liger-kernel-nightly 0.6.0.dev20250718050840__tar.gz → 0.6.0.dev20250719041120__tar.gz

{liger_kernel_nightly-0.6.0.dev20250718050840 → liger_kernel_nightly-0.6.0.dev20250719041120}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.0.dev20250718050840
+Version: 0.6.0.dev20250719041120
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.0.dev20250718050840 → liger_kernel_nightly-0.6.0.dev20250719041120}/benchmark/data/all_benchmark_data.csv

@@ -1493,3 +1493,75 @@ distill_cosine_loss,torch,full,memory,MB,BT,B x T,1024,7566.2822265625,7566.2822
 distill_cosine_loss,torch,full,memory,MB,BT,B x T,2048,11590.3134765625,11590.3134765625,11590.3134765625,"{""H"": 4096, ""V"": 128256, ""mode"": ""forward"", ""dtype"": ""torch.bfloat16"", ""bias"": false, ""weight_hard_loss"": 0.5, ""weight_soft_loss"": 0.5, ""ignore_index"": -100}",NVIDIA A100-SXM4-80GB,2025-06-27 09:23:28,0.5.10
 distill_cosine_loss,torch,full,memory,MB,BT,B x T,4096,19654.375,19654.375,19654.375,"{""H"": 4096, ""V"": 128256, ""mode"": ""forward"", ""dtype"": ""torch.bfloat16"", ""bias"": false, ""weight_hard_loss"": 0.5, ""weight_soft_loss"": 0.5, ""ignore_index"": -100}",NVIDIA A100-SXM4-80GB,2025-06-27 09:23:28,0.5.10
 distill_cosine_loss,torch,full,memory,MB,BT,B x T,8192,35782.5,35782.5,35782.5,"{""H"": 4096, ""V"": 128256, ""mode"": ""forward"", ""dtype"": ""torch.bfloat16"", ""bias"": false, ""weight_hard_loss"": 0.5, ""weight_soft_loss"": 0.5, ""ignore_index"": -100}",NVIDIA A100-SXM4-80GB,2025-06-27 09:23:28,0.5.10
+fused_add_rms_norm,liger_fused_add_rms_norm,forward,speed,ms,H,hidden size,1024,0.01759999990463257,0.017311999574303627,0.017920000478625298,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:20,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,forward,speed,ms,H,hidden size,2048,0.02924799919128418,0.028863999992609024,0.029983999207615852,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:20,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,forward,speed,ms,H,hidden size,4096,0.05129599943757057,0.050624001771211624,0.05209600180387497,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:20,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,forward,speed,ms,H,hidden size,8192,0.09344000369310379,0.09296000003814697,0.09382399916648865,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:20,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,forward,speed,ms,H,hidden size,16384,0.1791680008172989,0.17814399302005768,0.1796800047159195,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:20,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,forward,speed,ms,H,hidden size,32768,0.43830400705337524,0.43744000792503357,0.43929600715637207,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:20,0.6.0
+fused_add_rms_norm,huggingface,forward,speed,ms,H,hidden size,1024,0.060095999389886856,0.059808000922203064,0.06054399907588959,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:23,0.6.0
+fused_add_rms_norm,huggingface,forward,speed,ms,H,hidden size,2048,0.09084799885749817,0.09027200192213058,0.09161599725484848,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:23,0.6.0
+fused_add_rms_norm,huggingface,forward,speed,ms,H,hidden size,4096,0.17820799350738525,0.17744000256061554,0.17897599935531616,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:23,0.6.0
+fused_add_rms_norm,huggingface,forward,speed,ms,H,hidden size,8192,0.312608003616333,0.3118720054626465,0.31324800848960876,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:23,0.6.0
+fused_add_rms_norm,huggingface,forward,speed,ms,H,hidden size,16384,0.574944019317627,0.5740479826927185,0.5756288051605225,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:23,0.6.0
+fused_add_rms_norm,huggingface,forward,speed,ms,H,hidden size,32768,1.0943039655685425,1.0934272289276123,1.0951999425888062,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:23,0.6.0
+fused_add_rms_norm,liger_rms_norm,forward,speed,ms,H,hidden size,1024,0.0352960005402565,0.03481600061058998,0.03811199963092804,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:26,0.6.0
+fused_add_rms_norm,liger_rms_norm,forward,speed,ms,H,hidden size,2048,0.05430399999022484,0.05392000079154968,0.05503999814391136,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:26,0.6.0
