Upload folder using huggingface_hub
Browse files- .gitattributes +1 -0
- README.md +127 -3
- added_tokens.json +28 -0
- chat_template.jinja +89 -0
- global_step740/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt +3 -0
- global_step740/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt +3 -0
- global_step740/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt +3 -0
- global_step740/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt +3 -0
- global_step740/bf16_zero_pp_rank_4_mp_rank_00_optim_states.pt +3 -0
- global_step740/bf16_zero_pp_rank_5_mp_rank_00_optim_states.pt +3 -0
- global_step740/bf16_zero_pp_rank_6_mp_rank_00_optim_states.pt +3 -0
- global_step740/bf16_zero_pp_rank_7_mp_rank_00_optim_states.pt +3 -0
- global_step740/zero_pp_rank_0_mp_rank_00_model_states.pt +3 -0
- global_step740/zero_pp_rank_1_mp_rank_00_model_states.pt +3 -0
- global_step740/zero_pp_rank_2_mp_rank_00_model_states.pt +3 -0
- global_step740/zero_pp_rank_3_mp_rank_00_model_states.pt +3 -0
- global_step740/zero_pp_rank_4_mp_rank_00_model_states.pt +3 -0
- global_step740/zero_pp_rank_5_mp_rank_00_model_states.pt +3 -0
- global_step740/zero_pp_rank_6_mp_rank_00_model_states.pt +3 -0
- global_step740/zero_pp_rank_7_mp_rank_00_model_states.pt +3 -0
- latest +1 -0
- merges.txt +0 -0
- pytorch_model.bin +3 -0
- rng_state_0.pth +3 -0
- rng_state_1.pth +3 -0
- rng_state_2.pth +3 -0
- rng_state_3.pth +3 -0
- rng_state_4.pth +3 -0
- rng_state_5.pth +3 -0
- rng_state_6.pth +3 -0
- rng_state_7.pth +3 -0
- scheduler.pt +3 -0
- sft_qwen_var_classifier.py +725 -0
- special_tokens_map.json +31 -0
- tokenizer.json +3 -0
- tokenizer_config.json +239 -0
- trainer_state.json +1884 -0
- training_args.bin +3 -0
- vocab.json +0 -0
- zero_to_fp32.py +760 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
|
@@ -1,3 +1,127 @@
|
|
| 1 |
-
---
|
| 2 |
-
license:
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model: Qwen/Qwen3-0.6B
|
| 4 |
+
tags:
|
| 5 |
+
- SAT
|
| 6 |
+
- combinatorial-optimization
|
| 7 |
+
- classification
|
| 8 |
+
- cube-and-conquer
|
| 9 |
+
- branching-heuristics
|
| 10 |
+
language:
|
| 11 |
+
- en
|
| 12 |
+
pipeline_tag: text-classification
|
| 13 |
+
datasets:
|
| 14 |
+
- custom
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# Qwen3-0.6B-SAT-VarSelector-Sym
|
| 18 |
+
|
| 19 |
+
A lightweight Qwen3-0.6B model fine-tuned for **SAT branching variable selection** in Cube-and-Conquer (CnC) solvers.
|
| 20 |
+
|
| 21 |
+
## Model Description
|
| 22 |
+
|
| 23 |
+
This model predicts which variable to branch/cube on next, given a SAT CNF formula state. Instead of generating text, it outputs a **classification over variable IDs** (1-600).
|
| 24 |
+
|
| 25 |
+
### Architecture
|
| 26 |
+
|
| 27 |
+
- **Base**: `Qwen/Qwen3-0.6B` (causal language model)
|
| 28 |
+
- **Head**: LayerNorm → Linear(hidden_size, 601)
|
| 29 |
+
- **Pooling**: Last non-pad token hidden state
|
| 30 |
+
- **Masking**: Invalid variables (not in CNF) are masked to -10000 before softmax
|
| 31 |
+
- **Size**: ~1.2GB (bfloat16)
|
| 32 |
+
|
| 33 |
+
### Training Details
|
| 34 |
+
|
| 35 |
+
| Parameter | Value |
|
| 36 |
+
|-----------|-------|
|
| 37 |
+
| Dataset | 8,110 training / 902 validation samples |
|
| 38 |
+
| Task | Predict expert-selected branching variable |
|
| 39 |
+
| Epochs | ~6 (best checkpoint at epoch 5.83) |
|
| 40 |
+
| Hardware | 8×H100 GPUs |
|
| 41 |
+
| Framework | DeepSpeed ZeRO-3 |
|
| 42 |
+
| Learning Rate | 5e-6 |
|
| 43 |
+
| Max Sequence Length | 8192 tokens |
|
| 44 |
+
| Max Variables | 600 |
|
| 45 |
+
|
| 46 |
+
### Performance
|
| 47 |
+
|
| 48 |
+
| Metric | Value |
|
| 49 |
+
|--------|-------|
|
| 50 |
+
| **Top-1 Accuracy** | **23.95%** |
|
| 51 |
+
| Eval Loss | 3.30 |
|
| 52 |
+
| Inference Speed | ~45ms/sample (H100) |
|
| 53 |
+
|
| 54 |
+
## Usage
|
| 55 |
+
|
| 56 |
+
```python
|
| 57 |
+
import torch
|
| 58 |
+
from transformers import AutoTokenizer
|
| 59 |
+
from sft_qwen_var_classifier import QwenVarClassifier, cnf_valid_mask
|
| 60 |
+
|
| 61 |
+
# Load model
|
| 62 |
+
model = QwenVarClassifier("Qwen/Qwen3-0.6B", max_vars=600)
|
| 63 |
+
state_dict = torch.load("pytorch_model.bin", map_location="cpu")
|
| 64 |
+
model.load_state_dict(state_dict, strict=False)
|
| 65 |
+
model = model.to("cuda", dtype=torch.bfloat16)
|
| 66 |
+
model.eval()
|
| 67 |
+
|
| 68 |
+
# Load tokenizer
|
| 69 |
+
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
|
| 70 |
+
|
| 71 |
+
# Prepare CNF input
|
| 72 |
+
cnf_text = """p cnf 100 250
|
| 73 |
+
1 -2 3 0
|
| 74 |
+
-1 2 -4 0
|
| 75 |
+
...
|
| 76 |
+
"""
|
| 77 |
+
|
| 78 |
+
# Tokenize
|
| 79 |
+
inputs = tokenizer(cnf_text, return_tensors="pt", truncation=True, max_length=8192)
|
| 80 |
+
inputs = {k: v.to("cuda") for k, v in inputs.items()}
|
| 81 |
+
|
| 82 |
+
# Get valid variable mask
|
| 83 |
+
valid_mask = torch.tensor([cnf_valid_mask(cnf_text, max_vars=600)], dtype=torch.bool, device="cuda")
|
| 84 |
+
|
| 85 |
+
# Predict
|
| 86 |
+
with torch.no_grad():
|
| 87 |
+
outputs = model(**inputs)
|
| 88 |
+
logits = outputs["logits"]
|
| 89 |
+
logits = logits.masked_fill(~valid_mask, -1e4)
|
| 90 |
+
predicted_var = logits.argmax(dim=-1).item()
|
| 91 |
+
|
| 92 |
+
print(f"Predicted branching variable: {predicted_var}")
|
| 93 |
+
```
|
| 94 |
+
|
| 95 |
+
## Files
|
| 96 |
+
|
| 97 |
+
| File | Description |
|
| 98 |
+
|------|-------------|
|
| 99 |
+
| `pytorch_model.bin` | Model weights (~1.2GB, bfloat16) |
|
| 100 |
+
| `sft_qwen_var_classifier.py` | Model class definition (required for loading) |
|
| 101 |
+
| `tokenizer.json` | Tokenizer configuration |
|
| 102 |
+
| `tokenizer_config.json` | Tokenizer settings |
|
| 103 |
+
|
| 104 |
+
## Use Cases
|
| 105 |
+
|
| 106 |
+
- **Cube-and-Conquer SAT Solving**: Guide cube splitting decisions
|
| 107 |
+
- **Research**: Study neural branching heuristics
|
| 108 |
+
- **Benchmarking**: Compare against traditional heuristics (VSIDS, etc.)
|
| 109 |
+
|
| 110 |
+
## Limitations
|
| 111 |
+
|
| 112 |
+
- Maximum 600 variables supported
|
| 113 |
+
- Maximum 8192 tokens for CNF input
|
| 114 |
+
- Trained on specific CNF distribution
|
| 115 |
+
- Best used as a starting point for further fine-tuning on your specific problem domain
|
| 116 |
+
|
| 117 |
+
## Related Models
|
| 118 |
+
|
| 119 |
+
- [Yale-ROSE/Qwen3-0.6B-SAT-VarSelector](https://huggingface.co/Yale-ROSE/Qwen3-0.6B-SAT-VarSelector) - Earlier version
|
| 120 |
+
|
| 121 |
+
## Citation
|
| 122 |
+
|
| 123 |
+
If you use this model, please cite the Transformer-CnC paper.
