backfill batch 4 (400)

eval/ppl_block_mask/logs/pps1_temp0.96_gpu6.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ' and Scully, the advantages of being free, freely accessible; requiring full-time FCC technician to participate, and vid'
+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": 4.629389670978973,
+  "median_nll": 4.035597801208496,
+  "ppl": 102.4515158915416,
+  "acc": 0.26156040551525284,
+  "tokens": 65201,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.96_gpu6.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.96_gpu6.json
+Diversity metrics for pps1_temp0.96_gpu6.json:
+  distinct_1: 0.3016
+  repetition_1: 0.3350
+  d1: 0.3016
+  r1: 0.3350
+  distinct_2: 0.8052
+  repetition_2: 0.0336
+  d2: 0.8052
+  r2: 0.0336
+  distinct_4: 0.9982
+  repetition_4: 0.0006
+  d4: 0.9982
+  r4: 0.0006
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.96_gpu6.json

eval/ppl_block_mask/logs/pps1_temp0.96_gpu7.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: " much the exact original experience- optical DSL's existing network networking technologies, which are, reasonably, spee"
+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": 4.56603759567158,
+  "median_nll": 3.9685707092285156,
+  "ppl": 96.16231987689436,
+  "acc": 0.2598432443210577,
+  "tokens": 65197,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.96_gpu7.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.96_gpu7.json
+Diversity metrics for pps1_temp0.96_gpu7.json:
+  distinct_1: 0.2924
+  repetition_1: 0.3371
+  d1: 0.2924
+  r1: 0.3371
+  distinct_2: 0.7993
+  repetition_2: 0.0331
+  d2: 0.7993
+  r2: 0.0331
+  distinct_4: 0.9983
+  repetition_4: 0.0005
+  d4: 0.9983
+  r4: 0.0005
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.96_gpu7.json

eval/ppl_block_mask/logs/pps1_temp0.98_gpu0.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.98
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ' This was the tip of second depositedary unitary group on land within 100 years of soil in Hong Kong.\n\nIt is likely that'
+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": 4.890374490358972,
+  "median_nll": 4.346218824386597,
+  "ppl": 133.00337320702567,
+  "acc": 0.24291100824622175,
+  "tokens": 65242,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.98_gpu0.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.98_gpu0.json
+Diversity metrics for pps1_temp0.98_gpu0.json:
+  distinct_1: 0.3202
+  repetition_1: 0.3138
+  d1: 0.3202
+  r1: 0.3138
+  distinct_2: 0.8258
+  repetition_2: 0.0256
+  d2: 0.8258
+  r2: 0.0256
+  distinct_4: 0.9984
+  repetition_4: 0.0005
+  d4: 0.9984
+  r4: 0.0005
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.98_gpu0.json

eval/ppl_block_mask/logs/pps1_temp0.98_gpu1.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.98
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: '\nWhat is his main purpose?\n\nRedAnon (Sound) is just a guy travel because he is a MP with a bike:\n\nHe *also* is MP becaus'
+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": 4.864414271531302,
+  "median_nll": 4.3348588943481445,
+  "ppl": 129.59500889077177,
+  "acc": 0.24135604655085174,
+  "tokens": 65219,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.98_gpu1.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.98_gpu1.json
+Diversity metrics for pps1_temp0.98_gpu1.json:
+  distinct_1: 0.3057
+  repetition_1: 0.3274
+  d1: 0.3057
+  r1: 0.3274
+  distinct_2: 0.8132
+  repetition_2: 0.0284
+  d2: 0.8132
+  r2: 0.0284
+  distinct_4: 0.9983
+  repetition_4: 0.0005
+  d4: 0.9983
+  r4: 0.0005
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.98_gpu1.json

eval/ppl_block_mask/logs/pps1_temp0.98_gpu2.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.98
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: 'Following Newell’s promotion to Formula 1 lead, he has since been punted to a long list of standing entires panel chaire'
+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": 4.7924794058696385,
+  "median_nll": 4.230315208435059,
+  "ppl": 120.60001466213323,
+  "acc": 0.2505748536039489,
+  "tokens": 65234,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.98_gpu2.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.98_gpu2.json
+Diversity metrics for pps1_temp0.98_gpu2.json:
+  distinct_1: 0.3094
+  repetition_1: 0.3164
+  d1: 0.3094
+  r1: 0.3164
+  distinct_2: 0.8196
+  repetition_2: 0.0274
+  d2: 0.8196
+  r2: 0.0274
+  distinct_4: 0.9984
+  repetition_4: 0.0005
+  d4: 0.9984
+  r4: 0.0005
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.98_gpu2.json

eval/ppl_block_mask/logs/pps1_temp0.98_gpu3.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.98
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ' protest will be raised over the fact they know “there are some dead Haramamba”, but there’s only one that would be funn'
+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": 4.9274660803634465,
+  "median_nll": 4.4067909717559814,
+  "ppl": 138.02931365139068,
+  "acc": 0.2390797188200264,
+  "tokens": 65154,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.98_gpu3.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.98_gpu3.json
+Diversity metrics for pps1_temp0.98_gpu3.json:
+  distinct_1: 0.3234
+  repetition_1: 0.3121
+  d1: 0.3234
+  r1: 0.3121
+  distinct_2: 0.8299
+  repetition_2: 0.0280
+  d2: 0.8299
+  r2: 0.0280
+  distinct_4: 0.9977
+  repetition_4: 0.0012
+  d4: 0.9977
+  r4: 0.0012
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.98_gpu3.json

