llama : pad KV cache size #4280

ggerganov · 2023-12-01T09:01:12Z

Should result in slight TG speedup

ggml_init_cublas: GGML_CUDA_FORCE_MMQ: no
ggml_init_cublas: CUDA_USE_TENSOR_CORES: yes
ggml_init_cublas: found 1 CUDA devices:
Device 0: Tesla V100-PCIE-16GB, compute capability 7.0

model	size	backend	ngl	test	t/s	t/s	speedup
llama 7B F16	12.55 GiB	CUDA	99	pp 512	3457.68 ± 89.53	3444.83 ± 126.66	1.000
llama 7B F16	12.55 GiB	CUDA	99	tg 128	53.01 ± 0.05	53.14 ± 0.07	1.002
llama 7B F16	12.55 GiB	CUDA	99	tg 256	52.56 ± 0.03	52.84 ± 0.07	1.005
llama 7B F16	12.55 GiB	CUDA	99	tg 512	51.83 ± 0.06	52.43 ± 0.06	1.012
llama 7B Q8_0	6.67 GiB	CUDA	99	pp 512	2513.48 ± 31.58	2516.75 ± 34.50	1.000
llama 7B Q8_0	6.67 GiB	CUDA	99	tg 128	78.91 ± 0.15	79.25 ± 0.23	1.004
llama 7B Q8_0	6.67 GiB	CUDA	99	tg 256	78.16 ± 0.22	78.91 ± 0.08	1.010
llama 7B Q8_0	6.67 GiB	CUDA	99	tg 512	76.54 ± 0.14	77.83 ± 0.16	1.017
llama 7B Q4_0	3.56 GiB	CUDA	99	pp 512	2720.43 ± 34.16	2716.96 ± 32.91	1.000
llama 7B Q4_0	3.56 GiB	CUDA	99	tg 128	115.78 ± 0.34	116.34 ± 0.95	1.005
llama 7B Q4_0	3.56 GiB	CUDA	99	tg 256	114.17 ± 0.37	115.70 ± 0.31	1.013
llama 7B Q4_0	3.56 GiB	CUDA	99	tg 512	111.05 ± 0.27	113.79 ± 0.23	1.025

build: 75ba5ba (1594)

Also, try to improve the batched decoding performance for quantum models on Apple Silicon.

make -j batched-bench && ./batched-bench ./models/llama-7b-v2/ggml-model-q8_0.gguf 10880 1 99 0 128,2048 128 1,2,4,8

master

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.122	1049.89	1.899	67.39	2.021	126.65
128	128	2	384	0.120	1062.68	7.212	35.50	7.332	52.37
128	128	4	640	0.120	1067.81	7.207	71.04	7.327	87.35
128	128	8	1152	0.120	1065.87	7.380	138.75	7.500	153.59
2048	128	1	2176	1.665	1229.84	2.293	55.82	3.958	549.71
2048	128	2	2304	1.659	1234.55	7.568	33.83	9.227	249.72
2048	128	4	2560	1.658	1235.25	7.791	65.72	9.449	270.93
2048	128	8	3072	1.659	1234.77	8.187	125.08	9.845	312.03

PR

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.120	1062.85	1.898	67.42	2.019	126.81
128	128	2	384	0.121	1060.96	2.614	97.92	2.735	140.40
128	128	4	640	0.120	1069.25	3.983	128.54	4.103	155.98
128	128	8	1152	0.120	1066.80	6.810	150.37	6.930	166.24
2048	128	1	2176	1.665	1229.74	2.290	55.89	3.956	550.11
2048	128	2	2304	1.662	1232.19	3.085	82.98	4.747	485.35
2048	128	4	2560	1.658	1235.08	4.462	114.74	6.120	418.27
2048	128	8	3072	1.657	1235.68	7.299	140.29	8.956	343.00

make -j batched-bench && ./batched-bench ./models/llama-7b-v2/ggml-model-q4_0.gguf 10880 1 99 0 128,2048 128 1,2,4,8

master

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.121	1060.20	1.335	95.90	1.455	175.89
128	128	2	384	0.120	1065.36	7.042	36.35	7.162	53.62
128	128	4	640	0.118	1083.25	7.062	72.50	7.180	89.13
128	128	8	1152	0.118	1080.54	7.322	139.85	7.441	154.83
2048	128	1	2176	1.675	1222.93	1.730	73.99	3.405	639.12
2048	128	2	2304	1.674	1223.57	7.451	34.36	9.124	252.51
2048	128	4	2560	1.670	1225.99	7.656	66.88	9.326	274.50
2048	128	8	3072	1.670	1226.32	8.036	127.43	9.706	316.50

