backfill batch 13 (400)

eval/top3_latest_test/logs/pps4_lam0.6_temp0.1_gpu5.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.1
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 128)
+  sampled 128/128 (steps this call: 128)
+Average denoising rounds per sample: 128.00
+First sample preview: ' said.\n\n"The the process is going to be there,” Lehman said.\n\nIn the first phase of the new plan, the City Council will '
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 0.6509510393194246,
+  "median_nll": 0.03751843422651291,
+  "ppl": 1.9173634502150882,
+  "acc": 0.8601345708387829,
+  "tokens": 63461,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 128.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.1
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.1_gpu5.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.1_gpu5.json
+Diversity metrics for pps4_lam0.6_temp0.1_gpu5.json:
+  distinct_1: 0.0353
+  repetition_1: 0.9000
+  distinct_2: 0.0966
+  repetition_2: 0.8364
+  distinct_4: 0.2286
+  repetition_4: 0.7112
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.1_gpu5.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.1_gpu6.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.1
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 128)
+  sampled 128/128 (steps this call: 128)
+Average denoising rounds per sample: 128.00
+First sample preview: ' FBI to reopen the case.\n\nBut he said, that is not the case.\n\nThe report said, which was released Tuesday, that the FBI '
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 0.6457859240297437,
+  "median_nll": 0.031859297305345535,
+  "ppl": 1.9074855790746221,
+  "acc": 0.868684457599338,
+  "tokens": 62841,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 128.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.1
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.1_gpu6.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.1_gpu6.json
+Diversity metrics for pps4_lam0.6_temp0.1_gpu6.json:
+  distinct_1: 0.0377
+  repetition_1: 0.9133
+  distinct_2: 0.1010
+  repetition_2: 0.8568
+  distinct_4: 0.2161
+  repetition_4: 0.7502
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.1_gpu6.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.1_gpu7.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.1
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 128)
+  sampled 128/128 (steps this call: 128)
+Average denoising rounds per sample: 128.00
+First sample preview: '\n\n\nof the\n\nof the\n\n\n\n\n\n of the\n\nand of the\n\n\n\n\n\n\n\n\n\nof the\n\nof\n\nof\n\n\n\nof\n\nof\n\nof\n\nof\n\nof\n\nof\n\nof\n\nof\n\nof\n\nof\n\nof\n\nof the'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 0.6516889086104449,
+  "median_nll": 0.0424233041703701,
+  "ppl": 1.9187787359084396,
+  "acc": 0.8566443038579828,
+  "tokens": 62753,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 128.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.1
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.1_gpu7.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.1_gpu7.json
+Diversity metrics for pps4_lam0.6_temp0.1_gpu7.json:
+  distinct_1: 0.0319
+  repetition_1: 0.8956
+  distinct_2: 0.0932
+  repetition_2: 0.8163
+  distinct_4: 0.2333
+  repetition_4: 0.6856
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.1_gpu7.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.4_gpu0.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.4
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 131)
+  sampled 128/128 (steps this call: 134)
+Average denoising rounds per sample: 132.50
+First sample preview: 'BBR Inc., Inc.BBR, Inc. is.\n\n(USA, Canada, Australia, New York, USA)\n\nThe company BBR Inc. , Inc.\n\nBBR Inc. is.\n\nBBR Inc'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 1.1519089559516735,
+  "median_nll": 0.17637917399406433,
+  "ppl": 3.1642275189102804,
+  "acc": 0.7490172314672384,
+  "tokens": 64359,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 132.5,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.4
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.4_gpu0.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.4_gpu0.json
+Diversity metrics for pps4_lam0.6_temp0.4_gpu0.json:
+  distinct_1: 0.0671
+  repetition_1: 0.8292
+  distinct_2: 0.2072
+  repetition_2: 0.6862
+  distinct_4: 0.4266
+  repetition_4: 0.4948
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.4_gpu0.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.4_gpu1.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.4
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 131)
+  sampled 128/128 (steps this call: 129)
+Average denoising rounds per sample: 130.00
+First sample preview: 'Dear friends,\n\n\nsum’ to been a great year to us, the Reddit community, and the the community of Reddit it.\n\nWe’ve had a '
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 1.1469716861952688,
+  "median_nll": 0.14914770424365997,
+  "ppl": 3.1486433772917226,
+  "acc": 0.7549979663986485,
+  "tokens": 63926,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 130.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.4
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.4_gpu1.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.4_gpu1.json
+Diversity metrics for pps4_lam0.6_temp0.4_gpu1.json:
+  distinct_1: 0.0805
+  repetition_1: 0.7894
+  distinct_2: 0.2360
+  repetition_2: 0.6363
+  distinct_4: 0.4586
+  repetition_4: 0.4540
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.4_gpu1.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.4_gpu2.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.4
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 136)
+  sampled 128/128 (steps this call: 132)
+Average denoising rounds per sample: 134.00
+First sample preview: ' the some of the tutorial on how to make the the..\n\nThe design of a the. is not the same as the design of a the. .\n\nThe '
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 1.147666217670304,
+  "median_nll": 0.1741468757390976,
+  "ppl": 3.1508309688085583,
+  "acc": 0.7498316340114959,
+  "tokens": 63849,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 134.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.4
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.4_gpu2.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.4_gpu2.json
+Diversity metrics for pps4_lam0.6_temp0.4_gpu2.json:
+  distinct_1: 0.0745
+  repetition_1: 0.8154
+  distinct_2: 0.2242
+  repetition_2: 0.6672
+  distinct_4: 0.4587
+  repetition_4: 0.4786
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.4_gpu2.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.4_gpu3.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.4
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 133)
+  sampled 128/128 (steps this call: 132)
