Add new RCPs for v4.0

mlcommons · May 16, 2024 · 7849322 · 7849322
1 parent 267bfc4
commit 7849322
Show file tree

Hide file tree

Showing 4 changed files with 81 additions and 1 deletion.
diff --git a/mlperf_logging/rcp_checker/training_4.0.0/rcps_gnn.json b/mlperf_logging/rcp_checker/training_4.0.0/rcps_gnn.json
@@ -67,6 +67,24 @@
 	  1.20,1.20,1.15,1.25,1.20,1.15,
 	  1.10,1.15
     ]	  
+  },
+
+  "gnn_ref_262144":
+  {
+    "Benchmark": "gnn",
+    "Creator": "NVIDIA",
+    "When": "Reference RCPs before v4.0",
+    "Platform": "128xDGX-H100",
+    "BS": 262144,
+    "Hyperparams": {
+      "opt_base_learning_rate": 0.005
+    },
+    "Epochs to converge": [
+      2.40,2.55,2.35,2.45,2.50,2.35,
+      2.45,2.60,2.35,2.55,2.60,2.40,
+      2.40,2.30,2.30,2.45,2.60,2.50,
+      2.75,2.45
+    ]
   }
 }
 
diff --git a/mlperf_logging/rcp_checker/training_4.0.0/rcps_llama2_70b_lora.json b/mlperf_logging/rcp_checker/training_4.0.0/rcps_llama2_70b_lora.json
@@ -65,5 +65,27 @@
       5760,6528,6144,6528,5376,6528,5760,6144,6144,6528,
       6144,6144,6144,5760,5760,5760,5760,5760,6144,5760
     ]
-  }  
+  },
+  "llama2_70b_lora_ref_128":
+  {
+    "Benchmark": "llama2_70b_lora",
+    "Creator": "NVIDIA",
+    "When": "Prior to 4.0 submission",
+    "Platform": "TBD",
+    "BS": 128,
+    "Hyperparams": {
+      "opt_base_learning_rate": 1e-3,
+      "opt_max_grad_norm": 0.3,
+      "opt_learning_rate_warmup_epochs": 0,
+      "opt_learning_rate_decay_boundary_epochs": [],
+      "gradient_accumulation_steps": 1,
+      "lora_r": 16,
+      "lora_alpha": 32,
+      "max_steps": 1024
+    },
+    "samples to converge": [
+      11520,13056,10752,12672,12288,11136,10752,13056, 10752,9984,
+      11136,11136,11136,10752,11520,11136,11136,10752,11136,9984
+    ]
+  }
 }
diff --git a/mlperf_logging/rcp_checker/training_4.0.0/rcps_ssd.json b/mlperf_logging/rcp_checker/training_4.0.0/rcps_ssd.json
@@ -124,6 +124,24 @@
       8, 8, 8, 8, 8, 8, 8, 8, 8, 9]
   },
 
+  "ssd_ref_2560":
+  {
+    "Benchmark": "ssd",
+    "Creator": "NVIDIA",
+    "When": "Reference RCPs before v4.0",
+    "Platform": "20xDGX-H100",
+    "BS": 2560,
+    "Hyperparams": {
+      "opt_base_learning_rate": 0.000145,
+      "opt_learning_rate_warmup_factor": 1e-3,
+      "opt_learning_rate_warmup_epochs": 1,
+      "opt_weight_decay": 0
+    },
+    "Epochs to converge": [
+      8, 8, 8, 8, 8, 9, 9, 9, 9, 9,
+      9, 9, 9]
+  },
+
   "ssd_ref_4096":
   {
     "Benchmark": "ssd",

diff --git a/mlperf_logging/rcp_checker/training_4.0.0/rcps_stable_diffusion.json b/mlperf_logging/rcp_checker/training_4.0.0/rcps_stable_diffusion.json
@@ -1,5 +1,27 @@
 {
 
+  "sd_ref_384":
+  {
+    "Benchmark": "stable_diffusion",
+    "Creator": "NVIDIA",
+    "When": "Reference RCPs before v4.0",
+    "Platform": "16xDGX-H100",
+    "BS": 384,
+    "Hyperparams": {
+      "opt_adamw_beta_1": 0.9,
+      "opt_adamw_beta_2": 0.999,
+      "opt_adamw_epsilon": 1e-08,
+      "opt_adamw_weight_decay": 0.01,
+      "opt_base_learning_rate": 1.25e-7,
+      "opt_learning_rate_warmup_steps": 1000
+    },
+    "Epochs to converge": [
+      2049024, 2049024, 2049024, 2561280,
+      2561280, 2561280, 2561280, 2561280,
+      2561280, 2561280, 2561280, 2561280,
+      3073536, 3073536, 3073536]
+  },
+
   "sd_ref_512":
   {
     "Benchmark": "stable_diffusion",