|
| 124 |
+
|
| 125 |
+
## License
|
| 126 |
+
|
| 127 |
+
Apache 2.0
|
added_tokens.json
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"</think>": 151668,
|
| 3 |
+
"</tool_call>": 151658,
|
| 4 |
+
"</tool_response>": 151666,
|
| 5 |
+
"<think>": 151667,
|
| 6 |
+
"<tool_call>": 151657,
|
| 7 |
+
"<tool_response>": 151665,
|
| 8 |
+
"<|box_end|>": 151649,
|
| 9 |
+
"<|box_start|>": 151648,
|
| 10 |
+
"<|endoftext|>": 151643,
|
| 11 |
+
"<|file_sep|>": 151664,
|
| 12 |
+
"<|fim_middle|>": 151660,
|
| 13 |
+
"<|fim_pad|>": 151662,
|
| 14 |
+
"<|fim_prefix|>": 151659,
|
| 15 |
+
"<|fim_suffix|>": 151661,
|
| 16 |
+
"<|im_end|>": 151645,
|
| 17 |
+
"<|im_start|>": 151644,
|
| 18 |
+
"<|image_pad|>": 151655,
|
| 19 |
+
"<|object_ref_end|>": 151647,
|
| 20 |
+
"<|object_ref_start|>": 151646,
|
| 21 |
+
"<|quad_end|>": 151651,
|
| 22 |
+
"<|quad_start|>": 151650,
|
| 23 |
+
"<|repo_name|>": 151663,
|
| 24 |
+
"<|video_pad|>": 151656,
|
| 25 |
+
"<|vision_end|>": 151653,
|
| 26 |
+
"<|vision_pad|>": 151654,
|
| 27 |
+
"<|vision_start|>": 151652
|
| 28 |
+
}
|
chat_template.jinja
ADDED
|
@@ -0,0 +1,89 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{%- if tools %}
|
| 2 |
+
{{- '<|im_start|>system\n' }}
|
| 3 |
+
{%- if messages[0].role == 'system' %}
|
| 4 |
+
{{- messages[0].content + '\n\n' }}
|
| 5 |
+
{%- endif %}
|
| 6 |
+
{{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
|
| 7 |
+
{%- for tool in tools %}
|
| 8 |
+
{{- "\n" }}
|
| 9 |
+
{{- tool | tojson }}
|
| 10 |
+
{%- endfor %}
|
| 11 |
+
{{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
|
| 12 |
+
{%- else %}
|
| 13 |
+
{%- if messages[0].role == 'system' %}
|
| 14 |
+
{{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
|
| 15 |
+
{%- endif %}
|
| 16 |
+
{%- endif %}
|
| 17 |
+
{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
|
| 18 |
+
{%- for message in messages[::-1] %}
|
| 19 |
+
{%- set index = (messages|length - 1) - loop.index0 %}
|
| 20 |
+
{%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
|
| 21 |
+
{%- set ns.multi_step_tool = false %}
|
| 22 |
+
{%- set ns.last_query_index = index %}
|
| 23 |
+
{%- endif %}
|
| 24 |
+
{%- endfor %}
|
| 25 |
+
{%- for message in messages %}
|
| 26 |
+
{%- if message.content is string %}
|
| 27 |
+
{%- set content = message.content %}
|
| 28 |
+
{%- else %}
|
| 29 |
+
{%- set content = '' %}
|
| 30 |
+
{%- endif %}
|
| 31 |
+
{%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
|
| 32 |
+
{{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
|
| 33 |
+
{%- elif message.role == "assistant" %}
|
| 34 |
+
{%- set reasoning_content = '' %}
|
| 35 |
+
{%- if message.reasoning_content is string %}
|
| 36 |
+
{%- set reasoning_content = message.reasoning_content %}
|
| 37 |
+
{%- else %}
|
| 38 |
+
{%- if '</think>' in content %}
|
| 39 |
+
{%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
|
| 40 |
+
{%- set content = content.split('</think>')[-1].lstrip('\n') %}
|
| 41 |
+
{%- endif %}
|
| 42 |
+
{%- endif %}
|
| 43 |
+
{%- if loop.index0 > ns.last_query_index %}
|
| 44 |
+
{%- if loop.last or (not loop.last and reasoning_content) %}
|
| 45 |
+
{{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
|
| 46 |
+
{%- else %}
|
| 47 |
+
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 48 |
+
{%- endif %}
|
| 49 |
+
{%- else %}
|
| 50 |
+
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 51 |
+
{%- endif %}
|
| 52 |
+
{%- if message.tool_calls %}
|
| 53 |
+
{%- for tool_call in message.tool_calls %}
|
| 54 |
+
{%- if (loop.first and content) or (not loop.first) %}
|
| 55 |
+
{{- '\n' }}
|
| 56 |
+
{%- endif %}
|
| 57 |
+
{%- if tool_call.function %}
|
| 58 |
+
{%- set tool_call = tool_call.function %}
|
| 59 |
+
{%- endif %}
|
| 60 |
+
{{- '<tool_call>\n{"name": "' }}
|
| 61 |
+
{{- tool_call.name }}
|
| 62 |
+
{{- '", "arguments": ' }}
|
| 63 |
+
{%- if tool_call.arguments is string %}
|
| 64 |
+
{{- tool_call.arguments }}
|
| 65 |
+
{%- else %}
|
| 66 |
+
{{- tool_call.arguments | tojson }}
|
| 67 |
+
{%- endif %}
|
| 68 |
+
{{- '}\n</tool_call>' }}
|
| 69 |
+
{%- endfor %}
|
| 70 |
+
{%- endif %}
|
| 71 |
+
{{- '<|im_end|>\n' }}
|
| 72 |
+
{%- elif message.role == "tool" %}
|
| 73 |
+
{%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
|
| 74 |
+
{{- '<|im_start|>user' }}
|
| 75 |
+
{%- endif %}
|
| 76 |
+
{{- '\n<tool_response>\n' }}
|
| 77 |
+
{{- content }}
|
| 78 |
+
{{- '\n</tool_response>' }}
|
| 79 |
+
{%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
|
| 80 |
+
{{- '<|im_end|>\n' }}
|
| 81 |
+
{%- endif %}
|
| 82 |
+
{%- endif %}
|
| 83 |
+
{%- endfor %}
|
| 84 |
+
{%- if add_generation_prompt %}
|
| 85 |
+
{{- '<|im_start|>assistant\n' }}
|
| 86 |
+
{%- if enable_thinking is defined and enable_thinking is false %}
|
| 87 |
+
{{- '<think>\n\n</think>\n\n' }}
|
| 88 |
+
{%- endif %}
|
| 89 |
+
{%- endif %}
|
global_step740/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fbaff101fde6ddd15e6d7e431eb95a8f6a44fc41cdd05d490a7b340a72fcb373
|
| 3 |
+
size 895007429
|
global_step740/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ad39e9993b3436d918073c4ab98343a1025df8ca861ebff545b1fc9c33cdc410
|
| 3 |
+
size 895007429
|
global_step740/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1d6723b8ee706d73bb32f024c040630e3be78cfdb6dbb48bb600b871c0cfe208
|
| 3 |
+
size 895007429
|
global_step740/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ae9a790dc5805f3af047ea60cdc2e72768f434b2ca38c300ec814a462d3708a3
|
| 3 |
+
size 895007429
|
global_step740/bf16_zero_pp_rank_4_mp_rank_00_optim_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b89f1c3ae2077397a8811cb6f0582b53d0659c07ac0ffe6e9c07c5d512b18c15
|
| 3 |
+
size 895007429
|
global_step740/bf16_zero_pp_rank_5_mp_rank_00_optim_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ed64439325872dc60d255182980284a3aa32b884aef5685a2728a625e6b863ab
|
| 3 |
+
size 895007429
|
global_step740/bf16_zero_pp_rank_6_mp_rank_00_optim_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:aef0451b388c68640cb390362986e7bc112f81c11da4d55527dfc46d93e649c8
|
| 3 |
+
size 895007429
|
global_step740/bf16_zero_pp_rank_7_mp_rank_00_optim_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ff212d7092b446f8644d72e4804dfb413bbda03f18eb2ebb6cd9efd10f33bc2c
|
| 3 |
+
size 895007429
|
global_step740/zero_pp_rank_0_mp_rank_00_model_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f46a9f8f814a58cdd546081fc46bafbc8fff09e136c3b576ce84e92671e8465c
|
| 3 |
+
size 168971
|
global_step740/zero_pp_rank_1_mp_rank_00_model_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6a357e82be16a6f68af476bb99e3f3ee585d790f02384c125f40c2732554e7a6
|
| 3 |
+
size 168907
|
global_step740/zero_pp_rank_2_mp_rank_00_model_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:35a72b89bae1414cadb030af34710fa422f54eff93ebbc4174d2cfe0441ecb11
|
| 3 |
+
size 168907
|
global_step740/zero_pp_rank_3_mp_rank_00_model_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a40b158bb08e7a56c4ec6cbc54900770a0707df0fde7742af6b0a954da12da19
|
| 3 |
+
size 168907
|
global_step740/zero_pp_rank_4_mp_rank_00_model_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2e083b491852ba19d0ebaa221401562f87e90f0e4444b6e318dd9fb6db9abe29
|
| 3 |
+
size 168907
|
global_step740/zero_pp_rank_5_mp_rank_00_model_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2aee738ff58319ea96a85571aff489b53b96f84ea1a7c86521d98ae30b0907af
|
| 3 |
+
size 168907
|
global_step740/zero_pp_rank_6_mp_rank_00_model_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ee3e1bf4e60cc7745e9eeb2d46aaf6641c36c5f2166ddd058137d4ed97c8fe7c
|
| 3 |
+
size 168907
|
global_step740/zero_pp_rank_7_mp_rank_00_model_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e2fe73796fc97f4ce70ad4061c0e0ceaa3ffc535c1f43055eddf1f829fd34c23
|
| 3 |
+
size 168907
|
latest
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
global_step740
|
merges.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
pytorch_model.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a220725332b4cc2c9718697120389e6c2f6a104543c877e482f8e33b411f5aaa
|
| 3 |
+
size 1193440459
|
rng_state_0.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:eaa8338a3cc72cf287ec1b6a50ffbeeaf3d0881f2c66234bc3e77dca70e336ef
|
| 3 |
+
size 16325
|
rng_state_1.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6ccee6013843af1ec731c6a8c9204194b073c3849cb2afef0657cf1a9da0f737
|
| 3 |
+
size 16389
|
rng_state_2.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3e7c2d2f49f3a637e2692c849e3ce448026dff315d2b1827452e6f926a8ba1bc
|
| 3 |
+
size 16389
|
rng_state_3.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:4185c6c2f292e36f5f7af71c5307cbd0489d9778cea9eed9c449360ee47809e9
|
| 3 |
+
size 16389
|
rng_state_4.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ededdb3655c3cbb6938270f7310dc370a8d6a9c981f57735b5c64ef818ed027c
|
| 3 |
+
size 16389
|
rng_state_5.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:07bcd82eac2f7cf14ff6d25b668486523ae5f8f084cfdc0d120a15c9e31e2de3
|
| 3 |
+
size 16389
|
rng_state_6.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b195804802f1206b8b7f8786e3062e42d9bdd22ae61212684c3a6ab698db790c
|
| 3 |
+
size 16389
|
rng_state_7.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:465fb7aa776cb402325a6c00d12467959c48ecd84d8479cb2b538443ff58962e
|
| 3 |
+
size 16389
|
scheduler.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ca3afdfe233f944a3930c173fc7c2e07be5542dadbaf860071e6ec586cec9e8e
|
| 3 |
+
size 1465
|
sft_qwen_var_classifier.py
ADDED
|
@@ -0,0 +1,725 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Qwen Variable Classifier for SAT Cube-and-Conquer
|
| 3 |
+
|
| 4 |
+
This script trains a transformer-based policy to select the next branching variable
|
| 5 |
+
for SAT (Boolean Satisfiability) solving using the Cube-and-Conquer approach.
