Linear Attention & SSM

22 ops, 158 workloads.

One table per op, one row per workload. Ratio is the fastest other implementation's device time divided by ours, so green is faster than it, plain is level with it, red is slower. Times are in ms. How these numbers are taken.

Linear Attention / SSM

CBProducerFwd (2 workloads)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
mamba2-2p7b-b4-s2k-bfloat16	4.48×	0.0118	torch	0.0531	22.7
mamba2-780m-b1-s4k-float16	5.30×	0.00714	torch	0.0378	18.8

DeltaNetDecodeFwd (8 workloads · ✅)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
delta-decode-serving-b1-h16-d128-bfloat16	4.57×	0.0031	torch-compile torch	0.0142 0.0353	0.508
delta-decode-serving-b1-h32-d128-bfloat16	4.81×	0.00336	torch-compile torch	0.0162 0.0391	0.938
delta-decode-serving-b1-h48-d128-bfloat16	5.42×	0.00355	torch-compile torch	0.0192 0.0434	1.33
delta-decode-serving-b1-h64-d128-bfloat16	4.51×	0.00384	torch-compile torch	0.0173 0.0438	1.64
delta-decode-serving-b1-h8-d128-bfloat16	4.60×	0.00285	torch-compile torch	0.0131 0.0335	0.277
delta-decode-serving-b8-h32-d128-bfloat16	4.38×	0.0087	torch-compile torch	0.0381 0.0896	2.9
delta-decode-serving-b8-h48-d128-bfloat16	3.22×	0.0123	torch-compile torch	0.0396 0.1111	3.07
delta-decode-serving-b8-h64-d128-bfloat16	3.15×	0.0163	torch-compile torch	0.0515 0.1420	3.09

EngramGateConvBwd (3 workloads · ✅)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
bwd-b1-s128-d256-bfloat16	3.24×	0.0168	torch-compile torch	0.0544 0.1838	0.118
bwd-b1-s32-d256-float16	4.48×	0.0111	torch-compile torch	0.0498 0.1677	0.0444
bwd-b2-s64-d512-float16	3.01×	0.0198	torch-compile torch	0.0595 0.2003	0.199

GatedDeltaNetPrefillBTHDFwd (26 workloads)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
fallback-gdn-prefill-b1-s4k-h16-d128-bthd-bfloat16	2.49×	0.0794	fla	0.1974	108
fallback-gdn-prefill-b1-s4k-h16-d128-bthd-float16	2.48×	0.0792	fla	0.1960	109
qwen35-gdn-prefill-b1-s128k-h16-d128-bthd-bfloat16	4.50×	1.2814	fla	5.7677	215
qwen35-gdn-prefill-b1-s128k-h16-d128-bthd-float16	4.59×	1.2572	fla	5.7713	219
qwen35-gdn-prefill-b1-s128k-h32-d128-bthd-bfloat16	3.50×	2.5076	fla	8.7773	219
qwen35-gdn-prefill-b1-s128k-h32-d128-bthd-float16	3.58×	2.4458	fla	8.7477	225
qwen35-gdn-prefill-b1-s128k-h48-d128-bthd-bfloat16	3.30×	3.8074	fla	12.5781	217
qwen35-gdn-prefill-b1-s128k-h48-d128-bthd-float16	3.34×	3.7766	fla	12.5997	218
qwen35-gdn-prefill-b1-s128k-h64-d128-bthd-bfloat16	3.26×	4.7856	fla	15.5906	230
qwen35-gdn-prefill-b1-s128k-h64-d128-bthd-float16	3.35×	4.6645	fla	15.6272	236
qwen35-gdn-prefill-b1-s32k-h16-d128-bthd-bfloat16	3.96×	0.3711	fla	1.4701	185
qwen35-gdn-prefill-b1-s32k-h16-d128-bthd-float16	4.01×	0.3653	fla	1.4649	188
qwen35-gdn-prefill-b1-s32k-h32-d128-bthd-bfloat16	3.17×	0.7005	fla	2.2183	196
qwen35-gdn-prefill-b1-s32k-h32-d128-bthd-float16	3.23×	0.6862	fla	2.2146	200
qwen35-gdn-prefill-b1-s32k-h48-d128-bthd-bfloat16	2.97×	1.0652	fla	3.1612	194
qwen35-gdn-prefill-b1-s32k-h48-d128-bthd-float16	3.02×	1.0493	fla	3.1685	196
qwen35-gdn-prefill-b1-s32k-h64-d128-bthd-bfloat16	3.11×	1.2532	fla	3.8987	219
qwen35-gdn-prefill-b1-s32k-h64-d128-bthd-float16	3.19×	1.2242	fla	3.9055	225
qwen35-gdn-prefill-b1-s64k-h16-d128-bthd-bfloat16	4.13×	0.7069	fla	2.9162	194
qwen35-gdn-prefill-b1-s64k-h16-d128-bthd-float16	4.17×	0.6972	fla	2.9043	197
qwen35-gdn-prefill-b1-s64k-h32-d128-bthd-bfloat16	3.45×	1.2767	fla	4.4014	215
qwen35-gdn-prefill-b1-s64k-h32-d128-bthd-float16	3.53×	1.2458	fla	4.3951	221
qwen35-gdn-prefill-b1-s64k-h48-d128-bthd-bfloat16	3.25×	1.9437	fla	6.3077	212
qwen35-gdn-prefill-b1-s64k-h48-d128-bthd-float16	3.29×	1.9168	fla	6.3109	215
qwen35-gdn-prefill-b1-s64k-h64-d128-bthd-bfloat16	3.21×	2.4273	fla	7.7806	226
qwen35-gdn-prefill-b1-s64k-h64-d128-bthd-float16	3.29×	2.3772	fla	7.8149	231

