karantonis commited on
Commit
3df0e5b
·
verified ·
1 Parent(s): dee6de0

judge-imitation: artem2/llm-reasoning-rl-with-uhead-workdir

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/lpmean_selector.json +0 -0
  2. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/lpmin_selector.json +0 -0
  3. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/negent_selector.json +0 -0
  4. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/prm7b_selector.json +0 -0
  5. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/top1_selector.json +0 -0
  6. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/lpmean_selector.json +0 -0
  7. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/lpmean_selector.provenance.json +75 -0
  8. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/lpmin_selector.json +0 -0
  9. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/lpmin_selector.provenance.json +75 -0
  10. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/negent_selector.json +0 -0
  11. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/negent_selector.provenance.json +75 -0
  12. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/prm_cache.jsonl +0 -0
  13. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/top1_selector.json +0 -0
  14. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/top1_selector.provenance.json +75 -0
  15. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/logprob/shard0.jsonl +0 -0
  16. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/logprob/shard0.log +207 -0
  17. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/logprob/shard0.provenance.json +85 -0
  18. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/prm7b.log +165 -0
  19. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/base/lpmean_selector.json +0 -0
  20. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/base/lpmin_selector.json +0 -0
  21. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/base/negent_selector.json +0 -0
  22. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/base/top1_selector.json +0 -0
  23. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/lpmean_selector.json +0 -0
  24. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/lpmean_selector.provenance.json +73 -0
  25. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/lpmin_selector.json +0 -0
  26. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/lpmin_selector.provenance.json +73 -0
  27. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/negent_selector.json +0 -0
  28. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/negent_selector.provenance.json +73 -0
  29. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/top1_selector.json +0 -0
  30. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/top1_selector.provenance.json +73 -0
  31. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/logprob/shard0.jsonl +0 -0
  32. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/logprob/shard0.log +204 -0
  33. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/logprob/shard0.provenance.json +83 -0
  34. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/base/lpmean_selector.json +0 -0
  35. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/base/lpmin_selector.json +0 -0
  36. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/base/negent_selector.json +0 -0
  37. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/base/top1_selector.json +0 -0
  38. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/lpmean_selector.json +0 -0
  39. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/lpmean_selector.provenance.json +73 -0
  40. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/lpmin_selector.json +0 -0
  41. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/lpmin_selector.provenance.json +73 -0
  42. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/negent_selector.json +0 -0
  43. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/negent_selector.provenance.json +73 -0
  44. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/top1_selector.json +0 -0
  45. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/top1_selector.provenance.json +73 -0
  46. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/logprob/shard0.jsonl +0 -0
  47. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/logprob/shard0.log +205 -0
  48. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/logprob/shard0.provenance.json +83 -0
  49. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/laneLPX_g0.log +4 -0
  50. judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/laneLPX_g1.log +3 -0
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/lpmean_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/lpmin_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/negent_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/prm7b_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/top1_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/lpmean_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/lpmean_selector.provenance.json ADDED
@@ -0,0 +1,75 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T14:47:26+04:00",
3
+ "host": "ws006599",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/logprob_scores.py",
8
+ "--pool-dir",
9
+ "workdir/gpools/llama31_8b/chatpool",
10
+ "--model",
11
+ "meta-llama/Llama-3.1-8B-Instruct",
12
+ "--template",
13
+ "auto",
14
+ "--merge",
15
+ "workdir/gpools/llama31_8b/logprob/shard0.jsonl"
16
+ ],
17
+ "git": {
18
+ "commit": "6cae6bf292a53eff08d1ca10850479123a1aae47",
19
+ "branch": "HEAD",
20
+ "subject": "logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases",
21
+ "dirty": false,
22
+ "dirty_files": []
23
+ },
24
+ "gpus": {
25
+ "visible": "1",
26
+ "devices": [
27
+ {
28
+ "index": "0",
29
+ "name": "NVIDIA RTX 5880 Ada Generation",
30
+ "memory": "49140 MiB",
31
+ "driver": "580.178.04",
32
+ "uuid": "GPU-ea30b823-34f2-1633-1958-47244a40ac5d"
33
+ },
34
+ {
35
+ "index": "1",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-b843d197-f9cd-0190-cbf3-1ded8ca2f9a8"
40
+ }
41
+ ]
42
+ },
43
+ "packages": {
44
+ "python": "3.11.15",
45
+ "torch": "2.9.1",
46
+ "vllm": "0.15.1",
47
+ "transformers": "4.57.6",
48
+ "speculators": "0.4.0.dev49",
49
+ "peft": null,
50
+ "accelerate": "1.13.0",
51
+ "numpy": "2.2.6",
52
+ "huggingface_hub": "0.36.2",
53
+ "openai": "2.32.0",
54
+ "wandb": "0.25.1"
55
+ },
56
+ "env": {
57
+ "CUDA_VISIBLE_DEVICES": "1",
58
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
59
+ "HF_HUB_OFFLINE": "1",
60
+ "TRANSFORMERS_OFFLINE": "1",
61
+ "NCCL_P2P_DISABLE": "1",
62
+ "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"
63
+ },
64
+ "models": [],
65
+ "files": {
66
+ "pool_config": {
67
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/llama31_8b/chatpool/pool_config.json",
68
+ "bytes": 213,
69
+ "sha256": "2442b361015d3bdac6fd3c2c74e57151637e209015d0058bd595d0b955fd8d9e"
70
+ }
71
+ },
72
+ "finished_at": "2026-10-31T14:47:26+04:00",
73
+ "wall_secs": 0.3,
74
+ "chains": 10888
75
+ }
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/lpmin_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/lpmin_selector.provenance.json ADDED
@@ -0,0 +1,75 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T14:47:26+04:00",
3
+ "host": "ws006599",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/logprob_scores.py",
8
+ "--pool-dir",
9
+ "workdir/gpools/llama31_8b/chatpool",
10
+ "--model",
11
+ "meta-llama/Llama-3.1-8B-Instruct",
12
+ "--template",
13
+ "auto",
14
+ "--merge",
15
+ "workdir/gpools/llama31_8b/logprob/shard0.jsonl"
16
+ ],
17
+ "git": {
18
+ "commit": "6cae6bf292a53eff08d1ca10850479123a1aae47",
19
+ "branch": "HEAD",
20
+ "subject": "logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases",
21
+ "dirty": false,
22
+ "dirty_files": []
23
+ },
24
+ "gpus": {
25
+ "visible": "1",
26
+ "devices": [
27
+ {
28
+ "index": "0",
29
+ "name": "NVIDIA RTX 5880 Ada Generation",
30
+ "memory": "49140 MiB",
31
+ "driver": "580.178.04",
32
+ "uuid": "GPU-ea30b823-34f2-1633-1958-47244a40ac5d"
33
+ },
34
+ {
35
+ "index": "1",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-b843d197-f9cd-0190-cbf3-1ded8ca2f9a8"
40
+ }
41
+ ]
42
+ },
43
+ "packages": {
44
+ "python": "3.11.15",
45
+ "torch": "2.9.1",
46
+ "vllm": "0.15.1",
47
+ "transformers": "4.57.6",
48
+ "speculators": "0.4.0.dev49",
49
+ "peft": null,
50
+ "accelerate": "1.13.0",
51
+ "numpy": "2.2.6",
52
+ "huggingface_hub": "0.36.2",
53
+ "openai": "2.32.0",
54
+ "wandb": "0.25.1"
55
+ },
56
+ "env": {
57
+ "CUDA_VISIBLE_DEVICES": "1",
58
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
59
+ "HF_HUB_OFFLINE": "1",
60
+ "TRANSFORMERS_OFFLINE": "1",
61
+ "NCCL_P2P_DISABLE": "1",
62
+ "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"
63
+ },
64
+ "models": [],
65
+ "files": {
66
+ "pool_config": {
67
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/llama31_8b/chatpool/pool_config.json",
68
+ "bytes": 213,
69
+ "sha256": "2442b361015d3bdac6fd3c2c74e57151637e209015d0058bd595d0b955fd8d9e"
70
+ }
71
+ },
72
+ "finished_at": "2026-10-31T14:47:26+04:00",
73
+ "wall_secs": 0.4,
74
+ "chains": 10888
75
+ }
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/negent_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/negent_selector.provenance.json ADDED
@@ -0,0 +1,75 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T14:47:26+04:00",
3
+ "host": "ws006599",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/logprob_scores.py",
8
+ "--pool-dir",
9
+ "workdir/gpools/llama31_8b/chatpool",
10
+ "--model",
11
+ "meta-llama/Llama-3.1-8B-Instruct",
12
+ "--template",
13
+ "auto",
14
+ "--merge",
15
+ "workdir/gpools/llama31_8b/logprob/shard0.jsonl"
16
+ ],
17
+ "git": {
18
+ "commit": "6cae6bf292a53eff08d1ca10850479123a1aae47",
19
+ "branch": "HEAD",
20
+ "subject": "logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases",
21
+ "dirty": false,
22
+ "dirty_files": []
23
+ },
24
+ "gpus": {
25
+ "visible": "1",
26
+ "devices": [
27
+ {
28
+ "index": "0",
29
+ "name": "NVIDIA RTX 5880 Ada Generation",
30
+ "memory": "49140 MiB",
31
+ "driver": "580.178.04",
32
+ "uuid": "GPU-ea30b823-34f2-1633-1958-47244a40ac5d"
33
+ },
34
+ {
35
+ "index": "1",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-b843d197-f9cd-0190-cbf3-1ded8ca2f9a8"
40
+ }
41
+ ]
42
+ },
43
+ "packages": {
44
+ "python": "3.11.15",
45
+ "torch": "2.9.1",
46
+ "vllm": "0.15.1",
47
+ "transformers": "4.57.6",
48
+ "speculators": "0.4.0.dev49",
49
+ "peft": null,
50
+ "accelerate": "1.13.0",
51
+ "numpy": "2.2.6",
52
+ "huggingface_hub": "0.36.2",
53
+ "openai": "2.32.0",
54
+ "wandb": "0.25.1"
55
+ },
56
+ "env": {
57
+ "CUDA_VISIBLE_DEVICES": "1",
58
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
59
+ "HF_HUB_OFFLINE": "1",
60
+ "TRANSFORMERS_OFFLINE": "1",
61
+ "NCCL_P2P_DISABLE": "1",
62
+ "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"
63
+ },
64
+ "models": [],
65
+ "files": {
66
+ "pool_config": {
67
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/llama31_8b/chatpool/pool_config.json",
68
+ "bytes": 213,
69
+ "sha256": "2442b361015d3bdac6fd3c2c74e57151637e209015d0058bd595d0b955fd8d9e"
70
+ }
71
+ },
72
+ "finished_at": "2026-10-31T14:47:26+04:00",
73
+ "wall_secs": 0.4,
74
+ "chains": 10888
75
+ }
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/prm_cache.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/top1_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/chatpool/top1_selector.provenance.json ADDED
@@ -0,0 +1,75 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T14:47:26+04:00",
3
+ "host": "ws006599",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/logprob_scores.py",
8
+ "--pool-dir",
9
+ "workdir/gpools/llama31_8b/chatpool",
10
+ "--model",
11
+ "meta-llama/Llama-3.1-8B-Instruct",
12
+ "--template",
13
+ "auto",
14
+ "--merge",
15
+ "workdir/gpools/llama31_8b/logprob/shard0.jsonl"
16
+ ],
17
+ "git": {
18
+ "commit": "6cae6bf292a53eff08d1ca10850479123a1aae47",
19
+ "branch": "HEAD",
20
+ "subject": "logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases",
21
+ "dirty": false,
22
+ "dirty_files": []
23
+ },
24
+ "gpus": {
25
+ "visible": "1",
26
+ "devices": [
27
+ {
28
+ "index": "0",
29
+ "name": "NVIDIA RTX 5880 Ada Generation",
30
+ "memory": "49140 MiB",
31
+ "driver": "580.178.04",
32
+ "uuid": "GPU-ea30b823-34f2-1633-1958-47244a40ac5d"
33
+ },
34
+ {
35
+ "index": "1",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-b843d197-f9cd-0190-cbf3-1ded8ca2f9a8"
40
+ }
41
+ ]
42
+ },
43
+ "packages": {
44
+ "python": "3.11.15",
45
+ "torch": "2.9.1",
46
+ "vllm": "0.15.1",
47
+ "transformers": "4.57.6",
48
+ "speculators": "0.4.0.dev49",
49
+ "peft": null,
50
+ "accelerate": "1.13.0",
51
+ "numpy": "2.2.6",
52
+ "huggingface_hub": "0.36.2",
53
+ "openai": "2.32.0",
54
+ "wandb": "0.25.1"
55
+ },
56
+ "env": {
57
+ "CUDA_VISIBLE_DEVICES": "1",
58
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
59
+ "HF_HUB_OFFLINE": "1",
60
+ "TRANSFORMERS_OFFLINE": "1",
61
+ "NCCL_P2P_DISABLE": "1",
62
+ "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"
63
+ },
64
+ "models": [],
65
+ "files": {
66
+ "pool_config": {
67
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/llama31_8b/chatpool/pool_config.json",
68
+ "bytes": 213,
69
+ "sha256": "2442b361015d3bdac6fd3c2c74e57151637e209015d0058bd595d0b955fd8d9e"
70
+ }
71
+ },
72
+ "finished_at": "2026-10-31T14:47:26+04:00",
73
+ "wall_secs": 0.5,
74
+ "chains": 10888
75
+ }
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/logprob/shard0.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/logprob/shard0.log ADDED
@@ -0,0 +1,207 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [logprob] shard 0/1: 10888 chains over 1361 problems (0 already scored)
2
+ INFO 10-31 14:13:07 [arg_utils.py:602] HF_HUB_OFFLINE is True, replace model_id [meta-llama/Llama-3.1-8B-Instruct] to model_path [/mnt/data/users/vlad.smirnov/hf_cache/hub/models--meta-llama--Llama-3.1-8B-Instruct/snapshots/0e9e39f249a16976918f6564b8830bc894c89659]
3
+ INFO 10-31 14:13:07 [utils.py:261] non-default args: {'max_model_len': 8192, 'enable_prefix_caching': False, 'gpu_memory_utilization': 0.6, 'max_num_batched_tokens': 8192, 'disable_log_stats': True, 'model': '/mnt/data/users/vlad.smirnov/hf_cache/hub/models--meta-llama--Llama-3.1-8B-Instruct/snapshots/0e9e39f249a16976918f6564b8830bc894c89659'}
4
+ INFO 10-31 14:13:13 [model.py:541] Resolved architecture: LlamaForCausalLM
5
+ INFO 10-31 14:13:13 [model.py:1561] Using max model len 8192
6
+ INFO 10-31 14:13:13 [scheduler.py:226] Chunked prefill is enabled with max_num_batched_tokens=8192.