+Average denoising rounds per sample: 132.50
+First sample preview: 'U.S.S. is the best in the world, but not at the best of the world, to say according to the BBC. In a report released Tue'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 1.14529267073261,
+  "median_nll": 0.17469988763332367,
+  "ppl": 3.1433611920511138,
+  "acc": 0.7467620517136578,
+  "tokens": 64161,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 132.5,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.4
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.4_gpu3.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.4_gpu3.json
+Diversity metrics for pps4_lam0.6_temp0.4_gpu3.json:
+  distinct_1: 0.0750
+  repetition_1: 0.8150
+  distinct_2: 0.2269
+  repetition_2: 0.6603
+  distinct_4: 0.4797
+  repetition_4: 0.4461
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.4_gpu3.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.4_gpu4.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.4
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 133)
+  sampled 128/128 (steps this call: 132)
+Average denoising rounds per sample: 132.50
+First sample preview: '\n.\n\n,\n\n\n.\n\n.\n\n.\n\n\n.\n\n\n.\n\n.\n\n.\n\n\n.\n\n\n.\n\n.\n\n\n.\n\n.\n\n.\n\n.\n\n.\n\n\n\n.\n\n.\n\n.\n\n.\n\n. .\n\n.\n\n.\n\n.\n\n. .\n\n.\n\n.\n\n.\n\n.\n\n. .\n\n.\n\n.\n\n\n.\n\n.\n'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 1.0779396125068843,
+  "median_nll": 0.14342989027500153,
+  "ppl": 2.938618616228597,
+  "acc": 0.7666852834403798,
+  "tokens": 64458,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 132.5,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.4
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.4_gpu4.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.4_gpu4.json
+Diversity metrics for pps4_lam0.6_temp0.4_gpu4.json:
+  distinct_1: 0.0702
+  repetition_1: 0.8305
+  distinct_2: 0.2145
+  repetition_2: 0.6865
+  distinct_4: 0.4504
+  repetition_4: 0.4812
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.4_gpu4.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.4_gpu5.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.4
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 133)
+  sampled 128/128 (steps this call: 133)
+Average denoising rounds per sample: 133.00
+First sample preview: ' said.\n\n"The.C.R. is out of the game," he said.\n\nBut a company close to the company said the company is committed to wor'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 1.1531354440872796,
+  "median_nll": 0.16500601172447205,
+  "ppl": 3.16811078732511,
+  "acc": 0.7510950384237681,
+  "tokens": 64153,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 133.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.4
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.4_gpu5.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.4_gpu5.json
+Diversity metrics for pps4_lam0.6_temp0.4_gpu5.json:
+  distinct_1: 0.0797
+  repetition_1: 0.8124
+  distinct_2: 0.2423
+  repetition_2: 0.6581
+  distinct_4: 0.4821
+  repetition_4: 0.4636
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.4_gpu5.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.4_gpu6.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.4
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 130)
+  sampled 128/128 (steps this call: 131)
+Average denoising rounds per sample: 130.50
+First sample preview: ' was released for the North version of the game last year, following the release of Wii U.\n\nThere is a new version, howe'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 1.1192875892864118,
+  "median_nll": 0.16515083611011505,
+  "ppl": 3.062671545889563,
+  "acc": 0.7531603043892683,
+  "tokens": 63997,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 130.5,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.4
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.4_gpu6.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.4_gpu6.json
+Diversity metrics for pps4_lam0.6_temp0.4_gpu6.json:
+  distinct_1: 0.0701
+  repetition_1: 0.8150
+  distinct_2: 0.2151
+  repetition_2: 0.6697
+  distinct_4: 0.4521
+  repetition_4: 0.4745
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.4_gpu6.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.4_gpu7.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.4
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 132)
+  sampled 128/128 (steps this call: 132)
+Average denoising rounds per sample: 132.00
+First sample preview: ',” in the decision to appeal the decision.\n\nThe court gave the aisies the two days to make the decision on the case, the'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 1.1630522965856211,
+  "median_nll": 0.20357447117567062,
+  "ppl": 3.1996847732230664,
+  "acc": 0.7394515756269715,
+  "tokens": 64038,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 132.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.4
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.4_gpu7.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.4_gpu7.json
+Diversity metrics for pps4_lam0.6_temp0.4_gpu7.json:
+  distinct_1: 0.0792
+  repetition_1: 0.7923
+  distinct_2: 0.2420
+  repetition_2: 0.6238
+  distinct_4: 0.4965
+  repetition_4: 0.4170
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.4_gpu7.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.7_gpu0.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.7
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 180)
+  sampled 128/128 (steps this call: 183)
+Average denoising rounds per sample: 181.50
+First sample preview: ' and I’ve seen the victims of and the victims murdered in the Los Angeles area.” So what else…\n\nThe next day we responde'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 2.942398785483946,
+  "median_nll": 2.017829656600952,
+  "ppl": 18.961275835862214,
+  "acc": 0.44926435249065905,
+  "tokens": 64501,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 181.5,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.7
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.7_gpu0.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.7_gpu0.json
+Diversity metrics for pps4_lam0.6_temp0.7_gpu0.json:
+  distinct_1: 0.1963
+  repetition_1: 0.5336
+  distinct_2: 0.6266
+  repetition_2: 0.2041
+  distinct_4: 0.9265
+  repetition_4: 0.0655
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.7_gpu0.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.7_gpu1.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.7
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 183)
+  sampled 128/128 (steps this call: 174)
+Average denoising rounds per sample: 178.50
+First sample preview: 'Google Flipboard\n\nTwitter\n\nGoogle\n\n\nLinkedIn\n\nReddit Plus\n\ngoogle\n\n\n\nFacebook\n\n\n\nEmail\n\nGuest Note:\n\nThanks, Laura Lee, '
