liger-kernel-nightly 0.6.0.dev20250718080702__tar.gz → 0.6.0.dev20250719041256__tar.gz

{liger_kernel_nightly-0.6.0.dev20250718080702 → liger_kernel_nightly-0.6.0.dev20250719041256}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.0.dev20250718080702
+Version: 0.6.0.dev20250719041256
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.0.dev20250718080702 → liger_kernel_nightly-0.6.0.dev20250719041256}/benchmark/data/all_benchmark_data.csv

@@ -625,36 +625,6 @@ group_norm,huggingface,backward,memory,MB,C,num_channels,256,320.5078125,320.507
 group_norm,huggingface,backward,memory,MB,C,num_channels,512,641.015625,641.015625,641.015625,"{""M"": 128, ""H"": 512, ""channels_per_group"": 4, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:20:53,0.3.1
 group_norm,huggingface,backward,memory,MB,C,num_channels,1024,1282.03125,1282.03125,1282.03125,"{""M"": 128, ""H"": 512, ""channels_per_group"": 4, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:20:53,0.3.1
 group_norm,huggingface,backward,memory,MB,C,num_channels,2048,2564.0625,2564.0625,2564.0625,"{""M"": 128, ""H"": 512, ""channels_per_group"": 4, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:20:53,0.3.1
-layer_norm,liger,forward,speed,ms,N,hidden size,1024,0.035840000957250595,0.03481600061058998,0.035840000957250595,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:27:51,0.3.1
-layer_norm,liger,forward,speed,ms,N,hidden size,2048,0.05939200147986412,0.058368001133203506,0.060416001826524734,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:27:51,0.3.1
-layer_norm,liger,forward,speed,ms,N,hidden size,4096,0.10751999914646149,0.10751999914646149,0.1085439994931221,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:27:51,0.3.1
-layer_norm,liger,forward,speed,ms,N,hidden size,8192,0.20582400262355804,0.20479999482631683,0.20684799551963806,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:27:51,0.3.1
-layer_norm,liger,forward,speed,ms,N,hidden size,16384,0.3993600010871887,0.3983359932899475,0.40140798687934875,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:27:51,0.3.1
-layer_norm,huggingface,forward,speed,ms,N,hidden size,1024,0.03788800165057182,0.03788800165057182,0.03891199827194214,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:27:53,0.3.1
-layer_norm,huggingface,forward,speed,ms,N,hidden size,2048,0.0655359998345375,0.0655359998345375,0.06656000018119812,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:27:53,0.3.1
-layer_norm,huggingface,forward,speed,ms,N,hidden size,4096,0.14745600521564484,0.14643199741840363,0.14847999811172485,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:27:53,0.3.1
-layer_norm,huggingface,forward,speed,ms,N,hidden size,8192,0.31334400177001953,0.3123199939727783,0.31436800956726074,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:27:53,0.3.1
-layer_norm,huggingface,forward,speed,ms,N,hidden size,16384,0.6133760213851929,0.6123520135879517,0.6154239773750305,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:27:53,0.3.1
-layer_norm,liger,full,speed,ms,N,hidden size,1024,0.6860799789428711,0.6146048903465271,0.7049216032028198,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:02,0.3.1
-layer_norm,liger,full,speed,ms,N,hidden size,2048,0.6789119839668274,0.6737920045852661,0.6912000179290771,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:02,0.3.1
-layer_norm,liger,full,speed,ms,N,hidden size,4096,0.6686720252037048,0.6635519862174988,0.681984007358551,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:02,0.3.1
-layer_norm,liger,full,speed,ms,N,hidden size,8192,0.6789119839668274,0.5908480286598206,0.6932479739189148,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:02,0.3.1
-layer_norm,liger,full,speed,ms,N,hidden size,16384,6.071296215057373,5.331148624420166,6.08235502243042,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:02,0.3.1
-layer_norm,huggingface,full,speed,ms,N,hidden size,1024,0.13312000036239624,0.13209599256515503,0.13312000036239624,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:04,0.3.1
-layer_norm,huggingface,full,speed,ms,N,hidden size,2048,0.23244799673557281,0.2303999960422516,0.23347200453281403,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:04,0.3.1
-layer_norm,huggingface,full,speed,ms,N,hidden size,4096,0.5242879986763,0.5232639908790588,0.5263360142707825,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:04,0.3.1
-layer_norm,huggingface,full,speed,ms,N,hidden size,8192,1.0168319940567017,1.0147839784622192,1.018880009651184,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:04,0.3.1
-layer_norm,huggingface,full,speed,ms,N,hidden size,16384,1.994752049446106,1.9916800260543823,1.9967999458312988,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:04,0.3.1
-layer_norm,liger,full,memory,MB,N,hidden size,1024,80.90625,80.90625,80.90625,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:04,0.3.1