PR

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.121	1055.03	1.342	95.37	1.463	174.93
128	128	2	384	0.121	1060.52	1.846	138.67	1.967	195.24
128	128	4	640	0.119	1075.47	2.680	191.02	2.799	228.62
128	128	8	1152	0.120	1068.20	4.398	232.85	4.518	255.01
2048	128	1	2176	1.680	1218.71	1.735	73.77	3.416	637.06
2048	128	2	2304	1.676	1221.80	2.315	110.60	3.991	577.30
2048	128	4	2560	1.674	1223.71	3.159	162.08	4.832	529.75
2048	128	8	3072	1.672	1224.88	4.921	208.07	6.593	465.92

make -j batched-bench && ./batched-bench ./models/codellama-34b/ggml-model-q8_0.gguf 10880 1 99 0 128,2048 128 1,2,4,8

master

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.532	240.41	7.370	17.37	7.902	32.40
128	128	2	384	0.532	240.57	24.895	10.28	25.428	15.10
128	128	4	640	0.532	240.65	25.136	20.37	25.668	24.93
128	128	8	1152	0.532	240.71	25.648	39.93	26.180	44.00
2048	128	1	2176	7.400	276.77	8.464	15.12	15.863	137.17
2048	128	2	2304	7.393	277.01	25.747	9.94	33.141	69.52
2048	128	4	2560	7.391	277.09	26.215	19.53	33.606	76.18
2048	128	8	3072	7.392	277.07	27.027	37.89	34.419	89.25

PR

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.532	240.63	7.371	17.36	7.903	32.39
128	128	2	384	0.532	240.73	13.211	19.38	13.743	27.94
128	128	4	640	0.531	240.95	24.350	21.03	24.881	25.72
128	128	8	1152	0.530	241.30	46.880	21.84	47.410	24.30
2048	128	1	2176	7.398	276.82	8.457	15.14	15.855	137.24
2048	128	2	2304	7.397	276.86	14.142	18.10	21.539	106.97
2048	128	4	2560	7.392	277.06	25.311	20.23	32.703	78.28
2048	128	8	3072	7.392	277.05	47.932	21.36	55.325	55.53

make -j batched-bench && ./batched-bench ./models/codellama-34b/ggml-model-q4_0.gguf 10880 1 99 0 128,2048 128 1,2,4,8

`master

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.530	241.28	4.454	28.74	4.984	51.36
128	128	2	384	0.530	241.32	23.865	10.73	24.396	15.74
128	128	4	640	0.531	241.21	24.083	21.26	24.614	26.00
128	128	8	1152	0.530	241.60	24.660	41.53	25.190	45.73
2048	128	1	2176	7.475	274.00	5.536	23.12	13.010	167.26
2048	128	2	2304	7.466	274.32	24.595	10.41	32.061	71.86
2048	128	4	2560	7.465	274.34	25.176	20.34	32.641	78.43
2048	128	8	3072	7.464	274.37	26.181	39.11	33.645	91.31