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 2.9454587164628476,
+  "median_nll": 2.053232431411743,
+  "ppl": 19.01938489069995,
+  "acc": 0.44612356241668993,
+  "tokens": 64518,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 178.5,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.7
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.7_gpu1.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.7_gpu1.json
+Diversity metrics for pps4_lam0.6_temp0.7_gpu1.json:
+  distinct_1: 0.1902
+  repetition_1: 0.5426
+  distinct_2: 0.6052
+  repetition_2: 0.2125
+  distinct_4: 0.9222
+  repetition_4: 0.0645
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.7_gpu1.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.7_gpu2.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.7
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 175)
+  sampled 128/128 (steps this call: 170)
+Average denoising rounds per sample: 172.50
+First sample preview: ' the trip to the USA by the NNVE backing group to create it. The story was the template for NNVE.\n\nAnd the mysterious be'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 2.7603239347687127,
+  "median_nll": 1.7485754489898682,
+  "ppl": 15.804961895789319,
+  "acc": 0.48433919022154315,
+  "tokens": 64141,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 172.5,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.7
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.7_gpu2.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.7_gpu2.json
+Diversity metrics for pps4_lam0.6_temp0.7_gpu2.json:
+  distinct_1: 0.1953
+  repetition_1: 0.5519
+  distinct_2: 0.6067
+  repetition_2: 0.2377
+  distinct_4: 0.9070
+  repetition_4: 0.0826
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.7_gpu2.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.7_gpu3.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.7
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 173)
+  sampled 128/128 (steps this call: 175)
+Average denoising rounds per sample: 174.00
+First sample preview: 'If you are staying in hotel on Monday in UAMA, you have until at least 10:13 am Sunday on March 18 to get there. If a pl'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 2.8654510964581106,
+  "median_nll": 1.9266316890716553,
+  "ppl": 17.55697130610988,
+  "acc": 0.46492573145609717,
+  "tokens": 64563,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 174.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.7
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.7_gpu3.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.7_gpu3.json
+Diversity metrics for pps4_lam0.6_temp0.7_gpu3.json:
+  distinct_1: 0.1934
+  repetition_1: 0.5501
+  distinct_2: 0.6095
+  repetition_2: 0.2326
+  distinct_4: 0.9177
+  repetition_4: 0.0795
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.7_gpu3.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.7_gpu4.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.7
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 186)
+  sampled 128/128 (steps this call: 177)
+Average denoising rounds per sample: 181.50
+First sample preview: ' and, the electrician wanted to keep his eyes on the a Tesla. For a moment, if he was ever the same side with his driver'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 2.929448415530568,
+  "median_nll": 2.029412031173706,
+  "ppl": 18.717303474915926,
+  "acc": 0.44909500504932803,
+  "tokens": 64365,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 181.5,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.7
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.7_gpu4.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.7_gpu4.json
+Diversity metrics for pps4_lam0.6_temp0.7_gpu4.json:
+  distinct_1: 0.2037
+  repetition_1: 0.5318
+  distinct_2: 0.6382
+  repetition_2: 0.2019
+  distinct_4: 0.9355
+  repetition_4: 0.0559
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.7_gpu4.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.7_gpu5.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.7
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 174)
+  sampled 128/128 (steps this call: 176)
+Average denoising rounds per sample: 175.00
+First sample preview: 'Dell Vogts and Schmitz flew a three-year-old space shuttle, in orbit, but the company—fueled up to the and now active on'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 2.893385421106573,
+  "median_nll": 1.985037088394165,
+  "ppl": 18.05432775945453,
+  "acc": 0.45797362017390225,
+  "tokens": 64519,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 175.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.7
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.7_gpu5.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.7_gpu5.json
+Diversity metrics for pps4_lam0.6_temp0.7_gpu5.json:
+  distinct_1: 0.1917
+  repetition_1: 0.5309
+  distinct_2: 0.6122
+  repetition_2: 0.2127
+  distinct_4: 0.9240
+  repetition_4: 0.0702
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.7_gpu5.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.7_gpu6.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.7
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 174)
+  sampled 128/128 (steps this call: 180)
+Average denoising rounds per sample: 177.00
+First sample preview: '.”\n\nagainst\n\nthe hand, though, a bayonet is used to throw a punch. I had a claw handle and was\nduc using it,\n\nand he swu'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 2.8696047275553243,
+  "median_nll": 1.9438047409057617,
+  "ppl": 17.630048150161716,
+  "acc": 0.4609834731438588,
+  "tokens": 63896,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 177.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.7
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.7_gpu6.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.7_gpu6.json
+Diversity metrics for pps4_lam0.6_temp0.7_gpu6.json:
+  distinct_1: 0.1961
+  repetition_1: 0.5396
+  distinct_2: 0.6178
+  repetition_2: 0.2154
+  distinct_4: 0.9188
+  repetition_4: 0.0717
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.7_gpu6.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.7_gpu7.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.7
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 179)
+  sampled 128/128 (steps this call: 177)
+Average denoising rounds per sample: 178.00
+First sample preview: '“Not always in the past,” said Evron Morales.\n\nA quote from two-quite-assed-people was long ago an unspoken part of the '
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 2.9287707690608333,
+  "median_nll": 2.006821036338806,
+  "ppl": 18.704624056859846,
+  "acc": 0.4540028525362768,
+  "tokens": 64504,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 178.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.7