-layer_norm,liger,full,memory,MB,N,hidden size,2048,161.78125,161.78125,161.78125,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:04,0.3.1
-layer_norm,liger,full,memory,MB,N,hidden size,4096,323.53125,323.53125,323.53125,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:04,0.3.1
-layer_norm,liger,full,memory,MB,N,hidden size,8192,647.03125,647.03125,647.03125,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:04,0.3.1
-layer_norm,liger,full,memory,MB,N,hidden size,16384,1294.03125,1294.03125,1294.03125,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:04,0.3.1
-layer_norm,huggingface,full,memory,MB,N,hidden size,1024,80.0625,80.0625,80.0625,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:05,0.3.1
-layer_norm,huggingface,full,memory,MB,N,hidden size,2048,160.09375,160.09375,160.09375,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:05,0.3.1
-layer_norm,huggingface,full,memory,MB,N,hidden size,4096,320.15625,320.15625,320.15625,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:05,0.3.1
-layer_norm,huggingface,full,memory,MB,N,hidden size,8192,640.28125,640.28125,640.28125,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:05,0.3.1
-layer_norm,huggingface,full,memory,MB,N,hidden size,16384,1280.53125,1280.53125,1280.53125,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:28:05,0.3.1
 fused_linear_orpo_loss,liger,forward,speed,ms,B,B,2,116.00621032714844,116.00621032714844,116.00621032714844,"{""T"": 4096, ""H"": 4096, ""V"": 128256, ""mode"": ""forward"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2024-11-13 21:24:05,0.4.0
 fused_linear_orpo_loss,liger,forward,speed,ms,B,B,4,230.83609008789062,230.83609008789062,230.83609008789062,"{""T"": 4096, ""H"": 4096, ""V"": 128256, ""mode"": ""forward"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2024-11-13 21:24:05,0.4.0
 fused_linear_orpo_loss,liger,forward,speed,ms,B,B,8,461.9543151855469,461.9543151855469,461.9543151855469,"{""T"": 4096, ""H"": 4096, ""V"": 128256, ""mode"": ""forward"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2024-11-13 21:24:05,0.4.0
@@ -1493,6 +1463,46 @@ 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
+layer_norm,liger,forward,speed,ms,N,hidden size,1024,0.018848000094294548,0.018400000408291817,0.020102400332689285,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:11,0.6.0
+layer_norm,liger,forward,speed,ms,N,hidden size,2048,0.029152000322937965,0.02876799926161766,0.029823999851942062,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:11,0.6.0
+layer_norm,liger,forward,speed,ms,N,hidden size,4096,0.05104000121355057,0.05036799982190132,0.05177599936723709,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:11,0.6.0
+layer_norm,liger,forward,speed,ms,N,hidden size,8192,0.0947519987821579,0.09436800330877304,0.09507200121879578,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:11,0.6.0
+layer_norm,liger,forward,speed,ms,N,hidden size,16384,0.18476800620555878,0.18396799266338348,0.1852159947156906,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:11,0.6.0
+layer_norm,huggingface,forward,speed,ms,N,hidden size,1024,0.023584000766277313,0.023423999547958374,0.023840000852942467,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:14,0.6.0
+layer_norm,huggingface,forward,speed,ms,N,hidden size,2048,0.03734400123357773,0.03702399879693985,0.037811201065778746,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:14,0.6.0
+layer_norm,huggingface,forward,speed,ms,N,hidden size,4096,0.06617599725723267,0.06560000032186508,0.06678400188684464,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:14,0.6.0
+layer_norm,huggingface,forward,speed,ms,N,hidden size,8192,0.15267199277877808,0.15190400183200836,0.15347200632095337,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:14,0.6.0
+layer_norm,huggingface,forward,speed,ms,N,hidden size,16384,0.3067840039730072,0.3046143889427185,0.3081152021884918,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:14,0.6.0
+layer_norm,liger,backward,speed,ms,N,hidden size,1024,0.12006399780511856,0.11653760075569153,0.12467200309038162,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:16,0.6.0
+layer_norm,liger,backward,speed,ms,N,hidden size,2048,0.1207360029220581,0.1176128014922142,0.1256511986255646,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:16,0.6.0
+layer_norm,liger,backward,speed,ms,N,hidden size,4096,0.16630400717258453,0.16412800550460815,0.16838400065898895,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:16,0.6.0
+layer_norm,liger,backward,speed,ms,N,hidden size,8192,0.31279999017715454,0.31116798520088196,0.3145279884338379,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:16,0.6.0
+layer_norm,liger,backward,speed,ms,N,hidden size,16384,0.5776320099830627,0.5753471970558167,0.5798912048339844,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:16,0.6.0
