llama : per-layer KV cache #4309

ggerganov · 2023-12-03T15:57:37Z

This is an updated version of #3457 by @slaren

Also, adds support for quantum K cache: #4312

It's adapted to the latest code on master. Should work with all model arches

-ngl now goes up to n_layer + 1 instead of n_layer + 3
use --no-kv-offload or -nkvo to disable KV cache offloading to GPU

`llama.h` API Changes

struct llama_context_params
- add enum ggml_type type_k; // data type for K cache
- add enum ggml_type type_v; // data type for V cache
- add bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
- remove bool f16_kv
LLAMA_SESSION_VERSION bump to 3

Limitations

Context shift does not work with quantum K cache
V cache quantization is not yet supported

TODO:

add command-line args to disable KV offloading (enabled by default)
extensive testing with different models and different -ngl values
double-check that the KV shift on context overflow works correctly (initial tests indicate that it does)
update state save/load to the new KV cache and bump LLAMA_SESSION_VERSION

slaren · 2023-12-03T16:10:30Z

llama.cpp

-    OFFLOAD_FUNC_KQ,
-    OFFLOAD_FUNC_V,


In cparams it looks like it is still possible to offload K and V separately, but wouldn't this change break that?

Good point.

Do we ever want to offload them separately? If not, I will change cparams to just offload_kv

I am not sure, but looking at the difference in performance between 34 and 35 offloaded layers in the graphs in #3457 it is probably not worth supporting offloading V only.

On the koboldcpp fork, loading just one of the two KV cache layers (42/43 instead of 43/43 or 41/43 specifically was a problem) caused strange generation issues for me with seemingly random tokens appearing.
I am not sure if this is a bug unique to the koboldcpp fork, or if this is behavior caused by llama.cpp, but it helps support the argument that it's probably not worth supporting

Ok, since both ways are equally easy to implement, I will not close the door for now and support separate K and V offloading, even if we don't see a use case at the moment

I wonder if it makes a difference specifically in the context of models with/without GQA?

Ok, since both ways are equally easy to implement,

It's actually simpler to go with either both K and V offloaded or neither. Added --no-kv-offload cli arg

llama.cpp

oobabooga · 2023-12-03T18:10:33Z

Here is a test on my hardware (RTX 3090, 3200 MHz RAM).

Command:

./main \
  -m ../models/wizardlm-70b-v1.0.Q4_K_S.gguf \
  -e \
  -t 6 \
  -tb 12 \
  -n 256 \
  -c 4096 \
  -ngl 42 \
  --ignore-eos \
  -p "Because of the rapid pace of improvement of large language models, evaluation benchmarks have suffered from short lifespans, with state of the art models quickly saturating existing benchmarks, exceeding the performance of human annotators, leading to efforts to replace or augment the benchmark with more challenging tasks. In addition, there are cases of shortcut learning wherein AIs sometimes cheat on multiplechoice tests by using statistical correlations in superficial test question wording in order to guess the correct responses, without necessarily understanding the actual question being asked.\n\nSome datasets have been constructed adversarially, focusing on particular problems on which extant language models seem to have unusually poor performance compared to humans. One example is the TruthfulQA dataset, a question answering dataset consisting of  questions which language models are susceptible to answering incorrectly by mimicking falsehoods to which they were repeatedly exposed during training. For example, an LLM may answer No to the question Can you teach an old dog new tricks because of its exposure to the English idiom you cant teach an old dog new tricks, even though this is not literally true.\n\nAnother example of an adversarial evaluation dataset is Swag and its successor, HellaSwag, collections of problems in which one of multiple options must be selected to complete a text passage. The incorrect completions were generated by sampling from a language model and filtering with a set of classifiers. The resulting problems are trivial for humans but at the time the datasets were created state of the art language models had poor accuracy on them. For example:"

PR (`c80b8a2`)

Peak VRAM: 23049MiB

llama_print_timings:        load time =   30315,12 ms
llama_print_timings:      sample time =      66,38 ms /   256 runs   (    0,26 ms per token,  3856,82 tokens per second)
llama_print_timings: prompt eval time =    3365,29 ms /   332 tokens (   10,14 ms per token,    98,65 tokens per second)
llama_print_timings:        eval time =  118578,99 ms /   255 runs   (  465,02 ms per token,     2,15 tokens per second)
llama_print_timings:       total time =  122225,02 ms

Master

Peak VRAM: 22377MiB

llama_print_timings:        load time =   31612,69 ms
llama_print_timings:      sample time =      71,63 ms /   256 runs   (    0,28 ms per token,  3573,87 tokens per second)
llama_print_timings: prompt eval time =    4140,64 ms /   332 tokens (   12,47 ms per token,    80,18 tokens per second)
llama_print_timings:        eval time =  119670,77 ms /   255 runs   (  469,30 ms per token,     2,13 tokens per second)
llama_print_timings:       total time =  124069,09 ms

The output is stable and coherent in both cases. The conclusion is that the changes in this PR increase the tokens/second for prompt evaluation by +23% while increasing the VRAM usage by +672MB in this particular setup.

kalomaze · 2023-12-03T18:25:44Z

The prompt processing speed gains are a massive boon it would seem. 13b 4_K_M. Prompt processing was measured in tokens per second for the same exact prompt (~600 tokens) at different layer offload counts, with 4096 tokens ctx.

Generation tps as well, there's some milder gains there. If we focus on 20-35 it looks like:

CPU: i7-6700
GPU: RTX 3060

kalomaze · 2023-12-03T20:02:32Z

Something that surprised me when testing is that both thread counts seem to matter, perhaps equally now. In the past, it was said that you should focus primarily on batched threads instead of regular threads for CPU offloading... But if I set both to 7, I get better speeds overall. (My CPU has 4 cores and 8 threads)

On 13b, 6_K quantization, 31/41 offloaded layers on RTX 3060 (12GB VRAM):

Threads count 7, Batched threads count 7: 127 t/s prompt processing, 2.5 t/s generation speed
Threads count 2, Batched threads count 7: 137 t/s prompt processing, 1.3t/s generation speed
Threads count 8, Batched threads count 4: 108 t/s prompt processing, 2.09t/s generation speed
Threads count 8, Batched threads count 8: 135 t/s prompt processing, 2.76t/s generation speed

Seems to be optimal to set both to your total CPU thread count, or perhaps 1 layer less for both might more more stable overall.

* llama : support quantum K cache (wip) * metal : add F32 -> Q8_0 copy kernel * cuda : add F32 -> Q8_0 copy kernel ggml-ci * cuda : use mmv kernel for quantum cache ops * llama : pass KV cache type through API * llama : fix build ggml-ci * metal : add F32 -> Q4_0 copy kernel * metal : add F32 -> Q4_1 copy kernel * cuda : wip * cuda : add F32 -> Q4_0 and F32 -> Q4_1 copy kernels * llama-bench : support type_k/type_v * metal : use mm kernel only for quantum KV cache * cuda : add comment * llama : remove memory_f16 and kv_f16 flags --------- Co-authored-by: slaren <slarengh@gmail.com>

ggerganov · 2023-12-06T18:30:44Z

Plan to merge this tomorrow - haven't tested much yet. Hope it does not break something

The CUDA performance with quantum cache is a bit disappointing. Hopefully we will fix this in the future

kalomaze · 2023-12-07T01:05:44Z

Plan to merge this tomorrow - haven't tested much yet. Hope it does not break something

The CUDA performance with quantum cache is a bit disappointing. Hopefully we will fix this in the future

I have been testing my sampler experiments using it and I don't notice any breaking changes, but that was before you added the K layer quantization stuff apparently if I read the other PR right?

kalomaze · 2023-12-07T01:22:45Z

The CUDA performance with quantum cache is a bit disappointing.