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.7_gpu7.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.7_gpu7.json
+Diversity metrics for pps4_lam0.6_temp0.7_gpu7.json:
+  distinct_1: 0.1917
+  repetition_1: 0.5379
+  distinct_2: 0.6068
+  repetition_2: 0.2181
+  distinct_4: 0.9081
+  repetition_4: 0.0819
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.7_gpu7.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.8_gpu0.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 210)
+  sampled 128/128 (steps this call: 201)
+Average denoising rounds per sample: 205.50
+First sample preview: ' and I’ve had three victims dead four more victims murdered in the Los Angeles area.” So what else might there go on Ear'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 3.9368979006805738,
+  "median_nll": 3.2145655155181885,
+  "ppl": 51.25934283792319,
+  "acc": 0.32915863345415286,
+  "tokens": 64835,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 205.5,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.8_gpu0.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.8_gpu0.json
+Diversity metrics for pps4_lam0.6_temp0.8_gpu0.json:
+  distinct_1: 0.2465
+  repetition_1: 0.4214
+  distinct_2: 0.7338
+  repetition_2: 0.0834
+  distinct_4: 0.9898
+  repetition_4: 0.0081
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.8_gpu0.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.8_gpu1.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 206)
+  sampled 128/128 (steps this call: 204)
+Average denoising rounds per sample: 205.00
+First sample preview: 'Twenty-terrorist were killed during the clashes- the operation in Messed where eight ISIS sympathisers were being detain'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 3.9516018208961854,
+  "median_nll": 3.2193745374679565,
+  "ppl": 52.018624654934406,
+  "acc": 0.3302834533708386,
+  "tokens": 64702,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 205.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.8_gpu1.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.8_gpu1.json
+Diversity metrics for pps4_lam0.6_temp0.8_gpu1.json:
+  distinct_1: 0.2552
+  repetition_1: 0.4202
+  distinct_2: 0.7408
+  repetition_2: 0.0838
+  distinct_4: 0.9866
+  repetition_4: 0.0113
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.8_gpu1.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.8_gpu2.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 209)
+  sampled 128/128 (steps this call: 208)
+Average denoising rounds per sample: 208.50
+First sample preview: 'Im a big fan of this (BEVE specifically) myself.\n\nBELL is an engine for 3D 8 (No obvious). I used Spiv 2 models (I are 1'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 3.9009848416018436,
+  "median_nll": 3.143462657928467,
+  "ppl": 49.45112665854749,
+  "acc": 0.3373799545335045,
+  "tokens": 64663,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 208.5,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.8_gpu2.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.8_gpu2.json
+Diversity metrics for pps4_lam0.6_temp0.8_gpu2.json:
+  distinct_1: 0.2504
+  repetition_1: 0.4222
+  distinct_2: 0.7364
+  repetition_2: 0.0895
+  distinct_4: 0.9813
+  repetition_4: 0.0165
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.8_gpu2.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.8_gpu3.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 204)
+  sampled 128/128 (steps this call: 200)
+Average denoising rounds per sample: 202.00
+First sample preview: '\nHave he been attending Cancer Research on Aging (CAMA) in Chicago?\n\nIn the past week, though, he publicly revealed a fa'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 3.9212626953577887,
+  "median_nll": 3.1701507568359375,
+  "ppl": 50.464125381857606,
+  "acc": 0.334229930900153,
+  "tokens": 64689,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 202.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.8_gpu3.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.8_gpu3.json
+Diversity metrics for pps4_lam0.6_temp0.8_gpu3.json:
+  distinct_1: 0.2498
+  repetition_1: 0.4203
+  distinct_2: 0.7356
+  repetition_2: 0.0824
+  distinct_4: 0.9892
+  repetition_4: 0.0084
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.8_gpu3.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.8_gpu4.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 206)
+  sampled 128/128 (steps this call: 204)
+Average denoising rounds per sample: 205.00
+First sample preview: ' and-turned electrician Mark Anderson narrowly opened fire on the a Mercedes that came out of a home at 275 in a downtow'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 3.9627575752168345,
+  "median_nll": 3.2250115871429443,
+  "ppl": 52.60218060308741,
+  "acc": 0.3245725044755849,
+  "tokens": 64796,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 205.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.8_gpu4.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.8_gpu4.json
+Diversity metrics for pps4_lam0.6_temp0.8_gpu4.json:
+  distinct_1: 0.2448
+  repetition_1: 0.4107
+  distinct_2: 0.7334
+  repetition_2: 0.0736
+  distinct_4: 0.9900
+  repetition_4: 0.0071
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.8_gpu4.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.8_gpu5.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 200)
+  sampled 128/128 (steps this call: 202)
+Average denoising rounds per sample: 201.00
+First sample preview: 'ine W. Barker. While India’s aversion to Corydon is understandable, 1769 it parallels British resistance to Japanese, he'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 3.997569114385298,
+  "median_nll": 3.2696503400802612,
+  "ppl": 54.46558936083237,
+  "acc": 0.32421404061949366,
+  "tokens": 64698,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 201.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.8_gpu5.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.8_gpu5.json
+Diversity metrics for pps4_lam0.6_temp0.8_gpu5.json:
+  distinct_1: 0.2539
+  repetition_1: 0.4116
+  distinct_2: 0.7402
+  repetition_2: 0.0813
+  distinct_4: 0.9861
+  repetition_4: 0.0116
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.8_gpu5.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.8_gpu6.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 202)
+  sampled 128/128 (steps this call: 203)
+Average denoising rounds per sample: 202.50
+First sample preview: '\n“z okept e” noise, though a .” I used the “it. I had just unmasked the Hearst happened on him, which it was just “i….i\n'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 3.9174913685951633,
+  "median_nll": 3.1576714515686035,
+  "ppl": 50.27416709777929,
+  "acc": 0.33555683105037,
+  "tokens": 64463,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 202.5,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.8_gpu6.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.8_gpu6.json