EngramGateConvFwd (3 workloads · ✅)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
fwd-b1-s128-d256-bfloat16	2.60×	0.00448	torch-compile torch-ref	0.0116 0.0802	0.176
fwd-b1-s32-d256-float16	2.72×	0.004	torch-compile torch-ref	0.0109 0.0740	0.0493
fwd-b2-s64-d512-float16	2.46×	0.00506	torch-compile torch-ref	0.0124 0.0856	0.312

SSDDecodeFwd (4 workloads)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
mamba2-1p3b-decode-b1-bfloat16	2.30×	0.00397	torch-compile torch-ref	0.00912 0.0304	1.06
mamba2-1p3b-decode-b1-float16	2.28×	0.00397	torch-compile torch-ref	0.00906 0.0304	1.06
mamba2-2p7b-decode-b8-float16	1.84×	0.0163	torch-compile torch-ref	0.0301 0.1119	2.57
mamba2-780m-decode-b32-float16	1.92×	0.0361	torch-compile torch-ref	0.0695 0.2409	2.79

SSDStatePassingFwd (6 workloads · ✅)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
mamba2-1p3b-b1-s4k-dstate-bfloat16	2.10×	0.00195	torch-compile mamba torch-ref	0.0041 0.00835 0.1229	0.135
mamba2-1p3b-b1-s4k-dstate-float16	2.10×	0.00195	torch-compile mamba torch-ref	0.0041 0.00842 0.1215	0.135
mamba2-1p3b-b1-s4k-dstate-init-states-bfloat16	1.11×	0.00197	torch-compile mamba torch-ref	0.00218 0.0087 0.1220	0.134
mamba2-1p3b-b1-s4k-dstate-init-states-float16	1.71×	0.002	torch-compile mamba torch-ref	0.00342 0.0087 0.1206	0.132
mamba2-1p3b-b1-s4k-flat-init-states-float32	0.93×	0.0219	torch-compile mamba torch-ref	0.0204 0.0216 0.1270	0.765
mamba2-2p7b-b2-s32k-dstate-float16	5.65×	0.0106	mamba torch-compile torch-ref	0.0599 0.0912 1.1537	0.497

GatedDeltaNetDecodeFwd (8 workloads · ✅)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
gdn-decode-serving-b1-h16-d128-bfloat16	1.27×	0.0033	fla	0.00419	0.478
gdn-decode-serving-b1-h32-d128-bfloat16	1.30×	0.00362	fla	0.00469	0.872
gdn-decode-serving-b1-h48-d128-bfloat16	1.37×	0.00384	fla	0.00525	1.23
gdn-decode-serving-b1-h64-d128-bfloat16	1.38×	0.00419	fla	0.00579	1.5
gdn-decode-serving-b1-h8-d128-bfloat16	1.29×	0.0031	fla	0.004	0.254
gdn-decode-serving-b8-h32-d128-bfloat16	1.67×	0.00874	fla	0.0146	2.89
gdn-decode-serving-b8-h48-d128-bfloat16	1.56×	0.0124	fla	0.0193	3.05
gdn-decode-serving-b8-h64-d128-bfloat16	1.56×	0.0161	fla	0.0252	3.14