7
+ INFO 10-31 14:13:13 [vllm.py:624] Asynchronous scheduling is enabled.
8
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:13:14 [core.py:96] Initializing a V1 LLM engine (v0.15.1) with config: model='/mnt/data/users/vlad.smirnov/hf_cache/hub/models--meta-llama--Llama-3.1-8B-Instruct/snapshots/0e9e39f249a16976918f6564b8830bc894c89659', speculative_config=None, tokenizer='/mnt/data/users/vlad.smirnov/hf_cache/hub/models--meta-llama--Llama-3.1-8B-Instruct/snapshots/0e9e39f249a16976918f6564b8830bc894c89659', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=8192, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=0, served_model_name=/mnt/data/users/vlad.smirnov/hf_cache/hub/models--meta-llama--Llama-3.1-8B-Instruct/snapshots/0e9e39f249a16976918f6564b8830bc894c89659, enable_prefix_caching=False, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': <CompilationMode.VLLM_COMPILE: 3>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['none'], 'splitting_ops': ['vllm::unified_attention', 'vllm::unified_attention_with_output', 'vllm::unified_mla_attention', 'vllm::unified_mla_attention_with_output', 'vllm::mamba_mixer2', 'vllm::mamba_mixer', 'vllm::short_conv', 'vllm::linear_attention', 'vllm::plamo2_mamba_mixer', 'vllm::gdn_attention_core', 'vllm::kda_attention', 'vllm::sparse_attn_indexer', 'vllm::rocm_aiter_sparse_attn_indexer', 'vllm::unified_kv_cache_update'], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [8192], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.FULL_AND_PIECEWISE: (2, 1)>, 'cudagraph_num_of_warmups': 1, 'cudagraph_capture_sizes': [1, 2, 4, 8, 16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96, 104, 112, 120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232, 240, 248, 256, 272, 288, 304, 320, 336, 352, 368, 384, 400, 416, 432, 448, 464, 480, 496, 512], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': True, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False}, 'max_cudagraph_capture_size': 512, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': True}, 'local_cache_dir': None, 'static_all_moe_layers': []}
9
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:13:14 [parallel_state.py:1212] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.127.105.184:39187 backend=nccl
10
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:13:14 [parallel_state.py:1423] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A
11
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:13:14 [gpu_model_runner.py:4033] Starting to load model /mnt/data/users/vlad.smirnov/hf_cache/hub/models--meta-llama--Llama-3.1-8B-Instruct/snapshots/0e9e39f249a16976918f6564b8830bc894c89659...
12
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:13:15 [cuda.py:364] Using FLASH_ATTN attention backend out of potential backends: ('FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION')
13
+ (EngineCore_DP0 pid=3779824)
14
+ (EngineCore_DP0 pid=3779824)
15
+ (EngineCore_DP0 pid=3779824)
16
+ (EngineCore_DP0 pid=3779824)
17
+ (EngineCore_DP0 pid=3779824)
18
+ (EngineCore_DP0 pid=3779824)
19
+ (EngineCore_DP0 pid=3779824)
20
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:13:46 [default_loader.py:291] Loading weights took 30.42 seconds
21
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:13:46 [gpu_model_runner.py:4130] Model loading took 14.99 GiB memory and 30.877245 seconds
22
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:13:49 [backends.py:812] Using cache directory: /home/vlad.smirnov/.cache/vllm/torch_compile_cache/a0fdf2c2f6/rank_0_0/backbone for vLLM's torch.compile
23
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:13:49 [backends.py:872] Dynamo bytecode transform time: 2.65 s
24
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:13:53 [backends.py:302] Cache the graph of compile range (1, 8192) for later use
25
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:13:57 [backends.py:319] Compiling a graph for compile range (1, 8192) takes 6.10 s
26
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:13:57 [monitor.py:34] torch.compile takes 8.74 s in total
27
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:13:58 [gpu_worker.py:356] Available KV cache memory: 12.16 GiB
28
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:13:58 [kv_cache_utils.py:1307] GPU KV cache size: 99,632 tokens
29
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:13:58 [kv_cache_utils.py:1312] Maximum concurrency for 8,192 tokens per request: 12.16x
30
+ (EngineCore_DP0 pid=3779824)
31
+ (EngineCore_DP0 pid=3779824)
32
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:14:02 [gpu_model_runner.py:5063] Graph capturing finished in 4 secs, took 0.52 GiB
33
+ (EngineCore_DP0 pid=3779824) INFO 10-31 14:14:02 [core.py:272] init engine (profile, create kv cache, warmup model) took 15.68 seconds
34
+ INFO 10-31 14:14:03 [llm.py:343] Supported tasks: ['generate']
35
+ [logprob] 64/10888
36
+ [logprob] 128/10888
37
+ [logprob] 192/10888
38
+ [logprob] 256/10888
39
+ [logprob] 320/10888
40
+ [logprob] 384/10888
41
+ [logprob] 448/10888
42
+ [logprob] 512/10888
43
+ [logprob] 576/10888
44
+ [logprob] 640/10888
45
+ [logprob] 704/10888
46
+ [logprob] 768/10888
47
+ [logprob] 832/10888
48
+ [logprob] 896/10888
49
+ [logprob] 960/10888
50
+ [logprob] 1024/10888
51
+ [logprob] 1088/10888
52
+ [logprob] 1152/10888
53
+ [logprob] 1216/10888
54
+ [logprob] 1280/10888
55
+ [logprob] 1344/10888
56
+ [logprob] 1408/10888
57
+ [logprob] 1472/10888
58
+ [logprob] 1536/10888
59
+ [logprob] 1600/10888
60
+ [logprob] 1664/10888
61
+ [logprob] 1728/10888
62
+ [logprob] 1792/10888
63
+ [logprob] 1856/10888
64
+ [logprob] 1920/10888
65
+ [logprob] 1984/10888
66
+ [logprob] 2048/10888
67
+ [logprob] 2112/10888
68
+ [logprob] 2176/10888
69
+ [logprob] 2240/10888
70
+ [logprob] 2304/10888
71
+ [logprob] 2368/10888
72
+ [logprob] 2432/10888
73
+ [logprob] 2496/10888
74
+ [logprob] 2560/10888
75
+ [logprob] 2624/10888
76
+ [logprob] 2688/10888
77
+ [logprob] 2752/10888
78
+ [logprob] 2816/10888
79
+ [logprob] 2880/10888
80
+ [logprob] 2944/10888
81
+ [logprob] 3008/10888
82
+ [logprob] 3072/10888
83
+ [logprob] 3136/10888
84
+ [logprob] 3200/10888
85
+ [logprob] 3264/10888
86
+ [logprob] 3328/10888
87
+ [logprob] 3392/10888
88
+ [logprob] 3456/10888
89
+ [logprob] 3520/10888
90
+ [logprob] 3584/10888
91
+ [logprob] 3648/10888
92
+ [logprob] 3712/10888
93
+ [logprob] 3776/10888
94
+ [logprob] 3840/10888
95
+ [logprob] 3904/10888
96
+ [logprob] 3968/10888
97
+ [logprob] 4032/10888
98
+ [logprob] 4096/10888
99
+ [logprob] 4160/10888
100
+ [logprob] 4224/10888
101
+ [logprob] 4288/10888
102
+ [logprob] 4352/10888
103
+ [logprob] 4416/10888
104
+ [logprob] 4480/10888
105
+ [logprob] 4544/10888
106
+ [logprob] 4608/10888
107
+ [logprob] 4672/10888
108
+ [logprob] 4736/10888
109
+ [logprob] 4800/10888
110
+ [logprob] 4864/10888
111
+ [logprob] 4928/10888
112
+ [logprob] 4992/10888
113
+ [logprob] 5056/10888
114
+ [logprob] 5120/10888
115
+ [logprob] 5184/10888
116
+ [logprob] 5248/10888
117
+ [logprob] 5312/10888
118
+ [logprob] 5376/10888
119
+ [logprob] 5440/10888
120
+ [logprob] 5504/10888
121
+ [logprob] 5568/10888
122
+ [logprob] 5632/10888
123
+ [logprob] 5696/10888
124
+ [logprob] 5760/10888
125
+ [logprob] 5824/10888
126
+ [logprob] 5888/10888
127
+ [logprob] 5952/10888
128
+ [logprob] 6016/10888
129
+ [logprob] 6080/10888
130
+ [logprob] 6144/10888
131
+ [logprob] 6208/10888
132
+ [logprob] 6272/10888
133
+ [logprob] 6336/10888
134
+ [logprob] 6400/10888
135
+ [logprob] 6464/10888
136
+ [logprob] 6528/10888
137
+ [logprob] 6592/10888
138
+ [logprob] 6656/10888
139
+ [logprob] 6720/10888
140
+ [logprob] 6784/10888
141
+ [logprob] 6848/10888
142
+ [logprob] 6912/10888
143
+ [logprob] 6976/10888
144
+ [logprob] 7040/10888
145
+ [logprob] 7104/10888
146
+ [logprob] 7168/10888
147
+ [logprob] 7232/10888
148
+ [logprob] 7296/10888
149
+ [logprob] 7360/10888
150
+ [logprob] 7424/10888
151
+ [logprob] 7488/10888
152