+Diversity metrics for pps4_lam0.6_temp0.8_gpu6.json:
+  distinct_1: 0.2538
+  repetition_1: 0.4267
+  distinct_2: 0.7349
+  repetition_2: 0.0926
+  distinct_4: 0.9816
+  repetition_4: 0.0145
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.8_gpu6.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.8_gpu7.log

@@ -0,0 +1,50 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Loaded SAD checkpoint: outputs/sad_block_ar/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 64/128 (steps this call: 206)
+  sampled 128/128 (steps this call: 198)
+Average denoising rounds per sample: 202.00
+First sample preview: ' with reports from Politico and elsewhere. First, the New York Post reported that:\n\n\nReporting from Washington at the-ba'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 16/128
+  scored 32/128
+  scored 48/128
+  scored 64/128
+  scored 80/128
+  scored 96/128
+  scored 112/128
+  scored 128/128
+{
+  "avg_nll": 3.962866495961764,
+  "median_nll": 3.223786234855652,
+  "ppl": 52.60791038382395,
+  "acc": 0.3319229106183211,
+  "tokens": 64756,
+  "checkpoint": "outputs/sad_block_ar/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "level_lambdas": "0.6",
+  "avg_steps": 202.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/metrics/pps4_lam0.6_temp0.8_gpu7.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/block_ar_gen_ppl/samples/pps4_lam0.6_temp0.8_gpu7.json
+Diversity metrics for pps4_lam0.6_temp0.8_gpu7.json:
+  distinct_1: 0.2596
+  repetition_1: 0.4171
+  distinct_2: 0.7469
+  repetition_2: 0.0821
+  distinct_4: 0.9917
+  repetition_4: 0.0060
+  num_samples: 128.0000
+Saved -> eval/block_ar_gen_ppl/diversity/pps4_lam0.6_temp0.8_gpu7.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.94_gpu0.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.94
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 249)
+Average denoising rounds per sample: 249.00
+First sample preview: " back to the point that we can't Page or verify, but a core said or aired. Nor did the would need to guard that the judg"
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.543587042971091,
+  "median_nll": 5.149930000305176,
+  "ppl": 255.59318080251396,
+  "acc": 0.19897951342462386,
+  "tokens": 65067,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 249.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.94
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.94_gpu0.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.94_gpu0.json
+Diversity metrics for pps4_lam0.6_temp0.94_gpu0.json:
+  distinct_1: 0.3373
+  repetition_1: 0.3051
+  d1: 0.3373
+  r1: 0.3051
+  distinct_2: 0.8485
+  repetition_2: 0.0237
+  d2: 0.8485
+  r2: 0.0237
+  distinct_4: 0.9989
+  repetition_4: 0.0006
+  d4: 0.9989
+  r4: 0.0006
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.94_gpu0.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.94_gpu1.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.94
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 248)
+Average denoising rounds per sample: 248.00
+First sample preview: ', maybe a positive for their own transportation systems more efficient or acceptable. Personally, I believe that, after '
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.533697939093593,
+  "median_nll": 5.119769811630249,
+  "ppl": 253.07804997927,
+  "acc": 0.20001230277119922,
+  "tokens": 65026,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 248.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.94
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.94_gpu1.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.94_gpu1.json
+Diversity metrics for pps4_lam0.6_temp0.94_gpu1.json:
+  distinct_1: 0.3277
+  repetition_1: 0.3094
+  d1: 0.3277
+  r1: 0.3094
+  distinct_2: 0.8368
+  repetition_2: 0.0268
+  d2: 0.8368
+  r2: 0.0268
+  distinct_4: 0.9985
+  repetition_4: 0.0008
+  d4: 0.9985
+  r4: 0.0008
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.94_gpu1.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.94_gpu2.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.94
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 249)
+Average denoising rounds per sample: 249.00
+First sample preview: ' Altoona — A fifth-year physics student at an evangelical Alto State injured his self in Boulder after receiving death a'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.50760142278298,
+  "median_nll": 5.113685607910156,
+  "ppl": 246.55902638854954,
+  "acc": 0.20347219018621887,
+  "tokens": 65031,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 249.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.94
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.94_gpu2.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.94_gpu2.json
+Diversity metrics for pps4_lam0.6_temp0.94_gpu2.json:
+  distinct_1: 0.3302
+  repetition_1: 0.3070
+  d1: 0.3302
+  r1: 0.3070
+  distinct_2: 0.8423
+  repetition_2: 0.0257
+  d2: 0.8423
+  r2: 0.0257
+  distinct_4: 0.9991
+  repetition_4: 0.0003
+  d4: 0.9991
+  r4: 0.0003
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.94_gpu2.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.94_gpu3.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.94
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 244)
+Average denoising rounds per sample: 244.00
+First sample preview: "Microsoft's impressive Project Halo - as a cyber-punk augmented reality game brings a bit laser arsenal-like, thanks to "
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.539563543188986,
+  "median_nll": 5.134376525878906,
+  "ppl": 254.56686776493189,
+  "acc": 0.19772325205753405,
+  "tokens": 65005,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 244.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.94
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.94_gpu3.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.94_gpu3.json
+Diversity metrics for pps4_lam0.6_temp0.94_gpu3.json:
+  distinct_1: 0.3365
+  repetition_1: 0.3043
+  d1: 0.3365
+  r1: 0.3043
+  distinct_2: 0.8469
+  repetition_2: 0.0249
+  d2: 0.8469
+  r2: 0.0249
+  distinct_4: 0.9988
+  repetition_4: 0.0005
+  d4: 0.9988
+  r4: 0.0005
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.94_gpu3.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.94_gpu4.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.94
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 242)
+Average denoising rounds per sample: 242.00
+First sample preview: '’t think, “that there’s not all,” Give me some man, who proclaims the life and “s the earth like a leaf,””\n\nThe idea of '
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.478549847127144,
+  "median_nll": 5.078615188598633,
+  "ppl": 239.49914505266958,
+  "acc": 0.2047227136374125,