eval/ppl_block_mask/logs/pps1_temp0.98_gpu4.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.98
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ' sights as well as some enthusiasts and to some extent were put in. Something I saw just happened to surface when I was '
+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": 4.808717729415823,
+  "median_nll": 4.238682746887207,
+  "ppl": 122.5743432300199,
+  "acc": 0.24665746703465194,
+  "tokens": 65220,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.98_gpu4.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.98_gpu4.json
+Diversity metrics for pps1_temp0.98_gpu4.json:
+  distinct_1: 0.3117
+  repetition_1: 0.3212
+  d1: 0.3117
+  r1: 0.3212
+  distinct_2: 0.8141
+  repetition_2: 0.0284
+  d2: 0.8141
+  r2: 0.0284
+  distinct_4: 0.9988
+  repetition_4: 0.0005
+  d4: 0.9988
+  r4: 0.0005
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.98_gpu4.json

eval/ppl_block_mask/logs/pps1_temp0.98_gpu5.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.98
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ' that BNY Mellon needs more time to master its “beauty” process.\n\nBethow chief profitability strategist McCord said 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 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 4.897342011891062,
+  "median_nll": 4.3782196044921875,
+  "ppl": 133.93331299935363,
+  "acc": 0.24079785193709244,
+  "tokens": 65175,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.98_gpu5.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.98_gpu5.json
+Diversity metrics for pps1_temp0.98_gpu5.json:
+  distinct_1: 0.3191
+  repetition_1: 0.3116
+  d1: 0.3191
+  r1: 0.3116
+  distinct_2: 0.8256
+  repetition_2: 0.0275
+  d2: 0.8256
+  r2: 0.0275
+  distinct_4: 0.9988
+  repetition_4: 0.0006
+  d4: 0.9988
+  r4: 0.0006
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.98_gpu5.json

eval/ppl_block_mask/logs/pps1_temp0.98_gpu6.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.98
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ' and Scully, the advantages of being free, freely accessible; requiring full-time FCC technician to participate, and vid'
+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": 4.892898825781406,
+  "median_nll": 4.372953414916992,
+  "ppl": 133.33954245677452,
+  "acc": 0.24377407128937717,
+  "tokens": 65171,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.98_gpu6.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.98_gpu6.json
+Diversity metrics for pps1_temp0.98_gpu6.json:
+  distinct_1: 0.3154
+  repetition_1: 0.3164
+  d1: 0.3154
+  r1: 0.3164
+  distinct_2: 0.8234
+  repetition_2: 0.0283
+  d2: 0.8234
+  r2: 0.0283
+  distinct_4: 0.9983
+  repetition_4: 0.0005
+  d4: 0.9983
+  r4: 0.0005
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.98_gpu6.json

eval/ppl_block_mask/logs/pps1_temp0.98_gpu7.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.98
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: " much the exact original experience- optical DSL's existing network networking standards, which are, reasonably, speeds,"
+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": 4.835842140874659,
+  "median_nll": 4.286372184753418,
+  "ppl": 125.94460165636974,
+  "acc": 0.2447940626245898,
+  "tokens": 65214,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.98_gpu7.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.98_gpu7.json
+Diversity metrics for pps1_temp0.98_gpu7.json:
+  distinct_1: 0.3111
+  repetition_1: 0.3205
+  d1: 0.3111
+  r1: 0.3205
+  distinct_2: 0.8181
+  repetition_2: 0.0300
+  d2: 0.8181
+  r2: 0.0300
+  distinct_4: 0.9983
+  repetition_4: 0.0008
+  d4: 0.9983
+  r4: 0.0008
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.98_gpu7.json

eval/ppl_block_mask/logs/pps1_temp0.9_gpu0.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.9
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: "\n\nThe government of China is calling the holding of a referendum at the leader's office in Hong Kong.\n\nIt will, in the e"
+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": 3.968183796991617,
+  "median_nll": 3.2790353298187256,
+  "ppl": 52.88838750969809,
+  "acc": 0.308378415656465,
+  "tokens": 65251,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.9
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.9_gpu0.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.9_gpu0.json
+Diversity metrics for pps1_temp0.9_gpu0.json:
+  distinct_1: 0.2585
+  repetition_1: 0.3760
+  d1: 0.2585
+  r1: 0.3760
+  distinct_2: 0.7579
+  repetition_2: 0.0497
+  d2: 0.7579
+  r2: 0.0497
+  distinct_4: 0.9961
+  repetition_4: 0.0011
+  d4: 0.9961
+  r4: 0.0011
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.9_gpu0.json

eval/ppl_block_mask/logs/pps1_temp0.9_gpu1.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.9
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: '\nWhat is his main motivation?\n\n"Anon (i) is just a guy, because he\'s been together with a team in a positive environment'
+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": 3.885239225937816,
+  "median_nll": 3.2040858268737793,
+  "ppl": 48.6785862471334,
+  "acc": 0.3152870146188973,
+  "tokens": 65258,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.9
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.9_gpu1.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.9_gpu1.json
+Diversity metrics for pps1_temp0.9_gpu1.json:
+  distinct_1: 0.2457
+  repetition_1: 0.3913
+  d1: 0.2457
+  r1: 0.3913
+  distinct_2: 0.7418
+  repetition_2: 0.0554
+  d2: 0.7418
+  r2: 0.0554
+  distinct_4: 0.9944
+  repetition_4: 0.0024
+  d4: 0.9944
+  r4: 0.0024
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.9_gpu1.json

eval/ppl_block_mask/logs/pps1_temp0.9_gpu2.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.9
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ' was tagged Vikings’ first punt to this game, and has only since been punted twice as long as NFL games. However, the Li'
+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": 3.9296703042556294,
+  "median_nll": 3.2431626319885254,
+  "ppl": 50.89019662177709,
+  "acc": 0.31363887355052456,
+  "tokens": 65196,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.9
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.9_gpu2.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.9_gpu2.json
+Diversity metrics for pps1_temp0.9_gpu2.json:
+  distinct_1: 0.2582
+  repetition_1: 0.3780
+  d1: 0.2582
+  r1: 0.3780
+  distinct_2: 0.7573
+  repetition_2: 0.0493
+  d2: 0.7573
+  r2: 0.0493
+  distinct_4: 0.9954
+  repetition_4: 0.0022
+  d4: 0.9954
+  r4: 0.0022
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.9_gpu2.json