+layer_norm,huggingface,backward,speed,ms,N,hidden size,1024,0.0605119988322258,0.059647999703884125,0.061344001442193985,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:18,0.6.0
+layer_norm,huggingface,backward,speed,ms,N,hidden size,2048,0.09967999905347824,0.09849599748849869,0.10099200159311295,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:18,0.6.0
+layer_norm,huggingface,backward,speed,ms,N,hidden size,4096,0.17881600558757782,0.17795200645923615,0.17971199750900269,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:18,0.6.0
+layer_norm,huggingface,backward,speed,ms,N,hidden size,8192,0.33369600772857666,0.3328000009059906,0.33478400111198425,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:18,0.6.0
+layer_norm,huggingface,backward,speed,ms,N,hidden size,16384,0.6424000263214111,0.6412223815917969,0.643455982208252,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:18,0.6.0
+layer_norm,liger,full,speed,ms,N,hidden size,1024,0.26576000452041626,0.2629248082637787,0.2701759934425354,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:21,0.6.0
+layer_norm,liger,full,speed,ms,N,hidden size,2048,0.27427199482917786,0.26999040842056277,0.28091518878936766,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:21,0.6.0
+layer_norm,liger,full,speed,ms,N,hidden size,4096,0.27454400062561035,0.27004799246788025,0.2807359993457794,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:21,0.6.0
+layer_norm,liger,full,speed,ms,N,hidden size,8192,0.40556800365448,0.40403199195861816,0.40723198652267456,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:21,0.6.0
+layer_norm,liger,full,speed,ms,N,hidden size,16384,0.7608960270881653,0.7589311957359314,0.7631679773330688,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:21,0.6.0
+layer_norm,huggingface,full,speed,ms,N,hidden size,1024,0.08025600016117096,0.07942400127649307,0.08111999928951263,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:23,0.6.0
+layer_norm,huggingface,full,speed,ms,N,hidden size,2048,0.13315199315547943,0.13180799782276154,0.13468800485134125,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:23,0.6.0
+layer_norm,huggingface,full,speed,ms,N,hidden size,4096,0.2417600005865097,0.24089600145816803,0.24262399971485138,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:23,0.6.0
+layer_norm,huggingface,full,speed,ms,N,hidden size,8192,0.4832639992237091,0.48214399814605713,0.4843647956848145,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:23,0.6.0
+layer_norm,huggingface,full,speed,ms,N,hidden size,16384,0.950575977563858,0.9484800100326538,0.9528064012527466,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:23,0.6.0
+layer_norm,liger,full,memory,MB,N,hidden size,1024,80.0625,80.0625,80.0625,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:23,0.6.0
+layer_norm,liger,full,memory,MB,N,hidden size,2048,160.09375,160.09375,160.09375,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:23,0.6.0
+layer_norm,liger,full,memory,MB,N,hidden size,4096,320.15625,320.15625,320.15625,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:23,0.6.0
+layer_norm,liger,full,memory,MB,N,hidden size,8192,640.28125,640.28125,640.28125,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:23,0.6.0
+layer_norm,liger,full,memory,MB,N,hidden size,16384,1280.53125,1280.53125,1280.53125,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:23,0.6.0
+layer_norm,huggingface,full,memory,MB,N,hidden size,1024,80.0625,80.0625,80.0625,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:23,0.6.0
+layer_norm,huggingface,full,memory,MB,N,hidden size,2048,160.09375,160.09375,160.09375,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:23,0.6.0
+layer_norm,huggingface,full,memory,MB,N,hidden size,4096,320.15625,320.15625,320.15625,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:23,0.6.0
+layer_norm,huggingface,full,memory,MB,N,hidden size,8192,640.28125,640.28125,640.28125,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:23,0.6.0
+layer_norm,huggingface,full,memory,MB,N,hidden size,16384,1280.53125,1280.53125,1280.53125,"{""M"": 4096, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 80GB HBM3,2025-07-17 18:18:23,0.6.0
 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
@@ -1564,4 +1574,4 @@ fused_add_rms_norm,liger_rms_norm,full,memory,MB,H,hidden size,2048,208.06298828
 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
+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.dev20250718080702 → liger_kernel_nightly-0.6.0.dev20250719041256}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.6.0.dev20250718080702"
+version = "0.6.0.dev20250719041256"
 description = "Efficient Triton kernels for LLM Training"
 urls = { "Homepage" = "https://github.com/linkedin/Liger-Kernel" }
 readme = { file = "README.md", content-type = "text/markdown" }