SSDChunkScanFwd (4 workloads · ✅)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
mamba2-1p3b-b2-s32k-float16	1.39×	1.4677	mamba torch-compile torch-ref	2.0346 9.9818 40.0654	93.6
mamba2-2p7b-b4-s2k-bfloat16	1.30×	0.2372	mamba torch-compile torch-ref	0.3088 1.6394 6.5129	90.5
mamba2-780m-b1-s4k-bfloat16	1.34×	0.0759	mamba torch-compile torch-ref	0.1018 0.5090 1.9619	84.9
mamba2-780m-b1-s4k-float16	1.38×	0.0732	mamba torch-compile torch-ref	0.1007 0.5073 1.9596	88.1

GatedDeltaNetBTHDFwd (12 workloads)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
gdn-bthd-b1-s4k-h16-d128-bfloat16	1.12×	0.1742	fla	0.1948	58.6
gdn-bthd-b1-s4k-h16-d128-float16	1.11×	0.1750	fla	0.1935	58.3
gdn-bthd-b2-s16k-h4-d64-bfloat16	0.91×	0.4327	fla	0.3932	19.9
gdn-bthd-b2-s16k-h4-d64-float16	0.91×	0.4290	fla	0.3907	20
gdn-bthd-b2-s2k-h4-d64-bfloat16	1.02×	0.0664	fla	0.0675	16.2
gdn-bthd-b2-s2k-h4-d64-float16	1.00×	0.0669	fla	0.0672	16
gdn-bthd-b2-s32k-h4-d64-bfloat16	3.95×	0.1950	fla	0.7693	88.1
gdn-bthd-b2-s32k-h4-d64-float16	3.94×	0.1949	fla	0.7679	88.2
gdn-bthd-b2-s4k-h4-d64-bfloat16	0.95×	0.1145	fla	0.1083	18.8
gdn-bthd-b2-s4k-h4-d64-float16	0.94×	0.1151	fla	0.1077	18.7
gdn-bthd-b2-s8k-h4-d64-bfloat16	0.93×	0.2204	fla	0.2056	19.5
gdn-bthd-b2-s8k-h4-d64-float16	0.94×	0.2192	fla	0.2051	19.6

SSDChunkStateFwd (6 workloads · ✅)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
mamba2-1p3b-b2-s32k-seq-idx-float16	1.41×	0.4492	mamba torch-compile torch-ref	0.6315 16.6703 173.1990	154
mamba2-2p7b-b4-s2k-bfloat16	1.22×	0.0656	mamba torch-compile torch-ref	0.0802 2.4482 27.1195	164
mamba2-780m-b1-s4k-bfloat16	1.10×	0.0240	mamba torch-compile torch-ref	0.0265 0.6755 8.1408	135
mamba2-780m-b1-s4k-float16	1.05×	0.0238	mamba torch-compile torch-ref	0.0250 0.6331 8.1358	136
mamba2-780m-b1-s4k-seq-idx-bfloat16	1.01×	0.0290	mamba torch-compile torch-ref	0.0292 0.7870 8.1521	112
mamba2-780m-b1-s4k-seq-idx-float16	1.21×	0.0287	mamba torch-compile torch-ref	0.0347 0.7499 8.1473	113

Mamba2Fwd (8 workloads)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
mamba2-1p3b-b1-s8k-dt-bias-float16	1.08×	0.2904	mamba torch-compile torch-ref	0.3123 2.0138 6.9145	89.9
mamba2-1p3b-b1-s8k-dt-bias-init-states-float16	1.08×	0.2899	mamba torch-compile torch-ref	0.3125 2.0123 6.9059	90.1
mamba2-1p3b-b1-s8k-float16	1.08×	0.2909	mamba torch-compile torch-ref	0.3127 2.0180 6.9113	89.8
mamba2-1p3b-b1-s8k-init-states-float16	1.07×	0.2915	mamba torch-compile torch-ref	0.3126 2.0172 6.9060	89.6
mamba2-2p7b-b1-s2k-bfloat16	0.99×	0.1100	mamba torch-compile torch-ref	0.1093 0.6874 2.1548	74
mamba2-2p7b-b1-s2k-dt-bias-bfloat16	1.00×	0.1091	mamba torch-compile torch-ref	0.1096 0.6885 2.1567	74.6
mamba2-2p7b-b1-s2k-dt-bias-init-states-bfloat16	1.01×	0.1099	mamba torch-compile torch-ref	0.1106 0.6791 2.1555	74.1
mamba2-2p7b-b1-s2k-init-states-bfloat16	1.00×	0.1106	mamba torch-compile torch-ref	0.1108 0.6784 2.1521	73.6