eval/ppl_block_mask/logs/pps1_temp0.9_gpu3.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.9
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ',” said Castreshue. “There are some great people we already have, but there’s only one that will be able to take the plu'
+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": 3.916636537070942,
+  "median_nll": 3.2204318046569824,
+  "ppl": 50.23120951825262,
+  "acc": 0.3150145391325753,
+  "tokens": 64997,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.9
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.9_gpu3.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.9_gpu3.json
+Diversity metrics for pps1_temp0.9_gpu3.json:
+  distinct_1: 0.2609
+  repetition_1: 0.3813
+  d1: 0.2609
+  r1: 0.3813
+  distinct_2: 0.7531
+  repetition_2: 0.0562
+  d2: 0.7531
+  r2: 0.0562
+  distinct_4: 0.9937
+  repetition_4: 0.0040
+  d4: 0.9937
+  r4: 0.0040
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.9_gpu3.json

eval/ppl_block_mask/logs/pps1_temp0.9_gpu4.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.9
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ' sights as powerful as ever, thanks to some salts were put in. Something I found just happened to surface when I was try'
+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": 3.968989319620253,
+  "median_nll": 3.30350399017334,
+  "ppl": 52.931007465994284,
+  "acc": 0.30708093594169544,
+  "tokens": 65175,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.9
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.9_gpu4.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.9_gpu4.json
+Diversity metrics for pps1_temp0.9_gpu4.json:
+  distinct_1: 0.2601
+  repetition_1: 0.3775
+  d1: 0.2601
+  r1: 0.3775
+  distinct_2: 0.7572
+  repetition_2: 0.0492
+  d2: 0.7572
+  r2: 0.0492
+  distinct_4: 0.9952
+  repetition_4: 0.0023
+  d4: 0.9952
+  r4: 0.0023
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.9_gpu4.json

eval/ppl_block_mask/logs/pps1_temp0.9_gpu5.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.9
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ' that had "sought more time to space and more money," with a dozen of these employees. Development concluded their chief'
+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": 3.935485039766194,
+  "median_nll": 3.235534906387329,
+  "ppl": 51.18697165317862,
+  "acc": 0.317393637408969,
+  "tokens": 65225,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.9
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.9_gpu5.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.9_gpu5.json
+Diversity metrics for pps1_temp0.9_gpu5.json:
+  distinct_1: 0.2678
+  repetition_1: 0.3753
+  d1: 0.2678
+  r1: 0.3753
+  distinct_2: 0.7612
+  repetition_2: 0.0538
+  d2: 0.7612
+  r2: 0.0538
+  distinct_4: 0.9949
+  repetition_4: 0.0029
+  d4: 0.9949
+  r4: 0.0029
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.9_gpu5.json

eval/ppl_block_mask/logs/pps1_temp0.9_gpu6.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.9
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ' safety nets, bandwidth advantages, lower prices, and even lower costs have been evident in the rollout of XChat videoco'
+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": 3.9164174420410447,
+  "median_nll": 3.2135212421417236,
+  "ppl": 50.22020531542854,
+  "acc": 0.3128058039480344,
+  "tokens": 65197,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.9
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.9_gpu6.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.9_gpu6.json
+Diversity metrics for pps1_temp0.9_gpu6.json:
+  distinct_1: 0.2556
+  repetition_1: 0.3905
+  d1: 0.2556
+  r1: 0.3905
+  distinct_2: 0.7529
+  repetition_2: 0.0559
+  d2: 0.7529
+  r2: 0.0559
+  distinct_4: 0.9948
+  repetition_4: 0.0027
+  d4: 0.9948
+  r4: 0.0027
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.9_gpu6.json

eval/ppl_block_mask/logs/pps1_temp0.9_gpu7.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.9
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ' much the exact original base-haul required under existing law. Specifically, risks are, limits on speeds rising from or'
+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": 3.9011227771725787,
+  "median_nll": 3.18917977809906,
+  "ppl": 49.45794819838231,
+  "acc": 0.3208389591260587,
+  "tokens": 65176,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 0.9
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp0.9_gpu7.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp0.9_gpu7.json
+Diversity metrics for pps1_temp0.9_gpu7.json:
+  distinct_1: 0.2610
+  repetition_1: 0.3802
+  d1: 0.2610
+  r1: 0.3802
+  distinct_2: 0.7531
+  repetition_2: 0.0592
+  d2: 0.7531
+  r2: 0.0592
+  distinct_4: 0.9881
+  repetition_4: 0.0086
+  d4: 0.9881
+  r4: 0.0086
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp0.9_gpu7.json

eval/ppl_block_mask/logs/pps1_temp1_gpu0.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 1.0
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: "\n\nThe tip of an object called Fary Russian was found about 100 metres of soil in Hong Kong Bay\n\nIt's likely that folks 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": 5.172205562946662,
+  "median_nll": 4.715129137039185,
+  "ppl": 176.30325691568547,
+  "acc": 0.22221881421955034,
+  "tokens": 65206,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 1.0
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp1_gpu0.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp1_gpu0.json
+Diversity metrics for pps1_temp1_gpu0.json:
+  distinct_1: 0.3418
+  repetition_1: 0.2909
+  d1: 0.3418
+  r1: 0.2909
+  distinct_2: 0.8455
+  repetition_2: 0.0237
+  d2: 0.8455
+  r2: 0.0237
+  distinct_4: 0.9988
+  repetition_4: 0.0008
+  d4: 0.9988
+  r4: 0.0008
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp1_gpu0.json