+ [logprob] 7552/10888
153
+ [logprob] 7616/10888
154
+ [logprob] 7680/10888
155
+ [logprob] 7744/10888
156
+ [logprob] 7808/10888
157
+ [logprob] 7872/10888
158
+ [logprob] 7936/10888
159
+ [logprob] 8000/10888
160
+ [logprob] 8064/10888
161
+ [logprob] 8128/10888
162
+ [logprob] 8192/10888
163
+ [logprob] 8256/10888
164
+ [logprob] 8320/10888
165
+ [logprob] 8384/10888
166
+ [logprob] 8448/10888
167
+ [logprob] 8512/10888
168
+ [logprob] 8576/10888
169
+ [logprob] 8640/10888
170
+ [logprob] 8704/10888
171
+ [logprob] 8768/10888
172
+ [logprob] 8832/10888
173
+ [logprob] 8896/10888
174
+ [logprob] 8960/10888
175
+ [logprob] 9024/10888
176
+ [logprob] 9088/10888
177
+ [logprob] 9152/10888
178
+ [logprob] 9216/10888
179
+ [logprob] 9280/10888
180
+ [logprob] 9344/10888
181
+ [logprob] 9408/10888
182
+ [logprob] 9472/10888
183
+ [logprob] 9536/10888
184
+ [logprob] 9600/10888
185
+ [logprob] 9664/10888
186
+ [logprob] 9728/10888
187
+ [logprob] 9792/10888
188
+ [logprob] 9856/10888
189
+ [logprob] 9920/10888
190
+ [logprob] 9984/10888
191
+ [logprob] 10048/10888
192
+ [logprob] 10112/10888
193
+ [logprob] 10176/10888
194
+ [logprob] 10240/10888
195
+ [logprob] 10304/10888
196
+ [logprob] 10368/10888
197
+ [logprob] 10432/10888
198
+ [logprob] 10496/10888
199
+ [logprob] 10560/10888
200
+ [logprob] 10624/10888
201
+ [logprob] 10688/10888
202
+ [logprob] 10752/10888
203
+ [logprob] 10816/10888
204
+ [logprob] 10880/10888
205
+ [logprob] 10888/10888
206
+ [W1031 14:47:26.975227637 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
207
+ [logprob] merged 10888 chains into lpmean, lpmin, negent, top1 selector files
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/logprob/shard0.provenance.json ADDED
@@ -0,0 +1,85 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T14:12:54+04:00",
3
+ "host": "ws006599",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/logprob_scores.py",
8
+ "--pool-dir",
9
+ "workdir/gpools/llama31_8b/chatpool",
10
+ "--model",
11
+ "meta-llama/Llama-3.1-8B-Instruct",
12
+ "--template",
13
+ "auto",
14
+ "--chains",
15
+ "0",
16
+ "1",
17
+ "2",
18
+ "3",
19
+ "4",
20
+ "5",
21
+ "6",
22
+ "7",
23
+ "--resume",
24
+ "--out",
25
+ "workdir/gpools/llama31_8b/logprob/shard0.jsonl"
26
+ ],
27
+ "git": {
28
+ "commit": "6cae6bf292a53eff08d1ca10850479123a1aae47",
29
+ "branch": "HEAD",
30
+ "subject": "logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases",
31
+ "dirty": false,
32
+ "dirty_files": []
33
+ },
34
+ "gpus": {
35
+ "visible": "1",
36
+ "devices": [
37
+ {
38
+ "index": "0",
39
+ "name": "NVIDIA RTX 5880 Ada Generation",
40
+ "memory": "49140 MiB",
41
+ "driver": "580.178.04",
42
+ "uuid": "GPU-ea30b823-34f2-1633-1958-47244a40ac5d"
43
+ },
44
+ {
45
+ "index": "1",
46
+ "name": "NVIDIA RTX 5880 Ada Generation",
47
+ "memory": "49140 MiB",
48
+ "driver": "580.178.04",
49
+ "uuid": "GPU-b843d197-f9cd-0190-cbf3-1ded8ca2f9a8"
50
+ }
51
+ ]
52
+ },
53
+ "packages": {
54
+ "python": "3.11.15",
55
+ "torch": "2.9.1",
56
+ "vllm": "0.15.1",
57
+ "transformers": "4.57.6",
58
+ "speculators": "0.4.0.dev49",
59
+ "peft": null,
60
+ "accelerate": "1.13.0",
61
+ "numpy": "2.2.6",
62
+ "huggingface_hub": "0.36.2",
63
+ "openai": "2.32.0",
64
+ "wandb": "0.25.1"
65
+ },
66
+ "env": {
67
+ "CUDA_VISIBLE_DEVICES": "1",
68
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
69
+ "HF_HUB_OFFLINE": "1",
70
+ "TRANSFORMERS_OFFLINE": "1",
71
+ "NCCL_P2P_DISABLE": "1",
72
+ "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"
73
+ },
74
+ "models": [],
75
+ "files": {
76
+ "pool_config": {
77
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/llama31_8b/chatpool/pool_config.json",
78
+ "bytes": 213,
79
+ "sha256": "2442b361015d3bdac6fd3c2c74e57151637e209015d0058bd595d0b955fd8d9e"
80
+ }
81
+ },
82
+ "finished_at": "2026-10-31T14:47:25+04:00",
83
+ "wall_secs": 2070.9,
84
+ "shard": "0/1"
85
+ }
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/prm7b.log ADDED
@@ -0,0 +1,165 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-10-30 13:33:48,096 INFO prm_selector: pool: 1361 problems, chains/problem [(8, 1361)]
2
+ 2026-10-30 13:33:48,297 INFO prm_selector: PRM: 10869 answered chains, 10869 to score, 0 cached
3
+ `torch_dtype` is deprecated! Use `dtype` instead!
4
+
5
+ Some weights of the model checkpoint at Qwen/Qwen2.5-Math-PRM-7B were not used when initializing Qwen2ForProcessRewardModel: ['lm_head.weight']
6
+ - This IS expected if you are initializing Qwen2ForProcessRewardModel from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
7
+ - This IS NOT expected if you are initializing Qwen2ForProcessRewardModel from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
8
+ 2026-10-30 13:33:55,963 INFO prm_selector: loaded PRM Qwen/Qwen2.5-Math-PRM-7B (bf16) on cuda
9
+ 2026-10-30 13:34:16,430 INFO prm_selector: PRM scored 8/10869 chains (12s)
10
+ 2026-10-30 13:36:33,326 INFO prm_selector: PRM scored 168/10869 chains (149s)
11
+ 2026-10-30 13:38:38,313 INFO prm_selector: PRM scored 328/10869 chains (274s)
12
+ 2026-10-30 13:40:14,535 INFO prm_selector: PRM scored 488/10869 chains (370s)
13
+ 2026-10-30 13:41:32,409 INFO prm_selector: PRM scored 648/10869 chains (448s)
14
+ 2026-10-30 13:42:37,889 INFO prm_selector: PRM scored 808/10869 chains (514s)
15
+ 2026-10-30 13:43:35,461 INFO prm_selector: PRM scored 968/10869 chains (571s)
16
+ 2026-10-30 13:44:27,004 INFO prm_selector: PRM scored 1128/10869 chains (623s)
17
+ 2026-10-30 13:45:14,333 INFO prm_selector: PRM scored 1288/10869 chains (670s)
18
+ 2026-10-30 13:45:58,292 INFO prm_selector: PRM scored 1448/10869 chains (714s)
19
+ 2026-10-30 13:46:39,236 INFO prm_selector: PRM scored 1608/10869 chains (755s)
20
+ 2026-10-30 13:47:18,043 INFO prm_selector: PRM scored 1768/10869 chains (794s)
21
+ 2026-10-30 13:47:54,515 INFO prm_selector: PRM scored 1928/10869 chains (830s)
22
+ 2026-10-30 13:48:29,159 INFO prm_selector: PRM scored 2088/10869 chains (865s)
23
+ 2026-10-30 13:49:02,274 INFO prm_selector: PRM scored 2248/10869 chains (898s)
24
+ 2026-10-30 13:49:33,941 INFO prm_selector: PRM scored 2408/10869 chains (930s)
25
+ 2026-10-30 13:50:04,324 INFO prm_selector: PRM scored 2568/10869 chains (960s)
26
+ 2026-10-30 13:50:33,871 INFO prm_selector: PRM scored 2728/10869 chains (990s)
27
+ 2026-10-30 13:51:02,350 INFO prm_selector: PRM scored 2888/10869 chains (1018s)
28
+ 2026-10-30 13:51:29,986 INFO prm_selector: PRM scored 3048/10869 chains (1046s)
29
+ 2026-10-30 13:51:56,672 INFO prm_selector: PRM scored 3208/10869 chains (1072s)
30
+ 2026-10-30 13:52:22,717 INFO prm_selector: PRM scored 3368/10869 chains (1098s)
31
+ 2026-10-30 13:52:47,957 INFO prm_selector: PRM scored 3528/10869 chains (1124s)
32
+ 2026-10-30 13:53:12,392 INFO prm_selector: PRM scored 3688/10869 chains (1148s)
33
+ 2026-10-30 13:53:36,262 INFO prm_selector: PRM scored 3848/10869 chains (1172s)
34
+ 2026-10-30 13:53:59,380 INFO prm_selector: PRM scored 4008/10869 chains (1195s)
35
+ 2026-10-30 13:54:21,905 INFO prm_selector: PRM scored 4168/10869 chains (1218s)
36
+ 2026-10-30 13:54:43,974 INFO prm_selector: PRM scored 4328/10869 chains (1240s)
37
+ 2026-10-30 13:55:05,463 INFO prm_selector: PRM scored 4488/10869 chains (1261s)
38
+ 2026-10-30 13:55:26,412 INFO prm_selector: PRM scored 4648/10869 chains (1282s)
39
+ 2026-10-30 13:55:46,831 INFO prm_selector: PRM scored 4808/10869 chains (1303s)
40
+ 2026-10-30 13:56:06,645 INFO prm_selector: PRM scored 4968/10869 chains (1322s)
41
+ 2026-10-30 13:56:26,040 INFO prm_selector: PRM scored 5128/10869 chains (1342s)
42
+ 2026-10-30 13:56:44,958 INFO prm_selector: PRM scored 5288/10869 chains (1361s)
43
+ 2026-10-30 13:57:03,391 INFO prm_selector: PRM scored 5448/10869 chains (1379s)
44
+ 2026-10-30 13:57:21,363 INFO prm_selector: PRM scored 5608/10869 chains (1397s)
45
+ 2026-10-30 13:57:39,005 INFO prm_selector: PRM scored 5768/10869 chains (1415s)
46
+ 2026-10-30 13:57:56,189 INFO prm_selector: PRM scored 5928/10869 chains (1432s)