PR

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.531	240.92	4.473	28.62	5.004	51.16
128	128	2	384	0.529	241.79	7.421	34.50	7.951	48.30
128	128	4	640	0.530	241.66	13.086	39.13	13.616	47.00
128	128	8	1152	0.530	241.64	24.512	41.78	25.041	46.00
2048	128	1	2176	7.471	274.11	5.564	23.00	13.036	166.93
2048	128	2	2304	7.473	274.07	8.379	30.55	15.851	145.35
2048	128	4	2560	7.464	274.40	14.202	36.05	21.666	118.16
2048	128	8	3072	7.472	274.10	25.890	39.55	33.362	92.08

slaren · 2023-12-01T12:17:42Z

3090 Ti/ WSL2

Device 0: NVIDIA GeForce RTX 3090 Ti, compute capability 8.6

model	size	backend	ngl	test	`8d6d9f0` t/s	PR `75ba5ba` t/s	speedup
llama 7B mostly F16	12.55 GiB	CUDA	99	pp 512	4982.86 ± 71.87	5069.56 ± 68.03	1.017
llama 7B mostly F16	12.55 GiB	CUDA	99	tg 128	54.58 ± 0.55	54.65 ± 0.70	1.001
llama 7B mostly F16	12.55 GiB	CUDA	99	tg 256	54.90 ± 0.20	53.97 ± 0.29	0.983
llama 7B mostly F16	12.55 GiB	CUDA	99	tg 512	53.99 ± 0.34	54.36 ± 0.10	1.006
llama 7B mostly Q8_0	6.67 GiB	CUDA	99	pp 512	3809.50 ± 77.34	3758.40 ± 87.26	0.986
llama 7B mostly Q8_0	6.67 GiB	CUDA	99	tg 128	86.86 ± 0.33	85.46 ± 1.18	0.983
llama 7B mostly Q8_0	6.67 GiB	CUDA	99	tg 256	86.44 ± 0.10	84.48 ± 1.04	0.977
llama 7B mostly Q8_0	6.67 GiB	CUDA	99	tg 512	84.50 ± 0.92	85.21 ± 0.72	1.008
llama 7B mostly Q4_0	3.56 GiB	CUDA	99	pp 512	3810.83 ± 84.89	3887.95 ± 11.28	1.020
llama 7B mostly Q4_0	3.56 GiB	CUDA	99	tg 128	126.26 ± 0.98	127.93 ± 0.41	1.013
llama 7B mostly Q4_0	3.56 GiB	CUDA	99	tg 256	126.53 ± 0.47	127.25 ± 0.18	1.005
llama 7B mostly Q4_0	3.56 GiB	CUDA	99	tg 512	124.33 ± 0.51	124.98 ± 0.27	1.005

ggerganov · 2023-12-01T21:19:48Z

Looking for Apple Silicon runs of the following 7B benches, master vs PR:

make -j batched-bench && ./batched-bench ./models/llama-7b-v2/ggml-model-q8_0.gguf 10880 1 99 0 128,2048 128 1,2,4,8

make -j batched-bench && ./batched-bench ./models/llama-7b-v2/ggml-model-q4_0.gguf 10880 1 99 0 128,2048 128 1,2,4,8

cc @slaren @jhen0409 and everyone else with a Mac

jhen0409 · 2023-12-02T01:48:17Z

M1 Max MBP (32c):

make -j batched-bench && ./batched-bench ./models/llama-7b-v2/ggml-model-q8_0.gguf 10880 1 99 0 128,2048 128 1,2,4,8

master

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.247	517.85	3.205	39.94	3.452	74.16
128	128	2	384	0.248	516.60	10.249	24.98	10.497	36.58
128	128	4	640	0.246	519.61	10.387	49.29	10.633	60.19
128	128	8	1152	0.246	520.45	10.785	94.94	11.031	104.43
2048	128	1	2176	3.901	525.00	3.707	34.53	7.608	286.01
2048	128	2	2304	3.895	525.76	11.045	23.18	14.940	154.22
2048	128	4	2560	3.891	526.37	11.457	44.69	15.348	166.80
2048	128	8	3072	3.889	526.55	12.349	82.92	16.238	189.18