eval/ppl_block_mask/logs/pps1_temp1_gpu1.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 1.0
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: '\nWhat is his traffic record?\n\n@AnonTV\n\nHere just a few examples of highlights when the MP passed a bike:\n\nHe passed 4.5 '
+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": 5.173155560964764,
+  "median_nll": 4.696434736251831,
+  "ppl": 176.47082424205192,
+  "acc": 0.2247539022969119,
+  "tokens": 65218,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 1.0
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp1_gpu1.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp1_gpu1.json
+Diversity metrics for pps1_temp1_gpu1.json:
+  distinct_1: 0.3357
+  repetition_1: 0.2994
+  d1: 0.3357
+  r1: 0.2994
+  distinct_2: 0.8421
+  repetition_2: 0.0246
+  d2: 0.8421
+  r2: 0.0246
+  distinct_4: 0.9983
+  repetition_4: 0.0010
+  d4: 0.9983
+  r4: 0.0010
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp1_gpu1.json

eval/ppl_block_mask/logs/pps1_temp1_gpu2.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 1.0
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: 'Following Newell’s promotion to Formula 1 lead, he has since been punted to a long list of standing entires panel chaire'
+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": 5.123290712363324,
+  "median_nll": 4.6181640625,
+  "ppl": 167.88692920498178,
+  "acc": 0.2252296462144796,
+  "tokens": 65209,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 1.0
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp1_gpu2.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp1_gpu2.json
+Diversity metrics for pps1_temp1_gpu2.json:
+  distinct_1: 0.3272
+  repetition_1: 0.3022
+  d1: 0.3272
+  r1: 0.3022
+  distinct_2: 0.8351
+  repetition_2: 0.0247
+  d2: 0.8351
+  r2: 0.0247
+  distinct_4: 0.9979
+  repetition_4: 0.0010
+  d4: 0.9979
+  r4: 0.0010
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp1_gpu2.json

eval/ppl_block_mask/logs/pps1_temp1_gpu3.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 1.0
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ' protest will be raised over the fact they know “there are some dead Haramamba”, but there’s only one that can be answer'
+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": 5.215278590476556,
+  "median_nll": 4.75036096572876,
+  "ppl": 184.0630920620313,
+  "acc": 0.22022868544240656,
+  "tokens": 65155,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 1.0
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp1_gpu3.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp1_gpu3.json
+Diversity metrics for pps1_temp1_gpu3.json:
+  distinct_1: 0.3457
+  repetition_1: 0.2920
+  d1: 0.3457
+  r1: 0.2920
+  distinct_2: 0.8461
+  repetition_2: 0.0227
+  d2: 0.8461
+  r2: 0.0227
+  distinct_4: 0.9987
+  repetition_4: 0.0005
+  d4: 0.9987
+  r4: 0.0005
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp1_gpu3.json

eval/ppl_block_mask/logs/pps1_temp1_gpu4.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 1.0
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ' sights as train cruises enthusiasts get to state capitol building in Kentucky. I was the first to surface when I was tr'
+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": 5.110161557616514,
+  "median_nll": 4.610766410827637,
+  "ppl": 165.69712234357613,
+  "acc": 0.22696949713987763,
+  "tokens": 65207,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 1.0
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp1_gpu4.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp1_gpu4.json
+Diversity metrics for pps1_temp1_gpu4.json:
+  distinct_1: 0.3318
+  repetition_1: 0.3007
+  d1: 0.3318
+  r1: 0.3007
+  distinct_2: 0.8388
+  repetition_2: 0.0235
+  d2: 0.8388
+  r2: 0.0235
+  distinct_4: 0.9990
+  repetition_4: 0.0004
+  d4: 0.9990
+  r4: 0.0004
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp1_gpu4.json

eval/ppl_block_mask/logs/pps1_temp1_gpu5.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 1.0
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ' that BNY Mellon needs more time to master its “beautiful” process.\n\nBethow chief profitability strategist McCord said t'
+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": 5.194060956245589,
+  "median_nll": 4.737077713012695,
+  "ppl": 180.19884871873523,
+  "acc": 0.2212113773741217,
+  "tokens": 65182,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 1.0
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp1_gpu5.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp1_gpu5.json
+Diversity metrics for pps1_temp1_gpu5.json:
+  distinct_1: 0.3406
+  repetition_1: 0.2954
+  d1: 0.3406
+  r1: 0.2954
+  distinct_2: 0.8454
+  repetition_2: 0.0208
+  d2: 0.8454
+  r2: 0.0208
+  distinct_4: 0.9987
+  repetition_4: 0.0008
+  d4: 0.9987
+  r4: 0.0008
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp1_gpu5.json

eval/ppl_block_mask/logs/pps1_temp1_gpu6.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 1.0
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: ' and Scully, the advantages of being free, freely accessible; requiring full-contained servers, true to use; videoconfer'
+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": 5.172385331791189,
+  "median_nll": 4.7170538902282715,
+  "ppl": 176.33495359742025,
+  "acc": 0.22272859795301447,
+  "tokens": 65169,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 1.0
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp1_gpu6.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp1_gpu6.json
+Diversity metrics for pps1_temp1_gpu6.json:
+  distinct_1: 0.3387
+  repetition_1: 0.3007
+  d1: 0.3387
+  r1: 0.3007
+  distinct_2: 0.8414
+  repetition_2: 0.0242
+  d2: 0.8414
+  r2: 0.0242
+  distinct_4: 0.9991
+  repetition_4: 0.0002
+  d4: 0.9991
+  r4: 0.0002
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp1_gpu6.json

eval/ppl_block_mask/logs/pps1_temp1_gpu7.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 1.0
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 512)
+Average denoising rounds per sample: 512.00
+First sample preview: " much the exact original experience- optical DSL's existing network networking standards, which are, reasonably, speeds,"
+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": 5.0978294895838125,
+  "median_nll": 4.60785436630249,
+  "ppl": 163.66628212531316,
+  "acc": 0.22229892463221193,
+  "tokens": 65187,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 512.0,
+  "positions_per_step": 1,
+  "leaf_temperature": 1.0
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps1_temp1_gpu7.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps1_temp1_gpu7.json
+Diversity metrics for pps1_temp1_gpu7.json:
+  distinct_1: 0.3184
+  repetition_1: 0.3044
+  d1: 0.3184
+  r1: 0.3044
+  distinct_2: 0.8276
+  repetition_2: 0.0239
+  d2: 0.8276
+  r2: 0.0239
+  distinct_4: 0.9984
+  repetition_4: 0.0005
+  d4: 0.9984
+  r4: 0.0005
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps1_temp1_gpu7.json