Thankfully, I'd argue "disappointing" is potentially misleading, looking at the benchmark provided in the other PR. It was my presumption that quantization actually reduces generation throughput because of the overhead of reducing precision, but if it means you can fit a model entirely into VRAM that wouldn't fit before, or more layers of a model into VRAM specifically (thanks to this wonderful spread KV layers PR), that alone is much more important than the actual compute being slower because VRAM is much, much faster.

TLDR; the speed gains from better VRAM utilization probably outpace the speed loss from having to quantize the KV cache if you're VRAM bound. My guess is that high compute low VRAM benefits the most (e.g. most consumer gaming GPUs!!!)

ggerganov · 2023-12-07T08:35:45Z

I have been testing my sampler experiments using it and I don't notice any breaking changes, but that was before you added the K layer quantization stuff apparently if I read the other PR right?

Yes, thank you for testing. The changes that I expect to break something are related to the "per-layer KV cache". The other changes from #4312 should be good I think.

"breaking change" means that the API has changed - i.e. it is not backwards compatible.

Thankfully, I'd argue "disappointing" is potentially misleading, looking at the benchmark provided in the other PR.

It's disappointing because it could be much better, but the quantum kernels are not well adapted for multi-head tensors (i.e. ne[2] > 1). In that case, we are doing unnecessary copies and launching separate kernels for each head, while it could be done more efficiently in a single pass. I tried to make a quick workaround for the single GPU case, but it's ugly.

ggml-ci

Dampfinchen · 2023-12-07T16:10:09Z

So per-layer KV cache has far exceeded my expectations. Prompt processing went down from 10 ms/t to 5.5 ms/t on a 13B model and 21 layers offloaded. Truly amazing.

However it's a different story with quantum k cache, I can't notice a memory relief when I set -ctk q8_0, like at all. Does it only work with non-k models? Because I have a Q4K_S model.

x4080 · 2023-12-07T21:13:57Z

Is this benefits apple silicon as well ?

CyborgArmy83 · 2023-12-07T22:56:00Z

Is this benefits apple silicon as well ?

Curious about this as well. Would be nice if it could speed up prompt processing on Apple Silicon.

LostRuins · 2023-12-08T12:53:22Z

@Dampfinchen hmm I've tried the per-layer KV offloading, and I've had a the opposite experience from you, I think that it feels like an overall downgrade (to me), at least for the 6GB RTX 2060.

I tested both approaches with the max number of GPU layers I could offload before going OOM (at 2048 ctx).

main: build = 1620 (fe680e3)

max 19 layers, per layer kv offload

.\main.exe -m e:\LLaMA\models\LLaMA2-13B-Tiefighter.Q4_K_M.gguf -ngl 19 -n 32 -f c:\test\test.txt -c 2048
llama_print_timings:        load time =    3418.61 ms
llama_print_timings:      sample time =       8.17 ms /    32 runs   (    0.26 ms per token,  3916.77 tokens per second)
llama_print_timings: prompt eval time =    1797.20 ms /   202 tokens (    8.90 ms per token,   112.40 tokens per second)
llama_print_timings:        eval time =    6757.10 ms /    31 runs   (  217.97 ms per token,     4.59 tokens per second)
llama_print_timings:       total time =    8597.05 ms

max 23 layers, no kv offload

.\main.exe -m e:\LLaMA\models\LLaMA2-13B-Tiefighter.Q4_K_M.gguf -ngl 23 -nkvo -n 32 -f c:\test\test.txt -c 2048
llama_print_timings:        load time =    3705.89 ms
llama_print_timings:      sample time =       8.25 ms /    32 runs   (    0.26 ms per token,  3880.20 tokens per second)
llama_print_timings: prompt eval time =    1993.56 ms /   202 tokens (    9.87 ms per token,   101.33 tokens per second)
llama_print_timings:        eval time =    6076.94 ms /    31 runs   (  196.03 ms per token,     5.10 tokens per second)
llama_print_timings:       total time =    8107.60 ms

I then switched to my KCPP build which sets GGML_CUDA_FORCE_MMQ to true, allowing me to offload even more layers without going OOM.

max 25 layers, mmq, no kv offload

ContextLimit: 234/2048, Processing:2.58s (12.8ms/T), Generation:5.67s (177.2ms/T), Total:8.25s (3.88T/s)

Overall, trading a ~40% increase PP speed for a 22% reduced generation speed does not really feel like a worthwhile trade to me, especially if we take into account the reduced need for rapid prompt processing while using context shifting. I think perhaps you were testing at -c 512 which has a much smaller difference in number of layers offloaded. At 4096 ctx, the difference will probably be even greater.

Edit: @Pyroserenus suggested that most of the gains are only apparently when comparing between nearly offloaded cases, so here's some 7B tests, which I am able to offload all layers for.

per-layer KV offload, 7B, 28 layers, cublas
ContextLimit: 1606/2048, Processing:3.36s (2.2ms/T), Generation:3.62s (56.6ms/T), Total:6.98s (9.17T/s)

per-layer KV offload, 7B, 28 layers, MMQ
ContextLimit: 1606/2048, Processing:5.02s (3.3ms/T), Generation:3.72s (58.1ms/T), Total:8.74s (7.32T/s)

no KV offload, 7B, 28 layers, cublas
ContextLimit: 1606/2048, Processing:6.84s (4.4ms/T), Generation:5.43s (84.8ms/T), Total:12.27s (5.22T/s)

no KV offload, 7B, 28 layers, MMQ
ContextLimit: 1606/2048, Processing:8.17s (5.3ms/T), Generation:5.06s (79.0ms/T), Total:13.22s (4.84T/s)

full offload, 7B, cublas
ContextLimit: 1606/2048, Processing:2.38s (1.5ms/T), Generation:2.18s (34.0ms/T), Total:4.56s (14.03T/s)

indeed for the last few layers of a model, the PP speed difference between no KV offload and per-layer KV offload does differ much more.

Ph0rk0z · 2023-12-08T15:18:48Z

So for full offload this will not make much difference? I think I have to test this via the built in server rather than python bindings. I would hope that 8bit KV would have helped something.

Dampfinchen · 2023-12-10T12:16:26Z

@LostRuins That's because you are just processing 200 tokens, while I am using a full 2K prompt and also generate 180 Tokens instead of 31. The difference offloaded KV cache makes, is huge, both in generation and prompt processing speed.

