-21 14:24:23 +02:00
agray3	41f477879f	Update CUDA graph on scale change plus clear nodes/params (#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	2024-09-21 02:41:07 +02:00
Georgi Gerganov	d13edb17ed	ggml : fix builds (#0) ... ggml-ci	2024-09-20 21:15:05 +03:00
Georgi Gerganov	27609c49b9	ggml : fix trailing whitespace (#0) ... ggml-ci	2024-09-20 21:15:05 +03:00
Johannes Gäßler	424c5d00a9	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>	2024-09-20 21:15:05 +03:00
Georgi Gerganov	a6809c6a2e	examples : add null threadpool args where needed (ggml/0) ... ggml-ci	2024-09-20 21:15:05 +03:00
Johannes Gäßler	5cb12f6839	CUDA: fix sum.cu compilation for CUDA < 11.7 (#9562)	2024-09-20 18:35:35 +02:00
slaren	64c6af3195	ggml : fix n_threads_cur initialization with one thread (#9538) ... * ggml : fix n_threads_cur initialization with one thread * Update ggml/src/ggml.c --------- Co-authored-by: Max Krasnyansky <quic_maxk@quicinc.com>	2024-09-18 10:13:08 -07:00
Max Krasnyansky	0226613853	threadpool : skip polling for unused threads (#9461) ... * threadpool: skip polling for unused threads Currently all threads do N polling rounds even if only 1 thread is active (n_threads_cur == 1). This commit adds a check to skip the polling for unused threads (ith >= n_threads_cur). n_threads_cur is now an atomic_int to explicitly tell thread sanitizer that it is written from one thread and read from other threads (not a race conditions). * threadpool: further simplify and improve ggml_barrier Avoid using strict memory order while polling, yet make sure that all threads go through full memory barrier (memory fence) on ggml_barrier entrace and exit. * threads: add simple barrier test This test does lots of small, parallel matmul ops where the barriers in between dominate the overhead. * threadpool: improve thread sync for new-graphs Using the same tricks as ggml_barrier. All the polling is done with relaxed memory order to keep it efficient, once the new graph is detected we do full fence using read-modify-write with strict memory order. * threadpool: improve abort handling Do not use threadpool->ec (exit code) to decide whether to exit the compute loop. threadpool->ec is not atomic which makes thread-sanitizer rightfully unhappy about it. Instead introduce atomic threadpool->abort flag used for this. This is consistent with how we handle threadpool->stop or pause. While at it add an explicit atomic_load for n_threads_cur for consistency. * test-barrier: release threadpool before releasing the context fixes use-after-free detected by gcc thread-sanitizer on x86-64 for some reason llvm sanitizer is not detecting this issue.	2024-09-17 11:19:46 +03:00
slaren	23e0d70bac	ggml : move common CPU backend impl to new header (#9509)	2024-09-16 16:22:07 +02:00
Michael Podvitskiy	a6a3a5c531	ggml : link MATH_LIBRARY not by its full path (#9339)	2024-09-16 14:06:50 +03:00
Georgi Gerganov	19514d632e	cmake : do not hide GGML options + rename option (#9465) ... * cmake : do not hide GGML options ggml-ci * build : rename flag GGML_CUDA_USE_GRAPHS -> GGML_CUDA_GRAPHS for consistency ggml-ci	2024-09-16 10:27:50 +03:00
Eve	5c3d0f1824	ggml : IQ4_NL sgemm + Q4_0 AVX optimization (#9422) ... * squashed readd my iq4_nl sgemm PR https://github.com/ggerganov/llama.cpp/pull/8049 have ggml_vec_dot_q4_0 do two blocks per loop for avx try out f16c ggml_vec_dot_iq4_nl, but it's not really faster. as per https://github.com/ggerganov/llama.cpp/pull/8549 we can calculate several blocks at a time with no issue * shuffle * remove f16c iq4_nl as i cant make it faster than before	2024-09-16 09:48:24 +03:00
Georgi Gerganov	c4965a64f7	metal : handle zero-sized allocs (#9466)	2024-09-16 09:05:56 +03:00
Georgi Gerganov	6262d13e0b	common : reimplement logging (#9418) ... https://github.com/ggerganov/llama.cpp/pull/9418	2024-09-15 20:46:12 +03:00
Michael Podvitskiy	6988da94a2	cmake : correct order of sycl flags (#9497)	2024-09-15 19:55:52 +03:00
Michael Podvitskiy	7596487beb	cmake : try to fix sycl+intel build (#9487)	2024-09-15 10:06:38 +03:00
Yuri Khrustalev	822b6322de	ggml : ggml_type_name return "NONE" for invalid values (#9458) ... When running on Windows, the quantization utility attempts to print the types that are not set which leads to a crash.	2024-09-14 12:54:37 +03:00
Georgi Gerganov	1f4111e540	cmake : use list(APPEND ...) instead of set() + dedup linker (#9463) ... * cmake : use list(APPEND ...) instead of set() + dedup linker ggml-ci * cmake : try fix sycl * cmake : try to fix sycl 2 * cmake : fix sycl build (#9469) * try fix sycl build * use CMAKE_CXX_FLAGS as a string variable --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * one more CMAKE_CXX_FLAGS fix (#9471) --------- Co-authored-by: Michael Podvitskiy <podvitskiymichael@gmail.com>	2024-09-14 10:55:05 +03:00
Dou Xinpeng	e6b7801bd1	cann: Add host buffer type for Ascend NPU (#9406) ... * feat: Add host buffer type for Ascend NPU(CANN backend) * fix some checking errors * Add a few comments	2024-09-12 19:46:43 +08:00