47
+ 2026-10-30 13:58:13,083 INFO prm_selector: PRM scored 6088/10869 chains (1449s)
48
+ 2026-10-30 13:58:29,878 INFO prm_selector: PRM scored 6248/10869 chains (1466s)
49
+ 2026-10-30 13:58:46,156 INFO prm_selector: PRM scored 6408/10869 chains (1482s)
50
+ 2026-10-30 13:59:02,034 INFO prm_selector: PRM scored 6568/10869 chains (1498s)
51
+ 2026-10-30 13:59:17,779 INFO prm_selector: PRM scored 6728/10869 chains (1514s)
52
+ 2026-10-30 13:59:33,253 INFO prm_selector: PRM scored 6888/10869 chains (1529s)
53
+ 2026-10-30 13:59:48,127 INFO prm_selector: PRM scored 7048/10869 chains (1544s)
54
+ 2026-10-30 14:00:02,751 INFO prm_selector: PRM scored 7208/10869 chains (1558s)
55
+ 2026-10-30 14:00:17,171 INFO prm_selector: PRM scored 7368/10869 chains (1573s)
56
+ 2026-10-30 14:00:31,039 INFO prm_selector: PRM scored 7528/10869 chains (1587s)
57
+ 2026-10-30 14:00:44,488 INFO prm_selector: PRM scored 7688/10869 chains (1600s)
58
+ 2026-10-30 14:00:57,774 INFO prm_selector: PRM scored 7848/10869 chains (1614s)
59
+ 2026-10-30 14:01:10,889 INFO prm_selector: PRM scored 8008/10869 chains (1627s)
60
+ 2026-10-30 14:01:24,036 INFO prm_selector: PRM scored 8168/10869 chains (1640s)
61
+ 2026-10-30 14:01:36,509 INFO prm_selector: PRM scored 8328/10869 chains (1652s)
62
+ 2026-10-30 14:01:48,645 INFO prm_selector: PRM scored 8488/10869 chains (1664s)
63
+ 2026-10-30 14:02:00,501 INFO prm_selector: PRM scored 8648/10869 chains (1676s)
64
+ 2026-10-30 14:02:12,103 INFO prm_selector: PRM scored 8808/10869 chains (1688s)
65
+ 2026-10-30 14:02:23,574 INFO prm_selector: PRM scored 8968/10869 chains (1699s)
66
+ 2026-10-30 14:02:34,902 INFO prm_selector: PRM scored 9128/10869 chains (1711s)
67
+ 2026-10-30 14:02:45,573 INFO prm_selector: PRM scored 9288/10869 chains (1721s)
68
+ 2026-10-30 14:02:56,056 INFO prm_selector: PRM scored 9448/10869 chains (1732s)
69
+ 2026-10-30 14:03:06,414 INFO prm_selector: PRM scored 9608/10869 chains (1742s)
70
+ 2026-10-30 14:03:16,440 INFO prm_selector: PRM scored 9768/10869 chains (1752s)
71
+ 2026-10-30 14:03:26,101 INFO prm_selector: PRM scored 9928/10869 chains (1762s)
72
+ 2026-10-30 14:03:35,272 INFO prm_selector: PRM scored 10088/10869 chains (1771s)
73
+ 2026-10-30 14:03:44,298 INFO prm_selector: PRM scored 10248/10869 chains (1780s)
74
+ 2026-10-30 14:03:52,955 INFO prm_selector: PRM scored 10408/10869 chains (1789s)
75
+ 2026-10-30 14:04:01,031 INFO prm_selector: PRM scored 10568/10869 chains (1797s)
76
+ 2026-10-30 14:04:08,433 INFO prm_selector: PRM scored 10728/10869 chains (1804s)
77
+ 2026-10-30 14:04:13,953 INFO prm_selector: PRM scored 10869/10869 chains (1810s)
78
+ 2026-10-30 14:04:14,651 WARNING prm_selector: 1 chains truncated at --max-length, 1 with fewer rewards than steps
79
+ 2026-10-30 14:04:14,673 WARNING online_rl_train: Failed to load ThinkBooster evaluator: No module named 'thinkbooster.evaluation.latex2sympy'; using simple fallback
80
+ 2026-10-30 14:04:14,673 WARNING judge_selector: --answer-eq math: ThinkBooster grader not loaded, falling back to string equality
81
+
82
+ k=2 n=1361 majority=0.273 pass@k=0.391 recoverable=0 majority-right=210
83
+ method acc strict gap_cl recov regret hybrid h_gap
84
+ min 0.353 0.353 0.683 None 0.000 0.353 0.683
85
+ min_csum 0.353 0.353 0.683 None 0.000 0.353 0.683
86
+ min_cmax 0.353 0.353 0.683 None 0.000 0.353 0.683
87
+ last 0.340 0.340 0.571 None 0.000 0.340 0.571
88
+ last_csum 0.340 0.340 0.571 None 0.000 0.340 0.571
89
+ last_cmax 0.340 0.340 0.571 None 0.000 0.340 0.571
90
+ prod 0.353 0.353 0.677 None 0.000 0.353 0.677
91
+ prod_csum 0.353 0.353 0.677 None 0.000 0.353 0.677
92
+ prod_cmax 0.353 0.353 0.677 None 0.000 0.353 0.677
93
+ mean 0.346 0.346 0.621 None 0.000 0.346 0.621
94
+ mean_csum 0.346 0.346 0.621 None 0.000 0.346 0.621
95
+ mean_cmax 0.346 0.346 0.621 None 0.000 0.346 0.621
96
+ within-problem AUC: min 0.842, last 0.786, prod 0.839, mean 0.811 (n_mixed=322)
97
+
98
+ k=4 n=1361 majority=0.339 pass@k=0.507 recoverable=63 majority-right=398
99
+ method acc strict gap_cl recov regret hybrid h_gap
100
+ min 0.409 0.409 0.418 0.413 0.055 0.409 0.418
101
+ min_csum 0.415 0.415 0.453 0.365 0.030 0.415 0.453
102
+ min_cmax 0.409 0.409 0.418 0.413 0.055 0.409 0.418
103
+ last 0.406 0.406 0.401 0.540 0.055 0.406 0.401
104
+ last_csum 0.400 0.400 0.366 0.270 0.030 0.400 0.366
105
+ last_cmax 0.406 0.406 0.401 0.540 0.055 0.406 0.401
106
+ prod 0.411 0.411 0.427 0.444 0.043 0.411 0.427
107
+ prod_csum 0.411 0.411 0.427 0.397 0.035 0.411 0.427
108
+ prod_cmax 0.411 0.411 0.427 0.444 0.043 0.411 0.427
109
+ mean 0.408 0.408 0.409 0.476 0.045 0.408 0.409
110
+ mean_csum 0.403 0.403 0.379 0.159 0.010 0.403 0.379
111
+ mean_cmax 0.408 0.408 0.409 0.476 0.045 0.408 0.409
112
+ within-problem AUC: min 0.827, last 0.799, prod 0.824, mean 0.809 (n_mixed=564)
113
+
114
+ k=8 n=1361 majority=0.407 pass@k=0.622 recoverable=177 majority-right=508
115
+ method acc strict gap_cl recov regret hybrid h_gap
116
+ min 0.461 0.461 0.255 0.435 0.096 0.461 0.255
117
+ min_csum 0.470 0.470 0.296 0.379 0.053 0.470 0.296
118
+ min_cmax 0.461 0.461 0.255 0.435 0.096 0.461 0.255
119
+ last 0.440 0.440 0.156 0.395 0.112 0.440 0.156
120
+ last_csum 0.447 0.447 0.190 0.226 0.043 0.447 0.190
121
+ last_cmax 0.440 0.440 0.156 0.395 0.112 0.440 0.156
122
+ prod 0.468 0.468 0.286 0.469 0.083 0.468 0.286
123
+ prod_csum 0.474 0.474 0.313 0.418 0.047 0.474 0.313
124
+ prod_cmax 0.468 0.468 0.286 0.469 0.083 0.468 0.286
125
+ mean 0.455 0.455 0.224 0.395 0.089 0.455 0.224
126
+ mean_csum 0.452 0.452 0.211 0.198 0.031 0.452 0.211
127
+ mean_cmax 0.455 0.455 0.224 0.395 0.089 0.455 0.224
128
+ within-problem AUC: min 0.803, last 0.780, prod 0.812, mean 0.793 (n_mixed=775)
129
+ 2026-10-30 14:04:19,091 INFO prm_selector: done in 1831s
130
+ 2026-10-30 14:04:19,597 INFO prm_selector: wrote workdir/gpools/llama31_8b/base/prm7b_selector.json
131
+ wandb: [wandb.login()] Loaded credentials for https://api.wandb.ai from /home/vlad.smirnov/.netrc.
132
+ wandb: Currently logged in as: karantonis (nlpresearch.group) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
133
+ wandb: WARNING Using a boolean value for 'reinit' is deprecated. Use 'return_previous' or 'finish_previous' instead.
134
+ wandb: setting up run 97zmb3i5
135
+ wandb: Tracking run with wandb version 0.25.1
136
+ wandb: Run data is saved locally in /mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/wandb/run-20261030_140422-97zmb3i5
137
+ wandb: Run `wandb offline` to turn off syncing.
138
+ wandb: Syncing run prm-sel-chatpool
139
+ wandb: ⭐️ View project at https://wandb.ai/mbzuai-research/llm-reasoning-rl-controller
140
+ wandb: 🚀 View run at https://wandb.ai/mbzuai-research/llm-reasoning-rl-controller/runs/97zmb3i5
141
+ wandb: WARNING The get_url method is deprecated and will be removed in a future release. Please use `run.url` instead.
142
+ 2026-10-30 14:04:25,395 INFO wandb_log: W&B run prm-sel-chatpool: https://wandb.ai/mbzuai-research/llm-reasoning-rl-controller/runs/97zmb3i5
143
+ wandb: updating run metadata; uploading console lines 0-1
144
+ wandb: uploading console lines 0-1
145
+ wandb: uploading wandb-summary.json; uploading config.yaml; uploading wandb-metadata.json; uploading requirements.txt; uploading output.log
146
+ wandb: uploading summary, console lines 1-1
147
+ wandb:
148
+ wandb: Run summary:
149
+ wandb: results/k2/majority_acc 0.27259
150
+ wandb: results/k2/majority_ties 1096
151
+ wandb: results/k2/methods/last/distinct_answers_hist/1 265
152
+ wandb: results/k2/methods/last/distinct_answers_hist/2 1096
153
+ wandb: results/k2/methods/last/gap_closed 0.57143
154
+ wandb: results/k2/methods/last/hybrid_acc 0.34019
155
+ wandb: results/k2/methods/last/hybrid_frac_called 0.80529
156
+ wandb: results/k2/methods/last/hybrid_gap_closed 0.57143
157
+ wandb: results/k2/methods/last/inconsistent_clusters 0
158
+ wandb: results/k2/methods/last/majority_acc 0.27259
159
+ wandb: +1,337 ...