22 Layers, per-layer KV cache offloaded, 6 GB 2060

25 Layers, per layer KV Cache disabled

commit 53b5ae02cb1b533b78302422951bcfdeca6e2738 Author: YellowRoseCx <80486540+YellowRoseCx@users.noreply.github.com> Date: Tue Dec 12 12:08:29 2023 -0600 mixtral fan service commit 168b1d74e26d0321e2e89358303b6c33e8d7d33e Merge: f13295b de15d4a6 Author: YellowRoseCx <80486540+YellowRoseCx@users.noreply.github.com> Date: Tue Dec 12 12:00:52 2023 -0600 Merge branch 'kcpp-rocm-mixtral2' into main2 commit de15d4a632939a685ec12fa17355298542facf15 Merge: 74acc54 ea4402b Author: YellowRoseCx <80486540+YellowRoseCx@users.noreply.github.com> Date: Tue Dec 12 11:45:19 2023 -0600 Merge branch 'mixtral' into kcpp-rocm-mixtral commit ea4402b Author: Georgi Gerganov <ggerganov@gmail.com> Date: Tue Dec 12 17:03:38 2023 +0200 test-backend-ops : add one more sum_rows test commit a51bc0c Author: Georgi Gerganov <ggerganov@gmail.com> Date: Tue Dec 12 15:55:42 2023 +0200 metal : fix binary ops for ne10 % 4 != 0 commit 08eb991 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Tue Dec 12 14:14:15 2023 +0200 metal : add cpy f16 -> f32 kernel commit a742d9f Author: slaren <slarengh@gmail.com> Date: Tue Dec 12 12:46:33 2023 +0100 gguf-py : bump version commit 6a419f4 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Tue Dec 12 13:04:33 2023 +0200 convert : support safetensors format commit 74acc54 Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Tue Dec 12 10:53:34 2023 +0800 Revert "Hide hipBLAS (ROCm) if CuBLAS exists - vice versa" This reverts commit 4b854d4. commit f1cbfab Author: slaren <slarengh@gmail.com> Date: Mon Dec 11 20:02:55 2023 +0100 convert : fix style commit 7dc75e3 Author: slaren <slarengh@gmail.com> Date: Mon Dec 11 20:00:28 2023 +0100 convert : use 1e6 rope_freq_base for mixtral commit 296c945 Author: slaren <slarengh@gmail.com> Date: Mon Dec 11 16:53:25 2023 +0100 cuda : fix mul_mat_id with multi gpu commit 33e50f1 Author: slaren <slarengh@gmail.com> Date: Mon Dec 11 12:27:48 2023 +0100 test-backend-ops : disable MOE test with thread sanitizer commit ffda94c Author: slaren <slarengh@gmail.com> Date: Mon Dec 11 12:15:31 2023 +0100 test-backend-ops : simplify and disable slow tests to avoid CI timeout commit 06581f2 Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Mon Dec 11 16:54:42 2023 +0800 perf endpoint lets you monitor if the embedded horde worker has issues commit fce971d Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Mon Dec 11 16:17:10 2023 +0800 do not build the clblast noavx2 binary if not on windows commit 8cbaed1 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Mon Dec 11 08:55:16 2023 +0200 llama : fix hard-coded number of experts commit 4b854d4 Author: YellowRoseCx <80486540+YellowRoseCx@users.noreply.github.com> Date: Sun Dec 10 22:49:35 2023 -0600 Hide hipBLAS (ROCm) if CuBLAS exists - vice versa commit b002981 Author: slaren <slarengh@gmail.com> Date: Mon Dec 11 02:43:52 2023 +0100 test-backend-ops : fix dequantize block offset commit f1380d7 Author: slaren <slarengh@gmail.com> Date: Sun Dec 10 22:58:31 2023 +0100 test-backend-ops : add cpy from f32 -> all types test commit 54d254b Author: slaren <slarengh@gmail.com> Date: Sun Dec 10 21:52:11 2023 +0100 test-backend-ops : cleanup, add moe test for batches commit e2cf3b7 Author: henk717 <henk@henk.tech> Date: Sun Dec 10 14:30:17 2023 +0100 koboldcpp.sh - The Mamba Multitool (LostRuins#554) * .sh script V1 * koboldcpp.sh polish * koboldcpp.sh dist generator * Include html's in dist * RWKV in Linux Dist * Lower dependency requirements * Eliminate wget dependency * More distinct binary name I know its technically amd64, but I don't want to cause confusion among nvidia users. * Use System OpenCL Unsure how this will behave in the pyinstaller build, but pocl ended up CPU only. With a bit of luck the pyinstaller uses the one from the actual system if compiled in a system without opencl, while conda now includes it for that specific system. * Add cblas dependency Missing this causes compile failures on some system's * ICD workaround Ideally we find a better solution, but conda forces ICD and needs this for the successful compile. However, pyinstaller then embeds the ICD causing it to be limited to the system it was compiled for. By temporarily removing the ICD pyinstaller can't find it and everything remains functional. Ideally we do this on a pyinstaller level, but I could not find any good options to do so yet. --------- Co-authored-by: root <root@DESKTOP-DQ1QRAG> commit 54ba263 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sun Dec 10 15:27:41 2023 +0200 test-backend-ops : make experts more evenly probable (test_moe) commit b0b83dd Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sun Dec 10 14:30:38 2023 +0200 metal : fix ggml_mul_mat_id for F32 commit 65923a8 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sun Dec 10 14:17:46 2023 +0200 convert : determine n_ctx correctly commit 8614aa7 Author: slaren <slarengh@gmail.com> Date: Sun Dec 10 13:12:11 2023 +0100 cuda : fix get_rows when ncols is odd commit cefebb3 Author: slaren <slarengh@gmail.com> Date: Sun Dec 10 13:11:39 2023 +0100 test-backend-ops : add moe test commit e640cbe Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sun Dec 10 13:57:54 2023 +0200 llama : add n_expert and n_expert_used to hparams + change quants commit d1259b7 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sun Dec 10 13:00:13 2023 +0200 llama : do not quantize expert gating tensors commit 6cfb31f Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sun Dec 10 10:59:13 2023 +0200 metal : add indirect mat-vec kernels for all quantization types commit 016f9bb Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sun Dec 10 09:38:21 2023 +0200 metal : fix ggml_get_rows to work with non-cont src1 commit 0710b0f Author: slaren <slarengh@gmail.com> Date: Sat Dec 9 23:29:47 2023 +0100 llama : offload missing ffn_moe_silu commit 62b95f9 Author: slaren <slarengh@gmail.com> Date: Sat Dec 9 22:39:34 2023 +0100 cuda : support non-contiguous src1 in get_rows commit 2e4db48 Author: slaren <slarengh@gmail.com> Date: Sat Dec 9 22:38:22 2023 +0100 ggml : update get_rows f16 and q commit ac3f7d8 Author: slaren <slarengh@gmail.com> Date: Sat Dec 9 19:19:03 2023 +0100 ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D commit 8c5b66e Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sat Dec 9 15:30:34 2023 +0200 metal : reduce the kernel launches for ggml_mul_mat_id commit 7e2006b Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sat Dec 9 14:24:58 2023 +0200 metal : add/mul/div use general kernel when src1 not cont commit 06dfde3 Author: slaren <slarengh@gmail.com> Date: Sat Dec 9 13:21:09 2023 +0100 llama : add basic support for offloading moe with CUDA commit 2cbcba8 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sat Dec 9 14:18:42 2023 +0200 metal : add more general support for ggml_get_rows + tests commit 9064b1c Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sat Dec 9 14:04:54 2023 +0200 ggml : fix ggml_get_rows to take into account ne02 / ne11 commit ee8fb39 Author: slaren <slarengh@gmail.com> Date: Sat Dec 9 12:42:25 2023 +0100 ggml : add n_as argument to ggml_mul_mat_id commit 7372b62 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sat Dec 9 13:18:58 2023 +0200 ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) commit 8b185b7 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sat Dec 9 13:01:42 2023 +0200 llama : fix expert weighting in the FFN commit 7ea3695 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sat Dec 9 12:45:15 2023 +0200 llama : first working version commit af1a096 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sat Dec 9 12:07:39 2023 +0200 llama : fix cur -> cur_expert commit aedfad1 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sat Dec 9 11:47:40 2023 +0200 llama : update graph to support MoE commit 861cd67 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sat Dec 9 11:19:46 2023 +0200 ggml : sync latest ggml_mul_mat_id commit a3eefe9 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sat Dec 9 11:14:03 2023 +0200 llama : model loading commit d38e41e Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sat Dec 9 10:59:37 2023 +0200 convert : fix n_ff typo commit dff8cbe Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sat Dec 9 10:51:58 2023 +0200 convert : support Mixtral as LLAMA arch commit 7a69152 Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Fri Dec 8 21:06:32 2023 +0800 lowvram var defaults commit 7418bca Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Fri Dec 8 19:20:30 2023 +0800 up ver commit c47bc28 Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Fri Dec 8 18:35:45 2023 +0800 slight refactor for noscript ui commit 7469f20 Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Fri Dec 8 18:16:14 2023 +0800 use lowvram flag for offload qkv commit ec21fa7 Merge: 930cdfb fe680e3 Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Fri Dec 8 17:42:26 2023 +0800 Merge branch 'master' into concedo_experimental # Conflicts: # .github/workflows/build.yml # .gitignore # CMakeLists.txt # Makefile # Package.swift # README.md # ggml-cuda.cu # llama.cpp # llama.h # scripts/sync-ggml.sh # tests/CMakeLists.txt commit 930cdfb Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Fri Dec 8 16:53:30 2023 +0800 updated lite, added patch that links to noscript mode commit fe680e3 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Thu Dec 7 22:26:54 2023 +0200 sync : ggml (new ops, tests, backend, etc.) (ggerganov#4359) * sync : ggml (part 1) * sync : ggml (part 2, CUDA) * sync : ggml (part 3, Metal) * ggml : build fixes ggml-ci * cuda : restore lost changes * cuda : restore lost changes (StableLM rope) * cmake : enable separable compilation for CUDA ggml-ci * ggml-cuda : remove device side dequantize * Revert "cmake : enable separable compilation for CUDA" This reverts commit 09e35d0. * cuda : remove assert for rope * tests : add test-backend-ops * ggml : fix bug