{liger_kernel_nightly-0.6.0.dev20250718080702 → liger_kernel_nightly-0.6.0.dev20250719041256}/src/liger_kernel/ops/layer_norm.py

@@ -1,4 +1,3 @@
-import math
 import operator
 
 import torch
@@ -43,30 +42,45 @@ def _layer_norm_forward_kernel(
     https://arxiv.org/abs/1607.06450
     https://github.com/karpathy/llm.c/blob/master/doc/layernorm/layernorm.md
     """
-    row_idx = tl.program_id(0)
+    row_idx = tl.program_id(0).to(tl.int64)
     col_offsets = tl.arange(0, BLOCK_SIZE)
     mask = col_offsets < n_cols
 
-    Y_ptr += row_idx * Y_row_stride
-    X_ptr += row_idx * X_row_stride
-    Mean_ptr += row_idx * Mean_row_stride
-    RSTD_ptr += row_idx * RSTD_row_stride
-
-    X_row = tl.load(X_ptr + col_offsets, mask=mask, other=0)
-    W_row = tl.load(W_ptr + col_offsets, mask=mask, other=0)
-    B_row = tl.load(B_ptr + col_offsets, mask=mask, other=0)
-
-    mean = tl.sum(X_row, axis=0) / n_cols
-    Xmm = tl.where(mask, X_row - mean, 0)
-    var = tl.sum(Xmm * Xmm, axis=0) / n_cols
+    # Pre-load weights and bias in fp32 to avoid repeated conversions
+    W_row = tl.load(W_ptr + col_offsets, mask=mask, other=0.0)
+    B_row = tl.load(B_ptr + col_offsets, mask=mask, other=0.0)
+    W_f32 = W_row.to(tl.float32)
+    B_f32 = B_row.to(tl.float32)
+
+    # Calculate pointers for this row
+    row_X_ptr = X_ptr + row_idx * X_row_stride
+    row_Y_ptr = Y_ptr + row_idx * Y_row_stride
+    row_Mean_ptr = Mean_ptr + row_idx * Mean_row_stride
+    row_RSTD_ptr = RSTD_ptr + row_idx * RSTD_row_stride
+
+    # Load input data and convert to fp32 for numerical stability
+    X_row = tl.load(row_X_ptr + col_offsets, mask=mask, other=0.0)
+    X_f32 = X_row.to(tl.float32)
+
+    # Compute statistics in fp32 for numerical stability
+    n_cols_f32 = n_cols.to(tl.float32)
+    mean = tl.sum(X_f32, axis=0) / n_cols_f32
+    X_centered = X_f32 - mean
+    # Apply mask to variance calculation to exclude contributions from masked elements
+    X_centered_masked = tl.where(mask, X_centered, 0.0)
+    var = tl.sum(X_centered_masked * X_centered_masked, axis=0) / n_cols_f32
     rstd = rsqrt(var + eps)
 
-    tl.store(Mean_ptr, mean)
-    tl.store(RSTD_ptr, rstd)
+    # Store statistics (convert back to original dtype only once)
+    tl.store(row_Mean_ptr, mean.to(X_row.dtype))
+    tl.store(row_RSTD_ptr, rstd.to(X_row.dtype))
 