|
| 6 |
+
|
| 7 |
+
== Problem Overview ==
|
| 8 |
+
In Cube-and-Conquer SAT solving, we split a hard SAT problem into subproblems ("cubes")
|
| 9 |
+
by choosing variables to branch on. The quality of variable selection significantly
|
| 10 |
+
affects solving performance. This model learns to predict good branching variables
|
| 11 |
+
from expert demonstrations.
|
| 12 |
+
|
| 13 |
+
== Architecture ==
|
| 14 |
+
- Backbone: Qwen3-4B (pretrained causal language model)
|
| 15 |
+
- Head: LayerNorm + Linear classifier over variable IDs (1 to max_vars)
|
| 16 |
+
- The model reads a CNF formula as text and outputs logits for each possible variable
|
| 17 |
+
|
| 18 |
+
== Training Approach ==
|
| 19 |
+
- Supervised Fine-Tuning (SFT) on expert variable choices
|
| 20 |
+
- Masked classification: only variables appearing in the CNF are valid choices
|
| 21 |
+
- Loss: Cross-entropy with invalid variable logits masked to -infinity
|
| 22 |
+
|
| 23 |
+
== Data Format ==
|
| 24 |
+
JSONL with fields:
|
| 25 |
+
- "cnf": DIMACS-format CNF text (e.g., "p cnf 100 200\n1 -2 3 0\n...")
|
| 26 |
+
- "label": integer variable ID to branch on (1 to max_vars)
|
| 27 |
+
"""
|
| 28 |
+
|
| 29 |
+
import os
|
| 30 |
+
import argparse
|
| 31 |
+
from dataclasses import dataclass
|
| 32 |
+
from typing import Any, Dict, List
|
| 33 |
+
import numpy as np
|
| 34 |
+
import torch
|
| 35 |
+
import torch.nn as nn
|
| 36 |
+
import torch.nn.functional as F
|
| 37 |
+
from datasets import load_dataset
|
| 38 |
+
from transformers import (
|
| 39 |
+
AutoConfig,
|
| 40 |
+
AutoTokenizer,
|
| 41 |
+
AutoModelForCausalLM,
|
| 42 |
+
TrainingArguments,
|
| 43 |
+
Trainer,
|
| 44 |
+
set_seed,
|
| 45 |
+
)
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
# =============================================================================
|
| 49 |
+
# DEBUG FLAG: Set to True to enable verbose debug output, False to disable
|
| 50 |
+
# Can also be controlled via environment variable: DEBUG_TRAINING=1
|
| 51 |
+
# =============================================================================
|
| 52 |
+
DEBUG_TRAINING = os.environ.get("DEBUG_TRAINING", "0") == "1"
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
# =============================================================================
|
| 56 |
+
# CNF PARSING: Extract valid variables from DIMACS CNF text
|
| 57 |
+
# =============================================================================
|
| 58 |
+
|
| 59 |
+
def cnf_valid_mask(cnf_text: str, max_vars: int) -> List[int]:
|
| 60 |
+
"""
|
| 61 |
+
Build a binary mask indicating which variable IDs appear in the CNF.
|
| 62 |
+
|
| 63 |
+
This is crucial for masked classification:
|
| 64 |
+
- A variable that doesn't appear in the (simplified) CNF cannot be branched on
|
| 65 |
+
- By masking invalid variables, we ensure the model only learns over valid choices
|
| 66 |
+
|
| 67 |
+
Args:
|
| 68 |
+
cnf_text: DIMACS-format CNF string. Format example:
|
| 69 |
+
p cnf 100 200 # header: 100 variables, 200 clauses
|
| 70 |
+
1 -2 3 0 # clause: (x1 OR NOT x2 OR x3)
|
| 71 |
+
-1 4 0 # clause: (NOT x1 OR x4)
|
| 72 |
+
...
|
| 73 |
+
max_vars: Maximum variable ID supported (typically 500)
|
| 74 |
+
|
| 75 |
+
Returns:
|
| 76 |
+
List of length (max_vars + 1) where:
|
| 77 |
+
- mask[0] = 0 (unused, variables are 1-indexed)
|
| 78 |
+
- mask[v] = 1 if variable v appears in any clause
|
| 79 |
+
- mask[v] = 0 if variable v does not appear
|
| 80 |
+
|
| 81 |
+
Note: We skip the header line "p cnf ..." to avoid capturing the clause count
|
| 82 |
+
as a valid variable (which was a bug in the original regex-based approach).
|
| 83 |
+
"""
|
| 84 |
+
mask = [0] * (max_vars + 1)
|
| 85 |
+
|
| 86 |
+
for line in cnf_text.split('\n'):
|
| 87 |
+
line = line.strip()
|
| 88 |
+
|
| 89 |
+
# Skip empty lines, comment lines (start with 'c'), and header line (starts with 'p')
|
| 90 |
+
# The header "p cnf <num_vars> <num_clauses>" would incorrectly add num_clauses as a variable
|
| 91 |
+
if not line or line.startswith('c') or line.startswith('p'):
|
| 92 |
+
continue
|
| 93 |
+
|
| 94 |
+
# Parse clause: space-separated integers ending with 0
|
| 95 |
+
# Each integer is a literal: positive = variable, negative = negated variable
|
| 96 |
+
# Example: "1 -2 3 0" means (x1 OR NOT x2 OR x3)
|
| 97 |
+
for tok in line.split():
|
| 98 |
+
try:
|
| 99 |
+
lit = int(tok)
|
| 100 |
+
v = abs(lit) # Variable ID is absolute value of literal
|
| 101 |
+
if 1 <= v <= max_vars:
|
| 102 |
+
mask[v] = 1
|
| 103 |
+
except ValueError:
|
| 104 |
+
continue # Skip non-integer tokens (shouldn't happen in valid DIMACS)
|
| 105 |
+
|
| 106 |
+
# Fallback: if no variables found (e.g., truncated/malformed input), allow all
|
| 107 |
+
# This prevents the model from having zero valid outputs
|
| 108 |
+
if sum(mask) == 0:
|
| 109 |
+
for v in range(1, max_vars + 1):
|
| 110 |
+
mask[v] = 1
|
| 111 |
+
|
| 112 |
+
return mask
|
| 113 |
+
|
| 114 |
+
|
| 115 |
+
# =============================================================================
|
| 116 |
+
# MODEL: Qwen backbone with classification head for variable selection
|
| 117 |
+
# =============================================================================
|
| 118 |
+
|
| 119 |
+
class QwenVarClassifier(nn.Module):
|
| 120 |
+
"""
|
| 121 |
+
Transformer-based variable classifier for SAT branching.
|
| 122 |
+
|
| 123 |
+
Architecture:
|
| 124 |
+
Input (CNF text)
|
| 125 |
+
→ Tokenize
|
| 126 |
+
→ Qwen3-4B backbone (frozen initially, fine-tuned with small LR)
|
| 127 |
+
→ Extract last token's hidden state (sequence pooling)
|
| 128 |
+
→ LayerNorm (stabilizes hidden state magnitude)
|
| 129 |
+
→ Linear head (hidden_dim → num_classes)
|
| 130 |
+
→ Logits for each variable ID
|
| 131 |
+
|
| 132 |
+
Why this architecture?
|
| 133 |
+
1. Pretrained LLM backbone understands text structure and can learn CNF patterns
|
| 134 |
+
2. Last-token pooling: the final token has attended to the entire input
|
| 135 |
+
3. LayerNorm: Qwen's hidden states have large magnitudes; normalizing prevents
|
| 136 |
+
exploding gradients when combined with randomly-initialized head
|
| 137 |
+
4. Single linear head: simple, interpretable, efficient
|
| 138 |
+
"""
|
| 139 |
+
|
| 140 |
+
def __init__(self, base_model_name: str, max_vars: int):
|
| 141 |
+
"""
|
| 142 |
+
Initialize the classifier.