GLADecodeFwd (8 workloads · ✅)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
gla-decode-serving-b1-h16-d128-bfloat16	0.90×	0.00742	torch-compile fla torch	0.00669 0.00704 0.0318	0.142
gla-decode-serving-b1-h32-d128-bfloat16	0.94×	0.00784	fla torch-compile torch	0.00736 0.00835 0.0358	0.269
gla-decode-serving-b1-h48-d128-bfloat16	1.13×	0.00794	fla torch-compile torch	0.00896 0.0103 0.0404	0.398
gla-decode-serving-b1-h64-d128-bfloat16	1.09×	0.00816	fla torch-compile torch	0.00886 0.00973 0.0419	0.516
gla-decode-serving-b1-h8-d128-bfloat16	0.81×	0.00742	torch-compile fla torch	0.00598 0.00672 0.0304	0.0709
gla-decode-serving-b8-h32-d128-bfloat16	1.10×	0.0159	fla torch-compile torch	0.0175 0.0211 0.0895	1.06
gla-decode-serving-b8-h48-d128-bfloat16	0.96×	0.0231	fla torch-compile torch	0.0223 0.0246 0.1208	1.09
gla-decode-serving-b8-h64-d128-bfloat16	0.89×	0.0304	fla torch-compile torch	0.0269 0.0320 0.1579	1.11

DeltaNetFwd (8 workloads)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
dn-b2-s16k-h4-d64-bfloat16	0.78×	0.4729	fla	0.3700	2.27
dn-b2-s16k-h4-d64-float16	0.77×	0.4726	fla	0.3661	2.27
dn-b2-s2k-h4-d64-bfloat16	0.99×	0.0629	fla	0.0624	2.13
dn-b2-s2k-h4-d64-float16	0.99×	0.0628	fla	0.0620	2.14
dn-b2-s4k-h4-d64-bfloat16	0.91×	0.1097	fla	0.0993	2.45
dn-b2-s4k-h4-d64-float16	0.90×	0.1095	fla	0.0985	2.45
dn-b2-s8k-h4-d64-bfloat16	0.81×	0.2346	fla	0.1910	2.29
dn-b2-s8k-h4-d64-float16	0.81×	0.2337	fla	0.1894	2.3

DeltaNetBwd (8 workloads)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
dn-bwd-b2-s16k-h4-d64-bfloat16	0.87×	1.0041	fla	0.8701	2.14
dn-bwd-b2-s16k-h4-d64-float16	0.87×	0.9924	fla	0.8642	2.16
dn-bwd-b2-s2k-h4-d64-bfloat16	0.87×	0.1316	fla	0.1145	2.04
dn-bwd-b2-s2k-h4-d64-float16	0.87×	0.1307	fla	0.1133	2.05
dn-bwd-b2-s4k-h4-d64-bfloat16	0.83×	0.2619	fla	0.2167	2.05
dn-bwd-b2-s4k-h4-d64-float16	0.83×	0.2595	fla	0.2144	2.07
dn-bwd-b2-s8k-h4-d64-bfloat16	0.85×	0.5107	fla	0.4362	2.1
dn-bwd-b2-s8k-h4-d64-float16	0.86×	0.5052	fla	0.4323	2.13

GatedDeltaNetPrefillBHTDFwd (4 workloads · ✅)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
bhtd-fallback-gdn-prefill-b1-s4k-h16-d128-bfloat16	0.78×	0.2526	fla	0.1979	34
bhtd-fallback-gdn-prefill-b1-s4k-h16-d128-float16	0.78×	0.2508	fla	0.1960	34.2
bhtd-qwen35-gdn-prefill-b1-s128k-h64-d128-bfloat16	0.89×	17.5479	fla	15.5794	62.7
bhtd-qwen35-gdn-prefill-b1-s128k-h64-d128-float16	0.90×	17.4198	fla	15.6294	63.1

GLABwd (8 workloads)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
gla-bwd-b2-s16k-h4-d64-bfloat16	0.75×	1.4487	fla	1.0809	1.48
gla-bwd-b2-s16k-h4-d64-float16	0.71×	1.5163	fla	1.0812	1.42
gla-bwd-b2-s2k-h4-d64-bfloat16	0.80×	0.1844	fla	0.1481	1.46
gla-bwd-b2-s2k-h4-d64-float16	0.81×	0.1829	fla	0.1482	1.47
gla-bwd-b2-s4k-h4-d64-bfloat16	0.79×	0.3648	fla	0.2879	1.47
gla-bwd-b2-s4k-h4-d64-float16	0.78×	0.3686	fla	0.2882	1.46
gla-bwd-b2-s8k-h4-d64-bfloat16	0.77×	0.7267	fla	0.5569	1.48
gla-bwd-b2-s8k-h4-d64-float16	0.75×	0.7446	fla	0.5576	1.44

