GPT-OSS performance

[ikawrakow]

I see currently 2 of the 3 pinned items on the mainline llama.cpp discussion page are devoted to GPT-OSS, so I thought it is useful to have one here too.

Support

Support for the GPT-OSS models was added to ik_llama.cpp in PR #689. Support for MAXFP4, the fp4 variant used in these models was added in #682. Improvements for the Harmony chat template and tool call support followed in #677 and #723.
Performance for these models was improved via SWA-specific optimizations in #692, #754 and #757. Hybrid inference PP performance was increased in #698 by offloading only the activated experts to the GPU (a feature that appears to be only useful for the GPT-OSS models).

Performance

There are two main advantages of ik_llama.cpp over llama.cpp: availability of higher quality quantization types and better performance. Hence, it makes sense to compare ik_llama.cpp and llama.cpp performance for GPT-OSS. I'll only provide results for the native MXFP4 quantization type. All results reported here are for the 20B variant.

The GPT-OSS models use sliding window attention (SWA) for every second layer, with a fairly small windows size of 128 tokens. There is a significant difference between ik_llama.cpp and llama.cpp how SWA layers are handled. In mainline, by default only the tokens required to compute the last u-batch are kept in the cache. This has the distinct advantage of reducing KV-cache size by nearly a factor of 2. The downside of this is that the KV-cache is not reusable. To be able to reuse the KV-cache, one needs to add --swa-full to the llama.cpp command line. This keeps the full KV-cache also for SWA layers, and corresponds to what is done in ik_llama.cpp. Hence, when comparing performance, I'll provide llama.cpp results with and without --swas-full, the latter being a more fair comparison (due to the usability limitations one gets without --swa-full).

CUDA performance, GPT-OSS-20B-MXFP4

Tests are run on an RTX-4080 GPU in a Ryzen-7950X rig.

Command line is

./bin/llama-sweep-bench -m $model -c 30720 -ub 2048 -t 1 -ngl 100 [-fa -fmoe]

(the arguments in square brackets only apply to ik_llama.cpp).

PP performance

Here ik_llama.cpp is ~40% faster at small N_KV, and 65% faster at 30k tokens compared to the corresponding calculation with llama.cpp using --swa-full.

TG performance

Here performance is similar at low context length, but ik_llama.cpp is ~15% faster at 30k tokens compared to --swa-full. llama.cpp manages to outperform ik_llama.cpp by a small margin (4.5% at 30k tokens) for large N_KV when not using --swa-full. I think this is simply due to the much smaller KQ mask for the SWA layers that has to be copied to the GPU for each token prediction. The ik_llama.cpp mask is 4 MiB at 32k tokens, which takes 0.27 ms for a 15 GiB/s PCI-E, which is ~3.5% of the time taken per token generation at 130 t/s. Without --swa-full the llama.cpp KQ SWA mask is ~15X smaller, so time to copy to the GPU is negligible.

llama.cpp with --swa-full

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
2048	512	0	0.314	6530.99	2.751	186.09
2048	512	2048	0.300	6827.90	2.988	171.33
2048	512	4096	0.326	6275.32	3.203	159.87
2048	512	6144	0.353	5799.25	3.421	149.65
2048	512	8192	0.385	5322.84	3.316	154.43
2048	512	10240	0.415	4931.17	3.444	148.67
2048	512	12288	0.449	4560.65	3.599	142.27
2048	512	14336	0.473	4327.45	3.693	138.63
2048	512	16384	0.506	4044.18	3.798	134.81
2048	512	18432	0.534	3832.20	3.907	131.05
2048	512	20480	0.563	3634.44	4.018	127.44
2048	512	22528	0.587	3485.98	4.131	123.94
2048	512	24576	0.614	3334.87	4.261	120.15
2048	512	26624	0.652	3142.10	4.351	117.69
2048	512	28672	0.677	3023.10	4.463	114.71

llama.cpp without --swa-full

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
2048	512	0	0.312	6560.28	2.751	186.13
2048	512	2048	0.285	7181.66	2.984	171.59
2048	512	4096	0.303	6762.65	3.091	165.66
2048	512	6144	0.315	6504.91	3.204	159.80
2048	512	8192	0.335	6111.65	3.155	162.28
2048	512	10240	0.347	5910.30	3.210	159.48
2048	512	12288	0.365	5616.79	3.288	155.74
2048	512	14336	0.385	5316.97	3.330	153.73
2048	512	16384	0.402	5092.58	3.388	151.10
2048	512	18432	0.411	4977.99	3.449	148.44
2048	512	20480	0.434	4724.01	3.499	146.32
2048	512	22528	0.450	4553.29	3.559	143.86
2048	512	24576	0.462	4434.42	3.623	141.31
2048	512	26624	0.482	4252.86	3.664	139.76
2048	512	28672	0.498	4109.87	3.724	137.49

ik_llama.cpp

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
2048	512	0	0.205	10003.96	2.871	178.33
2048	512	2048	0.208	9826.36	2.964	172.75
2048	512	4096	0.230	8917.26	3.033	168.80
2048	512	6144	0.246	8330.35	3.122	164.01
2048	512	8192	0.263	7787.72	3.196	160.22
2048	512	10240	0.276	7409.15	3.274	156.40
2048	512	12288	0.292	7014.40	3.345	153.09
2048	512	14336	0.307	6661.16	3.412	150.08
2048	512	16384	0.322	6361.24	3.473	147.41
2048	512	18432	0.336	6102.96	3.543	144.53
2048	512	20480	0.351	5832.50	3.615	141.63
2048	512	22528	0.367	5586.18	3.685	138.95
2048	512	24576	0.381	5373.61	3.762	136.11
2048	512	26624	0.396	5174.20	3.833	133.59
2048	512	28672	0.410	4991.07	3.889	131.67