in ggml_concat * ggml : restore `ggml_get_n_tasks()` logic in `ggml_graph_plan()` * ci : try to fix macOS * ggml-backend : remove backend self-registration * ci : disable Metal for macOS cmake build ggml-ci * metal : fix "supports family" call * metal : fix assert * metal : print resource path ggml-ci --------- Co-authored-by: slaren <slarengh@gmail.com> commit bcc0eb4 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Thu Dec 7 13:03:17 2023 +0200 llama : per-layer KV cache + quantum K cache (ggerganov#4309) * per-layer KV * remove unnecessary copies * less code duplication, offload k and v separately * llama : offload KV cache per-layer * llama : offload K shift tensors * llama : offload for rest of the model arches * llama : enable offload debug temporarily * llama : keep the KV related layers on the device * llama : remove mirrors, perform Device -> Host when partial offload * common : add command-line arg to disable KV cache offloading * llama : update session save/load * llama : support quantum K cache (ggerganov#4312) * llama : support quantum K cache (wip) * metal : add F32 -> Q8_0 copy kernel * cuda : add F32 -> Q8_0 copy kernel ggml-ci * cuda : use mmv kernel for quantum cache ops * llama : pass KV cache type through API * llama : fix build ggml-ci * metal : add F32 -> Q4_0 copy kernel * metal : add F32 -> Q4_1 copy kernel * cuda : wip * cuda : add F32 -> Q4_0 and F32 -> Q4_1 copy kernels * llama-bench : support type_k/type_v * metal : use mm kernel only for quantum KV cache * cuda : add comment * llama : remove memory_f16 and kv_f16 flags --------- Co-authored-by: slaren <slarengh@gmail.com> * readme : add API change notice --------- Co-authored-by: slaren <slarengh@gmail.com> commit 81bc921 Author: Hongyu Ouyang <96765450+casavaca@users.noreply.github.com> Date: Thu Dec 7 02:25:22 2023 -0800 train : fix ggerganov#4227 (double free in examples/train-text-from-scratch/train-text-from-scratch.cpp) (ggerganov#4351) On commit b1108 (44c117f) xaedes added ggml_allocr * alloc = NULL; ... (many lines in between) if (alloc) { ggml_allocr_free(alloc); } Which is correct, but it's easy to lose context after many lines in between. On commit b1287 (0e76a899) xaedes made a big change. From here on, alloc is freed eagerly. alloc = ggml_allocr_new(...) ... (short lines of code) ggml_allocr_free(alloc) This happens a few times, but alloc is never set to NULL, and many lines below, we still have if (alloc) { ggml_allocr_free(alloc); } which causes a double-free. commit 05cd6e5 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Wed Dec 6 20:21:59 2023 +0200 server : recognize cache_prompt parameter in OAI API (ggerganov#4347) commit c751152 Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Thu Dec 7 00:52:25 2023 +0800 noscript mode is done commit 12002d8 Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Wed Dec 6 17:51:08 2023 +0800 very basic noscript mode commit caa9249 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Wed Dec 6 10:41:03 2023 +0200 common : fix compile warning commit da5eaef Author: stduhpf <stephduh@live.fr> Date: Wed Dec 6 09:08:17 2023 +0100 speculative : support `--color` (ggerganov#4343) * speculative: add some colors * minor : add braces --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> commit 5f6e0c0 Author: Marcus Dunn <51931484+MarcusDunn@users.noreply.github.com> Date: Tue Dec 5 10:55:12 2023 -1000 grammar : pre-computed pieces + reserve mem + less string copies (ggerganov#4330) * reserve space for codepoints * improvement for the appended 0 * used precomputed token text for grammar sample * reserve canidates_decoded * reserve canidates_grammar * remove candidates_decoded * Revert "remove candidates_decoded" This reverts commit 3773328. * changed decode_utf8 to take src by ref commit 5aa365d Author: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com> Date: Tue Dec 5 10:19:18 2023 -0700 llama : allow overriding GGUF metadata when loading model (ggerganov#4092) * feat: Allow overriding GGUF metadata when loading model * Fix the one time GCC is stricter than clang about something * Step1 * Refactor... basically everything! * Nuke obsolete GetArrayLen struct * simplify std::string specialization * Various cleanups Add informational output when overrides are applied Warn user when an override with the wrong type is specified * Fix broken logic for parsing bool KV overrides Fix issue where overrides didn't apply when key missing in GGUF metadata Resolve merge changes * llama : rearrange model params * Update new GET_KEY call Add note that metadata KV overrides aren't reflected in initial metadata KV info dump --------- Co-authored-by: cebtenzzre <cebtenzzre@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> commit b6f952f Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Tue Dec 5 21:08:10 2023 +0800 improved exit logic commit 52c8bc3 Author: MaggotHATE <clay1326@gmail.com> Date: Tue Dec 5 15:05:51 2023 +0500 sampling : custom samplers order (ggerganov#4285) * Samplers sequence order w parameter * Cleaned commented code * Fixed formatting * Rewrote with unordered_map * Revert and rewrite, too many problems and safeguards would be needed * Fixed code style * Code style fixes according to review * More readable samplers input string, fixed help * Style fix in sampler_queue * Formatting fixes * Fixing whitespaces commit e4b76bb Author: kchro3 <62481661+kchro3@users.noreply.github.com> Date: Mon Dec 4 23:29:46 2023 -0800 swift : revert compiler checks for swift package (ggerganov#4332) commit 23b5e12 Author: Daniel Bevenius <daniel.bevenius@gmail.com> Date: Mon Dec 4 17:04:21 2023 +0100 simple : update error message for KV cache check (ggerganov#4324) This commit updates the error message that is printed when the KV cache is not big enough to hold all the prompt and generated tokens. Specifically it removes the reference to n_parallel and replaces it with n_len. Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com> commit d208995 Author: Miwa / Ensan <63481257+ensan-hcl@users.noreply.github.com> Date: Tue Dec 5 01:03:49 2023 +0900 swift : fix concatenation method to avoid invalid UTF8 stringfication (ggerganov#4325) commit 5c9f90c Author: Miwa / Ensan <63481257+ensan-hcl@users.noreply.github.com> Date: Mon Dec 4 22:43:45 2023 +0900 swift : fix prompt tokenization logic (ggerganov#4321) commit a5a5839 Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Mon Dec 4 21:10:42 2023 +0800 handle accidentally selecting a kcpps file as model instead commit 4fa44e8 Author: Ikko Eltociear Ashimine <eltociear@gmail.com> Date: Mon Dec 4 16:57:35 2023 +0900 grammar-parser : fix typo (ggerganov#4318) preceeding -> preceding commit 8602f5a Merge: ac36aee fbbc428 Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Sun Dec 3 22:00:14 2023 +0800 Merge branch 'master' into concedo_experimental commit fbbc428 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sun Dec 3 15:56:35 2023 +0200 ggml : reuse ggml_get_n_tasks() in ggml_graph_plan() (ggerganov#4308) * ggml : fix soft max out-of-bounds access ggml-ci * ggml : reuse ggml_get_n_tasks() in ggml_graph_plan() ggml-ci commit ac36aee Merge: 48544cd 33e171d Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Sun Dec 3 21:56:29 2023 +0800 Merge branch 'master' into concedo_experimental # Conflicts: # CMakeLists.txt # Makefile commit adf3de4 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sun Dec 3 15:56:22 2023 +0200 ggml : fix soft max out-of-bounds access (ggerganov#4307) ggml-ci commit 48544cd Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Sun Dec 3 21:46:50 2023 +0800 Revert "Revert "ggml : add ggml_soft_max_ext (ggerganov#4256)"" This reverts commit a8e66ef. commit 33e171d Author: Ed Lee <edilee@mozilla.com> Date: Sun Dec 3 01:10:43 2023 -0800 server : fix OpenAI API `stop` field to be optional (ggerganov#4299) (cherry picked from commit Mozilla-Ocho/llamafile@e8c92bc) commit 6949b50 Author: Rickard Edén <rickardeden@gmail.com> Date: Sun Dec 3 10:03:25 2023 +0100 py : add grammar to oai like api (ggerganov#4294) commit d7b800b Author: Georgi Gerganov <ggerganov@gmail.com> Date: Sun Dec 3 10:58:16 2023 +0200 llama : pad KV cache size (ggerganov#4280) * llama : pad KV cache size to 32 * metal : try to improve batched decoding commit 6570a20 Author: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Sun Dec 3 15:44:53 2023 +0800 token count includes ids commit 5a7d312 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Fri Dec 1 20:39:12 2023 +0200 llama : avoid using "optional" keyword (ggerganov#4283) commit d5a1cbd Author: Georgi Gerganov <ggerganov@gmail.com> Date: Fri Dec 1 20:35:03 2023 +0200 llama : support optional tensors (ggerganov#4283) commit b220222 Author: Miwa / Ensan <63481257+ensan-hcl@users.noreply.github.com> Date: Sat Dec 2 03:19:45 2023 +0900 swift : fix token_to_piece implementation (ggerganov#4278) * Fix token_to_piece implementation in Swift * Fix errors commit 511f52c Author: Jared Van Bortel <jared@nomic.ai> Date: Fri Dec 1 13:18:35 2023 -0500 build : enable libstdc++ assertions for debug builds (ggerganov#4275) commit 03562f3 Author: CausalLM <148736309+CausalLM@users.noreply.github.com> Date: Sat Dec 2 02:17:06 2023 +0800 llama : support attention bias on LLaMA architecture (ggerganov#4283) * Support attention_bias on LLaMA architecture QKVO bias, should fix InternLM (ggerganov#3133) and works for LLaMAfied Qwen models (ggerganov#3743 (comment)). * check existence of qkvo bias while loading llama models Tested on LLaMA2, CUDA and CPU. * Update llama.cpp commit 37c746d Author: Shijie <821898965@qq.com> Date: Sat Dec 2 02:16:31 2023 +0800 llama : add Qwen support (ggerganov#4281) * enable qwen to llama.cpp * llama : do not GPU split bias tensors --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> commit 880f579 Author: Georgi Gerganov <ggerganov@gmail.com> Date: Fri Dec 1 18:42:11 2023 +0200 llama : fix integer overflow during quantization (ggerganov#4284) happens with multi-threaded quantization of Qwen-72B ggml-ci