+fused_add_rms_norm,liger_rms_norm,forward,speed,ms,H,hidden size,4096,0.10592000186443329,0.1054655984044075,0.10630399733781815,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:26,0.6.0
+fused_add_rms_norm,liger_rms_norm,forward,speed,ms,H,hidden size,8192,0.19679999351501465,0.19631999731063843,0.19724799692630768,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:26,0.6.0
+fused_add_rms_norm,liger_rms_norm,forward,speed,ms,H,hidden size,16384,0.37436801195144653,0.3733760118484497,0.3752320110797882,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:26,0.6.0
+fused_add_rms_norm,liger_rms_norm,forward,speed,ms,H,hidden size,32768,0.7376000285148621,0.7361343741416931,0.7391359806060791,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:26,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,full,speed,ms,H,hidden size,1024,0.3147200047969818,0.30796160697937014,0.32764801383018494,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:30,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,full,speed,ms,H,hidden size,2048,0.3089919984340668,0.30374398827552795,0.3226880133152008,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:30,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,full,speed,ms,H,hidden size,4096,0.30691200494766235,0.3023296058177948,0.3205504059791565,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:30,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,full,speed,ms,H,hidden size,8192,0.3246079981327057,0.3185984075069428,0.33656961321830753,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:30,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,full,speed,ms,H,hidden size,16384,0.6010559797286987,0.5996800065040588,0.6026239991188049,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:30,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,full,speed,ms,H,hidden size,32768,1.8402559757232666,1.8322880268096924,1.8461120128631592,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:30,0.6.0
+fused_add_rms_norm,huggingface,full,speed,ms,H,hidden size,1024,0.23878400027751923,0.23545600473880768,0.2507520020008087,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:33,0.6.0
+fused_add_rms_norm,huggingface,full,speed,ms,H,hidden size,2048,0.34513600170612335,0.34377598762512207,0.34678399562835693,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:33,0.6.0
+fused_add_rms_norm,huggingface,full,speed,ms,H,hidden size,4096,0.6330879926681519,0.631712019443512,0.6345599889755249,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:33,0.6.0
+fused_add_rms_norm,huggingface,full,speed,ms,H,hidden size,8192,1.1185599565505981,1.1172800064086914,1.1196800470352173,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:33,0.6.0
+fused_add_rms_norm,huggingface,full,speed,ms,H,hidden size,16384,2.0697600841522217,2.0678528785705566,2.0713536739349365,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:33,0.6.0
+fused_add_rms_norm,huggingface,full,speed,ms,H,hidden size,32768,3.9561920166015625,3.953824043273926,3.9581120014190674,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:33,0.6.0
+fused_add_rms_norm,liger_rms_norm,full,speed,ms,H,hidden size,1024,0.38916800916194916,0.3824320137500763,0.4037184059619903,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:36,0.6.0
+fused_add_rms_norm,liger_rms_norm,full,speed,ms,H,hidden size,2048,0.3890720009803772,0.38193280100822447,0.4032831907272339,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:36,0.6.0
+fused_add_rms_norm,liger_rms_norm,full,speed,ms,H,hidden size,4096,0.39715200662612915,0.3928639888763428,0.41097599267959595,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:36,0.6.0
+fused_add_rms_norm,liger_rms_norm,full,speed,ms,H,hidden size,8192,0.6275200247764587,0.6259520053863525,0.6287999749183655,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:36,0.6.0
+fused_add_rms_norm,liger_rms_norm,full,speed,ms,H,hidden size,16384,1.202239990234375,1.199679970741272,1.2048959732055664,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:36,0.6.0