CPU performance, GPT-OSS-20B-MXFP4

Tests are run on a Ryzen-7950X CPU. The command line used is

./bin/llama-sweep-bench -m $model -c 32768 -t 16 -ub 1024 -ctk q8_0 [-fa -fmoe]

PP performance

Not much comment is needed here. ik_llama.cpp is 3.6X faster at zero context, and ~8.8X faster at 25k tokens (I did not have the patience to wait for the llama.cpp results up two 32k tokens). There is no difference with and without --swa-full when running llama.cpp on the CPU

TG performance

Here performance is about the same at zero context length. At 25k tokens ik_llama.cpp is about 1.5X faster.

llama.cpp with --swa-full

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
1024	256	0	7.280	140.67	11.803	21.69
1024	256	1024	8.072	126.86	12.175	21.03
1024	256	2048	8.910	114.93	12.525	20.44
1024	256	3072	9.808	104.40	12.809	19.99
1024	256	4096	10.789	94.91	13.162	19.45
1024	256	5120	11.825	86.60	13.543	18.90
1024	256	6144	12.874	79.54	14.165	18.07
1024	256	7168	13.907	73.63	14.675	17.44
1024	256	8192	14.834	69.03	15.153	16.89
1024	256	9216	15.838	64.66	15.408	16.62
1024	256	10240	16.862	60.73	15.600	16.41
1024	256	11264	17.842	57.39	16.012	15.99
1024	256	12288	18.768	54.56	16.931	15.12
1024	256	13312	19.768	51.80	16.885	15.16
1024	256	14336	20.800	49.23	17.444	14.68
1024	256	15360	21.984	46.58	18.309	13.98
1024	256	16384	23.195	44.15	17.529	14.60
1024	256	17408	24.539	41.73	17.947	14.26
1024	256	18432	26.365	38.84	19.978	12.81
1024	256	19456	28.280	36.21	20.000	12.80
1024	256	20480	29.668	34.51	20.748	12.34
1024	256	21504	32.048	31.95	20.688	12.37
1024	256	22528	34.715	29.50	21.426	11.95
1024	256	23552	37.818	27.08	22.884	11.19
1024	256	24576	39.904	25.66	24.546	10.43

llama.cpp without --swa-full

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
1024	256	0	7.157	143.07	11.798	21.70
1024	256	1024	8.034	127.46	12.169	21.04
1024	256	2048	8.826	116.02	12.497	20.48
1024	256	3072	9.702	105.54	12.860	19.91
1024	256	4096	10.569	96.89	13.327	19.21
1024	256	5120	11.512	88.95	13.649	18.76
1024	256	6144	12.491	81.98	13.963	18.33
1024	256	7168	13.439	76.20	14.555	17.59
1024	256	8192	14.417	71.03	14.805	17.29
1024	256	9216	15.420	66.41	15.125	16.93
1024	256	10240	16.291	62.86	15.595	16.42
1024	256	11264	17.155	59.69	16.004	16.00
1024	256	12288	18.151	56.42	15.943	16.06
1024	256	13312	19.160	53.44	17.927	14.28
1024	256	14336	20.046	51.08	18.445	13.88
1024	256	15360	21.089	48.56	18.030	14.20
1024	256	16384	22.117	46.30	19.256	13.29
1024	256	17408	23.523	43.53	17.822	14.36
1024	256	18432	25.182	40.66	19.401	13.20
1024	256	19456	27.343	37.45	19.626	13.04
1024	256	20480	29.099	35.19	19.577	13.08
1024	256	21504	31.572	32.43	20.196	12.68
1024	256	22528	32.758	31.26	20.886	12.26
1024	256	23552	36.318	28.20	22.498	11.38
1024	256	24576	38.426	26.65	23.330	10.97

ik_llama.cpp

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
1024	256	0	1.993	513.76	11.842	21.62
1024	256	1024	2.053	498.86	12.038	21.27
1024	256	2048	2.133	480.17	12.245	20.91
1024	256	3072	2.232	458.81	12.430	20.59
1024	256	4096	2.314	442.56	12.599	20.32
1024	256	5120	2.404	425.97	12.746	20.08
1024	256	6144	2.497	410.12	12.900	19.84
1024	256	7168	2.587	395.89	13.040	19.63
1024	256	8192	2.682	381.83	13.180	19.42
1024	256	9216	2.774	369.18	13.319	19.22
1024	256	10240	2.894	353.79	13.594	18.83
1024	256	11264	3.172	322.85	13.663	18.74
1024	256	12288	3.271	313.05	13.779	18.58
1024	256	13312	3.251	314.96	13.777	18.58
1024	256	14336	3.624	282.56	14.228	17.99
1024	256	15360	3.511	291.66	14.155	18.08
1024	256	16384	3.504	292.25	14.518	17.63
1024	256	17408	3.793	269.98	14.338	17.85
1024	256	18432	3.972	257.79	14.488	17.67
1024	256	19456	4.127	248.12	14.776	17.33
1024	256	20480	3.798	269.60	15.039	17.02
1024	256	21504	4.409	232.24	14.810	17.29
1024	256	22528	4.194	244.14	15.571	16.44
1024	256	23552	4.313	237.42	15.386	16.64
1024	256	24576	4.526	226.24	15.575	16.44
1024	256	25600	4.482	228.45	15.735	16.27
1024	256	26624	4.482	228.49	16.177	15.82
1024	256	27648	4.537	225.69	16.216	15.79
1024	256	28672	4.756	215.33	16.305	15.70
1024	256	29696	4.735	216.27	16.765	15.27
1024	256	30720	5.095	200.97	16.782	15.25
1024	256	31744	5.113	200.28	16.992	15.07