-    Y_row = Xmm * rstd * W_row + B_row
+    # Fused normalization and affine transformation
+    # Y = (X - mean) * rstd * W + B = X_centered * rstd * W + B
+    Y_f32 = X_centered * rstd * W_f32 + B_f32
 
-    tl.store(Y_ptr + col_offsets, Y_row, mask=mask)
+    # Store output (single conversion back to original dtype)
+    tl.store(row_Y_ptr + col_offsets, Y_f32.to(X_row.dtype), mask=mask)
 
 
 @triton.jit
@@ -81,73 +95,87 @@ def _layer_norm_backward_kernel(
     DY_ptr,  # pointer to output grad, shape (n_rows, n_cols)
     stride_x,  # stride of each row in input
     stride_dx,  # stride of each row in input grad
-    stride_dw,  # stride of each row in weights grad
-    stride_db,  # stride of each row in bias grad
     stride_dy,  # stride of each row in output grad
-    n_rows,
     n_cols,
-    rows_per_program: tl.constexpr,
     BLOCK_SIZE: tl.constexpr,
     dtype: tl.constexpr,
+    atomic_dtype: tl.constexpr,
 ):
     """
     References:
     https://arxiv.org/abs/1607.06450
     https://github.com/karpathy/llm.c/blob/master/doc/layernorm/layernorm.md
-    https://triton-lang.org/main/getting-started/tutorials/05-layer-norm.html
-    https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/ops/triton/layer_norm.py
     """
-    row_block_id = tl.program_id(0)
-    row_start = row_block_id * rows_per_program
-    row_end = min((row_block_id + 1) * rows_per_program, n_rows)
+    row_idx = tl.program_id(0).to(tl.int64)
     cols = tl.arange(0, BLOCK_SIZE)
     mask = cols < n_cols
 
-    dw_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
-    db_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
-
-    X_ptr += row_start * stride_x
-    Mean_ptr += row_start
-    RSTD_ptr += row_start
-    DX_ptr += row_start * stride_dx
-    DY_ptr += row_start * stride_dy
-
-    for _ in range(row_start, row_end):
-        x = tl.load(X_ptr + cols, mask=mask, other=0.0)
-        w = tl.load(W_ptr + cols, mask=mask, other=0.0)
-        dy = tl.load(DY_ptr + cols, mask=mask, other=0.0)
-        mean = tl.load(Mean_ptr)
-        rstd = tl.load(RSTD_ptr)
-
-        x_hat = (x - mean) * rstd
-        wdy = w * dy
-        c1 = tl.sum(x_hat * wdy, axis=0) / n_cols
-        c2 = tl.sum(wdy, axis=0) / n_cols
-        dx = (wdy - (x_hat * c1 + c2)) * rstd
-        tl.store(DX_ptr + cols, dx.to(dtype), mask=mask)
-
-        dw_row += dy * x_hat
-        db_row += dy
-
-        X_ptr += stride_x
-        Mean_ptr += 1
-        RSTD_ptr += 1
-        DX_ptr += stride_dx
-        DY_ptr += stride_dy
-
-    tl.store(DW_ptr + row_block_id * stride_dw + cols, dw_row.to(dtype), mask=mask)
-    tl.store(DB_ptr + row_block_id * stride_db + cols, db_row.to(dtype), mask=mask)
+    # Pre-load weights once (same optimization as forward pass)
+    w = tl.load(W_ptr + cols, mask=mask, other=0.0)
+    w_f32 = w.to(tl.float32)
+    n_cols_f32 = n_cols.to(tl.float32)
+
+    # Calculate pointers for this specific row
+    row_X_ptr = X_ptr + row_idx * stride_x
+    row_DX_ptr = DX_ptr + row_idx * stride_dx
+    row_DY_ptr = DY_ptr + row_idx * stride_dy
+    row_Mean_ptr = Mean_ptr + row_idx
+    row_RSTD_ptr = RSTD_ptr + row_idx
+
+    # Load data for this row
+    x = tl.load(row_X_ptr + cols, mask=mask, other=0.0)
+    dy = tl.load(row_DY_ptr + cols, mask=mask, other=0.0)
+    mean = tl.load(row_Mean_ptr)
+    rstd = tl.load(row_RSTD_ptr)
+
+    # Convert to fp32 for numerical stability
+    x_f32 = x.to(tl.float32)
+    dy_f32 = dy.to(tl.float32)
+    mean_f32 = mean.to(tl.float32)
+    rstd_f32 = rstd.to(tl.float32)
+
+    # Compute backward pass for this row
+    x_hat = (x_f32 - mean_f32) * rstd_f32
+    wdy = w_f32 * dy_f32
+    c1 = tl.sum(x_hat * wdy, axis=0) / n_cols_f32
+    c2 = tl.sum(wdy, axis=0) / n_cols_f32
+    dx = (wdy - (x_hat * c1 + c2)) * rstd_f32
+
+    # Store input gradient
+    tl.store(row_DX_ptr + cols, dx.to(dtype), mask=mask)
+
+    # Accumulate weight and bias gradients using atomic operations
+    dw = dy_f32 * x_hat
+    db = dy_f32
+    tl.atomic_add(DW_ptr + cols, dw.to(atomic_dtype), mask=mask)
+    tl.atomic_add(DB_ptr + cols, db.to(atomic_dtype), mask=mask)
 