PR

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.248	515.87	3.191	40.11	3.439	74.43
128	128	2	384	0.248	516.49	5.758	44.46	6.006	63.94
128	128	4	640	0.247	518.64	10.320	49.61	10.567	60.57
128	128	8	1152	0.246	519.55	19.319	53.00	19.565	58.88
2048	128	1	2176	3.906	524.27	3.695	34.64	7.602	286.25
2048	128	2	2304	3.898	525.39	6.510	39.32	10.408	221.36
2048	128	4	2560	3.891	526.30	11.049	46.34	14.940	171.35
2048	128	8	3072	3.891	526.30	20.171	50.77	24.062	127.67

make -j batched-bench && ./batched-bench ./models/llama-7b-v2/ggml-model-q4_0.gguf 10880 1 99 0 128,2048 128 1,2,4,8

master

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.249	513.28	2.092	61.19	2.341	109.35
128	128	2	384	0.249	514.12	10.313	24.82	10.562	36.36
128	128	4	640	0.248	516.87	10.473	48.89	10.720	59.70
128	128	8	1152	0.247	518.07	10.871	94.20	11.118	103.62
2048	128	1	2176	3.978	514.80	2.620	48.85	6.599	329.76
2048	128	2	2304	3.972	515.62	11.055	23.16	15.027	153.33
2048	128	4	2560	3.968	516.12	11.518	44.45	15.486	165.31
2048	128	8	3072	3.967	516.27	12.458	82.20	16.425	187.04

PR

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.255	501.64	2.124	60.27	2.379	107.61
128	128	2	384	0.249	515.07	3.555	72.01	3.803	100.96
128	128	4	640	0.248	515.99	6.051	84.61	6.299	101.60
128	128	8	1152	0.247	517.46	11.131	92.00	11.378	101.25
2048	128	1	2176	3.982	514.36	2.620	48.85	6.602	329.60
2048	128	2	2304	3.979	514.73	4.363	58.67	8.342	276.19
2048	128	4	2560	3.974	515.33	6.885	74.36	10.859	235.74
2048	128	8	3072	3.975	515.21	11.925	85.87	15.900	193.21

q8_0 drops significantly at B=8.

slaren · 2023-12-02T11:57:42Z

M3 Max

make -j batched-bench && ./batched-bench ./models/llama-7b-v2/ggml-model-q8_0.gguf 10880 1 99 0 128,2048 128 1,2,4,8

master:

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.182	703.44	2.997	42.71	3.179	80.53
128	128	2	384	0.181	705.39	8.238	31.07	8.420	45.61
128	128	4	640	0.181	706.57	8.285	61.80	8.466	75.60
128	128	8	1152	0.185	691.87	8.621	118.78	8.806	130.82
2048	128	1	2176	3.327	615.60	3.851	33.24	7.178	303.15
2048	128	2	2304	3.190	641.97	9.112	28.10	12.302	187.29
2048	128	4	2560	3.430	597.02	9.608	53.29	13.038	196.35
2048	128	8	3072	3.613	566.77	10.194	100.45	13.808	222.49

PR:

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.183	700.89	3.008	42.55	3.191	80.23
128	128	2	384	0.182	705.07	3.203	79.94	3.384	113.47
128	128	4	640	0.181	706.02	4.287	119.42	4.469	143.22
128	128	8	1152	0.186	687.37	8.373	122.30	8.559	134.60
2048	128	1	2176	3.171	645.93	3.833	33.40	7.003	310.71
2048	128	2	2304	3.156	648.94	4.256	60.15	7.412	310.85
2048	128	4	2560	3.319	617.07	6.085	84.14	9.404	272.22
2048	128	8	3072	3.492	586.42	9.784	104.66	13.276	231.39

make -j batched-bench && ./batched-bench ./models/llama-7b-v2/ggml-model-q4_0.gguf 10880 1 99 0 128,2048 128 1,2,4,8

master:

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.182	703.85	1.969	65.00	2.151	119.01
128	128	2	384	0.182	705.21	8.237	31.08	8.418	45.61
128	128	4	640	0.181	706.64	8.289	61.77	8.470	75.56
128	128	8	1152	0.198	646.09	8.831	115.95	9.029	127.59
2048	128	1	2176	3.558	575.60	2.736	46.79	6.294	345.73
2048	128	2	2304	3.584	571.38	9.400	27.24	12.984	177.45
2048	128	4	2560	3.534	579.60	9.520	53.78	13.054	196.12
2048	128	8	3072	3.504	584.48	9.930	103.12	13.434	228.68