GLAFwd (8 workloads)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
gla-init-b2-s16k-h4-d64-bfloat16	0.74×	0.6112	fla	0.4522	1.76
gla-init-b2-s16k-h4-d64-float16	0.75×	0.6180	fla	0.4652	1.74
gla-init-b2-s4k-h4-d64-bfloat16	0.76×	0.1564	fla	0.1191	1.72
gla-init-b2-s4k-h4-d64-float16	0.80×	0.1569	fla	0.1254	1.71
gla-noinit-b2-s2k-h4-d64-bfloat16	0.68×	0.0969	fla	0.0659	1.39
gla-noinit-b2-s2k-h4-d64-float16	0.71×	0.0986	fla	0.0702	1.36
gla-noinit-b2-s8k-h4-d64-bfloat16	0.71×	0.3114	fla	0.2200	1.72
gla-noinit-b2-s8k-h4-d64-float16	0.79×	0.3132	fla	0.2476	1.71

GatedDeltaNetBHTDFwd (8 workloads)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
gdn-bhtd-b2-s16k-h4-d64-bfloat16	0.62×	0.6370	fla	0.3936	13.5
gdn-bhtd-b2-s16k-h4-d64-float16	0.61×	0.6363	fla	0.3908	13.5
gdn-bhtd-b2-s2k-h4-d64-bfloat16	0.78×	0.0866	fla	0.0675	12.4
gdn-bhtd-b2-s2k-h4-d64-float16	0.78×	0.0866	fla	0.0672	12.4
gdn-bhtd-b2-s4k-h4-d64-bfloat16	0.75×	0.1445	fla	0.1082	14.9
gdn-bhtd-b2-s4k-h4-d64-float16	0.72×	0.1488	fla	0.1077	14.4
gdn-bhtd-b2-s8k-h4-d64-bfloat16	0.65×	0.3166	fla	0.2056	13.6
gdn-bhtd-b2-s8k-h4-d64-float16	0.65×	0.3138	fla	0.2051	13.7

GatedDeltaNetBwd (8 workloads)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
gdn-bwd-b2-s16k-h4-d64-bfloat16	0.65×	1.4680	fla	0.9536	1.46
gdn-bwd-b2-s16k-h4-d64-float16	0.65×	1.4237	fla	0.9204	1.51
gdn-bwd-b2-s2k-h4-d64-bfloat16	0.68×	0.2049	fla	0.1401	1.31
gdn-bwd-b2-s2k-h4-d64-float16	0.66×	0.2017	fla	0.1337	1.33
gdn-bwd-b2-s4k-h4-d64-bfloat16	0.67×	0.3875	fla	0.2584	1.39
gdn-bwd-b2-s4k-h4-d64-float16	0.66×	0.3809	fla	0.2495	1.41
gdn-bwd-b2-s8k-h4-d64-bfloat16	0.67×	0.7504	fla	0.5018	1.43
gdn-bwd-b2-s8k-h4-d64-float16	0.67×	0.7228	fla	0.4867	1.49

DaCumsumFwd (5 workloads · ✅)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
mamba2-1p3b-b8-s2k-bfloat16	0.42×	0.0154	mamba torch-compile torch-ref	0.00643 0.0114 0.0912	0.477
mamba2-1p3b-b8-s2k-dt-bias-bfloat16	0.43×	0.0148	mamba torch-compile torch-ref	0.00637 0.0115 0.0964	0.565
mamba2-2p7b-b2-s32k-dt-bias-float16	0.37×	0.0599	mamba torch-compile torch-ref	0.0223 0.0346 0.2369	0.7
mamba2-780m-b1-s4k-dt-bias-float16	0.81×	0.00429	mamba torch-compile torch-ref	0.00346 0.0048 0.0742	0.367
mamba2-780m-b1-s4k-float16	0.65×	0.00509	mamba torch-compile torch-ref	0.00333 0.00477 0.0713	0.27

EngramDecodeFwd (3 workloads · ✅)

Workload	Ratio	Device time	Alternatives		Throughput
	alt / ours	ms	name	ms	TFLOP/s
decode-b1-dmem512-d256-float16	0.40×	0.0330	torch-compile torch-ref	0.0133 0.0946	0.0161
decode-b4-dmem1024-d512-float16	0.31×	0.0827	torch-compile torch-ref	0.0256 0.1212	0.102
decode-b8-dmem512-d256-bfloat16	0.63×	0.0334	torch-compile torch-ref	0.0212 0.1114	0.127