+fused_add_rms_norm,liger_rms_norm,full,speed,ms,H,hidden size,32768,2.7738559246063232,2.7705343723297116,2.777868890762329,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:36,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,backward,speed,ms,H,hidden size,1024,0.15619200468063354,0.15376000106334686,0.1661248028278351,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:39,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,backward,speed,ms,H,hidden size,2048,0.15825600177049637,0.15600000321865082,0.16911999881267548,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:39,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,backward,speed,ms,H,hidden size,4096,0.16700799763202667,0.16502399742603302,0.1709440052509308,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:39,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,backward,speed,ms,H,hidden size,8192,0.1712000072002411,0.1700800061225891,0.17215999960899353,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:39,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,backward,speed,ms,H,hidden size,16384,0.42505601048469543,0.4233280122280121,0.42691200971603394,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:39,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,backward,speed,ms,H,hidden size,32768,1.4057759642601013,1.3944000005722046,1.4099839925765991,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:39,0.6.0
+fused_add_rms_norm,huggingface,backward,speed,ms,H,hidden size,1024,0.1520960032939911,0.15136000514030457,0.1528960019350052,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:42,0.6.0
+fused_add_rms_norm,huggingface,backward,speed,ms,H,hidden size,2048,0.2533760070800781,0.2524160146713257,0.25436800718307495,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:42,0.6.0
+fused_add_rms_norm,huggingface,backward,speed,ms,H,hidden size,4096,0.4551039934158325,0.4540799856185913,0.45612800121307373,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:42,0.6.0
+fused_add_rms_norm,huggingface,backward,speed,ms,H,hidden size,8192,0.8053439855575562,0.8038079738616943,0.806656002998352,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:42,0.6.0
+fused_add_rms_norm,huggingface,backward,speed,ms,H,hidden size,16384,1.4933120012283325,1.492095947265625,1.49452805519104,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:42,0.6.0
+fused_add_rms_norm,huggingface,backward,speed,ms,H,hidden size,32768,2.8600640296936035,2.8583295822143557,2.8612607955932616,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:42,0.6.0
+fused_add_rms_norm,liger_rms_norm,backward,speed,ms,H,hidden size,1024,0.20175999402999878,0.199072003364563,0.2154303938150406,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_rms_norm,backward,speed,ms,H,hidden size,2048,0.20263999700546265,0.20000000298023224,0.21675519943237304,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_rms_norm,backward,speed,ms,H,hidden size,4096,0.25276800990104675,0.2515519857406616,0.2539199888706207,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_rms_norm,backward,speed,ms,H,hidden size,8192,0.4322720021009445,0.43088001012802124,0.4336000084877014,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_rms_norm,backward,speed,ms,H,hidden size,16384,0.8288000226020813,0.8266303777694701,0.8311295866966247,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_rms_norm,backward,speed,ms,H,hidden size,32768,2.03987193107605,2.0360767364501955,2.0436416149139403,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,full,memory,MB,H,hidden size,1024,72.546875,72.546875,72.546875,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,full,memory,MB,H,hidden size,2048,145.0859375,145.0859375,145.0859375,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,full,memory,MB,H,hidden size,4096,290.1640625,290.1640625,290.1640625,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,full,memory,MB,H,hidden size,8192,580.3203125,580.3203125,580.3203125,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,full,memory,MB,H,hidden size,16384,1160.6328125,1160.6328125,1160.6328125,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_fused_add_rms_norm,full,memory,MB,H,hidden size,32768,2321.2578125,2321.2578125,2321.2578125,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,huggingface,full,memory,MB,H,hidden