|
| 143 |
+
|
| 144 |
+
Args:
|
| 145 |
+
base_model_name: HuggingFace model ID (e.g., "Qwen/Qwen3-4B")
|
| 146 |
+
max_vars: Maximum variable ID to classify (e.g., 500)
|
| 147 |
+
Output dimension will be max_vars + 1 (index 0 unused)
|
| 148 |
+
"""
|
| 149 |
+
super().__init__()
|
| 150 |
+
self.max_vars = max_vars
|
| 151 |
+
|
| 152 |
+
# Load Qwen configuration and enable hidden state output
|
| 153 |
+
cfg = AutoConfig.from_pretrained(base_model_name)
|
| 154 |
+
cfg.output_hidden_states = True # We need hidden states, not just logits
|
| 155 |
+
|
| 156 |
+
# Load pretrained Qwen model
|
| 157 |
+
# Using bfloat16 for memory efficiency on modern GPUs (H100, A100)
|
| 158 |
+
self.backbone = AutoModelForCausalLM.from_pretrained(
|
| 159 |
+
base_model_name,
|
| 160 |
+
config=cfg,
|
| 161 |
+
torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
|
| 162 |
+
)
|
| 163 |
+
|
| 164 |
+
hidden = self.backbone.config.hidden_size # e.g., 2560 for Qwen3-4B
|
| 165 |
+
|
| 166 |
+
# LayerNorm to normalize hidden states before classification
|
| 167 |
+
# This is critical for stable training:
|
| 168 |
+
# - Qwen's hidden states can have large magnitude (std >> 1)
|
| 169 |
+
# - Randomly initialized linear head expects normalized inputs
|
| 170 |
+
# - Without LayerNorm, initial logits can be huge → high loss → exploding gradients
|
| 171 |
+
self.head_ln = nn.LayerNorm(hidden)
|
| 172 |
+
|
| 173 |
+
# Classification head: maps hidden state to variable logits
|
| 174 |
+
# Output shape: [batch, max_vars + 1]
|
| 175 |
+
# Index 0 is unused (variables are 1-indexed in DIMACS)
|
| 176 |
+
self.head = nn.Linear(hidden, max_vars + 1)
|
| 177 |
+
|
| 178 |
+
# Initialize head with standard small weights
|
| 179 |
+
# LayerNorm ensures the input has unit variance, so this init is appropriate
|
| 180 |
+
nn.init.normal_(self.head.weight, std=0.02)
|
| 181 |
+
nn.init.zeros_(self.head.bias)
|
| 182 |
+
|
| 183 |
+
# Expose backbone config for DeepSpeed compatibility
|
| 184 |
+
# DeepSpeed checks model.config.hidden_size for auto-configuration
|
| 185 |
+
self.config = self.backbone.config
|
| 186 |
+
|
| 187 |
+
def forward(self, input_ids, attention_mask, **kwargs):
|
| 188 |
+
"""
|
| 189 |
+
Forward pass: CNF tokens → variable logits.
|
| 190 |
+
|
| 191 |
+
Args:
|
| 192 |
+
input_ids: [batch, seq_len] token IDs from tokenizer
|
| 193 |
+
attention_mask: [batch, seq_len] binary mask (1 = real token, 0 = padding)
|
| 194 |
+
**kwargs: ignored (allows passing 'labels' without error during eval)
|
| 195 |
+
|
| 196 |
+
Returns:
|
| 197 |
+
dict with "logits": [batch, max_vars + 1] raw classification logits
|
| 198 |
+
"""
|
| 199 |
+
# Run through Qwen backbone
|
| 200 |
+
out = self.backbone(
|
| 201 |
+
input_ids=input_ids,
|
| 202 |
+
attention_mask=attention_mask,
|
| 203 |
+
output_hidden_states=True, # Need hidden states, not LM logits
|
| 204 |
+
use_cache=False, # Disable KV cache (not needed for training)
|
| 205 |
+
)
|
| 206 |
+
|
| 207 |
+
# Get hidden states from the last transformer layer
|
| 208 |
+
# Shape: [batch, seq_len, hidden_dim]
|
| 209 |
+
h = out.hidden_states[-1]
|
| 210 |
+
|
| 211 |
+
# Pool by taking the last non-padding token's hidden state
|
| 212 |
+
# This is the standard approach for causal LMs (like using [CLS] for BERT)
|
| 213 |
+
#
|
| 214 |
+
# Why last token?
|
| 215 |
+
# - In causal attention, each token only sees previous tokens
|
| 216 |
+
# - The last token has attended to the entire input sequence
|
| 217 |
+
# - It's a natural "summary" of the input
|
| 218 |
+
#
|
| 219 |
+
# Compute index of last real token: sum of attention mask minus 1
|
| 220 |
+
last_idx = attention_mask.sum(dim=1) - 1 # [batch]
|
| 221 |
+
last_idx = last_idx.clamp(min=0) # Safety: ensure non-negative
|
| 222 |
+
|
| 223 |
+
# Gather hidden state at the last token position for each batch element
|
| 224 |
+
b = torch.arange(h.size(0), device=h.device)
|
| 225 |
+
pooled = h[b, last_idx] # [batch, hidden_dim]
|
| 226 |
+
|
| 227 |
+
# DEBUG: Check hidden state stats
|
| 228 |
+
if DEBUG_TRAINING:
|
| 229 |
+
if not hasattr(self, '_debug_count'):
|
| 230 |
+
self._debug_count = 0
|
| 231 |
+
if self._debug_count < 3:
|
| 232 |
+
print(f"[DEBUG {self._debug_count}] pooled dtype={pooled.dtype}, mean={pooled.float().mean():.2f}, std={pooled.float().std():.2f}")
|
| 233 |
+
self._debug_count += 1
|
| 234 |
+
|
| 235 |
+
# Normalize hidden states for stable classification
|
| 236 |
+
pooled = self.head_ln(pooled)
|
| 237 |
+
|
| 238 |
+
# DEBUG: Check after LayerNorm
|
| 239 |
+
if DEBUG_TRAINING and hasattr(self, '_debug_count') and self._debug_count <= 3:
|
| 240 |
+
print(f"[DEBUG] after LN: dtype={pooled.dtype}, mean={pooled.float().mean():.4f}, std={pooled.float().std():.4f}")
|
| 241 |
+
|
| 242 |
+
# Project to variable logits
|
| 243 |
+
logits = self.head(pooled) # [batch, max_vars + 1]
|
| 244 |
+
|
| 245 |
+
# DEBUG: Check logits
|
| 246 |
+
if DEBUG_TRAINING and hasattr(self, '_debug_count') and self._debug_count <= 3:
|
| 247 |
+
print(f"[DEBUG] logits: dtype={logits.dtype}, mean={logits.float().mean():.2f}, std={logits.float().std():.2f}, min={logits.float().min():.2f}, max={logits.float().max():.2f}")
|
| 248 |
+
|
| 249 |
+
return {"logits": logits}
|
| 250 |
+
|
| 251 |
+
|
| 252 |
+
# =============================================================================
|
| 253 |
+
# DATA COLLATOR: Batch preparation with padding and mask handling
|
| 254 |
+
# =============================================================================
|
| 255 |
+
|
| 256 |
+
@dataclass
|
| 257 |
+
class Collator:
|
| 258 |
+
"""
|
| 259 |
+
Custom data collator for variable classification.
|
| 260 |
+
|
| 261 |
+
Responsibilities:
|
| 262 |
+
1. Pad variable-length token sequences to the same length within a batch
|
| 263 |
+
2. Stack labels and valid_mask tensors
|
| 264 |
+
3. Create proper attention masks for padded sequences
|
| 265 |
+
|
| 266 |
+
Why custom collator?
|
| 267 |
+
- We have custom fields (valid_mask) that need special handling
|
| 268 |
+
- Standard HF collators don't know about our mask format
|
| 269 |
+
"""
|
| 270 |
+
tokenizer: Any # Tokenizer for padding configuration
|
| 271 |
+
|
| 272 |
+
def __call__(self, features: List[Dict[str, Any]]) -> Dict[str, torch.Tensor]:
|
| 273 |
+
"""
|
| 274 |
+
Collate a list of examples into a batch.
|
| 275 |
+
|
| 276 |
+
Args:
|
| 277 |
+
features: List of dicts, each with:
|
| 278 |
+
- input_ids: List[int] - token IDs
|
| 279 |
+
- attention_mask: List[int] - attention mask
|
| 280 |
+
- label: int - target variable ID
|
| 281 |
+
- valid_mask: List[int] - binary mask of valid variables
|
| 282 |
+
|
| 283 |
+
Returns:
|
| 284 |
+
Dict with batched tensors:
|
| 285 |
+
- input_ids: [batch, max_seq_len]
|
| 286 |
+
- attention_mask: [batch, max_seq_len]
|
| 287 |
+
- labels: [batch]
|
| 288 |
+
- valid_mask: [batch, max_vars + 1]
|
| 289 |
+
"""
|
| 290 |
+
# Convert to tensors
|
| 291 |
+
input_ids = [torch.tensor(f["input_ids"], dtype=torch.long) for f in features]
|
| 292 |
+
attention_mask = [torch.tensor(f["attention_mask"], dtype=torch.long) for f in features]
|
| 293 |
+
labels = torch.tensor([f["label"] for f in features], dtype=torch.long)
|
| 294 |
+
valid_mask = torch.tensor([f["valid_mask"] for f in features], dtype=torch.bool)
|
| 295 |
+
|
| 296 |
+
# Pad sequences to same length within batch
|
| 297 |
+
# Using pad_sequence pads shorter sequences with padding_value
|
| 298 |
+
input_ids = torch.nn.utils.rnn.pad_sequence(
|
| 299 |
+
input_ids,
|
| 300 |
+
batch_first=True,
|
| 301 |
+
padding_value=self.tokenizer.pad_token_id
|
| 302 |
+
)
|
| 303 |
+
attention_mask = torch.nn.utils.rnn.pad_sequence(
|
| 304 |
+
attention_mask,
|
| 305 |
+
batch_first=True,
|
| 306 |
+
padding_value=0 # Padding positions get 0 attention
|
| 307 |
+
)
|
| 308 |
+
|
| 309 |
+
return {
|
| 310 |
+
"input_ids": input_ids,
|
| 311 |
+
"attention_mask": attention_mask,
|
| 312 |
+
"labels": labels,
|
| 313 |
+
"valid_mask": valid_mask,
|
| 314 |
+
}
|
| 315 |
+
|
| 316 |
+
|
| 317 |
+
# =============================================================================
|
| 318 |
+
# TRAINER: Custom loss computation with variable masking
|
| 319 |
+
# =============================================================================
|
| 320 |
+
|
| 321 |
+
class MaskedVarTrainer(Trainer):
|
| 322 |
+
"""
|
| 323 |
+
Custom HuggingFace Trainer with masked cross-entropy loss.