PR:

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
128	128	1	256	0.182	702.75	1.982	64.58	2.164	118.29
128	128	2	384	0.181	705.49	2.263	113.13	2.444	157.09
128	128	4	640	0.181	706.66	3.698	138.44	3.880	164.97
128	128	8	1152	0.182	703.27	7.024	145.79	7.206	159.87
2048	128	1	2176	3.169	646.36	2.755	46.45	5.924	367.33
2048	128	2	2304	3.199	640.28	3.182	80.44	6.381	361.07
2048	128	4	2560	3.279	624.66	4.917	104.13	8.196	312.36
2048	128	8	3072	3.333	614.53	8.237	124.31	11.570	265.51

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

I found that parallel (in examples/parallel) was unusable when -np > 1. I bisected the issue down to d7b800b I don't really understand anything about kv-cache, just that the change caused parallel to emit nonsense on my M2 Mac Studio (Apple M2 Max, macOS 14.1.2 (23B92)). The comments around it say kv_self.n is a heuristic (and seems to have comments suggesting other possible values for assignment), so I presume that it shouldn't be a problem to remove the GGML_PAD(). Empirically it seems to work fine. That said, it does sound like the bug could run deeper, but it is beyond my ability to understand what the root cause might be. It apparently reproduces on various models not only tinyllama, but this one is small so it should be more convenient to reproduce. While tinyllama isn't known for the quality of its output, there is still an obvious difference between the nonsense output and the normal output. Reproduction: `./parallel -c 99999 -n 30 -ns 10 -np 2 -m ~/Downloads/tinyllama-1.1b-chat-v1.0.Q5_K_M.gguf` Before fix (example bad output): ``` Input: If you could have any superpower, what would it be? Response: . In the 812 ``` After fix (example expected output): ``` Input: If you could have any superpower, what would it be? Response: I would choose the power of being able to control time. The power ``` ---- After typing the above I realized when running larger models, the problem is less apparent, but still exists. For example, I tried mixtral: `./parallel -c 99999 -n 50 -ns 50 -np 2 -m ~/Downloads/mixtral-8x7b-instruct-v0.1.Q5_K_M.gguf` And one of the outputs was: ``` Input: Recommend some interesting books to read. Response: I recommend the book "Surelyourecommend would suggest starting with the book "The Foundation for Self-Help by Dr. Micahelle myself to anywhere in the world" ``` The problem with the above response is obvious, but here's another that isn't so obvious if you just glance at it: ``` Input: I want to learn how to play the piano. Response: That's great! I could recommend a personalize piano lessons with a piano teacher. This will allow you to learn at your own pace. You can practice scales and chords, ``` Note that "a personalize piano lessons" is not grammatical English, a mistake that mixtral should not make. I didn't notice any such errors when testing with this patch applied.

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>

llama : pad KV cache size to 32

3e68df8

ggerganov force-pushed the gg/pad-kv-cache branch from 75ba5ba to 3e68df8 Compare December 1, 2023 18:41

metal : try to improve batched decoding

3cb1c34

ggerganov force-pushed the gg/pad-kv-cache branch from 10cb459 to 3cb1c34 Compare December 1, 2023 20:02

ggerganov merged commit d7b800b into master Dec 3, 2023
37 checks passed

AriX mentioned this pull request Mar 20, 2024

Regression: llama.cpp produces nonsensical outputs when using batched decoding on Metal #6173

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

llama : pad KV cache size #4280

llama : pad KV cache size #4280

ggerganov commented Dec 1, 2023 •

edited

Loading

slaren commented Dec 1, 2023

ggerganov commented Dec 1, 2023

jhen0409 commented Dec 2, 2023

slaren commented Dec 2, 2023

llama : pad KV cache size #4280

llama : pad KV cache size #4280

Conversation

ggerganov commented Dec 1, 2023 • edited Loading

slaren commented Dec 1, 2023

ggerganov commented Dec 1, 2023

jhen0409 commented Dec 2, 2023

slaren commented Dec 2, 2023

ggerganov commented Dec 1, 2023 •

edited

Loading