size,1024,104.03173828125,104.03173828125,104.03173828125,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,huggingface,full,memory,MB,H,hidden size,2048,208.05517578125,208.05517578125,208.05517578125,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,huggingface,full,memory,MB,H,hidden size,4096,416.10205078125,416.10205078125,416.10205078125,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,huggingface,full,memory,MB,H,hidden size,8192,832.19580078125,832.19580078125,832.19580078125,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,huggingface,full,memory,MB,H,hidden size,16384,1664.3125,1664.3125,1664.3125,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,huggingface,full,memory,MB,H,hidden size,32768,3328.625,3328.625,3328.625,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_rms_norm,full,memory,MB,H,hidden size,1024,104.03564453125,104.03564453125,104.03564453125,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_rms_norm,full,memory,MB,H,hidden size,2048,208.06298828125,208.06298828125,208.06298828125,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_rms_norm,full,memory,MB,H,hidden size,4096,416.11767578125,416.11767578125,416.11767578125,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_rms_norm,full,memory,MB,H,hidden size,8192,832.22705078125,832.22705078125,832.22705078125,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_rms_norm,full,memory,MB,H,hidden size,16384,1544.44580078125,1544.44580078125,1544.44580078125,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_add_rms_norm,liger_rms_norm,full,memory,MB,H,hidden size,32768,2960.8837890625,2960.8837890625,2960.8837890625,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0

liger_kernel_nightly-0.6.0.dev20250719041120/benchmark/scripts/benchmark_fused_add_rms_norm.py

@@ -0,0 +1,201 @@
+import torch
+import torch.nn as nn
+import triton
+
+from utils import QUANTILES
+from utils import SingleBenchmarkRunInput
+from utils import SingleBenchmarkRunOutput
+from utils import _test_memory
+from utils import parse_benchmark_script_args
+from utils import run_benchmarks
+
+from liger_kernel.transformers.fused_add_rms_norm import LigerFusedAddRMSNorm
+from liger_kernel.transformers.rms_norm import LigerRMSNorm
+from liger_kernel.utils import infer_device
+
+device = infer_device()
+
+
+class NaiveAddRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Naive implementation of the add residual rms norm.
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states, residual):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        residual = residual.to(torch.float32)
+        hidden_states = hidden_states + residual
+        residual = hidden_states
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype), residual.to(input_dtype)
+
+
+class AddLigerRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        AddLigerRMSNorm is equivalent to NaiveAddRMSNorm class above, but uses the LigerRMSNorm kernel.
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+        self.rms_norm = LigerRMSNorm(hidden_size, eps, in_place=False)
+
+    def forward(self, hidden_states, residual):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        residual = residual.to(torch.float32)
+        hidden_states = hidden_states + residual
+        residual = hidden_states
+        hidden_states = self.rms_norm(hidden_states)
+        return self.weight * hidden_states.to(input_dtype), residual.to(input_dtype)
+
+
+def bench_speed_fused_residual_rms_norm(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunOutput:
+    N = input.x
+    provider = input.kernel_provider
+    mode = input.kernel_operation_mode
+
+    extra_benchmark_config = input.extra_benchmark_config
+    M = extra_benchmark_config["M"]
+    eps = extra_benchmark_config["eps"]
+    dtype = extra_benchmark_config["dtype"]
+
+    x_shape = (M, N)
+
+    # Fused Add RMS Norm
+    fused_add_rms_norm = LigerFusedAddRMSNorm(hidden_size=N, eps=eps).to(device)
+    # Naive implementation
+    naive_rms_norm = NaiveAddRMSNorm(hidden_size=N, eps=eps).to(device)
+    # LigerRMSNorm without fused residual addition
+    liger_rms_norm = AddLigerRMSNorm(hidden_size=N, eps=eps).to(device)
+