 
 def layer_norm_forward(X, W, B, eps):
+    """
+    Args:
+        X: Input tensor of shape (..., hidden_size)
+        W: Weight tensor of shape (hidden_size,)
+        B: Bias tensor of shape (hidden_size,)
+        eps: Small constant for numerical stability
+
+    Returns:
+        Tuple of (output, input, mean, rstd, block_size, num_warps)
+    """
     shape = X.shape
     dim = shape[-1]
     X = X.view(-1, dim)
     n_rows, n_cols = X.shape
+
+    # Calculate optimal block size and warp configuration
     BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+
+    # Allocate output tensors
     Y = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
     Mean = torch.empty(n_rows, dtype=X.dtype, device=X.device)
     RSTD = torch.empty(n_rows, dtype=X.dtype, device=X.device)
+
+    # Validate input dimensions
     if X.shape[1] != W.shape[0]:
         raise ValueError(
             f"Incompatible dimensions: input feature size (X.shape[1]={X.shape[1]}) "
@@ -159,7 +187,9 @@ def layer_norm_forward(X, W, B, eps):
     if X.device.type == "xpu":
         kernel_args["grf_mode"] = "large"
 
-    _layer_norm_forward_kernel[(n_rows,)](
+    # Launch kernel with one thread block per row for optimal performance
+    grid = (n_rows,)
+    _layer_norm_forward_kernel[grid](
         Y,
         Y.stride(0),
         X,
@@ -176,35 +206,43 @@ def layer_norm_forward(X, W, B, eps):
         eps,
         BLOCK_SIZE=BLOCK_SIZE,
         num_warps=num_warps,
-        **kernel_args,  # XPU-specific optimization
+        **kernel_args,
     )
+
     return Y.view(*shape), X, Mean, RSTD, BLOCK_SIZE, num_warps
 
 
 def layer_norm_backward(dY, X, W, B, Mean, RSTD):
+    """
+    Args:
+        dY: Gradient of output
+        X: Input tensor
+        W: Weight tensor
+        B: Bias tensor
+        Mean: Pre-computed mean
+        RSTD: Pre-computed reciprocal standard deviation
+
+    Returns:
+        Tuple of (input_grad, weight_grad, bias_grad)
+    """
     shape = dY.shape
     dim = shape[-1]
     dY = dY.view(-1, dim)
     n_rows, n_cols = dY.shape
 
-    sm_count = 1
-    if X.device.type == "cuda":
-        sm_count = torch.cuda.get_device_properties(X.device).multi_processor_count
-    elif X.device.type == "xpu":
-        sm_count = torch.xpu.get_device_properties(X.device).gpu_eu_count
-
+    # Allocate gradient tensors
     DX = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
-    _DW = torch.empty((sm_count, n_cols), dtype=W.dtype, device=W.device)
-    _DB = torch.empty((sm_count, n_cols), dtype=W.dtype, device=W.device)
+    # Use float32 for weight/bias gradients if bfloat16 (due to atomic_add limitation)
+    grad_dtype = torch.float32 if W.dtype == torch.bfloat16 else W.dtype
+    DW = torch.zeros(n_cols, dtype=grad_dtype, device=W.device)
+    DB = torch.zeros(n_cols, dtype=grad_dtype, device=W.device)
 
+    # Calculate optimal block size and warp configuration
     BLOCK_SIZE, num_warps = calculate_settings(n_cols)
     if n_cols > BLOCK_SIZE:
-        raise RuntimeError(
-            f"Feature dimension {n_cols} exceeds maximum supported size of {BLOCK_SIZE}. Consider using a smaller feature dimension."
-        )
+        raise RuntimeError(f"Feature dimension {n_cols} exceeds maximum supported size of {BLOCK_SIZE}.")
 