+  "tokens": 65005,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 242.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.94
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.94_gpu4.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.94_gpu4.json
+Diversity metrics for pps4_lam0.6_temp0.94_gpu4.json:
+  distinct_1: 0.3377
+  repetition_1: 0.3060
+  d1: 0.3377
+  r1: 0.3060
+  distinct_2: 0.8462
+  repetition_2: 0.0236
+  d2: 0.8462
+  r2: 0.0236
+  distinct_4: 0.9990
+  repetition_4: 0.0006
+  d4: 0.9990
+  r4: 0.0006
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.94_gpu4.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.94_gpu5.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.94
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 246)
+Average denoising rounds per sample: 246.00
+First sample preview: '\'(Amused" faces : not because of himself vs her) I\'re sure that\'s or anything could relate.\n\n/No, this one.. I mean, I\'m'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.5353770273285425,
+  "median_nll": 5.149424076080322,
+  "ppl": 253.50334731146614,
+  "acc": 0.19396186929693632,
+  "tokens": 64987,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 246.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.94
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.94_gpu5.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.94_gpu5.json
+Diversity metrics for pps4_lam0.6_temp0.94_gpu5.json:
+  distinct_1: 0.3386
+  repetition_1: 0.3035
+  d1: 0.3386
+  r1: 0.3035
+  distinct_2: 0.8467
+  repetition_2: 0.0230
+  d2: 0.8467
+  r2: 0.0230
+  distinct_4: 0.9992
+  repetition_4: 0.0003
+  d4: 0.9992
+  r4: 0.0003
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.94_gpu5.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.94_gpu6.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.94
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 245)
+Average denoising rounds per sample: 245.00
+First sample preview: 'Attitude is probably the simplest and most important venue to establish emotional attachment. I\'m terrified of saying "s'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.497107732198619,
+  "median_nll": 5.089093208312988,
+  "ppl": 243.9852401313613,
+  "acc": 0.19989539909858636,
+  "tokens": 65009,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 245.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.94
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.94_gpu6.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.94_gpu6.json
+Diversity metrics for pps4_lam0.6_temp0.94_gpu6.json:
+  distinct_1: 0.3360
+  repetition_1: 0.3111
+  d1: 0.3360
+  r1: 0.3111
+  distinct_2: 0.8476
+  repetition_2: 0.0266
+  d2: 0.8476
+  r2: 0.0266
+  distinct_4: 0.9986
+  repetition_4: 0.0008
+  d4: 0.9986
+  r4: 0.0008
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.94_gpu6.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.94_gpu7.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.94
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 248)
+Average denoising rounds per sample: 248.00
+First sample preview: '\nThe Happ tech application that officials and and developers at the Department of Safety and Human Services are set to s'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.409098334819644,
+  "median_nll": 4.98369288444519,
+  "ppl": 223.4300377436419,
+  "acc": 0.2115686214213229,
+  "tokens": 65038,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 248.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.94
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.94_gpu7.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.94_gpu7.json
+Diversity metrics for pps4_lam0.6_temp0.94_gpu7.json:
+  distinct_1: 0.3313
+  repetition_1: 0.3136
+  d1: 0.3313
+  r1: 0.3136
+  distinct_2: 0.8402
+  repetition_2: 0.0263
+  d2: 0.8402
+  r2: 0.0263
+  distinct_4: 0.9987
+  repetition_4: 0.0008
+  d4: 0.9987
+  r4: 0.0008
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.94_gpu7.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.96_gpu0.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 253)
+Average denoising rounds per sample: 253.00
+First sample preview: " new to 'cas' - wasn't Page or Mozilla, or a Newsgram executive course. Nor a pay watch - same thing of that slavery in "
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.746680855705214,
+  "median_nll": 5.41691780090332,
+  "ppl": 313.14954491001765,
+  "acc": 0.18630933161123667,
+  "tokens": 65037,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 253.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.96_gpu0.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.96_gpu0.json
+Diversity metrics for pps4_lam0.6_temp0.96_gpu0.json:
+  distinct_1: 0.3528
+  repetition_1: 0.2864
+  d1: 0.3528
+  r1: 0.2864
+  distinct_2: 0.8623
+  repetition_2: 0.0203
+  d2: 0.8623
+  r2: 0.0203
+  distinct_4: 0.9995
+  repetition_4: 0.0002
+  d4: 0.9995
+  r4: 0.0002
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.96_gpu0.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.96_gpu1.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 253)
+Average denoising rounds per sample: 253.00
+First sample preview: ', maybe annos for their own transportation systems more efficient than acceptable. Personally, I believe that, after see'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.8007773466924935,
+  "median_nll": 5.507569789886475,
+  "ppl": 330.55641700064734,
+  "acc": 0.17892665529062707,
+  "tokens": 65049,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 253.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.96_gpu1.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.96_gpu1.json
+Diversity metrics for pps4_lam0.6_temp0.96_gpu1.json:
+  distinct_1: 0.3488
+  repetition_1: 0.2891
+  d1: 0.3488
+  r1: 0.2891
+  distinct_2: 0.8615
+  repetition_2: 0.0198
+  d2: 0.8615
+  r2: 0.0198
+  distinct_4: 0.9996
+  repetition_4: 0.0001
+  d4: 0.9996
+  r4: 0.0001
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.96_gpu1.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.96_gpu2.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 251)
+Average denoising rounds per sample: 251.00
+First sample preview: ' a form of scummy self-indulgence at an ego tulips at the self, instead of at other," said Washington political analyst '
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.672595226476284,
+  "median_nll": 5.31312894821167,
+  "ppl": 290.7882172602754,
+  "acc": 0.191125388091359,
+  "tokens": 65062,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 251.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.96_gpu2.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.96_gpu2.json
+Diversity metrics for pps4_lam0.6_temp0.96_gpu2.json:
+  distinct_1: 0.3354
+  repetition_1: 0.2976
+  d1: 0.3354