|
| 324 |
+
|
| 325 |
+
The key modification: before computing cross-entropy, we mask out logits
|
| 326 |
+
for invalid variables (those not appearing in the CNF). This ensures:
|
| 327 |
+
1. The model cannot predict invalid variables
|
| 328 |
+
2. No gradient flows to invalid variable logits
|
| 329 |
+
3. Training focuses only on distinguishing valid choices
|
| 330 |
+
|
| 331 |
+
NOTE on displayed metrics:
|
| 332 |
+
- 'loss' shown by Trainer is summed across GPUs (loss × world_size)
|
| 333 |
+
We add 'true_loss' which is the actual per-sample loss
|
| 334 |
+
- 'grad_norm' is the L2 norm across ALL ~4B parameters BEFORE clipping
|
| 335 |
+
Values of 100-200 are normal for large models; it gets clipped to max_grad_norm
|
| 336 |
+
"""
|
| 337 |
+
|
| 338 |
+
def __init__(self, *args, max_vars: int, **kwargs):
|
| 339 |
+
"""
|
| 340 |
+
Args:
|
| 341 |
+
max_vars: Maximum variable ID (for sanity checking labels)
|
| 342 |
+
*args, **kwargs: Passed to parent Trainer
|
| 343 |
+
"""
|
| 344 |
+
super().__init__(*args, **kwargs)
|
| 345 |
+
self.max_vars = max_vars
|
| 346 |
+
self._accumulated_loss = 0.0
|
| 347 |
+
self._loss_count = 0
|
| 348 |
+
|
| 349 |
+
def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
|
| 350 |
+
"""
|
| 351 |
+
Compute masked cross-entropy loss for variable classification.
|
| 352 |
+
|
| 353 |
+
Algorithm:
|
| 354 |
+
1. Extract labels and valid_mask from inputs
|
| 355 |
+
2. Forward pass to get logits
|
| 356 |
+
3. Set logits for invalid variables to -inf (or -1e4 for bf16 stability)
|
| 357 |
+
4. Compute cross-entropy loss
|
| 358 |
+
|
| 359 |
+
Args:
|
| 360 |
+
model: The QwenVarClassifier
|
| 361 |
+
inputs: Dict with input_ids, attention_mask, labels, valid_mask
|
| 362 |
+
return_outputs: If True, return (loss, outputs) tuple
|
| 363 |
+
num_items_in_batch: Unused (for API compatibility)
|
| 364 |
+
|
| 365 |
+
Returns:
|
| 366 |
+
loss: Scalar loss value, or (loss, outputs) tuple if return_outputs=True
|
| 367 |
+
"""
|
| 368 |
+
# Get labels and mask (don't pop - prediction_loop needs labels for compute_metrics)
|
| 369 |
+
labels = inputs.get("labels") # [batch]
|
| 370 |
+
valid_mask = inputs.get("valid_mask") # [batch, max_vars + 1] boolean
|
| 371 |
+
|
| 372 |
+
# Remove from inputs for model.forward (which doesn't expect them)
|
| 373 |
+
model_inputs = {k: v for k, v in inputs.items() if k not in ["labels", "valid_mask"]}
|
| 374 |
+
|
| 375 |
+
# Forward pass
|
| 376 |
+
outputs = model(**model_inputs)
|
| 377 |
+
logits = outputs["logits"] # [batch, max_vars + 1]
|
| 378 |
+
|
| 379 |
+
# DEBUG: Check if label is in valid_mask
|
| 380 |
+
if DEBUG_TRAINING:
|
| 381 |
+
if not hasattr(self, '_loss_debug_count'):
|
| 382 |
+
self._loss_debug_count = 0
|
| 383 |
+
if self._loss_debug_count < 5:
|
| 384 |
+
for i, (lbl, vmask) in enumerate(zip(labels, valid_mask)):
|
| 385 |
+
label_in_mask = vmask[lbl].item()
|
| 386 |
+
valid_count = vmask.sum().item()
|
| 387 |
+
logit_at_label = logits[i, lbl].item()
|
| 388 |
+
print(f"[LOSS DEBUG {self._loss_debug_count}] label={lbl.item()}, in_mask={label_in_mask}, valid_vars={valid_count}, logit_at_label={logit_at_label:.2f}")
|
| 389 |
+
self._loss_debug_count += 1
|
| 390 |
+
|
| 391 |
+
# Mask invalid variables by setting their logits to a large negative value
|
| 392 |
+
# After softmax, these will have probability ≈ 0
|
| 393 |
+
#
|
| 394 |
+
# Why -1e4 instead of -inf or -1e9?
|
| 395 |
+
# - bfloat16 has limited dynamic range
|
| 396 |
+
# - -1e9 can cause NaN issues when computing softmax/cross-entropy
|
| 397 |
+
# - -1e4 is small enough to give ~0 probability while staying numerically stable
|
| 398 |
+
logits = logits.masked_fill(~valid_mask.to(logits.device), -1e4)
|
| 399 |
+
|
| 400 |
+
# Sanity check: labels must be valid variable IDs (1 to max_vars)
|
| 401 |
+
# This catches data bugs early
|
| 402 |
+
if torch.any(labels <= 0) or torch.any(labels > self.max_vars):
|
| 403 |
+
bad = labels[(labels <= 0) | (labels > self.max_vars)].detach().cpu().tolist()
|
| 404 |
+
raise ValueError(f"Out-of-range labels detected (showing up to 20): {bad[:20]}")
|
| 405 |
+
|
| 406 |
+
# DEBUG: Check logit at label after masking
|
| 407 |
+
if DEBUG_TRAINING and hasattr(self, '_loss_debug_count') and self._loss_debug_count <= 5:
|
| 408 |
+
for i, lbl in enumerate(labels):
|
| 409 |
+
masked_logit = logits[i, lbl].item()
|
| 410 |
+
print(f"[LOSS DEBUG] after mask: logit_at_label={masked_logit:.2f}")
|
| 411 |
+
|
| 412 |
+
# Standard cross-entropy loss
|
| 413 |
+
# PyTorch's cross_entropy expects logits, not probabilities
|
| 414 |
+
loss = F.cross_entropy(logits, labels.to(logits.device))
|
| 415 |
+
|
| 416 |
+
# Track true loss for accurate logging
|
| 417 |
+
self._accumulated_loss += loss.item()
|
| 418 |
+
self._loss_count += 1
|
| 419 |
+
|
| 420 |
+
# DEBUG: Print loss
|
| 421 |
+
if DEBUG_TRAINING and hasattr(self, '_loss_debug_count') and self._loss_debug_count <= 5:
|
| 422 |
+
print(f"[LOSS DEBUG] loss={loss.item():.2f}")
|
| 423 |
+
|
| 424 |
+
# Return masked logits in outputs (so compute_metrics gets properly masked predictions)
|
| 425 |
+
masked_outputs = {"logits": logits}
|
| 426 |
+
return (loss, masked_outputs) if return_outputs else loss
|
| 427 |
+
|
| 428 |
+
def prediction_step(self, model, inputs, prediction_loss_only, ignore_keys=None):
|
| 429 |
+
"""
|
| 430 |
+
Override prediction_step to properly return loss and logits for evaluation.
|
| 431 |
+
|
| 432 |
+
The default HF Trainer prediction_step doesn't work well with custom compute_loss,
|
| 433 |
+
so we implement our own that properly computes masked loss and returns logits.
|
| 434 |
+
"""
|
| 435 |
+
model.eval()
|
| 436 |
+
|
| 437 |
+
with torch.no_grad():
|
| 438 |
+
# Get labels and mask
|
| 439 |
+
labels = inputs.get("labels")
|
| 440 |
+
valid_mask = inputs.get("valid_mask")
|
| 441 |
+
|
| 442 |
+
# Forward pass
|
| 443 |
+
model_inputs = {k: v for k, v in inputs.items() if k not in ["labels", "valid_mask"]}
|
| 444 |
+
outputs = model(**model_inputs)
|
| 445 |
+
logits = outputs["logits"]
|
| 446 |
+
|
| 447 |
+
# Mask invalid variables
|
| 448 |
+
logits = logits.masked_fill(~valid_mask.to(logits.device), -1e4)
|
| 449 |
+
|
| 450 |
+
# Compute loss
|
| 451 |
+
loss = F.cross_entropy(logits, labels.to(logits.device))
|
| 452 |
+
|
| 453 |
+
# Return (loss, logits, labels) - this is what compute_metrics expects
|
| 454 |
+
return (loss, logits.detach(), labels.detach())
|
| 455 |
+
|
| 456 |
+
def log(self, logs: Dict[str, float], start_time: float = None) -> None:
|
| 457 |
+
"""
|
| 458 |
+
Override log to add true_loss and ensure eval metrics are logged to W&B.