eval/ppl_block_mask/logs/pps2_temp0.8_gpu0.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: 'Today is feast day for the King of Hong Kong. In August of 2015, Jon & Hong Kong dedicated their legendary dogs to the 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": 3.3397223812194596,
+  "median_nll": 2.5761420726776123,
+  "ppl": 28.211293633219025,
+  "acc": 0.37587203048649315,
+  "tokens": 65078,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.8_gpu0.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.8_gpu0.json
+Diversity metrics for pps2_temp0.8_gpu0.json:
+  distinct_1: 0.2141
+  repetition_1: 0.4551
+  d1: 0.2141
+  r1: 0.4551
+  distinct_2: 0.6845
+  repetition_2: 0.1042
+  d2: 0.6845
+  r2: 0.1042
+  distinct_4: 0.9817
+  repetition_4: 0.0136
+  d4: 0.9817
+  r4: 0.0136
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.8_gpu0.json

eval/ppl_block_mask/logs/pps2_temp0.8_gpu1.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: 'What is the Google Pixel?\n\nPolyAnon owners can now told just how the company was announced, and announced the new Google'
+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": 3.395374689174047,
+  "median_nll": 2.6400675773620605,
+  "ppl": 29.825826795953585,
+  "acc": 0.3669434680501619,
+  "tokens": 65149,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.8_gpu1.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.8_gpu1.json
+Diversity metrics for pps2_temp0.8_gpu1.json:
+  distinct_1: 0.2154
+  repetition_1: 0.4504
+  d1: 0.2154
+  r1: 0.4504
+  distinct_2: 0.6861
+  repetition_2: 0.0988
+  d2: 0.6861
+  r2: 0.0988
+  distinct_4: 0.9819
+  repetition_4: 0.0134
+  d4: 0.9819
+  r4: 0.0134
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.8_gpu1.json

eval/ppl_block_mask/logs/pps2_temp0.8_gpu2.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: 'If you’re not familiar with a new Apple iOS device or camera, you may not long to find out how to spend your time on a 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 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 3.3294132645714942,
+  "median_nll": 2.5770978927612305,
+  "ppl": 27.921954096410403,
+  "acc": 0.37611733988634616,
+  "tokens": 65110,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.8_gpu2.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.8_gpu2.json
+Diversity metrics for pps2_temp0.8_gpu2.json:
+  distinct_1: 0.2167
+  repetition_1: 0.4502
+  d1: 0.2167
+  r1: 0.4502
+  distinct_2: 0.6903
+  repetition_2: 0.1020
+  d2: 0.6903
+  r2: 0.1020
+  distinct_4: 0.9810
+  repetition_4: 0.0166
+  d4: 0.9810
+  r4: 0.0166
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.8_gpu2.json

eval/ppl_block_mask/logs/pps2_temp0.8_gpu3.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: 'TAMPA — There’s been a lot of good news for the Leafs roster, whether it’s return in the form of the 5-year-1 defenceman'
+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": 3.350425038918358,
+  "median_nll": 2.572860598564148,
+  "ppl": 28.514850989834876,
+  "acc": 0.37952029520295205,
+  "tokens": 65040,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.8_gpu3.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.8_gpu3.json
+Diversity metrics for pps2_temp0.8_gpu3.json:
+  distinct_1: 0.2249
+  repetition_1: 0.4519
+  d1: 0.2249
+  r1: 0.4519
+  distinct_2: 0.6928
+  repetition_2: 0.1059
+  d2: 0.6928
+  r2: 0.1059
+  distinct_4: 0.9805
+  repetition_4: 0.0147
+  d4: 0.9805
+  r4: 0.0147
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.8_gpu3.json

eval/ppl_block_mask/logs/pps2_temp0.8_gpu4.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: ' sights as such a state short of an awkward and expensive one in the political world. The problem is that when you are t'
+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": 3.411889458919376,
+  "median_nll": 2.65773344039917,
+  "ppl": 30.322483253448148,
+  "acc": 0.3663361770809315,
+  "tokens": 65055,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.8_gpu4.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.8_gpu4.json
+Diversity metrics for pps2_temp0.8_gpu4.json:
+  distinct_1: 0.2189
+  repetition_1: 0.4453
+  d1: 0.2189
+  r1: 0.4453
+  distinct_2: 0.6909
+  repetition_2: 0.0995
+  d2: 0.6909
+  r2: 0.0995
+  distinct_4: 0.9786
+  repetition_4: 0.0147
+  d4: 0.9786
+  r4: 0.0147
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.8_gpu4.json

eval/ppl_block_mask/logs/pps2_temp0.8_gpu5.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: " have had to find out more about why it isn’t, or if I had forgotten his. Instead of a r blog post that I get here, I'll"
+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": 3.409145855587747,
+  "median_nll": 2.6582645177841187,
+  "ppl": 30.239404407184114,
+  "acc": 0.36817189757462115,
+  "tokens": 65062,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.8_gpu5.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.8_gpu5.json
+Diversity metrics for pps2_temp0.8_gpu5.json:
+  distinct_1: 0.2193
+  repetition_1: 0.4535
+  d1: 0.2193
+  r1: 0.4535
+  distinct_2: 0.6969
+  repetition_2: 0.0976
+  d2: 0.6969
+  r2: 0.0976
+  distinct_4: 0.9887
+  repetition_4: 0.0074
+  d4: 0.9887
+  r4: 0.0074
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.8_gpu5.json