+  r1: 0.2976
+  distinct_2: 0.8513
+  repetition_2: 0.0230
+  d2: 0.8513
+  r2: 0.0230
+  distinct_4: 0.9994
+  repetition_4: 0.0002
+  d4: 0.9994
+  r4: 0.0002
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.96_gpu2.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.96_gpu3.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 254)
+Average denoising rounds per sample: 254.00
+First sample preview: "Microsoft's savage Project Incredible VR is a cyber-punk virtual reality game which functions are psychological arsenal-"
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.729456543464223,
+  "median_nll": 5.379550933837891,
+  "ppl": 307.801945936048,
+  "acc": 0.18857046216291012,
+  "tokens": 65042,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 254.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.96_gpu3.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.96_gpu3.json
+Diversity metrics for pps4_lam0.6_temp0.96_gpu3.json:
+  distinct_1: 0.3476
+  repetition_1: 0.2958
+  d1: 0.3476
+  r1: 0.2958
+  distinct_2: 0.8540
+  repetition_2: 0.0243
+  d2: 0.8540
+  r2: 0.0243
+  distinct_4: 0.9989
+  repetition_4: 0.0007
+  d4: 0.9989
+  r4: 0.0007
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.96_gpu3.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.96_gpu4.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 251)
+Average denoising rounds per sample: 251.00
+First sample preview: '’t think, “that there’s not all,” Give me this whole book it proclaims the life and ’s blood of the real monster, NASCAR'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.7241878748270025,
+  "median_nll": 5.377181529998779,
+  "ppl": 306.18450410108153,
+  "acc": 0.18794402583423037,
+  "tokens": 65030,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 251.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.96_gpu4.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.96_gpu4.json
+Diversity metrics for pps4_lam0.6_temp0.96_gpu4.json:
+  distinct_1: 0.3521
+  repetition_1: 0.2910
+  d1: 0.3521
+  r1: 0.2910
+  distinct_2: 0.8587
+  repetition_2: 0.0219
+  d2: 0.8587
+  r2: 0.0219
+  distinct_4: 0.9992
+  repetition_4: 0.0005
+  d4: 0.9992
+  r4: 0.0005
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.96_gpu4.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.96_gpu5.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 250)
+Average denoising rounds per sample: 250.00
+First sample preview: "'(AmWarning to work..... not because Tikta is interesting or I're doing that' or anything, narunas thought/No, this one "
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.725282199884571,
+  "median_nll": 5.394169807434082,
+  "ppl": 306.51975287827094,
+  "acc": 0.1891035013466718,
+  "tokens": 64975,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 250.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.96_gpu5.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.96_gpu5.json
+Diversity metrics for pps4_lam0.6_temp0.96_gpu5.json:
+  distinct_1: 0.3553
+  repetition_1: 0.2871
+  d1: 0.3553
+  r1: 0.2871
+  distinct_2: 0.8593
+  repetition_2: 0.0212
+  d2: 0.8593
+  r2: 0.0212
+  distinct_4: 0.9991
+  repetition_4: 0.0006
+  d4: 0.9991
+  r4: 0.0006
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.96_gpu5.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.96_gpu6.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 251)
+Average denoising rounds per sample: 251.00
+First sample preview: 'Attendingment inside the panel") is a venue to establish emotional attachment to the things left of you (they\'re now any'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.746712427606919,
+  "median_nll": 5.3975830078125,
+  "ppl": 313.15943179274166,
+  "acc": 0.18455819097947646,
+  "tokens": 65096,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 251.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.96_gpu6.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.96_gpu6.json
+Diversity metrics for pps4_lam0.6_temp0.96_gpu6.json:
+  distinct_1: 0.3455
+  repetition_1: 0.2982
+  d1: 0.3455
+  r1: 0.2982
+  distinct_2: 0.8554
+  repetition_2: 0.0226
+  d2: 0.8554
+  r2: 0.0226
+  distinct_4: 0.9989
+  repetition_4: 0.0006
+  d4: 0.9989
+  r4: 0.0006
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.96_gpu6.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.96_gpu7.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 254)
+Average denoising rounds per sample: 254.00
+First sample preview: '\nThe Happ tech application that officials and and developers at the Department of Safety and Human Services are set to s'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.675058493158582,
+  "median_nll": 5.313899517059326,
+  "ppl": 291.5053891175974,
+  "acc": 0.19025290183718963,
+  "tokens": 65045,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 254.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.96_gpu7.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.96_gpu7.json
+Diversity metrics for pps4_lam0.6_temp0.96_gpu7.json:
+  distinct_1: 0.3437
+  repetition_1: 0.2939
+  d1: 0.3437
+  r1: 0.2939
+  distinct_2: 0.8529
+  repetition_2: 0.0223
+  d2: 0.8529
+  r2: 0.0223
+  distinct_4: 0.9988
+  repetition_4: 0.0006
+  d4: 0.9988
+  r4: 0.0006
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.96_gpu7.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.98_gpu0.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.98
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 258)
+Average denoising rounds per sample: 258.00
+First sample preview: " new to 'cas' - wasn't Page or Mozilla, or a Newsgram search report. Nor a pay watch - same tech of that investigations "
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.965963974092866,
+  "median_nll": 5.685207843780518,
+  "ppl": 389.92872815117715,
+  "acc": 0.1719249692496925,
+  "tokens": 65040,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 258.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.98_gpu0.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.98_gpu0.json
+Diversity metrics for pps4_lam0.6_temp0.98_gpu0.json:
+  distinct_1: 0.3546
+  repetition_1: 0.2820
+  d1: 0.3546
+  r1: 0.2820
+  distinct_2: 0.8656
+  repetition_2: 0.0175
+  d2: 0.8656
+  r2: 0.0175
+  distinct_4: 0.9995
+  repetition_4: 0.0001
+  d4: 0.9995
+  r4: 0.0001
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.98_gpu0.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.98_gpu1.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.98
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: ', maybe annos for their own transportation systems more efficient than acceptable. Personally, I believe that, after see'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 6.022894606221046,
+  "median_nll": 5.761094093322754,
+  "ppl": 412.77167961127736,