160
+ wandb:
161
+ wandb: 🚀 View run prm-sel-chatpool at: https://wandb.ai/mbzuai-research/llm-reasoning-rl-controller/runs/97zmb3i5
162
+ wandb: ⭐️ View project at: https://wandb.ai/mbzuai-research/llm-reasoning-rl-controller
163
+ wandb: Synced 5 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
164
+ wandb: Find logs at: ./wandb/run-20261030_140422-97zmb3i5/logs
165
+ [W1030 14:04:28.311846690 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/base/lpmean_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/base/lpmin_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/base/negent_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/base/top1_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/lpmean_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/lpmean_selector.provenance.json ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T15:05:48+04:00",
3
+ "host": "ws006599",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/logprob_scores.py",
8
+ "--pool-dir",
9
+ "workdir/gpools/qwen3_1.7b/chatpool",
10
+ "--model",
11
+ "Qwen/Qwen3-1.7B",
12
+ "--template",
13
+ "qwen",
14
+ "--merge",
15
+ "workdir/gpools/qwen3_1.7b/logprob/shard0.jsonl"
16
+ ],
17
+ "git": {
18
+ "commit": "6cae6bf292a53eff08d1ca10850479123a1aae47",
19
+ "branch": "HEAD",
20
+ "subject": "logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases",
21
+ "dirty": false,
22
+ "dirty_files": []
23
+ },
24
+ "gpus": {
25
+ "visible": "0",
26
+ "devices": [
27
+ {
28
+ "index": "0",
29
+ "name": "NVIDIA RTX 5880 Ada Generation",
30
+ "memory": "49140 MiB",
31
+ "driver": "580.178.04",
32
+ "uuid": "GPU-ea30b823-34f2-1633-1958-47244a40ac5d"
33
+ },
34
+ {
35
+ "index": "1",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-b843d197-f9cd-0190-cbf3-1ded8ca2f9a8"
40
+ }
41
+ ]
42
+ },
43
+ "packages": {
44
+ "python": "3.11.15",
45
+ "torch": "2.9.1",
46
+ "vllm": "0.15.1",
47
+ "transformers": "4.57.6",
48
+ "speculators": "0.4.0.dev49",
49
+ "peft": null,
50
+ "accelerate": "1.13.0",
51
+ "numpy": "2.2.6",
52
+ "huggingface_hub": "0.36.2",
53
+ "openai": "2.32.0",
54
+ "wandb": "0.25.1"
55
+ },
56
+ "env": {
57
+ "CUDA_VISIBLE_DEVICES": "0",
58
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
59
+ "NCCL_P2P_DISABLE": "1",
60
+ "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"
61
+ },
62
+ "models": [],
63
+ "files": {
64
+ "pool_config": {
65
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/qwen3_1.7b/chatpool/pool_config.json",
66
+ "bytes": 213,
67
+ "sha256": "7fd98e1bdd36c2ec95d3123f1b8ad66d74d1fb31ab98fea34112204b305ef6fe"
68
+ }
69
+ },
70
+ "finished_at": "2026-10-31T15:05:48+04:00",
71
+ "wall_secs": 0.3,
72
+ "chains": 10888
73
+ }
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/lpmin_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/lpmin_selector.provenance.json ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T15:05:48+04:00",
3
+ "host": "ws006599",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/logprob_scores.py",
8
+ "--pool-dir",
9
+ "workdir/gpools/qwen3_1.7b/chatpool",
10
+ "--model",
11
+ "Qwen/Qwen3-1.7B",
12
+ "--template",
13
+ "qwen",
14
+ "--merge",
15
+ "workdir/gpools/qwen3_1.7b/logprob/shard0.jsonl"
16
+ ],
17
+ "git": {
18
+ "commit": "6cae6bf292a53eff08d1ca10850479123a1aae47",
19
+ "branch": "HEAD",
20
+ "subject": "logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases",
21
+ "dirty": false,
22
+ "dirty_files": []
23
+ },
24
+ "gpus": {
25
+ "visible": "0",
26
+ "devices": [
27
+ {
28
+ "index": "0",
29
+ "name": "NVIDIA RTX 5880 Ada Generation",
30
+ "memory": "49140 MiB",
31
+ "driver": "580.178.04",
32
+ "uuid": "GPU-ea30b823-34f2-1633-1958-47244a40ac5d"
33
+ },
34
+ {
35
+ "index": "1",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-b843d197-f9cd-0190-cbf3-1ded8ca2f9a8"
40
+ }
41
+ ]
42
+ },
43
+ "packages": {
44
+ "python": "3.11.15",
45
+ "torch": "2.9.1",
46
+ "vllm": "0.15.1",
47
+ "transformers": "4.57.6",
48
+ "speculators": "0.4.0.dev49",
49
+ "peft": null,
50
+ "accelerate": "1.13.0",
51
+ "numpy": "2.2.6",
52
+ "huggingface_hub": "0.36.2",
53
+ "openai": "2.32.0",
54
+ "wandb": "0.25.1"
55
+ },
56
+ "env": {
57
+ "CUDA_VISIBLE_DEVICES": "0",
58
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
59
+ "NCCL_P2P_DISABLE": "1",
60
+ "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"
61
+ },
62
+ "models": [],
63
+ "files": {
64
+ "pool_config": {
65
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/qwen3_1.7b/chatpool/pool_config.json",
66
+ "bytes": 213,
67
+ "sha256": "7fd98e1bdd36c2ec95d3123f1b8ad66d74d1fb31ab98fea34112204b305ef6fe"
68
+ }
69
+ },
70
+ "finished_at": "2026-10-31T15:05:48+04:00",
71
+ "wall_secs": 0.4,
72
+ "chains": 10888
73
+ }
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/negent_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/negent_selector.provenance.json ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T15:05:48+04:00",
3
+ "host": "ws006599",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/logprob_scores.py",
8
+ "--pool-dir",
9
+ "workdir/gpools/qwen3_1.7b/chatpool",
10
+ "--model",
11
+ "Qwen/Qwen3-1.7B",
12
+ "--template",
13
+ "qwen",
14
+ "--merge",
15
+ "workdir/gpools/qwen3_1.7b/logprob/shard0.jsonl"
16
+ ],
17
+ "git": {
18
+ "commit": "6cae6bf292a53eff08d1ca10850479123a1aae47",
19
+ "branch": "HEAD",
20
+ "subject": "logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases",
21
+ "dirty": false,
22
+ "dirty_files": []
23
+ },
24
+ "gpus": {
25
+ "visible": "0",
26
+ "devices": [
27
+ {
28
+ "index": "0",
29
+ "name": "NVIDIA RTX 5880 Ada Generation",
30
+ "memory": "49140 MiB",
31
+ "driver": "580.178.04",
32
+ "uuid": "GPU-ea30b823-34f2-1633-1958-47244a40ac5d"
33
+ },
34
+ {
35
+ "index": "1",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-b843d197-f9cd-0190-cbf3-1ded8ca2f9a8"
40
+ }
41
+ ]
42
+ },
43
+ "packages": {
44
+ "python": "3.11.15",
45
+ "torch": "2.9.1",
46
+ "vllm": "0.15.1",
47
+ "transformers": "4.57.6",
48
+ "speculators": "0.4.0.dev49",
49
+ "peft": null,
50
+ "accelerate": "1.13.0",
51
+ "numpy": "2.2.6",
52
+ "huggingface_hub": "0.36.2",
53
+ "openai": "2.32.0",
54
+ "wandb": "0.25.1"
55
+ },
56
+ "env": {
57
+ "CUDA_VISIBLE_DEVICES": "0",
58
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
59
+ "NCCL_P2P_DISABLE": "1",
60
+ "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"
61
+ },
62
+ "models": [],
63
+ "files": {
64
+ "pool_config": {
65
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/qwen3_1.7b/chatpool/pool_config.json",
66
+ "bytes": 213,
67
+ "sha256": "7fd98e1bdd36c2ec95d3123f1b8ad66d74d1fb31ab98fea34112204b305ef6fe"
68
+ }
69
+ },
70
+ "finished_at": "2026-10-31T15:05:48+04:00",
71
+ "wall_secs": 0.5,
72
+ "chains": 10888
73
+ }
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/top1_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/chatpool/top1_selector.provenance.json ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T15:05:48+04:00",
3
+ "host": "ws006599",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/logprob_scores.py",
8
+ "--pool-dir",
9
+ "workdir/gpools/qwen3_1.7b/chatpool",
10
+ "--model",
11
+ "Qwen/Qwen3-1.7B",
12
+ "--template",
13
+ "qwen",
14
+ "--merge",
15
+ "workdir/gpools/qwen3_1.7b/logprob/shard0.jsonl"
16
+ ],
17
+ "git": {
18
+ "commit": "6cae6bf292a53eff08d1ca10850479123a1aae47",
19
+ "branch": "HEAD",
20
+ "subject": "logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases",
21
+ "dirty": false,
22
+ "dirty_files": []
23
+ },
24
+ "gpus": {
25
+ "visible": "0",
26
+ "devices": [
27
+ {
28
+ "index": "0",
29
+ "name": "NVIDIA RTX 5880 Ada Generation",
30
+ "memory": "49140 MiB",
31
+ "driver": "580.178.04",
32
+ "uuid": "GPU-ea30b823-34f2-1633-1958-47244a40ac5d"
33
+ },
34
+ {
35
+ "index": "1",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-b843d197-f9cd-0190-cbf3-1ded8ca2f9a8"
40
+ }
41
+ ]
42
+ },
43
+ "packages": {
44
+ "python": "3.11.15",
45
+ "torch": "2.9.1",
46
+ "vllm": "0.15.1",
47
+ "transformers": "4.57.6",
48
+ "speculators": "0.4.0.dev49",
49
+ "peft": null,
50
+ "accelerate": "1.13.0",
51
+ "numpy": "2.2.6",
52
+ "huggingface_hub": "0.36.2",
53
+ "openai": "2.32.0",
54
+ "wandb": "0.25.1"
55
+ },
56
+ "env": {
57
+ "CUDA_VISIBLE_DEVICES": "0",
58
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
59
+ "NCCL_P2P_DISABLE": "1",
60
+ "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"
61
+ },
62
+ "models": [],
63
+ "files": {
64
+ "pool_config": {
65
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/qwen3_1.7b/chatpool/pool_config.json",
66
+ "bytes": 213,
67
+ "sha256": "7fd98e1bdd36c2ec95d3123f1b8ad66d74d1fb31ab98fea34112204b305ef6fe"
68
+ }
69
+ },
70
+ "finished_at": "2026-10-31T15:05:48+04:00",
71
+ "wall_secs": 0.6,
72
+ "chains": 10888
73
+ }
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/logprob/shard0.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/logprob/shard0.log ADDED
@@ -0,0 +1,204 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [logprob] shard 0/1: 10888 chains over 1361 problems (0 already scored)
2
+ INFO 10-31 14:42:41 [utils.py:261] non-default args: {'max_model_len': 8192, 'enable_prefix_caching': False, 'gpu_memory_utilization': 0.6, 'max_num_batched_tokens': 8192, 'disable_log_stats': True, 'model': 'Qwen/Qwen3-1.7B'}
3
+ INFO 10-31 14:42:42 [model.py:541] Resolved architecture: Qwen3ForCausalLM
4
+ INFO 10-31 14:42:42 [model.py:1561] Using max model len 8192
5
+ INFO 10-31 14:42:42 [scheduler.py:226] Chunked prefill is enabled with max_num_batched_tokens=8192.
6
+ INFO 10-31 14:42:42 [vllm.py:624] Asynchronous scheduling is enabled.
7
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:42:44 [core.py:96] Initializing a V1 LLM engine (v0.15.1) with config: model='Qwen/Qwen3-1.7B', speculative_config=None, tokenizer='Qwen/Qwen3-1.7B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=8192, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=0, served_model_name=Qwen/Qwen3-1.7B, enable_prefix_caching=False, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': <CompilationMode.VLLM_COMPILE: 3>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['none'], 'splitting_ops': ['vllm::unified_attention', 'vllm::unified_attention_with_output', 'vllm::unified_mla_attention', 'vllm::unified_mla_attention_with_output', 'vllm::mamba_mixer2', 'vllm::mamba_mixer', 'vllm::short_conv', 'vllm::linear_attention', 'vllm::plamo2_mamba_mixer', 'vllm::gdn_attention_core', 'vllm::kda_attention', 'vllm::sparse_attn_indexer', 'vllm::rocm_aiter_sparse_attn_indexer', 'vllm::unified_kv_cache_update'], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [8192], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.FULL_AND_PIECEWISE: (2, 1)>, 'cudagraph_num_of_warmups': 1, 'cudagraph_capture_sizes': [1, 2, 4, 8, 16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96, 104, 112, 120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232, 240, 248, 256, 272, 288, 304, 320, 336, 352, 368, 384, 400, 416, 432, 448, 464, 480, 496, 512], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': True, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False}, 'max_cudagraph_capture_size': 512, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': True}, 'local_cache_dir': None, 'static_all_moe_layers': []}
8
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:42:46 [parallel_state.py:1212] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.127.105.184:42489 backend=nccl
9
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:42:46 [parallel_state.py:1423] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A
10
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:42:46 [gpu_model_runner.py:4033] Starting to load model Qwen/Qwen3-1.7B...
