void.1 / performance.json
appvoid's picture
Qwen90M step 930000; complete resumable checkpoint
596994f verified
Raw History Blame Contribute Delete
6.05 kB
{
"cache_key": {
"version": 1,
"kind": "gpu",
"gpu": {
"name": "NVIDIA RTX PRO 6000 Blackwell Server Edition",
"total_memory": 101975851008,
"capability": [
12,
0
]
},
"torch": "2.8.0+cu128",
"cuda": "12.8",
"python": "3.13.11",
"packages": {
"transformers": "4.57.1",
"numpy": "2.2.6",
"pyarrow": "21.0.0",
"huggingface_hub": "0.36.0"
},
"settings": {
"context": 2048,
"global_batch": 90,
"microbatch_candidates": [
8,
16,
32,
48,
64,
72,
76,
80,
86,
90
],
"batch_selection": "max_batch",
"gradient_checkpointing": false,
"compile_model": true,
"autotune_microbatch": true,
"attention_backend": "flash",
"vram_fraction": 0.98
},
"source": "d52752a64023cf29dc386d43835cf3173c5318ee093c3b375a812733c74023a1",
"requested_microbatch_ceiling": 90
},
"cache_reused": true,
"attention": "PyTorch FlashAttention (forced; forward/backward verified)",
"selected": {
"microbatch": 90,
"mode": "compiled",
"microbatch_seconds": 0.9440444560000287,
"peak_gb": 89.37790203094482,
"fits_budget": true,
"estimated_update_seconds": 0.9440444560000287,
"input_tokens_per_second": 195245.04257031978
},
"measurements": [
{
"microbatch": 8,
"mode": "eager",
"microbatch_seconds": 0.15785124100000303,
"peak_gb": 14.764304637908936,
"fits_budget": true,
"estimated_update_seconds": 1.7943123250000212,
"input_tokens_per_second": 103793.92582665654
},
{
"microbatch": 16,
"mode": "eager",
"microbatch_seconds": 0.3454984619999948,
"peak_gb": 28.084960460662842,
"fits_budget": true,
"estimated_update_seconds": 1.92592258199997,
"input_tokens_per_second": 94842.67979172798
},
{
"microbatch": 32,
"mode": "eager",
"microbatch_seconds": 0.7818738260000089,
"peak_gb": 54.827155113220215,
"fits_budget": true,
"estimated_update_seconds": 2.1863095480000254,
"input_tokens_per_second": 83819.15063620413
},
{
"microbatch": 48,
"mode": "eager",
"microbatch_seconds": 1.171907979999986,
"peak_gb": 81.63183498382568,
"fits_budget": true,
"estimated_update_seconds": 2.1970569614999818,
"input_tokens_per_second": 83883.71926608194
},
{
"microbatch": 8,
"mode": "compiled",
"microbatch_seconds": 0.0880895730000475,
"peak_gb": 9.113307476043701,
"fits_budget": true,
"estimated_update_seconds": 1.0067526180005189,
"input_tokens_per_second": 185992.50106469658
},
{
"microbatch": 16,
"mode": "compiled",
"microbatch_seconds": 0.1701795650000122,
"peak_gb": 16.87287473678589,
"fits_budget": true,
"estimated_update_seconds": 0.9559394930000451,
"input_tokens_per_second": 192549.55787434086
},
{
"microbatch": 32,
"mode": "compiled",
"microbatch_seconds": 0.34178767000003063,
"peak_gb": 32.49641752243042,
"fits_budget": true,
"estimated_update_seconds": 0.950241911000063,
"input_tokens_per_second": 191744.77534544803
},
{
"microbatch": 48,
"mode": "compiled",
"microbatch_seconds": 0.5061955770000282,
"peak_gb": 48.123157024383545,
"fits_budget": true,
"estimated_update_seconds": 0.9457143725000208,
"input_tokens_per_second": 194201.61784620755
},
{
"microbatch": 64,
"mode": "compiled",
"microbatch_seconds": 0.6595429280000076,
"peak_gb": 63.757601737976074,
"fits_budget": true,
"estimated_update_seconds": 0.9208904890000156,
"input_tokens_per_second": 198731.56762890573
},
{
"microbatch": 72,
"mode": "compiled",
"microbatch_seconds": 0.741346276999991,
"peak_gb": 71.65071392059326,
"fits_budget": true,
"estimated_update_seconds": 0.9243867199999727,
"input_tokens_per_second": 198903.00197730918
},
{
"microbatch": 76,
"mode": "compiled",
"microbatch_seconds": 0.7816785220000497,
"peak_gb": 75.5242338180542,
"fits_budget": true,
"estimated_update_seconds": 0.9251490005000278,
"input_tokens_per_second": 199120.22093398403
},
{
"microbatch": 80,
"mode": "compiled",
"microbatch_seconds": 0.8258502979999776,
"peak_gb": 79.40062236785889,
"fits_budget": true,
"estimated_update_seconds": 0.9283437199999867,
"input_tokens_per_second": 198389.46646478592
},
{
"microbatch": 86,
"mode": "compiled",
"microbatch_seconds": 0.8883623469999975,
"peak_gb": 85.4687147140503,
"fits_budget": true,
"estimated_update_seconds": 0.9419139469999607,
"input_tokens_per_second": 198261.44207347915
},
{
"microbatch": 90,
"mode": "compiled",
"microbatch_seconds": 0.9299377820000245,
"peak_gb": 89.3772840499878,
"fits_budget": true,
"estimated_update_seconds": 0.9299377820000245,
"input_tokens_per_second": 198206.80863571487
}
],
"failures": [
{
"microbatch": 64,
"mode": "eager",
"reason": "oom"
},
{
"microbatch": 72,
"mode": "eager",
"reason": "oom"
},
{
"microbatch": 76,
"mode": "eager",
"reason": "oom"
},
{
"microbatch": 80,
"mode": "eager",
"reason": "oom"
},
{
"microbatch": 86,
"mode": "eager",
"reason": "oom"
},
{
"microbatch": 90,
"mode": "eager",
"reason": "oom"
}
],
"selection_policy": "max_batch",
"budget_gb": 92.52170804977416,
"reserve_gb": 2.1791439056396484,
"note": "Forward/backward throughput; excludes optimizer, data, logging and checkpoint time.",
"legacy_seed_verified": false
}