|
| 459 |
+
|
| 460 |
+
The default 'loss' in HF Trainer is summed across GPUs in DDP/DeepSpeed.
|
| 461 |
+
We track the actual per-sample loss and report it as 'true_loss'.
|
| 462 |
+
"""
|
| 463 |
+
if self._loss_count > 0:
|
| 464 |
+
# Calculate true average loss on this device
|
| 465 |
+
true_loss = self._accumulated_loss / self._loss_count
|
| 466 |
+
logs["true_loss"] = round(true_loss, 4)
|
| 467 |
+
|
| 468 |
+
# Reset for next logging interval
|
| 469 |
+
self._accumulated_loss = 0.0
|
| 470 |
+
self._loss_count = 0
|
| 471 |
+
|
| 472 |
+
# Let HF Trainer handle W&B logging - it manages step ordering correctly
|
| 473 |
+
super().log(logs, start_time)
|
| 474 |
+
|
| 475 |
+
|
| 476 |
+
def compute_metrics(eval_pred):
|
| 477 |
+
"""
|
| 478 |
+
Compute accuracy for evaluation.
|
| 479 |
+
|
| 480 |
+
Args:
|
| 481 |
+
eval_pred: (logits, labels) from Trainer's prediction_loop
|
| 482 |
+
- logits: [num_samples, max_vars + 1] (already masked with -1e4 for invalid vars)
|
| 483 |
+
- labels: [num_samples]
|
| 484 |
+
|
| 485 |
+
Returns:
|
| 486 |
+
Dict with "accuracy" (Trainer will prefix with "eval_")
|
| 487 |
+
|
| 488 |
+
Note: eval_loss is computed automatically by Trainer from prediction_step's loss.
|
| 489 |
+
We don't need to compute it here.
|
| 490 |
+
|
| 491 |
+
Since invalid variables have logits ≈ -1e4, argmax will naturally avoid them.
|
| 492 |
+
"""
|
| 493 |
+
logits, labels = eval_pred
|
| 494 |
+
|
| 495 |
+
# Accuracy: argmax prediction vs true label
|
| 496 |
+
preds = np.argmax(logits, axis=-1)
|
| 497 |
+
accuracy = float((preds == labels).mean())
|
| 498 |
+
|
| 499 |
+
return {"accuracy": accuracy}
|
| 500 |
+
|
| 501 |
+
|
| 502 |
+
def get_wandb_report_to():
|
| 503 |
+
"""
|
| 504 |
+
Determine if this process should log to W&B.
|
| 505 |
+
|
| 506 |
+
Only the main process (rank 0) should log to W&B to avoid creating multiple runs.
|
| 507 |
+
Other ranks should not log to any external service.
|
| 508 |
+
|
| 509 |
+
Returns:
|
| 510 |
+
["wandb"] for rank 0, [] for other ranks
|
| 511 |
+
"""
|
| 512 |
+
local_rank = int(os.environ.get("LOCAL_RANK", 0))
|
| 513 |
+
|
| 514 |
+
if local_rank == 0:
|
| 515 |
+
return ["wandb"]
|
| 516 |
+
else:
|
| 517 |
+
return []
|
| 518 |
+
|
| 519 |
+
|
| 520 |
+
# =============================================================================
|
| 521 |
+
# MAIN: Training pipeline
|
| 522 |
+
# =============================================================================
|
| 523 |
+
|
| 524 |
+
def main():
|
| 525 |
+
"""
|
| 526 |
+
Main training function.
|
| 527 |
+
|
| 528 |
+
Pipeline:
|
| 529 |
+
1. Parse command line arguments
|
| 530 |
+
2. Load tokenizer and datasets
|
| 531 |
+
3. Preprocess: tokenize CNF text, compute valid masks
|
| 532 |
+
4. Initialize model with pretrained backbone + new classification head
|
| 533 |
+
5. Configure training (optimizer, scheduler, logging, etc.)
|
| 534 |
+
6. Train and evaluate
|
| 535 |
+
"""
|
| 536 |
+
ap = argparse.ArgumentParser(
|
| 537 |
+
description="Train a Qwen-based variable classifier for SAT branching"
|
| 538 |
+
)
|
| 539 |
+
|
| 540 |
+
# Model and data arguments
|
| 541 |
+
ap.add_argument("--model_name", type=str, default="Qwen/Qwen3-4B",
|
| 542 |
+
help="HuggingFace model ID for the backbone")
|
| 543 |
+
ap.add_argument("--train_jsonl", type=str, required=True,
|
| 544 |
+
help="Path to training data (JSONL with 'cnf' and 'label' fields)")
|
| 545 |
+
ap.add_argument("--valid_jsonl", type=str, required=True,
|
| 546 |
+
help="Path to validation data (same format)")
|
| 547 |
+
ap.add_argument("--output_dir", type=str, default="./out_qwen_var_sft",
|
| 548 |
+
help="Directory for checkpoints and logs")
|
| 549 |
+
ap.add_argument("--max_vars", type=int, default=500,
|
| 550 |
+
help="Maximum variable ID (determines output dimension)")
|
| 551 |
+
ap.add_argument("--max_length", type=int, default=8192,
|
| 552 |
+
help="Maximum sequence length in tokens (truncates longer CNFs)")
|
| 553 |
+
ap.add_argument("--seed", type=int, default=0,
|
| 554 |
+
help="Random seed for reproducibility")
|
| 555 |
+
|
| 556 |
+
# Training hyperparameters
|
| 557 |
+
ap.add_argument("--per_device_train_batch_size", type=int, default=1,
|
| 558 |
+
help="Batch size per GPU for training")
|
| 559 |
+
ap.add_argument("--per_device_eval_batch_size", type=int, default=1,
|
| 560 |
+
help="Batch size per GPU for evaluation")
|
| 561 |
+
ap.add_argument("--gradient_accumulation_steps", type=int, default=8,
|
| 562 |
+
help="Accumulate gradients over this many steps (effective batch = this * batch_size * num_gpus)")
|
| 563 |
+
ap.add_argument("--learning_rate", type=float, default=5e-6,
|
| 564 |
+
help="Peak learning rate (after warmup). Lower than typical fine-tuning due to classification head")
|
| 565 |
+
ap.add_argument("--num_train_epochs", type=float, default=3.0,
|
| 566 |
+
help="Total training epochs")
|
| 567 |
+
ap.add_argument("--warmup_ratio", type=float, default=0.03,
|
| 568 |
+
help="Fraction of training steps for learning rate warmup")
|
| 569 |
+
ap.add_argument("--weight_decay", type=float, default=0.0,
|
| 570 |
+
help="Weight decay (L2 regularization)")
|
| 571 |
+
ap.add_argument("--logging_steps", type=int, default=10,
|
| 572 |
+
help="Log training metrics every N steps")
|
| 573 |
+
ap.add_argument("--eval_steps", type=int, default=200,
|
| 574 |
+
help="Evaluate every N steps")
|
| 575 |
+
ap.add_argument("--save_steps", type=int, default=200,
|
| 576 |
+
help="Save checkpoint every N steps")
|
| 577 |
+
ap.add_argument("--report_to", type=str, default="wandb",
|
| 578 |
+
choices=["wandb", "tensorboard", "none"],
|
| 579 |
+
help="Logging backend")
|
| 580 |
+
ap.add_argument("--deepspeed", type=str, default=None,
|
| 581 |
+
help="Path to DeepSpeed config JSON for distributed training")
|
| 582 |
+
|
| 583 |
+
args = ap.parse_args()
|
| 584 |
+
|
| 585 |
+
# Set random seeds for reproducibility
|
| 586 |
+
set_seed(args.seed)
|
| 587 |
+
|
| 588 |
+
# Load tokenizer
|
| 589 |
+
# Qwen uses a byte-level BPE tokenizer
|
| 590 |
+
tok = AutoTokenizer.from_pretrained(args.model_name, use_fast=True)
|
| 591 |
+
if tok.pad_token is None:
|
| 592 |
+
# Qwen doesn't have a dedicated pad token; use eos as pad
|
| 593 |
+
tok.pad_token = tok.eos_token
|
| 594 |
+
|
| 595 |
+
# Load datasets from JSONL files
|
| 596 |
+
ds = load_dataset(
|
| 597 |
+
"json",
|
| 598 |
+
data_files={"train": args.train_jsonl, "validation": args.valid_jsonl},
|
| 599 |
+
)
|
| 600 |
+
|
| 601 |
+
def preprocess(ex):
|
| 602 |
+
"""
|
| 603 |
+
Preprocess a single example.