11
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:42:47 [cuda.py:364] Using FLASH_ATTN attention backend out of potential backends: ('FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION')
12
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:43:34 [weight_utils.py:527] Time spent downloading weights for Qwen/Qwen3-1.7B: 46.344710 seconds
13
+ (EngineCore_DP0 pid=3800666)
14
+ (EngineCore_DP0 pid=3800666)
15
+ (EngineCore_DP0 pid=3800666)
16
+ (EngineCore_DP0 pid=3800666)
17
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:43:35 [default_loader.py:291] Loading weights took 0.46 seconds
18
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:43:35 [gpu_model_runner.py:4130] Model loading took 3.21 GiB memory and 48.370296 seconds
19
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:43:38 [backends.py:812] Using cache directory: /home/vlad.smirnov/.cache/vllm/torch_compile_cache/306b753730/rank_0_0/backbone for vLLM's torch.compile
20
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:43:38 [backends.py:872] Dynamo bytecode transform time: 2.83 s
21
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:43:46 [backends.py:302] Cache the graph of compile range (1, 8192) for later use
22
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:43:50 [backends.py:319] Compiling a graph for compile range (1, 8192) takes 10.47 s
23
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:43:50 [monitor.py:34] torch.compile takes 13.30 s in total
24
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:43:51 [gpu_worker.py:356] Available KV cache memory: 23.76 GiB
25
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:43:51 [kv_cache_utils.py:1307] GPU KV cache size: 222,400 tokens
26
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:43:51 [kv_cache_utils.py:1312] Maximum concurrency for 8,192 tokens per request: 27.15x
27
+ (EngineCore_DP0 pid=3800666)
28
+ (EngineCore_DP0 pid=3800666)
29
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:43:54 [gpu_model_runner.py:5063] Graph capturing finished in 2 secs, took 0.45 GiB
30
+ (EngineCore_DP0 pid=3800666) INFO 10-31 14:43:54 [core.py:272] init engine (profile, create kv cache, warmup model) took 18.50 seconds
31
+ INFO 10-31 14:43:56 [llm.py:343] Supported tasks: ['generate']
32
+ [logprob] 64/10888
33
+ [logprob] 128/10888
34
+ [logprob] 192/10888
35
+ [logprob] 256/10888
36
+ [logprob] 320/10888
37
+ [logprob] 384/10888
38
+ [logprob] 448/10888
39
+ [logprob] 512/10888
40
+ [logprob] 576/10888
41
+ [logprob] 640/10888
42
+ [logprob] 704/10888
43
+ [logprob] 768/10888
44
+ [logprob] 832/10888
45
+ [logprob] 896/10888
46
+ [logprob] 960/10888
47
+ [logprob] 1024/10888
48
+ [logprob] 1088/10888
49
+ [logprob] 1152/10888
50
+ [logprob] 1216/10888
51
+ [logprob] 1280/10888
52
+ [logprob] 1344/10888
53
+ [logprob] 1408/10888
54
+ [logprob] 1472/10888
55
+ [logprob] 1536/10888
56
+ [logprob] 1600/10888
57
+ [logprob] 1664/10888
58
+ [logprob] 1728/10888
59
+ [logprob] 1792/10888
60
+ [logprob] 1856/10888
61
+ [logprob] 1920/10888
62
+ [logprob] 1984/10888
63
+ [logprob] 2048/10888
64
+ [logprob] 2112/10888
65
+ [logprob] 2176/10888
66
+ [logprob] 2240/10888
67
+ [logprob] 2304/10888
68
+ [logprob] 2368/10888
69
+ [logprob] 2432/10888
70
+ [logprob] 2496/10888
71
+ [logprob] 2560/10888
72
+ [logprob] 2624/10888
73
+ [logprob] 2688/10888
74
+ [logprob] 2752/10888
75
+ [logprob] 2816/10888
76
+ [logprob] 2880/10888
77
+ [logprob] 2944/10888
78
+ [logprob] 3008/10888
79
+ [logprob] 3072/10888
80
+ [logprob] 3136/10888
81
+ [logprob] 3200/10888
82
+ [logprob] 3264/10888
83
+ [logprob] 3328/10888
84
+ [logprob] 3392/10888
85
+ [logprob] 3456/10888
86
+ [logprob] 3520/10888
87
+ [logprob] 3584/10888
88
+ [logprob] 3648/10888
89
+ [logprob] 3712/10888
90
+ [logprob] 3776/10888
91
+ [logprob] 3840/10888
92
+ [logprob] 3904/10888
93
+ [logprob] 3968/10888
94
+ [logprob] 4032/10888
95
+ [logprob] 4096/10888
96
+ [logprob] 4160/10888
97
+ [logprob] 4224/10888
98
+ [logprob] 4288/10888
99
+ [logprob] 4352/10888
100
+ [logprob] 4416/10888
101
+ [logprob] 4480/10888
102
+ [logprob] 4544/10888
103
+ [logprob] 4608/10888
104
+ [logprob] 4672/10888
105
+ [logprob] 4736/10888
106
+ [logprob] 4800/10888
107
+ [logprob] 4864/10888
108
+ [logprob] 4928/10888
109
+ [logprob] 4992/10888
110
+ [logprob] 5056/10888
111
+ [logprob] 5120/10888
112
+ [logprob] 5184/10888
113
+ [logprob] 5248/10888
114
+ [logprob] 5312/10888
115
+ [logprob] 5376/10888
116
+ [logprob] 5440/10888
117
+ [logprob] 5504/10888
118
+ [logprob] 5568/10888
119
+ [logprob] 5632/10888
120
+ [logprob] 5696/10888
121
+ [logprob] 5760/10888
122
+ [logprob] 5824/10888
123
+ [logprob] 5888/10888
124
+ [logprob] 5952/10888
125
+ [logprob] 6016/10888
126
+ [logprob] 6080/10888
127
+ [logprob] 6144/10888
128
+ [logprob] 6208/10888
129
+ [logprob] 6272/10888
130
+ [logprob] 6336/10888
131
+ [logprob] 6400/10888
132
+ [logprob] 6464/10888
133
+ [logprob] 6528/10888
134
+ [logprob] 6592/10888
135
+ [logprob] 6656/10888
136
+ [logprob] 6720/10888
137
+ [logprob] 6784/10888
138
+ [logprob] 6848/10888
139
+ [logprob] 6912/10888
140
+ [logprob] 6976/10888
141
+ [logprob] 7040/10888
142
+ [logprob] 7104/10888
143
+ [logprob] 7168/10888
144
+ [logprob] 7232/10888
145
+ [logprob] 7296/10888
146
+ [logprob] 7360/10888
147
+ [logprob] 7424/10888
148
+ [logprob] 7488/10888
149
+ [logprob] 7552/10888
150
+ [logprob] 7616/10888
151
+ [logprob] 7680/10888
152
+ [logprob] 7744/10888
153
+ [logprob] 7808/10888
154
+ [logprob] 7872/10888
155
+ [logprob] 7936/10888
156
+ [logprob] 8000/10888
157
+ [logprob] 8064/10888
158
+ [logprob] 8128/10888
159
+ [logprob] 8192/10888
160
+ [logprob] 8256/10888
161
+ [logprob] 8320/10888
162
+ [logprob] 8384/10888
163
+ [logprob] 8448/10888
164
+ [logprob] 8512/10888
165
+ [logprob] 8576/10888
166
+ [logprob] 8640/10888
167
+ [logprob] 8704/10888
168
+ [logprob] 8768/10888
169
+ [logprob] 8832/10888
170
+ [logprob] 8896/10888
171
+ [logprob] 8960/10888
172
+ [logprob] 9024/10888
173
+ [logprob] 9088/10888
174
+ [logprob] 9152/10888
175
+ [logprob] 9216/10888
176
+ [logprob] 9280/10888
177
+ [logprob] 9344/10888
178
+ [logprob] 9408/10888
179
+ [logprob] 9472/10888
180
+ [logprob] 9536/10888
181
+ [logprob] 9600/10888
182
+ [logprob] 9664/10888
183
+ [logprob] 9728/10888
184
+ [logprob] 9792/10888
185
+ [logprob] 9856/10888
186
+ [logprob] 9920/10888
187
+ [logprob] 9984/10888
188
+ [logprob] 10048/10888
189
+ [logprob] 10112/10888
190
+ [logprob] 10176/10888
191
+ [logprob] 10240/10888
192
+ [logprob] 10304/10888
193
+ [logprob] 10368/10888
194
+ [logprob] 10432/10888
195
+ [logprob] 10496/10888
196
+ [logprob] 10560/10888
197
+ [logprob] 10624/10888
198
+ [logprob] 10688/10888
199
+ [logprob] 10752/10888
200
+ [logprob] 10816/10888
201
+ [logprob] 10880/10888
202
+ [logprob] 10888/10888
203
+ [W1031 15:05:48.919635456 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
204
+ [logprob] merged 10888 chains into lpmean, lpmin, negent, top1 selector files
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/logprob/shard0.provenance.json ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T14:42:16+04:00",
3
+ "host": "ws006599",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/logprob_scores.py",
8
+ "--pool-dir",
9
+ "workdir/gpools/qwen3_1.7b/chatpool",
10
+ "--model",
11
+ "Qwen/Qwen3-1.7B",
12
+ "--template",
13
+ "qwen",
14
+ "--chains",
15
+ "0",
16
+ "1",
17
+ "2",
18
+ "3",
19
+ "4",
20
+ "5",
21
+ "6",
22
+ "7",
23
+ "--resume",
24
+ "--out",
25
+ "workdir/gpools/qwen3_1.7b/logprob/shard0.jsonl"
26
+ ],
27
+ "git": {
28
+ "commit": "6cae6bf292a53eff08d1ca10850479123a1aae47",
29
+ "branch": "HEAD",
30
+ "subject": "logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases",
31
+ "dirty": false,
32
+ "dirty_files": []
33
+ },
34
+ "gpus": {
35
+ "visible": "0",
36
+ "devices": [
37
+ {
38
+ "index": "0",
39
+ "name": "NVIDIA RTX 5880 Ada Generation",
40
+ "memory": "49140 MiB",
41
+ "driver": "580.178.04",
42
+ "uuid": "GPU-ea30b823-34f2-1633-1958-47244a40ac5d"
43
+ },
44
+ {
45
+ "index": "1",
46
+ "name": "NVIDIA RTX 5880 Ada Generation",
47
+ "memory": "49140 MiB",
48
+ "driver": "580.178.04",
49
+ "uuid": "GPU-b843d197-f9cd-0190-cbf3-1ded8ca2f9a8"
50
+ }
51
+ ]
52
+ },
53
+ "packages": {
54
+ "python": "3.11.15",
55
+ "torch": "2.9.1",
56
+ "vllm": "0.15.1",
57
+ "transformers": "4.57.6",
58
+ "speculators": "0.4.0.dev49",
59
+ "peft": null,
60
+ "accelerate": "1.13.0",
61
+ "numpy": "2.2.6",
62
+ "huggingface_hub": "0.36.2",
63
+ "openai": "2.32.0",
64
+ "wandb": "0.25.1"
65
+ },
66
+ "env": {
67
+ "CUDA_VISIBLE_DEVICES": "0",
68
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
69
+ "NCCL_P2P_DISABLE": "1",
70
+ "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"
71
+ },
72
+ "models": [],
73
+ "files": {
74
+ "pool_config": {
75
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/qwen3_1.7b/chatpool/pool_config.json",
76
+ "bytes": 213,
77
+ "sha256": "7fd98e1bdd36c2ec95d3123f1b8ad66d74d1fb31ab98fea34112204b305ef6fe"
78
+ }
79
+ },
80
+ "finished_at": "2026-10-31T15:05:47+04:00",
81
+ "wall_secs": 1411.0,
82
+ "shard": "0/1"
83
+ }
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/base/lpmean_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/base/lpmin_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/base/negent_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/base/top1_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/lpmean_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/lpmean_selector.provenance.json ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T14:42:15+04:00",
3
+ "host": "ws006599",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/logprob_scores.py",
8
+ "--pool-dir",
9
+ "workdir/gpools/qwen3_4b/chatpool",
10
+ "--model",
11
+ "Qwen/Qwen3-4B",
12
+ "--template",
13
+ "qwen",
14
+ "--merge",
15
+ "workdir/gpools/qwen3_4b/logprob/shard0.jsonl"
16
+ ],
17
+ "git": {
18
+ "commit": "6cae6bf292a53eff08d1ca10850479123a1aae47",
19
+ "branch": "HEAD",
20
+ "subject": "logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases",
21
+ "dirty": false,
22
+ "dirty_files": []
23
+ },
24
+ "gpus": {
25
+ "visible": "0",
26
+ "devices": [
27
+ {
28
+ "index": "0",
29