llama : restore prefix space in llama tokenizer (ggerganov#4081) gguf : fix potential infinite loops while parsing (ggerganov#4100) Co-authored-by: Bernhard Gstrein <gstrein@cs.uni-freiburg.de> Respect tokenizer.ggml.add_bos_token value when tokenizing (ggerganov#4040) * gguf-py: gguf-dump: Respect --no-tensor flag in JSON mode. * Respect add_bos_token GGUF metadata value * gguf-py: Try to fix SpecialVocab giving up too easily for the Nth time llama : fix data units (ggerganov#4101) * llama : fix data units ggml-ci * Revert "llama : fix data units" This reverts commit f5feac8. * llama : disambiguate data units ggml-ci cuda : get_row_rounding F32 (ggerganov#4095) * Fix ggerganov#4017 * Update ggml-cuda.cu Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> * Update ggml-cuda.cu Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> --------- Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> finetune : zero the loraB initial vectors (ggerganov#4082) * finetune : zero the loraB initial vectors Without this, the first iteration is starting out far from the base model, instead of exactly on it. Zeroing loraB is what the paper recommends. loralib also zeroes at least one of the init vector pairs (though it departs from the paper in using a different distribution for the other vector, in some cases). * tabs to spaces * Use ggml_set_zero instead of adding a new function finetune : speed-up ggml_compute_forward_out_prod_f32 via BLAS (ggerganov#4079) * Remove logically superfluous assertions and order by dimension * Use cblas_sgemm() to implement ggml_compute_forward_out_prod() * Remove ggml_compute_forward_out_prod_use_blas(), fix compiling errors on cmake/zig, remove trailing whitespace * Add openBLAS support for sgemm() in compute_forward_out_prod() llama : add functions to get the model's metadata (ggerganov#4013) * llama : add functions to get the model's metadata * format -> std::to_string * better documentation train : move number of gpu layers argument parsing to common/train.cpp (ggerganov#4074) - introduces help entry for the argument - cuts '--gpu-layers' form in order to simplify usage and documentation. Signed-off-by: Jiri Podivin <jpodivin@gmail.com> Co-authored-by: Jiri Podivin <jpodivin@redhat.com> py : remove superfluous import statements (ggerganov#4076) Signed-off-by: Jiri Podivin <jpodivin@gmail.com> Co-authored-by: Jiri Podivin <jpodivin@redhat.com> llava : fix compilation warning that fread return value is not used (ggerganov#4069) common : improve yaml log escaping (ggerganov#4080) * logging: improve escaping in yaml output * logging: include review feedback py : Falcon HF compatibility (ggerganov#4104) Falcon HF compatibility convert : use 'model' value if it exists. This allows karpathy/tinyllamas to load (ggerganov#4089) Co-authored-by: Don Mahurin <@> examples : add tokenize (ggerganov#4039) tokenize : fix trailing whitespace build : support ppc64le build for make and CMake (ggerganov#3963) * build: support ppc64le build for make and CMake * build: keep __POWER9_VECTOR__ ifdef and extend with __powerpc64__ Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> llama : increase max nodes (ggerganov#4115) Clean up ggml-cuda.cu warnings when compiling with clang (for ROCM) (ggerganov#4124) * ggml-cuda.cu: Clean up warnings when compiling with clang * ggml-cuda.cu: Move static items into anonymous namespace * ggml-cuda.cu: Fix use of namespace start macro * Revert "ggml-cuda.cu: Fix use of namespace start macro" This reverts commit 26c1149. * Revert "ggml-cuda.cu: Move static items into anonymous namespace" This reverts commit e29757e. scripts : Remove missed baichuan convert script (ggerganov#4127) tokenize example: Respect normal add BOS token behavior (ggerganov#4126) Allow building with Makefile gguf-py : export chat templates (ggerganov#4125) * gguf-py : export chat templates * llama.cpp : escape new lines in gguf kv info prints * gguf-py : bump version * gguf-py : check chat_template type * gguf-py : initialize chat_template gitignore : tokenize common : comma should be semicolon (ggerganov#4137) server : relay error messages (ggerganov#4131) finetune : add --n-gpu-layers flag info to --help (ggerganov#4128) Revert "finetune : add --n-gpu-layers flag info to --help (ggerganov#4128)" This reverts commit 05e8301. speculative : fix prompt tokenization in speculative example (ggerganov#4025) * Support special tokens and not adding BOS to prompt in speculative * Adapt to new should_add_bos function * Ensure tgt and dft have same add_bos setting ci : add flake8 to github actions (python linting) (ggerganov#4129) Disabled rules: * E203 Whitespace before ':' - disabled because we often use 'C' Style where values are aligned * E211 Whitespace before '(' (E211) - disabled because we often use 'C' Style where values are aligned * E221 Multiple spaces before operator - disabled because we often use 'C' Style where values are aligned * E225 Missing whitespace around operator - disabled because it's broken so often it seems like a standard * E231 Missing whitespace after ',', ';', or ':' - disabled because we often use 'C' Style where values are aligned * E241 Multiple spaces after ',' - disabled because we often use 'C' Style where values are aligned * E251 Unexpected spaces around keyword / parameter equals - disabled because it's broken so often it seems like a standard * E261 At least two spaces before inline comment - disabled because it's broken so often it seems like a standard * E266 Too many leading '#' for block comment - sometimes used as "section" separator * E501 Line too long - disabled because it's broken so often it seems like a standard * E701 Multiple statements on one line (colon) - broken only in convert.py when defining abstract methods (we can use# noqa instead) * E704 Multiple statements on one line - broken only in convert.py when defining abstract methods (we can use# noqa instead) main : Add ChatML functionality to main example (ggerganov#4046) Co-authored-by: Sebastian Cramond <sebby37@users.noreply.github.com> readme : update ROCm Windows instructions (ggerganov#4122) * Update README.md * Update README.md Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> --------- Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> finetune - update readme to mention llama support only (ggerganov#4148) stablelm : simplify + speedup generation (ggerganov#4153) docs : add llama-star arch idea examples : fix typo in parallel example doc comment (ggerganov#4181) Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com> readme : update hot topics llama : KV cache view API + better KV cache management (ggerganov#4170) * llama : keep track of used KV cells + better KV cache management * llama : zero KV cache used upon clear ggml-ci * llama : allow exporting a view of the KV cache (ggerganov#4180) * Allow exporting a view of the KV cache * Allow dumping the sequences per cell in common * Track max contiguous cells value and position as well * Fix max contiguous empty cells index calculation Make dump functions deal with lengths or sequences counts > 10 better * Fix off by one error in dump_kv_cache_view * Add doc comments for KV cache view functions Eliminate cell sequence struct; use llama_seq_id directly Minor cleanups * common : add -dkvc arg for enabling kv cache dumps --------- Co-authored-by: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com> Fix incorrect format strings and uninitialized variables. (ggerganov#4133) * Fix incorrect format strings and uninitialized variables. * Address comments * Add the missing include statement readme : use PATH for Windows ROCm (ggerganov#4195) * Update README.md to use PATH for Windows ROCm * Update README.md * Update README.md main.swift : fix eos checking (ggerganov#4197) llama_token_eos(const struct llama_model *) is currently getting struct llama_context type variable context as a parameter. convert : fix tensors using grad in some models (ggerganov#4173) ggml-cuda : support stablelm rope (ggerganov#4156) * ggml-cuda : support stablelm rope * remove unused freq_base kernel parameter * add n_dims parameter to llm_build_k_shift, default to n_rot via overload * llama : fix llm_build_k_shift args --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> llama : set metal log callback correctly (ggerganov#4204) server : OAI API compatibility (ggerganov#4198) * Add openai-compatible POST /v1/chat/completions API endpoint to server example * fix code style * Update server README.md * Improve server README.md * Fix server.cpp code style according to review * server : some style changes * server : indentation * server : enable special tokens during tokenization by default * server : minor code style * server : change random string generator * straightforward /v1/models endpoint --------- Co-authored-by: kir-gadjello <111190790+kir-gadjello@users.noreply.github.com> Co-authored-by: Tobi Lütke <tobi@Tobis-MacBook-Pro.local> readme : update hot topics Update docs for yarn_ext_factor <0.0 as unspecified instead of NaN (ggerganov#4189) llama : grammar `reserve` space in `decode_utf8` (ggerganov#4210) * reserve space for codepoints * improvement for the appended 0 scripts : Use mmap in torch load (ggerganov#4202) * Use mmap in torch load, prefer .bin files when loading * Revert .bin > .safetensors preference metal : fix yarn (ggerganov#4220) get the correct n_orig_ctx in metal lookahead : add example for lookahead decoding (ggerganov#4207) * lookahead : init * lookahead : generate and store n-grams * lookahead : use loop instead recursion to generate n-grams * lookahead : initial working implementation * lookahead : filter repeating n-grams * lookahead : use deterministic init * lookahead : add to Makefile * lookahead : fix a bug in the seq_id of the lookahead tokens * lookahead : add comments --------- Co-authored-by: slaren <slarengh@gmail.com> readme : update hot topics lookahead : support `-n -1` infinite generation ggml : fix -Warray-bounds warning with gcc (ggerganov#4231) examples : iOS example with swift ui (ggerganov#4159) * copy to llama.cpp as subdir * attempt enabling metal, fails * ggml metal compiles! * Update README.md * initial conversion to new format, utf8 errors? * bug fixes, but now has an invalid memory access :( * added O3, now has insufficient memory access * begin sync with master * update to match latest code, new errors * fixed it! * fix for loop conditionals, increase result size * fix current workflow errors * attempt a llama.swiftui workflow * Update .github/workflows/build.yml Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> readme : add Amica to UI list (ggerganov#4230) cmake : fix issue with version info not getting baked into LlamaConfig.cmake (ggerganov#3970) * Split CPP generation from build-info query * Remove blank lines * Add BUILD_SHARED_LIBS option ggml : re-enable BLAS for CPU when src0 != F32 + remove redundant full offload checks in llama.cpp (ggerganov#4240) * ggml : use blas even if src0 is not F32 * llama : use n_threads_batch only when n_tokens >= 32 ggml-ci * llama : revert n_threads_batch logic ggml-ci ggml : restore abort() in GGML_ASSERT (ggerganov#4242) readme : add FreeChat (ggerganov#4248) examples : add readme files py : fix oai proxy (ggerganov#3972) * fix oai proxy fix generation not stoped while bot stop talking in chat mode fix possible `slot_id` not exist response for cors (and pre flight) * oai proxy: workaround for some client (such as Chatbox) * use stop as separator to replace hardcoded `\n` llama : fix typical sampling (ggerganov#4261) Typical sampling was broken because after copying new_candidates into canditates, the "sorted" bool is left at "true", but the new data is no longer sorted according to probability. Patch to set "sorted" to false. Test: Generating with temp=0.0001 (approx. argmax) should generate the same sequence at typical>=1.0 and typical=0.9999 (approx. disabled, but enters the typical sampling codepath). convert.py : fix llama/llama2 conversion due to vocab_size=-1 (ggerganov#4258) llama : fix alignment of general.name in print meta (ggerganov#4254) * llama: fix alignment of general.name in print meta This commit fixes the alignment of the general.name field in the llm_load_print_meta function. Currently the output looks like this: ```console llm_load_print_meta: model ftype = mostly Q4_0 llm_load_print_meta: model params = 13.02 B llm_load_print_meta: model size = 6.86 GiB (4.53 BPW) llm_load_print_meta: general.name = LLaMA v2 ``` And with this commit it looks like this: ```console llm_load_print_meta: model ftype = mostly Q4_0 llm_load_print_meta: model params = 13.02 B llm_load_print_meta: model size = 6.86 GiB (4.53 BPW) llm_load_print_meta: general.name = LLaMA v2 ``` Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com> * llama: fix alignment of special tokens Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com> --------- Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com> readme : fix typo (ggerganov#4253) llama.cpp uses GitHub Actions, not Gitlab Actions. cmake : fix the metal file foder path (ggerganov#4217) batched.swift : update README.md (ggerganov#4214) docs: update how to run docker : add finetune option (ggerganov#4211) readme : fix (ggerganov#4135) * fix: readme * chore: resolve comments * chore: resolve comments main : pass LOG_TEE callback to llama.cpp log (ggerganov#4033) * main : Call llama_log_set to use LOG_TEE * tabs to spaces llava : ShareGPT4V compatibility (vision encoder only loading) (ggerganov#4172) * ShareGPT4 compatibility (vision encoder only loading) Load only a CLIP vision encoder (as supplied by ShareGPT finetunes) Corrects the argument parsing for --img_mean and --img_std (which were previously not parsed but attempted to access) Defines defaults for img_mean and img_std which are equal to the llava 1.5 CLIP encoder, so you do not have to provide them * Update convert-image-encoder-to-gguf.py build : fix build info generation and cleanup Makefile (ggerganov#3920) * cmake : fix joining of REAL_GIT_DIR * fix includes with help from include-what-you-use * make : remove unneeded deps and add test-rope target * fix C includes in C++ source files * Revert "fix includes with help from include-what-you-use" This reverts commit 635e9fa. make : fix Apple clang determination bug (ggerganov#4272) Co-authored-by: Will Findley <findley@gmail.com> server : add single-client multi-prompt support (ggerganov#4232) * * add multiprompt support * * cleanup * * more cleanup * * remove atomicity of id_gen, and change lock_guard to unique_lock on completion requests * * remove all references to mutex_multitasks * Update examples/server/server.cpp Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> * Update examples/server/server.cpp Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> * Update examples/server/server.cpp Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> * Update examples/server/server.cpp Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> * * change to set --------- Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> server : add --log-disable to disable logging to file (ggerganov#4260) * * add --log-disable to disable logging to file in the server example * * typo fix ggml : add ggml_soft_max_ext (ggerganov#4256) * metal : implement soft_max_ext * cuda : implement soft_max_ext * ggml : implement soft_max_ext (CPU) * batched-bench : print threads ggml-ci * metal : simplify soft_max encoding ggml-ci * cuda : use 512 threads for soft_max instead of 32 * ggml : update soft max cpu * cuda : do warp-based block reduce * cuda : increase max block size to 1024 * cuda : fix warp reduction initialization of shared mem * metal : warp-based reduction for soft max kernel * metal : warp-based reduce for rms_norm * metal : simplify soft max kernel ggml-ci * alloc : fix build with debug py : add requirements file for convert-hf-to-gguf.py (ggerganov#4277) This commit adds a requirements file for the convert-hf-to-gguf.py script, and also add the torch and transformers packages to it. The motivation for this is that currently running convert-hf-to-gguf.py will produce the following error: ```console $ python3 -m venv venv $ source venv/bin/activate (venv) $ pip install -r requirements.txt Collecting numpy==1.24.4 Collecting sentencepiece==0.1.98 Collecting gguf>=0.1.0 Installing collected packages: sentencepiece, numpy, gguf Successfully installed gguf-0.5.1 numpy-1.24.4 sentencepiece-0.1.98 (venv) $ python convert-hf-to-gguf.py --help Traceback (most recent call last): File "llama.cpp/convert-hf-to-gguf.py", line 16, in <module> import torch ModuleNotFoundError: No module named 'torch' ``` With this commit, and using requirements-hf-to-gguf.txt instead of requirements.txt, the script can be run and shows the help output. Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com> llama : fix integer overflow during quantization (ggerganov#4284) happens with multi-threaded quantization of Qwen-72B ggml-ci llama : add Qwen support (ggerganov#4281) * enable qwen to llama.cpp * llama : do not GPU split bias tensors --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> llama : support attention bias on LLaMA architecture (ggerganov#4283) * Support attention_bias on LLaMA architecture QKVO bias, should fix InternLM (ggerganov#3133) and works for LLaMAfied Qwen models (ggerganov#3743 (comment)). * check existence of qkvo bias while loading llama models Tested on LLaMA2, CUDA and CPU. * Update llama.cpp build : enable libstdc++ assertions for debug builds (ggerganov#4275) swift : fix token_to_piece implementation (ggerganov#4278) * Fix token_to_piece implementation in Swift * Fix errors llama : support optional tensors (ggerganov#4283) llama : avoid using "optional" keyword (ggerganov#4283) llama : pad KV cache size (ggerganov#4280) * llama : pad KV cache size to 32 * metal : try to improve batched decoding py : add grammar to oai like api (ggerganov#4294) server : fix OpenAI API `stop` field to be optional (ggerganov#4299) (cherry picked from commit Mozilla-Ocho/llamafile@e8c92bc) ggml : fix soft max out-of-bounds access (ggerganov#4307) ggml-ci ggml : reuse ggml_get_n_tasks() in ggml_graph_plan() (ggerganov#4308) * ggml : fix soft max out-of-bounds access ggml-ci * ggml : reuse ggml_get_n_tasks() in ggml_graph_plan() ggml-ci grammar-parser : fix typo (ggerganov#4318) preceeding -> preceding swift : fix prompt tokenization logic (ggerganov#4321) swift : fix concatenation method to avoid invalid UTF8 stringfication (ggerganov#4325) simple : update error message for KV cache check (ggerganov#4324) This commit updates the error message that is printed when the KV cache is not big enough to hold all the prompt and generated tokens. Specifically it removes the reference to n_parallel and replaces it with n_len. Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com> swift : revert compiler checks for swift package (ggerganov#4332) sampling : custom samplers order (ggerganov#4285) * Samplers sequence order w parameter * Cleaned commented code * Fixed formatting * Rewrote with unordered_map * Revert and rewrite, too many problems and safeguards would be needed * Fixed code style * Code style fixes according to review * More readable samplers input string, fixed help * Style fix in sampler_queue * Formatting fixes * Fixing whitespaces llama : allow overriding GGUF metadata when loading model (ggerganov#4092) * feat: Allow overriding GGUF metadata when loading model * Fix the one time GCC is stricter than clang about something * Step1 * Refactor... basically everything! * Nuke obsolete GetArrayLen struct * simplify std::string specialization * Various cleanups Add informational output when overrides are applied Warn user when an override with the wrong type is specified * Fix broken logic for parsing bool KV overrides Fix issue where overrides didn't apply when key missing in GGUF metadata Resolve merge changes * llama : rearrange model params * Update new GET_KEY call Add note that metadata KV overrides aren't reflected in initial metadata KV info dump --------- Co-authored-by: cebtenzzre <cebtenzzre@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> grammar : pre-computed pieces + reserve mem + less string copies (ggerganov#4330) * reserve space for codepoints * improvement for the appended 0 * used precomputed token text for grammar sample * reserve canidates_decoded * reserve canidates_grammar * remove candidates_decoded * Revert "remove candidates_decoded" This reverts commit 3773328. * changed decode_utf8 to take src by ref speculative : support `--color` (ggerganov#4343) * speculative: add some colors * minor : add braces --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> common : fix compile warning server : recognize cache_prompt parameter in OAI API (ggerganov#4347) train : fix ggerganov#4227 (double free in examples/train-text-from-scratch/train-text-from-scratch.cpp) (ggerganov#4351) On commit b1108 (44c117f) xaedes added ggml_allocr * alloc = NULL; ... (many lines in between) if (alloc) { ggml_allocr_free(alloc); } Which is correct, but it's easy to lose context after many lines in between. On commit b1287 (0e76a899) xaedes made a big change. From here on, alloc is freed eagerly. alloc = ggml_allocr_new(...) ... (short lines of code) ggml_allocr_free(alloc) This happens a few times, but alloc is never set to NULL, and many lines below, we still have if (alloc) { ggml_allocr_free(alloc); } which causes a double-free. llama : per-layer KV cache + quantum K cache (ggerganov#4309) * per-layer KV * remove unnecessary copies * less code duplication, offload k and v separately * llama : offload KV cache per-layer * llama : offload K shift tensors * llama : offload for rest of the model arches * llama : enable offload debug temporarily * llama : keep the KV related layers on the device * llama : remove mirrors, perform Device -> Host when partial offload * common : add command-line arg to disable KV cache offloading * llama : update session save/load * llama : support quantum K cache (ggerganov#4312) * llama : support quantum K cache (wip) * metal : add F32 -> Q8_0 copy kernel * cuda : add F32 -> Q8_0 copy kernel ggml-ci * cuda : use mmv kernel for quantum cache ops * llama : pass KV cache type through API * llama : fix build ggml-ci * metal : add F32 -> Q4_0 copy kernel * metal : add F32 -> Q4_1 copy kernel * cuda : wip * cuda : add F32 -> Q4_0 and F32 -> Q4_1 copy kernels * llama-bench : support type_k/type_v * metal : use mm kernel only for quantum KV cache * cuda : add comment * llama : remove memory_f16 and kv_f16 flags --------- Co-authored-by: slaren <slarengh@gmail.com> * readme : add API change notice --------- Co-authored-by: slaren <slarengh@gmail.com> sync : ggml (new ops, tests, backend, etc.) (ggerganov#4359) * sync : ggml (part 1) * sync : ggml (part 2, CUDA) * sync : ggml (part 3, Metal) * ggml : build fixes ggml-ci * cuda : restore lost changes * cuda : restore lost changes (StableLM rope) * cmake : enable separable compilation for CUDA ggml-ci * ggml-cuda : remove device side dequantize * Revert "cmake : enable separable compilation for CUDA" This reverts commit 09e35d0. * cuda : remove assert for rope * tests : add test-backend-ops * ggml : fix bug in ggml_concat * ggml : restore `ggml_get_n_tasks()` logic in `ggml_graph_plan()` * ci : try to fix macOS * ggml-backend : remove backend self-registration * ci : disable Metal for macOS cmake build ggml-ci * metal : fix "supports family" call * metal : fix assert * metal : print resource path ggml-ci --------- Co-authored-by: slaren <slarengh@gmail.com> grammar : revert the replacement of llama_token_to_piece with id_to_token (ggerganov#4396) Update README.md (ggerganov#4388) Fix small typo. ggml : increased GGML_MAX_PARAMS to allow finetuning of 70b models (ggerganov#4424) server : fix local model name in server (ggerganov#4420) llama : document logits_all deprecation (ggerganov#4418) llama_context_params.logits_all is a parameter for controlling llama_eval. This documents that logits_all should not be used with llama_decode and llama_batch. build : target Windows 8 for standard mingw-w64 (ggerganov#4405) * build : target Windows 8 for standard mingw-w64 * make : fix missing console.o deps This was causing a link error with `make all` on Windows. english : use `typos` to fix comments and logs (ggerganov#4354) server : tweak default sampling parameters (ggerganov#4367) * Set a more typical Top P setting as the default * Update temp max llama : add Mixtral support (ggerganov#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (ggerganov#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: ggerganov/ggml@1914017 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>

