Repository navigation
Expand file tree
/
Copy pathrtx5070ti-capacity.json
More file actions
198 lines (198 loc) · 6.14 KB
/
Copy pathrtx5070ti-capacity.json
File metadata and controls
198 lines (198 loc) · 6.14 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
{
"schema_version": 1,
"evidence_kind": "synthetic_full_parameter_adamw_capacity",
"sources": [
{
"path": "results/grpo_local_memory_probe/base_longest512_20260905.json",
"sha256": "840233c3527786ef1649249ff5c13bbfa5de0ddfb54fe079be2dfe94cd94edc5"
},
{
"sha256": "add5c915a83be061ef742dcaa1bd0d691cf6cc32bf1484eefd2b04480fc5182e",
"path": "docs/grpo-local-low-memory.md",
"role": "historical_hardware_and_software_notes"
}
],
"source_note": "Relative source paths and SHA256 hashes identify the local evidence used for this sanitized extract. Raw prompts, model weights, and machine-specific paths are omitted.",
"reported_environment": {
"gpu": "NVIDIA GeForce RTX 5070 Ti",
"advertised_vram_gb": 16,
"operating_environment": "WSL Ubuntu-24.04",
"pytorch": "2.10.0+cu129",
"provenance": "Historical hardware/software notes; the capacity JSON records CUDA device and byte counters but not the GPU product name."
},
"model": {
"id": "allenai/OLMo-2-0425-1B",
"revision": "a1847dff35000b4271fa70afc5db10fd29fedbdf",
"parameter_count": 1484916736,
"parameter_bytes": 2969833472,
"dtype": "torch.bfloat16",
"full_parameter_updates": true
},
"configuration": {
"algorithm": "offpolicy_gspo",
"seed": 0,
"learning_rate": 0.00001,
"attention_implementation": "sdpa",
"gradient_checkpointing": true,
"logits_chunk_size": 32,
"batch_size": 8,
"group_size": 8,
"microbatch_size": 1,
"gradient_accumulation_steps": 8,
"prompt_token_lengths": [
305,
305,
305,
305,
305,
305,
305,
305
],
"response_token_lengths": [
512,
512,
512,
512,
512,
512,
512,
512
],
"actual_input_width": 816,
"nonzero_advantages": 8,
"synthetic_rewards": [
1,
0,
1,
0,
1,
0,
1,
0
],
"optimizer": {
"foreach": false,
"initial_state_tensors": 0,
"learning_rate": 0.00001,
"name": "AdamW"
},
"old_policy_cache": "Rescored before each synthetic update; held fixed through that update."
},
"status": "passed",
"updates": [
{
"step": 1,
"status": "passed",
"elapsed_seconds": 4.152674449000017,
"gradient_checks": {
"all_finite": true,
"nonzero_tensors": 179
},
"old_log_probs_dtype": "torch.float32",
"optimizer_step": 1,
"gradient_norm": 4.21875,
"parameter_delta_max": 0.0000152587890625,
"parameter_delta_sample_count": 91648,
"optimizer_state": {
"all_moments_finite": true,
"all_moments_match_parameter_dtype": true,
"fp32_master_parameters": false,
"initialized_parameter_tensors": 179,
"moment_bytes": 5939666944,
"moment_dtypes": [
"torch.bfloat16"
],
"step_max": 1,
"step_min": 1
},
"memory_bytes": {
"allocated_bytes": 8926572544,
"cuda": true,
"driver_free_bytes": 1759481856,
"driver_total_bytes": 17094475776,
"peak_allocated_bytes": 12718557696,
"peak_reserved_bytes": 12924747776,
"reserved_bytes": 12924747776
}
},
{
"step": 2,
"status": "passed",
"elapsed_seconds": 3.2282285089999903,
"gradient_checks": {
"all_finite": true,
"nonzero_tensors": 179
},
"old_log_probs_dtype": "torch.float32",
"optimizer_step": 1,
"gradient_norm": 17.75,
"parameter_delta_max": 0.0000152587890625,
"parameter_delta_sample_count": 91648,
"optimizer_state": {
"all_moments_finite": true,
"all_moments_match_parameter_dtype": true,
"fp32_master_parameters": false,
"initialized_parameter_tensors": 179,
"moment_bytes": 5939666944,
"moment_dtypes": [
"torch.bfloat16"
],
"step_max": 2,
"step_min": 2
},
"memory_bytes": {
"allocated_bytes": 8926572544,
"cuda": true,
"driver_free_bytes": 1122013184,
"driver_total_bytes": 17094475776,
"peak_allocated_bytes": 13746177536,
"peak_reserved_bytes": 13994295296,
"reserved_bytes": 13994295296
}
}
],
"generation_between_updates": {
"requested_max_tokens": 512,
"response_count": 8,
"retokenized_response_lengths": [
76,
178,
75,
201,
224,
396,
510,
38
],
"optimizer_state_parameter_tensors": 179,
"same_policy_object": true,
"training_restored": true,
"elapsed_seconds": 5.933126256000008,
"memory_bytes": {
"allocated_bytes": 8926572544,
"cuda": true,
"driver_free_bytes": 6651117568,
"driver_total_bytes": 17094475776,
"peak_allocated_bytes": 9803807232,
"peak_reserved_bytes": 12924747776,
"reserved_bytes": 9019850752
}
},
"derived_peak_gib": {
"definition": "bytes / 1073741824",
"second_update_allocated": 12.802125453948975,
"second_update_reserved": 13.033203125
},
"caveats": [
"Synthetic mixed rewards establish that two finite, nonzero optimizer updates fit for this workload; they are not GSM8K accuracy measurements.",
"Each synthetic training response is 512 tokens; generation between updates produced lengths 38-510, not eight simultaneous 512-token completions.",
"Old-policy scores were recomputed before each synthetic update. This is not a complete off-policy rollout experiment.",
"Allocated memory is contained within reserved memory; do not add them. These CUDA allocator counters are not the entire GPU's memory usage.",
"Parameter change detection samples all 179 parameter tensors; it does not compare every parameter element.",
"The second update includes initialized native BF16 AdamW moments, with no separate FP32 master parameters.",
"No controlled before-versus-after VRAM saving is established by this evidence.",
"Longer prompts, larger batches, other GPU applications, and long-running memory variation can change capacity.",
"Synchronized wall time includes finite checks and CPU parameter samples."
]
}