|
| 604 |
+
|
| 605 |
+
Steps:
|
| 606 |
+
1. Tokenize the CNF text
|
| 607 |
+
2. Compute valid variable mask
|
| 608 |
+
3. Return features for training
|
| 609 |
+
|
| 610 |
+
Args:
|
| 611 |
+
ex: Dict with 'cnf' (str) and 'label' (int)
|
| 612 |
+
|
| 613 |
+
Returns:
|
| 614 |
+
Dict with input_ids, attention_mask, label, valid_mask
|
| 615 |
+
"""
|
| 616 |
+
cnf = ex["cnf"]
|
| 617 |
+
label = int(ex["label"])
|
| 618 |
+
|
| 619 |
+
# Tokenize CNF text
|
| 620 |
+
# No special prompt/instruction - the model learns to interpret raw CNF
|
| 621 |
+
enc = tok(
|
| 622 |
+
cnf,
|
| 623 |
+
truncation=True,
|
| 624 |
+
max_length=args.max_length,
|
| 625 |
+
padding=False # We handle padding in the collator
|
| 626 |
+
)
|
| 627 |
+
|
| 628 |
+
return {
|
| 629 |
+
"input_ids": enc["input_ids"],
|
| 630 |
+
"attention_mask": enc["attention_mask"],
|
| 631 |
+
"label": label,
|
| 632 |
+
"valid_mask": cnf_valid_mask(cnf, args.max_vars),
|
| 633 |
+
}
|
| 634 |
+
|
| 635 |
+
# Apply preprocessing to all examples
|
| 636 |
+
# remove_columns drops original fields (cnf, label) since we've extracted what we need
|
| 637 |
+
ds = ds.map(preprocess, remove_columns=ds["train"].column_names)
|
| 638 |
+
|
| 639 |
+
# Initialize model
|
| 640 |
+
model = QwenVarClassifier(args.model_name, max_vars=args.max_vars)
|
| 641 |
+
|
| 642 |
+
# Enable gradient checkpointing to save memory on long sequences
|
| 643 |
+
# This trades compute for memory by recomputing activations during backward pass
|
| 644 |
+
model.backbone.gradient_checkpointing_enable()
|
| 645 |
+
|
| 646 |
+
# Configure W&B logging (only rank 0 logs to avoid duplicate runs)
|
| 647 |
+
report_to = get_wandb_report_to()
|
| 648 |
+
|
| 649 |
+
# Configure training
|
| 650 |
+
training_args = TrainingArguments(
|
| 651 |
+
output_dir=args.output_dir,
|
| 652 |
+
overwrite_output_dir=True,
|
| 653 |
+
|
| 654 |
+
# Precision settings for modern GPUs
|
| 655 |
+
bf16=True, # Use bfloat16 for training (good for H100/A100)
|
| 656 |
+
tf32=True, # Enable TF32 for faster matmuls on Ampere+
|
| 657 |
+
|
| 658 |
+
# Batch configuration
|
| 659 |
+
per_device_train_batch_size=args.per_device_train_batch_size,
|
| 660 |
+
per_device_eval_batch_size=args.per_device_eval_batch_size,
|
| 661 |
+
gradient_accumulation_steps=args.gradient_accumulation_steps,
|
| 662 |
+
|
| 663 |
+
# Optimizer settings
|
| 664 |
+
learning_rate=args.learning_rate,
|
| 665 |
+
warmup_ratio=args.warmup_ratio,
|
| 666 |
+
num_train_epochs=args.num_train_epochs,
|
| 667 |
+
weight_decay=args.weight_decay,
|
| 668 |
+
|
| 669 |
+
# Gradient clipping for training stability
|
| 670 |
+
# Clips gradient norm to this value if it exceeds it
|
| 671 |
+
# This prevents exploding gradients from destabilizing training
|
| 672 |
+
max_grad_norm=1.0,
|
| 673 |
+
|
| 674 |
+
# Logging and evaluation
|
| 675 |
+
logging_steps=args.logging_steps,
|
| 676 |
+
eval_strategy="steps",
|
| 677 |
+
eval_steps=args.eval_steps,
|
| 678 |
+
|
| 679 |
+
# Checkpointing - keep best checkpoints based on validation accuracy
|
| 680 |
+
save_strategy="steps",
|
| 681 |
+
save_steps=args.save_steps,
|
| 682 |
+
save_total_limit=3, # Keep best 3 checkpoints
|
| 683 |
+
load_best_model_at_end=True, # Load best checkpoint at end of training
|
| 684 |
+
metric_for_best_model="eval_accuracy", # Use validation accuracy to determine best
|
| 685 |
+
greater_is_better=True, # Higher accuracy is better
|
| 686 |
+
|
| 687 |
+
# Logging backend
|
| 688 |
+
report_to=report_to,
|
| 689 |
+
run_name=os.environ.get("WANDB_RUN_NAME", "qwen-var-sft") if args.report_to == "wandb" else None,
|
| 690 |
+
logging_dir=os.path.join(args.output_dir, "logs"),
|
| 691 |
+
|
| 692 |
+
# Important: don't remove valid_mask column (we need it in compute_loss)
|
| 693 |
+
remove_unused_columns=False,
|
| 694 |
+
|
| 695 |
+
# DDP settings (for multi-GPU)
|
| 696 |
+
ddp_find_unused_parameters=False,
|
| 697 |
+
|
| 698 |
+
# DeepSpeed for efficient distributed training
|
| 699 |
+
deepspeed=args.deepspeed,
|
| 700 |
+
|
| 701 |
+
# Use pickle format for saving (safetensors has issues with some weight tying configs)
|
| 702 |
+
save_safetensors=False,
|
| 703 |
+
)
|
| 704 |
+
|
| 705 |
+
# Create trainer with custom loss computation
|
| 706 |
+
trainer = MaskedVarTrainer(
|
| 707 |
+
model=model,
|
| 708 |
+
args=training_args,
|
| 709 |
+
train_dataset=ds["train"],
|
| 710 |
+
eval_dataset=ds["validation"],
|
| 711 |
+
tokenizer=tok,
|
| 712 |
+
data_collator=Collator(tok),
|
| 713 |
+
compute_metrics=compute_metrics,
|
| 714 |
+
max_vars=args.max_vars,
|
| 715 |
+
)
|
| 716 |
+
|
| 717 |
+
# Train!
|
| 718 |
+
trainer.train()
|
| 719 |
+
|
| 720 |
+
# Final evaluation
|
| 721 |
+
trainer.evaluate()
|
| 722 |
+
|
| 723 |
+
|
| 724 |
+
if __name__ == "__main__":
|
| 725 |
+
main()
|
special_tokens_map.json
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"additional_special_tokens": [
|
| 3 |
+
"<|im_start|>",
|
| 4 |
+
"<|im_end|>",
|
| 5 |
+
"<|object_ref_start|>",
|
| 6 |
+
"<|object_ref_end|>",
|
| 7 |
+
"<|box_start|>",
|
| 8 |
+
"<|box_end|>",
|
| 9 |
+
"<|quad_start|>",
|
| 10 |
+
"<|quad_end|>",
|
| 11 |
+
"<|vision_start|>",
|
| 12 |
+
"<|vision_end|>",
|
| 13 |
+
"<|vision_pad|>",
|
| 14 |
+
"<|image_pad|>",
|
| 15 |
+
"<|video_pad|>"
|
| 16 |
+
],
|
| 17 |
+
"eos_token": {
|
| 18 |
+
"content": "<|im_end|>",
|
| 19 |
+
"lstrip": false,
|
| 20 |
+
"normalized": false,
|
| 21 |
+
"rstrip": false,
|
| 22 |
+
"single_word": false
|
| 23 |
+
},
|
| 24 |
+
"pad_token": {
|
| 25 |
+
"content": "<|endoftext|>",
|
| 26 |
+
"lstrip": false,
|
| 27 |
+
"normalized": false,
|
| 28 |
+
"rstrip": false,
|
| 29 |
+
"single_word": false
|
| 30 |
+
}
|
| 31 |
+
}
|
tokenizer.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b7a6a993d40b42d517297bb247ff66679e5bc9dd7a5143be0620faf210b42861
|
| 3 |
+
size 11422753
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,239 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_bos_token": false,
|
| 3 |
+
"add_prefix_space": false,
|
| 4 |
+
"added_tokens_decoder": {
|
| 5 |
+
"151643": {
|
| 6 |
+
"content": "<|endoftext|>",
|
| 7 |
+
"lstrip": false,
|
| 8 |
+
"normalized": false,
|
| 9 |
+
"rstrip": false,
|
| 10 |
+
"single_word": false,
|
| 11 |
+
"special": true
|
| 12 |
+
},
|
| 13 |
+
"151644": {
|
| 14 |
+
"content": "<|im_start|>",
|
| 15 |
+
"lstrip": false,
|
| 16 |
+
"normalized": false,
|
| 17 |
+
"rstrip": false,
|
| 18 |
+
"single_word": false,
|
| 19 |
+
"special": true
|
| 20 |
+
},
|
| 21 |
+
"151645": {
|
| 22 |
+
"content": "<|im_end|>",
|
| 23 |
+
"lstrip": false,
|
| 24 |
+
"normalized": false,
|
| 25 |
+
"rstrip": false,
|
| 26 |
+
"single_word": false,
|
| 27 |
+
"special": true
|
| 28 |
+
},
|
| 29 |
+
"151646": {
|
| 30 |
+
"content": "<|object_ref_start|>",
|
| 31 |
+
"lstrip": false,
|
| 32 |
+
"normalized": false,
|
| 33 |
+
"rstrip": false,
|
| 34 |
+
"single_word": false,
|
| 35 |
+
"special": true
|
| 36 |
+
},
|
| 37 |
+
"151647": {
|
| 38 |