eval/ppl_block_mask/logs/pps2_temp0.8_gpu6.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: ' was that, he said, a hunger strike was likely to mean continued psychological isolation and would be more likely to die'
+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": 3.375300637586726,
+  "median_nll": 2.636515259742737,
+  "ppl": 29.233071020205696,
+  "acc": 0.3663658132252717,
+  "tokens": 65148,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.8_gpu6.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.8_gpu6.json
+Diversity metrics for pps2_temp0.8_gpu6.json:
+  distinct_1: 0.2077
+  repetition_1: 0.4608
+  d1: 0.2077
+  r1: 0.4608
+  distinct_2: 0.6767
+  repetition_2: 0.1046
+  d2: 0.6767
+  r2: 0.1046
+  distinct_4: 0.9842
+  repetition_4: 0.0105
+  d4: 0.9842
+  r4: 0.0105
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.8_gpu6.json

eval/ppl_block_mask/logs/pps2_temp0.8_gpu7.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.8
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: ' storage.\n\nIf your website spans an existing file (such28.pm), a separate external label or or a brand additional e.g. 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": 3.377343486019598,
+  "median_nll": 2.6435821056365967,
+  "ppl": 29.292850793245645,
+  "acc": 0.36265339661183554,
+  "tokens": 65109,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.8
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.8_gpu7.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.8_gpu7.json
+Diversity metrics for pps2_temp0.8_gpu7.json:
+  distinct_1: 0.2087
+  repetition_1: 0.4579
+  d1: 0.2087
+  r1: 0.4579
+  distinct_2: 0.6809
+  repetition_2: 0.1023
+  d2: 0.6809
+  r2: 0.1023
+  distinct_4: 0.9856
+  repetition_4: 0.0089
+  d4: 0.9856
+  r4: 0.0089
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.8_gpu7.json

eval/ppl_block_mask/logs/pps2_temp0.96_gpu0.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: 'HBO Arrested after City Suspects Lotters Young In Hotel Car, Club Sex & Hong Kong\n\nStar Wars Up: Skyler Freaked Out Joey'
+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": 5.011767065596034,
+  "median_nll": 4.491327524185181,
+  "ppl": 150.16986182922395,
+  "acc": 0.22912958075010742,
+  "tokens": 65164,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.96_gpu0.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.96_gpu0.json
+Diversity metrics for pps2_temp0.96_gpu0.json:
+  distinct_1: 0.3096
+  repetition_1: 0.3219
+  d1: 0.3096
+  r1: 0.3219
+  distinct_2: 0.8197
+  repetition_2: 0.0277
+  d2: 0.8197
+  r2: 0.0277
+  distinct_4: 0.9987
+  repetition_4: 0.0006
+  d4: 0.9987
+  r4: 0.0006
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.96_gpu0.json

eval/ppl_block_mask/logs/pps2_temp0.96_gpu1.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: 'Who is calling it that? The rule "Anon" can be told just to kick in your face.\n\nDear Koni: Koney poseo. Are you too? Con'
+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": 5.039804006316775,
+  "median_nll": 4.540257692337036,
+  "ppl": 154.43974284537305,
+  "acc": 0.22665172540832618,
+  "tokens": 65144,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.96_gpu1.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.96_gpu1.json
+Diversity metrics for pps2_temp0.96_gpu1.json:
+  distinct_1: 0.3097
+  repetition_1: 0.3230
+  d1: 0.3097
+  r1: 0.3230
+  distinct_2: 0.8135
+  repetition_2: 0.0286
+  d2: 0.8135
+  r2: 0.0286
+  distinct_4: 0.9981
+  repetition_4: 0.0008
+  d4: 0.9981
+  r4: 0.0008
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.96_gpu1.json

eval/ppl_block_mask/logs/pps2_temp0.96_gpu2.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: ' of last month’ legacy for Hitch with a new LP that features his entire gutted/record long UK catalogue.\n\nWith the 270mm'
+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": 5.042779049376842,
+  "median_nll": 4.519660949707031,
+  "ppl": 154.89989187271524,
+  "acc": 0.22927922396856582,
+  "tokens": 65152,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.96_gpu2.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.96_gpu2.json
+Diversity metrics for pps2_temp0.96_gpu2.json:
+  distinct_1: 0.3133
+  repetition_1: 0.3174
+  d1: 0.3133
+  r1: 0.3174
+  distinct_2: 0.8207
+  repetition_2: 0.0272
+  d2: 0.8207
+  r2: 0.0272
+  distinct_4: 0.9988
+  repetition_4: 0.0006
+  d4: 0.9988
+  r4: 0.0006
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.96_gpu2.json

eval/ppl_block_mask/logs/pps2_temp0.96_gpu3.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: 'LA JONG — China’s minor volcanoes are beginning to appear like the rest of the world’s massive volcanic eruptions.\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 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.057072785417947,
+  "median_nll": 4.544251203536987,
+  "ppl": 157.12988957157427,
+  "acc": 0.22661339889366933,
+  "tokens": 65080,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.96_gpu3.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.96_gpu3.json
+Diversity metrics for pps2_temp0.96_gpu3.json:
+  distinct_1: 0.3196
+  repetition_1: 0.3166
+  d1: 0.3196
+  r1: 0.3166
+  distinct_2: 0.8230
+  repetition_2: 0.0264
+  d2: 0.8230
+  r2: 0.0264
+  distinct_4: 0.9986
+  repetition_4: 0.0002
+  d4: 0.9986
+  r4: 0.0002
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.96_gpu3.json

eval/ppl_block_mask/logs/pps2_temp0.96_gpu4.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: ' sights as you exchange "concern, news and ideas" in a way the Tumbler would surface when Cinderella was trying out some'
+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": 5.052290129983268,
+  "median_nll": 4.525528430938721,
+  "ppl": 156.38018566803564,
+  "acc": 0.22608722349293095,
+  "tokens": 65143,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.96_gpu4.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.96_gpu4.json
+Diversity metrics for pps2_temp0.96_gpu4.json:
+  distinct_1: 0.3056
+  repetition_1: 0.3257
+  d1: 0.3056
+  r1: 0.3257
+  distinct_2: 0.8139
+  repetition_2: 0.0284
+  d2: 0.8139
+  r2: 0.0284
+  distinct_4: 0.9990
+  repetition_4: 0.0004
+  d4: 0.9990
+  r4: 0.0004
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.96_gpu4.json