+ "name": "NVIDIA RTX 5880 Ada Generation",
30
+ "memory": "49140 MiB",
31
+ "driver": "580.178.04",
32
+ "uuid": "GPU-ea30b823-34f2-1633-1958-47244a40ac5d"
33
+ },
34
+ {
35
+ "index": "1",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-b843d197-f9cd-0190-cbf3-1ded8ca2f9a8"
40
+ }
41
+ ]
42
+ },
43
+ "packages": {
44
+ "python": "3.11.15",
45
+ "torch": "2.9.1",
46
+ "vllm": "0.15.1",
47
+ "transformers": "4.57.6",
48
+ "speculators": "0.4.0.dev49",
49
+ "peft": null,
50
+ "accelerate": "1.13.0",
51
+ "numpy": "2.2.6",
52
+ "huggingface_hub": "0.36.2",
53
+ "openai": "2.32.0",
54
+ "wandb": "0.25.1"
55
+ },
56
+ "env": {
57
+ "CUDA_VISIBLE_DEVICES": "0",
58
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
59
+ "NCCL_P2P_DISABLE": "1",
60
+ "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"
61
+ },
62
+ "models": [],
63
+ "files": {
64
+ "pool_config": {
65
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/qwen3_4b/chatpool/pool_config.json",
66
+ "bytes": 211,
67
+ "sha256": "be24fd70b2ed8c5b0dc17d78ea9f6fbaa86a62becd6b2cfa8809fb6e38604285"
68
+ }
69
+ },
70
+ "finished_at": "2026-10-31T14:42:15+04:00",
71
+ "wall_secs": 0.4,
72
+ "chains": 10888
73
+ }
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/lpmin_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/lpmin_selector.provenance.json ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T14:42:15+04:00",
3
+ "host": "ws006599",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/logprob_scores.py",
8
+ "--pool-dir",
9
+ "workdir/gpools/qwen3_4b/chatpool",
10
+ "--model",
11
+ "Qwen/Qwen3-4B",
12
+ "--template",
13
+ "qwen",
14
+ "--merge",
15
+ "workdir/gpools/qwen3_4b/logprob/shard0.jsonl"
16
+ ],
17
+ "git": {
18
+ "commit": "6cae6bf292a53eff08d1ca10850479123a1aae47",
19
+ "branch": "HEAD",
20
+ "subject": "logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases",
21
+ "dirty": false,
22
+ "dirty_files": []
23
+ },
24
+ "gpus": {
25
+ "visible": "0",
26
+ "devices": [
27
+ {
28
+ "index": "0",
29
+ "name": "NVIDIA RTX 5880 Ada Generation",
30
+ "memory": "49140 MiB",
31
+ "driver": "580.178.04",
32
+ "uuid": "GPU-ea30b823-34f2-1633-1958-47244a40ac5d"
33
+ },
34
+ {
35
+ "index": "1",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-b843d197-f9cd-0190-cbf3-1ded8ca2f9a8"
40
+ }
41
+ ]
42
+ },
43
+ "packages": {
44
+ "python": "3.11.15",
45
+ "torch": "2.9.1",
46
+ "vllm": "0.15.1",
47
+ "transformers": "4.57.6",
48
+ "speculators": "0.4.0.dev49",
49
+ "peft": null,
50
+ "accelerate": "1.13.0",
51
+ "numpy": "2.2.6",
52
+ "huggingface_hub": "0.36.2",
53
+ "openai": "2.32.0",
54
+ "wandb": "0.25.1"
55
+ },
56
+ "env": {
57
+ "CUDA_VISIBLE_DEVICES": "0",
58
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
59
+ "NCCL_P2P_DISABLE": "1",
60
+ "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"
61
+ },
62
+ "models": [],
63
+ "files": {
64
+ "pool_config": {
65
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/qwen3_4b/chatpool/pool_config.json",
66
+ "bytes": 211,
67
+ "sha256": "be24fd70b2ed8c5b0dc17d78ea9f6fbaa86a62becd6b2cfa8809fb6e38604285"
68
+ }
69
+ },
70
+ "finished_at": "2026-10-31T14:42:15+04:00",
71
+ "wall_secs": 0.5,
72
+ "chains": 10888
73
+ }
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/negent_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/negent_selector.provenance.json ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T14:42:15+04:00",
3
+ "host": "ws006599",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/logprob_scores.py",
8
+ "--pool-dir",
9
+ "workdir/gpools/qwen3_4b/chatpool",
10
+ "--model",
11
+ "Qwen/Qwen3-4B",
12
+ "--template",
13
+ "qwen",
14
+ "--merge",
15
+ "workdir/gpools/qwen3_4b/logprob/shard0.jsonl"
16
+ ],
17
+ "git": {
18
+ "commit": "6cae6bf292a53eff08d1ca10850479123a1aae47",
19
+ "branch": "HEAD",
20
+ "subject": "logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases",
21
+ "dirty": false,
22
+ "dirty_files": []
23
+ },
24
+ "gpus": {
25
+ "visible": "0",
26
+ "devices": [
27
+ {
28
+ "index": "0",
29
+ "name": "NVIDIA RTX 5880 Ada Generation",
30
+ "memory": "49140 MiB",
31
+ "driver": "580.178.04",
32
+ "uuid": "GPU-ea30b823-34f2-1633-1958-47244a40ac5d"
33
+ },
34
+ {
35
+ "index": "1",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-b843d197-f9cd-0190-cbf3-1ded8ca2f9a8"
40
+ }
41
+ ]
42
+ },
43
+ "packages": {
44
+ "python": "3.11.15",
45
+ "torch": "2.9.1",
46
+ "vllm": "0.15.1",
47
+ "transformers": "4.57.6",
48
+ "speculators": "0.4.0.dev49",
49
+ "peft": null,
50
+ "accelerate": "1.13.0",
51
+ "numpy": "2.2.6",
52
+ "huggingface_hub": "0.36.2",
53
+ "openai": "2.32.0",
54
+ "wandb": "0.25.1"
55
+ },
56
+ "env": {
57
+ "CUDA_VISIBLE_DEVICES": "0",
58
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
59
+ "NCCL_P2P_DISABLE": "1",
60
+ "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"
61
+ },
62
+ "models": [],
63
+ "files": {
64
+ "pool_config": {
65
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/qwen3_4b/chatpool/pool_config.json",
66
+ "bytes": 211,
67
+ "sha256": "be24fd70b2ed8c5b0dc17d78ea9f6fbaa86a62becd6b2cfa8809fb6e38604285"
68
+ }
69
+ },
70
+ "finished_at": "2026-10-31T14:42:16+04:00",
71
+ "wall_secs": 0.5,
72
+ "chains": 10888
73
+ }
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/top1_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/chatpool/top1_selector.provenance.json ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T14:42:15+04:00",
3
+ "host": "ws006599",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/logprob_scores.py",
8
+ "--pool-dir",
9
+ "workdir/gpools/qwen3_4b/chatpool",
10
+ "--model",
11
+ "Qwen/Qwen3-4B",
12
+ "--template",
13
+ "qwen",
14
+ "--merge",
15
+ "workdir/gpools/qwen3_4b/logprob/shard0.jsonl"
16
+ ],
17
+ "git": {
18
+ "commit": "6cae6bf292a53eff08d1ca10850479123a1aae47",
19
+ "branch": "HEAD",
20
+ "subject": "logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases",
21
+ "dirty": false,
22
+ "dirty_files": []
23
+ },
24
+ "gpus": {
25
+ "visible": "0",
26
+ "devices": [
27
+ {
28
+ "index": "0",
29
+ "name": "NVIDIA RTX 5880 Ada Generation",
30
+ "memory": "49140 MiB",
31
+ "driver": "580.178.04",
32
+ "uuid": "GPU-ea30b823-34f2-1633-1958-47244a40ac5d"
33
+ },
34
+ {
35
+ "index": "1",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-b843d197-f9cd-0190-cbf3-1ded8ca2f9a8"
40
+ }
41
+ ]
42
+ },
43
+ "packages": {
44
+ "python": "3.11.15",
45
+ "torch": "2.9.1",
46
+ "vllm": "0.15.1",
47
+ "transformers": "4.57.6",
48
+ "speculators": "0.4.0.dev49",
49
+ "peft": null,
50
+ "accelerate": "1.13.0",
51
+ "numpy": "2.2.6",
52
+ "huggingface_hub": "0.36.2",
53
+ "openai": "2.32.0",
54
+ "wandb": "0.25.1"
55
+ },
56
+ "env": {
57
+ "CUDA_VISIBLE_DEVICES": "0",
58
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
59
+ "NCCL_P2P_DISABLE": "1",
60
+ "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"
61
+ },
62
+ "models": [],
63
+ "files": {
64
+ "pool_config": {
65
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/qwen3_4b/chatpool/pool_config.json",
66
+ "bytes": 211,
67
+ "sha256": "be24fd70b2ed8c5b0dc17d78ea9f6fbaa86a62becd6b2cfa8809fb6e38604285"
68
+ }
69
+ },
70
+ "finished_at": "2026-10-31T14:42:16+04:00",
71
+ "wall_secs": 0.6,
72
+ "chains": 10888
73
+ }
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/logprob/shard0.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/logprob/shard0.log ADDED
@@ -0,0 +1,205 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [logprob] shard 0/1: 10888 chains over 1361 problems (0 already scored)
2
+ INFO 10-31 14:10:46 [utils.py:261] non-default args: {'max_model_len': 8192, 'enable_prefix_caching': False, 'gpu_memory_utilization': 0.6, 'max_num_batched_tokens': 8192, 'disable_log_stats': True, 'model': 'Qwen/Qwen3-4B'}
3
+ INFO 10-31 14:10:55 [model.py:541] Resolved architecture: Qwen3ForCausalLM
4
+ INFO 10-31 14:10:55 [model.py:1561] Using max model len 8192
5
+ INFO 10-31 14:10:56 [scheduler.py:226] Chunked prefill is enabled with max_num_batched_tokens=8192.
6
+ INFO 10-31 14:10:56 [vllm.py:624] Asynchronous scheduling is enabled.
7
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:10:57 [core.py:96] Initializing a V1 LLM engine (v0.15.1) with config: model='Qwen/Qwen3-4B', speculative_config=None, tokenizer='Qwen/Qwen3-4B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=8192, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=0, served_model_name=Qwen/Qwen3-4B, enable_prefix_caching=False, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': <CompilationMode.VLLM_COMPILE: 3>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['none'], 'splitting_ops': ['vllm::unified_attention', 'vllm::unified_attention_with_output', 'vllm::unified_mla_attention', 'vllm::unified_mla_attention_with_output', 'vllm::mamba_mixer2', 'vllm::mamba_mixer', 'vllm::short_conv', 'vllm::linear_attention', 'vllm::plamo2_mamba_mixer', 'vllm::gdn_attention_core', 'vllm::kda_attention', 'vllm::sparse_attn_indexer', 'vllm::rocm_aiter_sparse_attn_indexer', 'vllm::unified_kv_cache_update'], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [8192], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.FULL_AND_PIECEWISE: (2, 1)>, 'cudagraph_num_of_warmups': 1, 'cudagraph_capture_sizes': [1, 2, 4, 8, 16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96, 104, 112, 120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232, 240, 248, 256, 272, 288, 304, 320, 336, 352, 368, 384, 400, 416, 432, 448, 464, 480, 496, 512], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': True, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False}, 'max_cudagraph_capture_size': 512, 'dynamic_shapes_config': {'type': <DynamicShapesType.BACKED: 'backed'>, 'evaluate_guards': False, 'assume_32_bit_indexing': True}, 'local_cache_dir': None, 'static_all_moe_layers': []}
8
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:10:58 [parallel_state.py:1212] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.127.105.184:37673 backend=nccl
9
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:10:58 [parallel_state.py:1423] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A
10
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:10:59 [gpu_model_runner.py:4033] Starting to load model Qwen/Qwen3-4B...