-    rows_per_program = math.ceil(n_rows / sm_count)
-    grid = (sm_count,)
+    # Determine dtype for triton operations
     triton_dtype = (
         tl.float32
         if X.dtype == torch.float32
@@ -212,41 +250,41 @@ def layer_norm_backward(dY, X, W, B, Mean, RSTD):
         if X.dtype == torch.bfloat16
         else tl.float16
         if X.dtype == torch.float16
-        else tl.float32  # fallback to float32 for other types
+        else tl.float32  # fallback
     )
 
+    # Use float32 for atomic operations if bfloat16 is not supported
+    atomic_dtype = tl.float32 if triton_dtype == tl.bfloat16 else triton_dtype
+
     # XPU-specific optimization
     kernel_args = {}
     if X.device.type == "xpu":
         kernel_args.update({"grf_mode": "large", "num_warps": 32, "num_stages": 4})
 
+    # Launch kernel with one thread block per row for optimal performance
+    grid = (n_rows,)
     _layer_norm_backward_kernel[grid](
         X,
         W,
         Mean,
         RSTD,
         DX,
-        _DW,
-        _DB,
+        DW,
+        DB,
         dY,
         X.stride(0),
         DX.stride(0),
-        _DW.stride(0),
-        _DB.stride(0),
         dY.stride(0),
-        n_rows,
         n_cols,
-        rows_per_program,
         BLOCK_SIZE=BLOCK_SIZE,
         dtype=triton_dtype,
-        **kernel_args,  # XPU-specific optimization
+        atomic_dtype=atomic_dtype,
+        num_warps=num_warps,
+        **kernel_args,
     )
 
-    DW = _DW.sum(dim=0).to(W.dtype)
-    DB = _DB.sum(dim=0).to(W.dtype)
-
     DX = DX.view(*shape)
-    return DX, DW, DB
+    return DX, DW.to(W.dtype), DB.to(W.dtype)
 
 
 class LigerLayerNormFunction(torch.autograd.Function):

{liger_kernel_nightly-0.6.0.dev20250718080702 → liger_kernel_nightly-0.6.0.dev20250719041256}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.0.dev20250718080702
+Version: 0.6.0.dev20250719041256
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.0.dev20250718080702 → liger_kernel_nightly-0.6.0.dev20250719041256}/test/transformers/test_layer_norm.py

@@ -16,12 +16,14 @@ device = infer_device()
         (4, 16, 128),
         (1, 1, 1023),  # Minimal batch/seq with near power-of-2 hidden
         (3, 7, 256),  # Prime numbers for batch/seq
+        (1, 1, 1500),
     ],
 )
 @pytest.mark.parametrize(
     "dtype, atol, rtol",
     [
         (torch.float32, 1e-5, 1e-5),
+        (torch.bfloat16, 2e-2, 2e-2),  # Relaxed tolerance for bfloat16 due to lower precision + atomic limitations
     ],
 )
 def test_liger_layer_norm(
@@ -72,6 +74,7 @@ def test_liger_layer_norm(
     "dtype, atol, rtol",
     [
         (torch.float32, 1e-5, 1e-5),
+        (torch.bfloat16, 2e-2, 2e-2),  # Relaxed tolerance for bfloat16 due to lower precision + atomic limitations
     ],
 )
 def test_liger_layer_norm_functional(

{liger_kernel_nightly-0.6.0.dev20250718080702 → liger_kernel_nightly-0.6.0.dev20250719041256}/test/transformers/test_monkey_patch.py

@@ -1634,6 +1634,7 @@ def test_apply_liger_kernel_to_instance_for_glm4():
             pytest.fail(f"An exception occured in extra_expr: {type(e).__name__} - {e}")
 
 
+@pytest.mark.skipif(not is_smollm3_available(), reason="smollm3 module not available")
 def test_apply_liger_kernel_to_instance_for_smollm3():
     # Ensure any monkey patching is cleaned up for subsequent tests
     with patch("transformers.models.smollm3.modeling_smollm3"):