slaren and others added 7 commits October 3, 2023 19:20

per-layer KV

e9bcf66

remove unnecessary copies

55f2f2f

less code duplication, offload k and v separately

f4f9367

Merge branch 'master' into per-layer-kv

c294c78

llama : offload KV cache per-layer

986b3da

llama : offload K shift tensors

f3dbfb9

llama : offload for rest of the model arches

3d3e6bd

ggerganov added performance Speed related topics need feedback Testing and feedback with results are needed labels Dec 3, 2023

ggerganov mentioned this pull request Dec 3, 2023

demo : per-layer KV / partial offloading of KV cache #3457

Closed

slaren reviewed Dec 3, 2023

View reviewed changes

ggerganov commented Dec 3, 2023

View reviewed changes

llama.cpp Outdated Show resolved Hide resolved

ggerganov added 3 commits December 3, 2023 18:36

llama : enable offload debug temporarily

1fa91a4

llama : keep the KV related layers on the device

c44bc1e

llama : remove mirrors, perform Device -> Host when partial offload

c80b8a2

ggerganov added 2 commits December 3, 2023 20:31

common : add command-line arg to disable KV cache offloading

e262947

llama : update session save/load

66aaac9

ggerganov mentioned this pull request Dec 5, 2023

llama : support quantum K cache #4312

Merged

12 tasks

ggerganov added the breaking change Changes that break ABIs, APIs, file formats, or other forms of backwards compatibility. label Dec 6, 2023

ggerganov mentioned this pull request Dec 7, 2023

test-backend-ops : add performance eval mode + improve CUDA repeat and binary broadcast ops performance ggerganov/ggml#636

Merged

ggerganov added 2 commits December 7, 2023 12:33

Merge branch 'master' into gg/per-layer-kv

680a99e

ggml-ci

readme : add API change notice

fc5f334

ggerganov merged commit bcc0eb4 into master Dec 7, 2023
37 checks passed

abetlen mentioned this pull request Dec 11, 2023

Add support for new KV Cache Offloading API abetlen/llama-cpp-python#995

Closed

CISC mentioned this pull request Dec 12, 2023

Quantizing V cache not working yet #4425

Closed

LostRuins mentioned this pull request Dec 14, 2023

"Not enough space in the context's memory pool" exception in 1.52 LostRuins/koboldcpp#563

Closed

kalomaze mentioned this pull request Dec 18, 2023

Possible ideas for speeding up CPU inference with Mixtral (KV cache prioritization) #4518

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

llama : per-layer KV cache #4309

llama : per-layer KV cache #4309

ggerganov commented Dec 3, 2023 •

edited

Loading

slaren Dec 3, 2023 •

edited

Loading

ggerganov Dec 3, 2023

slaren Dec 3, 2023

kalomaze Dec 3, 2023 •

edited

Loading

ggerganov Dec 3, 2023

kalomaze Dec 3, 2023 •

edited

Loading

ggerganov Dec 3, 2023 •

edited

Loading

oobabooga commented Dec 3, 2023

kalomaze commented Dec 3, 2023 •

edited

Loading

kalomaze commented Dec 3, 2023 •

edited

Loading

ggerganov commented Dec 6, 2023

kalomaze commented Dec 7, 2023 •

edited

Loading

kalomaze commented Dec 7, 2023 •

edited

Loading

ggerganov commented Dec 7, 2023

Dampfinchen commented Dec 7, 2023 •

edited

Loading

x4080 commented Dec 7, 2023

CyborgArmy83 commented Dec 7, 2023

LostRuins commented Dec 8, 2023 •

edited

Loading

Ph0rk0z commented Dec 8, 2023

Dampfinchen commented Dec 10, 2023

llama : per-layer KV cache #4309

llama : per-layer KV cache #4309

Conversation

ggerganov commented Dec 3, 2023 • edited Loading

llama.h API Changes

Limitations

slaren Dec 3, 2023 • edited Loading

Choose a reason for hiding this comment

ggerganov Dec 3, 2023

Choose a reason for hiding this comment

slaren Dec 3, 2023

Choose a reason for hiding this comment

kalomaze Dec 3, 2023 • edited Loading

Choose a reason for hiding this comment

ggerganov Dec 3, 2023

Choose a reason for hiding this comment

kalomaze Dec 3, 2023 • edited Loading

Choose a reason for hiding this comment

ggerganov Dec 3, 2023 • edited Loading

Choose a reason for hiding this comment

oobabooga commented Dec 3, 2023

PR (c80b8a2)

Master

kalomaze commented Dec 3, 2023 • edited Loading

kalomaze commented Dec 3, 2023 • edited Loading

ggerganov commented Dec 6, 2023

kalomaze commented Dec 7, 2023 • edited Loading

kalomaze commented Dec 7, 2023 • edited Loading

ggerganov commented Dec 7, 2023

Dampfinchen commented Dec 7, 2023 • edited Loading

x4080 commented Dec 7, 2023

CyborgArmy83 commented Dec 7, 2023

LostRuins commented Dec 8, 2023 • edited Loading

Ph0rk0z commented Dec 8, 2023

Dampfinchen commented Dec 10, 2023

ggerganov commented Dec 3, 2023 •

edited

Loading

`llama.h` API Changes

slaren Dec 3, 2023 •

edited

Loading

kalomaze Dec 3, 2023 •

edited

Loading

kalomaze Dec 3, 2023 •

edited

Loading

ggerganov Dec 3, 2023 •

edited

Loading

PR (`c80b8a2`)

kalomaze commented Dec 3, 2023 •

edited

Loading

kalomaze commented Dec 3, 2023 •

edited

Loading

kalomaze commented Dec 7, 2023 •

edited

Loading

kalomaze commented Dec 7, 2023 •

edited

Loading

Dampfinchen commented Dec 7, 2023 •

edited

Loading

LostRuins commented Dec 8, 2023 •

edited

Loading