eval/ppl_block_mask/logs/pps2_temp0.96_gpu5.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: ' that BNP has done more damage than making its labels "beautiful or respectful of liberalism forgotten".. Ukip\'s chief e'
+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": 5.022410725933941,
+  "median_nll": 4.503439903259277,
+  "ppl": 151.77675527549476,
+  "acc": 0.22870430058016392,
+  "tokens": 65154,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.96_gpu5.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.96_gpu5.json
+Diversity metrics for pps2_temp0.96_gpu5.json:
+  distinct_1: 0.3166
+  repetition_1: 0.3212
+  d1: 0.3166
+  r1: 0.3212
+  distinct_2: 0.8185
+  repetition_2: 0.0275
+  d2: 0.8185
+  r2: 0.0275
+  distinct_4: 0.9985
+  repetition_4: 0.0004
+  d4: 0.9985
+  r4: 0.0004
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.96_gpu5.json

eval/ppl_block_mask/logs/pps2_temp0.96_gpu6.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: ' financial aid, employee advantages, a planned strike and other employee benefits continued.\n\n"Our decision to terminate'
+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": 4.953173845376899,
+  "median_nll": 4.408069610595703,
+  "ppl": 141.62374322939493,
+  "acc": 0.23509531929274774,
+  "tokens": 65097,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.96_gpu6.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.96_gpu6.json
+Diversity metrics for pps2_temp0.96_gpu6.json:
+  distinct_1: 0.3058
+  repetition_1: 0.3313
+  d1: 0.3058
+  r1: 0.3313
+  distinct_2: 0.8147
+  repetition_2: 0.0325
+  d2: 0.8147
+  r2: 0.0325
+  distinct_4: 0.9956
+  repetition_4: 0.0036
+  d4: 0.9956
+  r4: 0.0036
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.96_gpu6.json

eval/ppl_block_mask/logs/pps2_temp0.96_gpu7.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+leaf_temperature = 0.96
+Generating 128 samples (L=512)...
+  sampled 128/128 (steps this call: 256)
+Average denoising rounds per sample: 256.00
+First sample preview: '.\n\nor (ASUS Automated Freedom of Oversight and Theft of Treasury), a proposal to label nearly 500,000 additional videos '
+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": 5.097687022988683,
+  "median_nll": 4.637754440307617,
+  "ppl": 163.64296680822753,
+  "acc": 0.22233142400835393,
+  "tokens": 65119,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.96
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.96_gpu7.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.96_gpu7.json
+Diversity metrics for pps2_temp0.96_gpu7.json:
+  distinct_1: 0.3082
+  repetition_1: 0.3267
+  d1: 0.3082
+  r1: 0.3267
+  distinct_2: 0.8168
+  repetition_2: 0.0278
+  d2: 0.8168
+  r2: 0.0278
+  distinct_4: 0.9987
+  repetition_4: 0.0004
+  d4: 0.9987
+  r4: 0.0004
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.96_gpu7.json

eval/ppl_block_mask/logs/pps2_temp0.98_gpu0.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+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: 'HBO Arrested after City Suspected Lotters of Inly Car Lot Club Sex & Hong Kong\n\nStar Wars Up: Skyler Freaked Out Joey 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 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.334710401993279,
+  "median_nll": 4.910536766052246,
+  "ppl": 207.41267381265482,
+  "acc": 0.20933265816569227,
+  "tokens": 65169,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.98_gpu0.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.98_gpu0.json
+Diversity metrics for pps2_temp0.98_gpu0.json:
+  distinct_1: 0.3321
+  repetition_1: 0.3044
+  d1: 0.3321
+  r1: 0.3044
+  distinct_2: 0.8391
+  repetition_2: 0.0235
+  d2: 0.8391
+  r2: 0.0235
+  distinct_4: 0.9992
+  repetition_4: 0.0005
+  d4: 0.9992
+  r4: 0.0005
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.98_gpu0.json

eval/ppl_block_mask/logs/pps2_temp0.98_gpu1.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+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: 'Who is calling it that? The rule "Anon" can be told just to kick in your face.\n\nDear Koni: Koney poseo. Are you too? Con'
+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": 5.3139040605767605,
+  "median_nll": 4.860509872436523,
+  "ppl": 203.14176002587092,
+  "acc": 0.21196494620685422,
+  "tokens": 65157,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.98_gpu1.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.98_gpu1.json
+Diversity metrics for pps2_temp0.98_gpu1.json:
+  distinct_1: 0.3301
+  repetition_1: 0.3082
+  d1: 0.3301
+  r1: 0.3082
+  distinct_2: 0.8351
+  repetition_2: 0.0258
+  d2: 0.8351
+  r2: 0.0258
+  distinct_4: 0.9991
+  repetition_4: 0.0005
+  d4: 0.9991
+  r4: 0.0005
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.98_gpu1.json

eval/ppl_block_mask/logs/pps2_temp0.98_gpu2.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+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: ' of last month’ legacy for Hitch with a new LP that features his entire gutted/ crappy long UK catalogue.\n\nWith the 270m'
+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": 5.280290978384353,
+  "median_nll": 4.832667112350464,
+  "ppl": 196.42702305483374,
+  "acc": 0.213838312505757,
+  "tokens": 65138,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.98_gpu2.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.98_gpu2.json
+Diversity metrics for pps2_temp0.98_gpu2.json:
+  distinct_1: 0.3287
+  repetition_1: 0.3018
+  d1: 0.3287
+  r1: 0.3018
+  distinct_2: 0.8387
+  repetition_2: 0.0226
+  d2: 0.8387
+  r2: 0.0226
+  distinct_4: 0.9992
+  repetition_4: 0.0003
+  d4: 0.9992
+  r4: 0.0003
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.98_gpu2.json