11
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:11:00 [cuda.py:364] Using FLASH_ATTN attention backend out of potential backends: ('FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION')
12
+ (EngineCore_DP0 pid=3777344)
13
+ (EngineCore_DP0 pid=3777344)
14
+ (EngineCore_DP0 pid=3777344)
15
+ (EngineCore_DP0 pid=3777344)
16
+ (EngineCore_DP0 pid=3777344)
17
+ (EngineCore_DP0 pid=3777344)
18
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:11:17 [default_loader.py:291] Loading weights took 14.75 seconds
19
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:11:17 [gpu_model_runner.py:4130] Model loading took 7.5 GiB memory and 17.596819 seconds
20
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:11:25 [backends.py:812] Using cache directory: /home/vlad.smirnov/.cache/vllm/torch_compile_cache/9026f263be/rank_0_0/backbone for vLLM's torch.compile
21
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:11:25 [backends.py:872] Dynamo bytecode transform time: 7.65 s
22
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:11:33 [backends.py:302] Cache the graph of compile range (1, 8192) for later use
23
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:11:39 [backends.py:319] Compiling a graph for compile range (1, 8192) takes 10.99 s
24
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:11:39 [monitor.py:34] torch.compile takes 18.65 s in total
25
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:11:40 [gpu_worker.py:356] Available KV cache memory: 19.46 GiB
26
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:11:40 [kv_cache_utils.py:1307] GPU KV cache size: 141,680 tokens
27
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:11:40 [kv_cache_utils.py:1312] Maximum concurrency for 8,192 tokens per request: 17.29x
28
+ (EngineCore_DP0 pid=3777344)
29
+ (EngineCore_DP0 pid=3777344)
30
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:11:43 [gpu_model_runner.py:5063] Graph capturing finished in 3 secs, took 0.53 GiB
31
+ (EngineCore_DP0 pid=3777344) INFO 10-31 14:11:43 [core.py:272] init engine (profile, create kv cache, warmup model) took 25.45 seconds
32
+ INFO 10-31 14:11:45 [llm.py:343] Supported tasks: ['generate']
33
+ [logprob] 64/10888
34
+ [logprob] 128/10888
35
+ [logprob] 192/10888
36
+ [logprob] 256/10888
37
+ [logprob] 320/10888
38
+ [logprob] 384/10888
39
+ [logprob] 448/10888
40
+ [logprob] 512/10888
41
+ [logprob] 576/10888
42
+ [logprob] 640/10888
43
+ [logprob] 704/10888
44
+ [logprob] 768/10888
45
+ [logprob] 832/10888
46
+ [logprob] 896/10888
47
+ [logprob] 960/10888
48
+ [logprob] 1024/10888
49
+ [logprob] 1088/10888
50
+ [logprob] 1152/10888
51
+ [logprob] 1216/10888
52
+ [logprob] 1280/10888
53
+ [logprob] 1344/10888
54
+ [logprob] 1408/10888
55
+ [logprob] 1472/10888
56
+ [logprob] 1536/10888
57
+ [logprob] 1600/10888
58
+ [logprob] 1664/10888
59
+ [logprob] 1728/10888
60
+ [logprob] 1792/10888
61
+ [logprob] 1856/10888
62
+ [logprob] 1920/10888
63
+ [logprob] 1984/10888
64
+ [logprob] 2048/10888
65
+ [logprob] 2112/10888
66
+ [logprob] 2176/10888
67
+ [logprob] 2240/10888
68
+ [logprob] 2304/10888
69
+ [logprob] 2368/10888
70
+ [logprob] 2432/10888
71
+ [logprob] 2496/10888
72
+ [logprob] 2560/10888
73
+ [logprob] 2624/10888
74
+ [logprob] 2688/10888
75
+ [logprob] 2752/10888
76
+ [logprob] 2816/10888
77
+ [logprob] 2880/10888
78
+ [logprob] 2944/10888
79
+ [logprob] 3008/10888
80
+ [logprob] 3072/10888
81
+ [logprob] 3136/10888
82
+ [logprob] 3200/10888
83
+ [logprob] 3264/10888
84
+ [logprob] 3328/10888
85
+ [logprob] 3392/10888
86
+ [logprob] 3456/10888
87
+ [logprob] 3520/10888
88
+ [logprob] 3584/10888
89
+ [logprob] 3648/10888
90
+ [logprob] 3712/10888
91
+ [logprob] 3776/10888
92
+ [logprob] 3840/10888
93
+ [logprob] 3904/10888
94
+ [logprob] 3968/10888
95
+ [logprob] 4032/10888
96
+ [logprob] 4096/10888
97
+ [logprob] 4160/10888
98
+ [logprob] 4224/10888
99
+ [logprob] 4288/10888
100
+ [logprob] 4352/10888
101
+ [logprob] 4416/10888
102
+ [logprob] 4480/10888
103
+ [logprob] 4544/10888
104
+ [logprob] 4608/10888
105
+ [logprob] 4672/10888
106
+ [logprob] 4736/10888
107
+ [logprob] 4800/10888
108
+ [logprob] 4864/10888
109
+ [logprob] 4928/10888
110
+ [logprob] 4992/10888
111
+ [logprob] 5056/10888
112
+ [logprob] 5120/10888
113
+ [logprob] 5184/10888
114
+ [logprob] 5248/10888
115
+ [logprob] 5312/10888
116
+ [logprob] 5376/10888
117
+ [logprob] 5440/10888
118
+ [logprob] 5504/10888
119
+ [logprob] 5568/10888
120
+ [logprob] 5632/10888
121
+ [logprob] 5696/10888
122
+ [logprob] 5760/10888
123
+ [logprob] 5824/10888
124
+ [logprob] 5888/10888
125
+ [logprob] 5952/10888
126
+ [logprob] 6016/10888
127
+ [logprob] 6080/10888
128
+ [logprob] 6144/10888
129
+ [logprob] 6208/10888
130
+ [logprob] 6272/10888
131
+ [logprob] 6336/10888
132
+ [logprob] 6400/10888
133
+ [logprob] 6464/10888
134
+ [logprob] 6528/10888
135
+ [logprob] 6592/10888
136
+ [logprob] 6656/10888
137
+ [logprob] 6720/10888
138
+ [logprob] 6784/10888
139
+ [logprob] 6848/10888
140
+ [logprob] 6912/10888
141
+ [logprob] 6976/10888
142
+ [logprob] 7040/10888
143
+ [logprob] 7104/10888
144
+ [logprob] 7168/10888
145
+ [logprob] 7232/10888
146
+ [logprob] 7296/10888
147
+ [logprob] 7360/10888
148
+ [logprob] 7424/10888
149
+ [logprob] 7488/10888
150
+ [logprob] 7552/10888
151
+ [logprob] 7616/10888
152
+ [logprob] 7680/10888
153
+ [logprob] 7744/10888
154
+ [logprob] 7808/10888
155
+ [logprob] 7872/10888
156
+ [logprob] 7936/10888
157
+ [logprob] 8000/10888
158
+ [logprob] 8064/10888
159
+ [logprob] 8128/10888
160
+ [logprob] 8192/10888
161
+ [logprob] 8256/10888
162
+ [logprob] 8320/10888
163
+ [logprob] 8384/10888
164
+ [logprob] 8448/10888
165
+ [logprob] 8512/10888
166
+ [logprob] 8576/10888
167
+ [logprob] 8640/10888
168
+ [logprob] 8704/10888
169
+ [logprob] 8768/10888
170
+ [logprob] 8832/10888
171
+ [logprob] 8896/10888
172
+ [logprob] 8960/10888
173
+ [logprob] 9024/10888
174
+ [logprob] 9088/10888
175
+ [logprob] 9152/10888
176
+ [logprob] 9216/10888
177
+ [logprob] 9280/10888
178
+ [logprob] 9344/10888
179
+ [logprob] 9408/10888
180
+ [logprob] 9472/10888
181
+ [logprob] 9536/10888
182
+ [logprob] 9600/10888
183
+ [logprob] 9664/10888
184
+ [logprob] 9728/10888
185
+ [logprob] 9792/10888
186
+ [logprob] 9856/10888
187
+ [logprob] 9920/10888
188
+ [logprob] 9984/10888
189
+ [logprob] 10048/10888
190
+ [logprob] 10112/10888
191
+ [logprob] 10176/10888
192
+ [logprob] 10240/10888
193
+ [logprob] 10304/10888
194
+ [logprob] 10368/10888
195
+ [logprob] 10432/10888
196
+ [logprob] 10496/10888
197
+ [logprob] 10560/10888
198
+ [logprob] 10624/10888
199
+ [logprob] 10688/10888
200
+ [logprob] 10752/10888
201
+ [logprob] 10816/10888
202
+ [logprob] 10880/10888
203
+ [logprob] 10888/10888
204
+ [W1031 14:42:15.016017704 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
205
+ [logprob] merged 10888 chains into lpmean, lpmin, negent, top1 selector files
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_4b/logprob/shard0.provenance.json ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T14:10:11+04:00",
3
+ "host": "ws006599",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/logprob_scores.py",
8
+ "--pool-dir",
9
+ "workdir/gpools/qwen3_4b/chatpool",
10
+ "--model",
11
+ "Qwen/Qwen3-4B",
12
+ "--template",
13
+ "qwen",
14
+ "--chains",
15
+ "0",
16
+ "1",
17
+ "2",
18
+ "3",
19
+ "4",
20
+ "5",
21
+ "6",
22
+ "7",
23
+ "--resume",
24
+ "--out",
25
+ "workdir/gpools/qwen3_4b/logprob/shard0.jsonl"
26
+ ],
27
+ "git": {
28
+ "commit": "6cae6bf292a53eff08d1ca10850479123a1aae47",
29
+ "branch": "HEAD",
30
+ "subject": "logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases",
31
+ "dirty": false,
32
+ "dirty_files": []
33
+ },
34
+ "gpus": {
35
+ "visible": "0",
36
+ "devices": [
37
+ {
38
+ "index": "0",
39
+ "name": "NVIDIA RTX 5880 Ada Generation",
40
+ "memory": "49140 MiB",
41
+ "driver": "580.178.04",
42
+ "uuid": "GPU-ea30b823-34f2-1633-1958-47244a40ac5d"
43
+ },
44
+ {
45
+ "index": "1",
46
+ "name": "NVIDIA RTX 5880 Ada Generation",
47
+ "memory": "49140 MiB",
48
+ "driver": "580.178.04",
49
+ "uuid": "GPU-b843d197-f9cd-0190-cbf3-1ded8ca2f9a8"
50
+ }
51
+ ]
52
+ },
53
+ "packages": {
54
+ "python": "3.11.15",
55
+ "torch": "2.9.1",
56
+ "vllm": "0.15.1",
57
+ "transformers": "4.57.6",
58
+ "speculators": "0.4.0.dev49",
59
+ "peft": null,
60
+ "accelerate": "1.13.0",
61
+ "numpy": "2.2.6",
62
+ "huggingface_hub": "0.36.2",
63
+ "openai": "2.32.0",
64
+ "wandb": "0.25.1"
65
+ },
66
+ "env": {
67
+ "CUDA_VISIBLE_DEVICES": "0",
68
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
69
+ "NCCL_P2P_DISABLE": "1",
70
+ "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"
71
+ },
72
+ "models": [],
73
+ "files": {
74
+ "pool_config": {
75
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/qwen3_4b/chatpool/pool_config.json",
76
+ "bytes": 211,
77
+ "sha256": "be24fd70b2ed8c5b0dc17d78ea9f6fbaa86a62becd6b2cfa8809fb6e38604285"
78
+ }
79
+ },
80
+ "finished_at": "2026-10-31T14:42:13+04:00",
81
+ "wall_secs": 1922.3,
82
+ "shard": "0/1"
83
+ }
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/laneLPX_g0.log ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ 2026-10-31 14:10:11 lane LPX gpu 0 start 6cae6bf2 logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases bases: qwen3_4b qwen3_1.7b
2
+ 14:42:16 qwen3_4b ok (10888 chains)
3
+ 15:05:48 qwen3_1.7b ok (10888 chains)
4
+ LANE_LPX_DONE
judge-imitation/artem2/llm-reasoning-rl-with-uhead-workdir/laneLPX_g1.log ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ 2026-10-31 14:12:53 lane LPX gpu 1 start 6cae6bf2 logprob_scores --template; lane LPX: same-model logit and entropy signals on the other bases bases: llama31_8b
2
+ 14:47:26 llama31_8b ok (10888 chains)
3
+ LANE_LPX_DONE