+    x = torch.randn(x_shape, dtype=dtype, device=device)
+    r = torch.randn(x_shape, dtype=dtype, device=device)
+    dy = torch.randn_like(x)
+    ds = torch.randn_like(r)
+    x.requires_grad_(True)
+    r.requires_grad_(True)
+    # utility functions
+
+    def y_fwd():
+        if provider == "liger_fused_add_rms_norm":
+            return fused_add_rms_norm(x, r)
+
+        if provider == "huggingface":
+            return naive_rms_norm(x, r)
+
+        if provider == "liger_rms_norm":
+            return liger_rms_norm(x, r)
+
+    if mode == "forward":
+        ms_50, ms_20, ms_80 = triton.testing.do_bench(
+            y_fwd,
+            grad_to_none=[x, r],
+            rep=500,
+            quantiles=QUANTILES,
+        )
+    elif mode == "backward":
+        y, s = y_fwd()
+        ms_50, ms_20, ms_80 = triton.testing.do_bench(
+            lambda: (torch.autograd.backward((y, s), (dy, ds), retain_graph=True)),
+            grad_to_none=[x, r],
+            rep=500,
+            quantiles=QUANTILES,
+        )
+    elif mode == "full":
+
+        def full():
+            y, s = y_fwd()
+            torch.autograd.backward((y, s), (dy, ds))
+
+        ms_50, ms_20, ms_80 = triton.testing.do_bench(
+            full,
+            grad_to_none=[x, r],
+            rep=500,
+            quantiles=QUANTILES,
+        )
+
+    return SingleBenchmarkRunOutput(
+        y_20=ms_20,
+        y_50=ms_50,
+        y_80=ms_80,
+    )
+
+
+def bench_memory_fused_residual_rms_norm(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunOutput:
+    N = input.x
+    provider = input.kernel_provider
+
+    extra_benchmark_config = input.extra_benchmark_config
+    M = extra_benchmark_config["M"]
+    eps = extra_benchmark_config["eps"]
+    dtype = extra_benchmark_config["dtype"]
+
+    x_shape = (M, N)
+
+    fused_add_rms_norm = LigerFusedAddRMSNorm(hidden_size=N, eps=eps).to(device)
+    naive_rms_norm = NaiveAddRMSNorm(hidden_size=N, eps=eps).to(device)
+    liger_rms_norm = AddLigerRMSNorm(hidden_size=N, eps=eps).to(device)
+
+    x = torch.randn(x_shape, dtype=dtype, device=device)
+    r = torch.randn(x_shape, dtype=dtype, device=device)
+    dy = torch.randn_like(x)
+    ds = torch.randn_like(r)
+    x.requires_grad_(True)
+    r.requires_grad_(True)
+
+    # utility functions
+    def y_fwd():
+        if provider == "liger_fused_add_rms_norm":
+            return fused_add_rms_norm(x, r)
+        if provider == "huggingface":
+            return naive_rms_norm(x, r)
+        if provider == "liger_rms_norm":
+            return liger_rms_norm(x, r)
+
+    def full():
+        y, s = y_fwd()
+        torch.autograd.backward((y, s), (dy, ds))
+
+    mem_50, mem_20, mem_80 = _test_memory(full, quantiles=QUANTILES)
+
+    return SingleBenchmarkRunOutput(
+        y_20=mem_20,
+        y_50=mem_50,
+        y_80=mem_80,
+    )
+
+
+if __name__ == "__main__":
+    args = parse_benchmark_script_args()
+
+    common_configs = {
+        "kernel_name": "fused_add_rms_norm",
+        "x_name": "H",
+        "x_label": "hidden size",
+        "x_values": [2**i for i in range(10, 16)],
+        "kernel_providers": ["liger_fused_add_rms_norm", "huggingface", "liger_rms_norm"],
+        "extra_benchmark_configs": [{"M": 2048, "dtype": torch.float32, "eps": 1e-6}],
+        "overwrite": args.overwrite,
+    }
+
+    run_benchmarks(
+        bench_test_fn=bench_speed_fused_residual_rms_norm,
+        kernel_operation_modes=["forward", "full", "backward"],
+        metric_name="speed",
+        metric_unit="ms",
+        **common_configs,
+    )
+    run_benchmarks(
+        bench_test_fn=bench_memory_fused_residual_rms_norm,
+        kernel_operation_modes=["full"],
+        metric_name="memory",
+        metric_unit="MB",
+        **common_configs,
+    )

{liger_kernel_nightly-0.6.0.dev20250718050840 → liger_kernel_nightly-0.6.0.dev20250719041120}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.6.0.dev20250718050840"
+version = "0.6.0.dev20250719041120"
 description = "Efficient Triton kernels for LLM Training"
 urls = { "Homepage" = "https://github.com/linkedin/Liger-Kernel" }
 readme = { file = "README.md", content-type = "text/markdown" }