eval/ppl_block_mask/logs/pps2_temp0.98_gpu3.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+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: 'LA JONG — China’s minor volcanoes are beginning to appear like becoming one of the world’s massive volcanic eruptions.\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 32/128
+  scored 64/128
+  scored 96/128
+  scored 128/128
+{
+  "avg_nll": 5.383843381810781,
+  "median_nll": 4.929405212402344,
+  "ppl": 217.85797982767744,
+  "acc": 0.2066544800995407,
+  "tokens": 65099,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.98_gpu3.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.98_gpu3.json
+Diversity metrics for pps2_temp0.98_gpu3.json:
+  distinct_1: 0.3418
+  repetition_1: 0.2943
+  d1: 0.3418
+  r1: 0.2943
+  distinct_2: 0.8476
+  repetition_2: 0.0223
+  d2: 0.8476
+  r2: 0.0223
+  distinct_4: 0.9986
+  repetition_4: 0.0009
+  d4: 0.9986
+  r4: 0.0009
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.98_gpu3.json

eval/ppl_block_mask/logs/pps2_temp0.98_gpu4.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+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: ' sights as you exchange "concernful news and ideas" in a way the Tumbler would surface when Cinderella was trying out so'
+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": 5.3368604071972605,
+  "median_nll": 4.8968517780303955,
+  "ppl": 207.85909186933148,
+  "acc": 0.20943767083320536,
+  "tokens": 65122,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.98_gpu4.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.98_gpu4.json
+Diversity metrics for pps2_temp0.98_gpu4.json:
+  distinct_1: 0.3295
+  repetition_1: 0.3037
+  d1: 0.3295
+  r1: 0.3037
+  distinct_2: 0.8386
+  repetition_2: 0.0224
+  d2: 0.8386
+  r2: 0.0224
+  distinct_4: 0.9991
+  repetition_4: 0.0003
+  d4: 0.9991
+  r4: 0.0003
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.98_gpu4.json

eval/ppl_block_mask/logs/pps2_temp0.98_gpu5.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+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: ' that BNP has done more damage than making its labels "beautiful or respectful of liberalism forgotten".. Ukip\'s chief e'
+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": 5.316916610706063,
+  "median_nll": 4.881314516067505,
+  "ppl": 203.7546574899438,
+  "acc": 0.21254029163468918,
+  "tokens": 65150,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.98_gpu5.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.98_gpu5.json
+Diversity metrics for pps2_temp0.98_gpu5.json:
+  distinct_1: 0.3310
+  repetition_1: 0.3034
+  d1: 0.3310
+  r1: 0.3034
+  distinct_2: 0.8355
+  repetition_2: 0.0245
+  d2: 0.8355
+  r2: 0.0245
+  distinct_4: 0.9985
+  repetition_4: 0.0008
+  d4: 0.9985
+  r4: 0.0008
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.98_gpu5.json

eval/ppl_block_mask/logs/pps2_temp0.98_gpu6.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+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: ' the noteworthy, tactical advantages of a direct strike.\n\n— continued industrial isolation\n\nOur decision to bypass the H'
+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": 5.289520576352939,
+  "median_nll": 4.822242736816406,
+  "ppl": 198.24835767168256,
+  "acc": 0.2129833023548749,
+  "tokens": 65099,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.98_gpu6.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.98_gpu6.json
+Diversity metrics for pps2_temp0.98_gpu6.json:
+  distinct_1: 0.3268
+  repetition_1: 0.3114
+  d1: 0.3268
+  r1: 0.3114
+  distinct_2: 0.8317
+  repetition_2: 0.0255
+  d2: 0.8317
+  r2: 0.0255
+  distinct_4: 0.9985
+  repetition_4: 0.0008
+  d4: 0.9985
+  r4: 0.0008
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.98_gpu6.json

eval/ppl_block_mask/logs/pps2_temp0.98_gpu7.log

@@ -0,0 +1,52 @@
+`torch_dtype` is deprecated! Use `dtype` instead!
+Using config from cli:/home/sunhc/diffusion/sad/configs/block_mask_owt_b32.yaml
+Loaded block-mask checkpoint: /home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/latest.pt (step=500000)
+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: '.\n\nor "Brewed", which means that not all pop musicians are literary figures. Each label in or in one additional series o'
+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": 5.3361520012438195,
+  "median_nll": 4.8864312171936035,
+  "ppl": 207.7118953947596,
+  "acc": 0.20866988114615645,
+  "tokens": 65122,
+  "checkpoint": "/home/sunhc/diffusion/sad/outputs/sad_block_mask_b32/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": "block_mask_generation",
+  "model_type": "block_mask",
+  "level_lambdas": null,
+  "avg_steps": 256.0,
+  "positions_per_step": 2,
+  "leaf_temperature": 0.98
+}
+Saved metrics → /home/sunhc/diffusion/sad/eval/ppl_block_mask/metrics/pps2_temp0.98_gpu7.json
+Saved samples  → /home/sunhc/diffusion/sad/eval/ppl_block_mask/samples/pps2_temp0.98_gpu7.json
+Diversity metrics for pps2_temp0.98_gpu7.json:
+  distinct_1: 0.3227
+  repetition_1: 0.3158
+  d1: 0.3227
+  r1: 0.3158
+  distinct_2: 0.8304
+  repetition_2: 0.0261
+  d2: 0.8304
+  r2: 0.0261
+  distinct_4: 0.9980
+  repetition_4: 0.0014
+  d4: 0.9980
+  r4: 0.0014
+  num_samples: 128.0000
+Saved -> /home/sunhc/diffusion/sad/eval/ppl_block_mask/diversity/pps2_temp0.98_gpu7.json