+  "acc": 0.16454527265171687,
+  "tokens": 64973,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 256.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.98_gpu1.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.98_gpu1.json
+Diversity metrics for pps4_lam0.6_temp0.98_gpu1.json:
+  distinct_1: 0.3711
+  repetition_1: 0.2737
+  d1: 0.3711
+  r1: 0.2737
+  distinct_2: 0.8719
+  repetition_2: 0.0174
+  d2: 0.8719
+  r2: 0.0174
+  distinct_4: 0.9994
+  repetition_4: 0.0002
+  d4: 0.9994
+  r4: 0.0002
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.98_gpu1.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.98_gpu2.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.98
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 258)
+Average denoising rounds per sample: 258.00
+First sample preview: ' a form of scrawled self-indulgence at an ego tulips onto the self, instead prompts written criticism, said Senate candi'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.992293243009679,
+  "median_nll": 5.717593193054199,
+  "ppl": 400.331615746006,
+  "acc": 0.16747580273467508,
+  "tokens": 65090,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 258.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.98_gpu2.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.98_gpu2.json
+Diversity metrics for pps4_lam0.6_temp0.98_gpu2.json:
+  distinct_1: 0.3511
+  repetition_1: 0.2760
+  d1: 0.3511
+  r1: 0.2760
+  distinct_2: 0.8615
+  repetition_2: 0.0172
+  d2: 0.8615
+  r2: 0.0172
+  distinct_4: 0.9994
+  repetition_4: 0.0001
+  d4: 0.9994
+  r4: 0.0001
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.98_gpu2.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.98_gpu3.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.98
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 252)
+Average denoising rounds per sample: 252.00
+First sample preview: "Imagine's revolutionary Blocks arcade mode as a cyber-punk augmented reality game which functions from laser arsenal-lik"
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.9575406407160765,
+  "median_nll": 5.658609390258789,
+  "ppl": 386.65802293967437,
+  "acc": 0.1708832531028437,
+  "tokens": 65021,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 252.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.98_gpu3.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.98_gpu3.json
+Diversity metrics for pps4_lam0.6_temp0.98_gpu3.json:
+  distinct_1: 0.3571
+  repetition_1: 0.2828
+  d1: 0.3571
+  r1: 0.2828
+  distinct_2: 0.8640
+  repetition_2: 0.0189
+  d2: 0.8640
+  r2: 0.0189
+  distinct_4: 0.9994
+  repetition_4: 0.0003
+  d4: 0.9994
+  r4: 0.0003
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.98_gpu3.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.98_gpu4.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.98
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 253)
+Average denoising rounds per sample: 253.00
+First sample preview: '’t think, “that there’s not all,” Give me this whole book it proclaims the life and ’s blood of the real monster, NASCAR'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.985462752583502,
+  "median_nll": 5.731670618057251,
+  "ppl": 397.6064721063864,
+  "acc": 0.171787537676078,
+  "tokens": 65028,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 253.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.98_gpu4.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.98_gpu4.json
+Diversity metrics for pps4_lam0.6_temp0.98_gpu4.json:
+  distinct_1: 0.3664
+  repetition_1: 0.2722
+  d1: 0.3664
+  r1: 0.2722
+  distinct_2: 0.8746
+  repetition_2: 0.0163
+  d2: 0.8746
+  r2: 0.0163
+  distinct_4: 0.9997
+  repetition_4: 0.0002
+  d4: 0.9997
+  r4: 0.0002
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.98_gpu4.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.98_gpu5.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.98
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 254)
+Average denoising rounds per sample: 254.00
+First sample preview: "'(AmWarning to work..... not because obta is interesting or I're doing that's or anything else narun titles related/No, "
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.945007449058054,
+  "median_nll": 5.640978813171387,
+  "ppl": 381.84220565183045,
+  "acc": 0.17253364090734333,
+  "tokens": 65025,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 254.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.98_gpu5.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.98_gpu5.json
+Diversity metrics for pps4_lam0.6_temp0.98_gpu5.json:
+  distinct_1: 0.3613
+  repetition_1: 0.2780
+  d1: 0.3613
+  r1: 0.2780
+  distinct_2: 0.8714
+  repetition_2: 0.0169
+  d2: 0.8714
+  r2: 0.0169
+  distinct_4: 0.9995
+  repetition_4: 0.0002
+  d4: 0.9995
+  r4: 0.0002
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.98_gpu5.json

eval/top3_latest_test/logs/pps4_lam0.6_temp0.98_gpu6.log

@@ -0,0 +1,53 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from checkpoint:/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt
+Loaded SAD checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt (step=500000)
+level_lambdas = [0.6]
+leaf_temperature = 0.98
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 253)
+Average denoising rounds per sample: 253.00
+First sample preview: ' Dynamo will acquire stadium licenses on camera video for every venue from experience of one.\n\nMembers of Vancouver-base'
+Loading eval LM:        /home/sunhc/diffusion/sad/models/gpt2-large
+Loading eval tokenizer: /home/sunhc/diffusion/sad/models/gpt2-large
+
+Eval LM loaded (774,030,080 params)
+Scoring samples under eval LM...
+  scored 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.938803984301763,
+  "median_nll": 5.663514137268066,
+  "ppl": 379.4807930294375,
+  "acc": 0.1718584315593736,
+  "tokens": 65071,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_ar_b32_top3/latest.pt",
+  "eval_model": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "eval_tokenizer": "/home/sunhc/diffusion/sad/models/gpt2-large",
+  "num_samples": 128,
+  "generated_seq_len": 512,
+  "mode": "sad_generation",
+  "model_type": "block",
+  "level_lambdas": "0.6",
+  "avg_steps": 253.0,
+  "positions_per_step": 4,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/top3_latest_test/metrics/pps4_lam0.6_temp0.98_gpu6.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/top3_latest_test/samples/pps4_lam0.6_temp0.98_gpu6.json
+Diversity metrics for pps4_lam0.6_temp0.98_gpu6.json:
+  distinct_1: 0.3631
+  repetition_1: 0.2781
+  d1: 0.3631
+  r1: 0.2781
+  distinct_2: 0.8696
+  repetition_2: 0.0176
+  d2: 0.8696
+  r2: 0.0176
+  distinct_4: 0.9995
+  repetition_4: 0.0001
+  d4: 0.9995
+  r4: 0.0001
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/top3_latest_test/diversity/pps4_lam0.6_temp0.98_gpu6.json