+
"content": "<|object_ref_end|>",
|
| 39 |
+
"lstrip": false,
|
| 40 |
+
"normalized": false,
|
| 41 |
+
"rstrip": false,
|
| 42 |
+
"single_word": false,
|
| 43 |
+
"special": true
|
| 44 |
+
},
|
| 45 |
+
"151648": {
|
| 46 |
+
"content": "<|box_start|>",
|
| 47 |
+
"lstrip": false,
|
| 48 |
+
"normalized": false,
|
| 49 |
+
"rstrip": false,
|
| 50 |
+
"single_word": false,
|
| 51 |
+
"special": true
|
| 52 |
+
},
|
| 53 |
+
"151649": {
|
| 54 |
+
"content": "<|box_end|>",
|
| 55 |
+
"lstrip": false,
|
| 56 |
+
"normalized": false,
|
| 57 |
+
"rstrip": false,
|
| 58 |
+
"single_word": false,
|
| 59 |
+
"special": true
|
| 60 |
+
},
|
| 61 |
+
"151650": {
|
| 62 |
+
"content": "<|quad_start|>",
|
| 63 |
+
"lstrip": false,
|
| 64 |
+
"normalized": false,
|
| 65 |
+
"rstrip": false,
|
| 66 |
+
"single_word": false,
|
| 67 |
+
"special": true
|
| 68 |
+
},
|
| 69 |
+
"151651": {
|
| 70 |
+
"content": "<|quad_end|>",
|
| 71 |
+
"lstrip": false,
|
| 72 |
+
"normalized": false,
|
| 73 |
+
"rstrip": false,
|
| 74 |
+
"single_word": false,
|
| 75 |
+
"special": true
|
| 76 |
+
},
|
| 77 |
+
"151652": {
|
| 78 |
+
"content": "<|vision_start|>",
|
| 79 |
+
"lstrip": false,
|
| 80 |
+
"normalized": false,
|
| 81 |
+
"rstrip": false,
|
| 82 |
+
"single_word": false,
|
| 83 |
+
"special": true
|
| 84 |
+
},
|
| 85 |
+
"151653": {
|
| 86 |
+
"content": "<|vision_end|>",
|
| 87 |
+
"lstrip": false,
|
| 88 |
+
"normalized": false,
|
| 89 |
+
"rstrip": false,
|
| 90 |
+
"single_word": false,
|
| 91 |
+
"special": true
|
| 92 |
+
},
|
| 93 |
+
"151654": {
|
| 94 |
+
"content": "<|vision_pad|>",
|
| 95 |
+
"lstrip": false,
|
| 96 |
+
"normalized": false,
|
| 97 |
+
"rstrip": false,
|
| 98 |
+
"single_word": false,
|
| 99 |
+
"special": true
|
| 100 |
+
},
|
| 101 |
+
"151655": {
|
| 102 |
+
"content": "<|image_pad|>",
|
| 103 |
+
"lstrip": false,
|
| 104 |
+
"normalized": false,
|
| 105 |
+
"rstrip": false,
|
| 106 |
+
"single_word": false,
|
| 107 |
+
"special": true
|
| 108 |
+
},
|
| 109 |
+
"151656": {
|
| 110 |
+
"content": "<|video_pad|>",
|
| 111 |
+
"lstrip": false,
|
| 112 |
+
"normalized": false,
|
| 113 |
+
"rstrip": false,
|
| 114 |
+
"single_word": false,
|
| 115 |
+
"special": true
|
| 116 |
+
},
|
| 117 |
+
"151657": {
|
| 118 |
+
"content": "<tool_call>",
|
| 119 |
+
"lstrip": false,
|
| 120 |
+
"normalized": false,
|
| 121 |
+
"rstrip": false,
|
| 122 |
+
"single_word": false,
|
| 123 |
+
"special": false
|
| 124 |
+
},
|
| 125 |
+
"151658": {
|
| 126 |
+
"content": "</tool_call>",
|
| 127 |
+
"lstrip": false,
|
| 128 |
+
"normalized": false,
|
| 129 |
+
"rstrip": false,
|
| 130 |
+
"single_word": false,
|
| 131 |
+
"special": false
|
| 132 |
+
},
|
| 133 |
+
"151659": {
|
| 134 |
+
"content": "<|fim_prefix|>",
|
| 135 |
+
"lstrip": false,
|
| 136 |
+
"normalized": false,
|
| 137 |
+
"rstrip": false,
|
| 138 |
+
"single_word": false,
|
| 139 |
+
"special": false
|
| 140 |
+
},
|
| 141 |
+
"151660": {
|
| 142 |
+
"content": "<|fim_middle|>",
|
| 143 |
+
"lstrip": false,
|
| 144 |
+
"normalized": false,
|
| 145 |
+
"rstrip": false,
|
| 146 |
+
"single_word": false,
|
| 147 |
+
"special": false
|
| 148 |
+
},
|
| 149 |
+
"151661": {
|
| 150 |
+
"content": "<|fim_suffix|>",
|
| 151 |
+
"lstrip": false,
|
| 152 |
+
"normalized": false,
|
| 153 |
+
"rstrip": false,
|
| 154 |
+
"single_word": false,
|
| 155 |
+
"special": false
|
| 156 |
+
},
|
| 157 |
+
"151662": {
|
| 158 |
+
"content": "<|fim_pad|>",
|
| 159 |
+
"lstrip": false,
|
| 160 |
+
"normalized": false,
|
| 161 |
+
"rstrip": false,
|
| 162 |
+
"single_word": false,
|
| 163 |
+
"special": false
|
| 164 |
+
},
|
| 165 |
+
"151663": {
|
| 166 |
+
"content": "<|repo_name|>",
|
| 167 |
+
"lstrip": false,
|
| 168 |
+
"normalized": false,
|
| 169 |
+
"rstrip": false,
|
| 170 |
+
"single_word": false,
|
| 171 |
+
"special": false
|
| 172 |
+
},
|
| 173 |
+
"151664": {
|
| 174 |
+
"content": "<|file_sep|>",
|
| 175 |
+
"lstrip": false,
|
| 176 |
+
"normalized": false,
|
| 177 |
+
"rstrip": false,
|
| 178 |
+
"single_word": false,
|
| 179 |
+
"special": false
|
| 180 |
+
},
|
| 181 |
+
"151665": {
|
| 182 |
+
"content": "<tool_response>",
|
| 183 |
+
"lstrip": false,
|
| 184 |
+
"normalized": false,
|
| 185 |
+
"rstrip": false,
|
| 186 |
+
"single_word": false,
|
| 187 |
+
"special": false
|
| 188 |
+
},
|
| 189 |
+
"151666": {
|
| 190 |
+
"content": "</tool_response>",
|
| 191 |
+
"lstrip": false,
|
| 192 |
+
"normalized": false,
|
| 193 |
+
"rstrip": false,
|
| 194 |
+
"single_word": false,
|
| 195 |
+
"special": false
|
| 196 |
+
},
|
| 197 |
+
"151667": {
|
| 198 |
+
"content": "<think>",
|
| 199 |
+
"lstrip": false,
|
| 200 |
+
"normalized": false,
|
| 201 |
+
"rstrip": false,
|
| 202 |
+
"single_word": false,
|
| 203 |
+
"special": false
|
| 204 |
+
},
|
| 205 |
+
"151668": {
|
| 206 |
+
"content": "</think>",
|
| 207 |
+
"lstrip": false,
|
| 208 |
+
"normalized": false,
|
| 209 |
+
"rstrip": false,
|
| 210 |
+
"single_word": false,
|
| 211 |
+
"special": false
|
| 212 |
+
}
|
| 213 |
+
},
|
| 214 |
+
"additional_special_tokens": [
|
| 215 |
+
"<|im_start|>",
|
| 216 |
+
"<|im_end|>",
|
| 217 |
+
"<|object_ref_start|>",
|
| 218 |
+
"<|object_ref_end|>",
|
| 219 |
+
"<|box_start|>",
|
| 220 |
+
"<|box_end|>",
|
| 221 |
+
"<|quad_start|>",
|
| 222 |
+
"<|quad_end|>",
|
| 223 |
+
"<|vision_start|>",
|
| 224 |
+
"<|vision_end|>",
|
| 225 |
+
"<|vision_pad|>",
|
| 226 |
+
"<|image_pad|>",
|
| 227 |
+
"<|video_pad|>"
|
| 228 |
+
],
|
| 229 |
+
"bos_token": null,
|
| 230 |
+
"clean_up_tokenization_spaces": false,
|
| 231 |
+
"eos_token": "<|im_end|>",
|
| 232 |
+
"errors": "replace",
|
| 233 |
+
"extra_special_tokens": {},
|
| 234 |
+
"model_max_length": 131072,
|
| 235 |
+
"pad_token": "<|endoftext|>",
|
| 236 |
+
"split_special_tokens": false,
|
| 237 |
+
"tokenizer_class": "Qwen2Tokenizer",
|
| 238 |
+
"unk_token": null
|
| 239 |
+
}
|
trainer_state.json
ADDED
|
@@ -0,0 +1,1884 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"best_global_step": 740,
|
| 3 |
+
"best_metric": 0.2394678492239468,
|
| 4 |
+
"best_model_checkpoint": "out_qwen_0.6b_sft_hq/checkpoint-740",
|
| 5 |
+
"epoch": 5.828402366863905,
|
| 6 |
+
"eval_steps": 10,
|
| 7 |
+
"global_step": 740,
|
| 8 |
+
"is_hyper_param_search": false,
|
| 9 |
+
"is_local_process_zero": true,
|
| 10 |
+
"is_world_process_zero": true,
|
| 11 |
+
"log_history": [
|
| 12 |
+
{
|
| 13 |
+
"epoch": 0.03944773175542406,
|
| 14 |
+
"grad_norm": 137.58168100959043,
|
| 15 |
+
"learning_rate": 1.9607843137254904e-07,
|
| 16 |
+
"loss": 33.8811,
|
| 17 |
+
"step": 5,
|
| 18 |
+
"true_loss": 4.4586
|
| 19 |
+
},
|
| 20 |
+
{
|
| 21 |
+
"epoch": 0.07889546351084813,
|
| 22 |
+
"grad_norm": 96.05399842775171,
|
| 23 |
+
"learning_rate": 4.4117647058823536e-07,
|
| 24 |
+
"loss": 33.9447,
|
| 25 |
+
"step": 10,
|
| 26 |
+
"true_loss": 4.3732
|
| 27 |
+
},
|
| 28 |
+
{
|
| 29 |
+
"epoch": 0.07889546351084813,
|
| 30 |
+
"eval_accuracy": 0.022172949002217297,
|
| 31 |
+
"eval_loss": 4.352970123291016,
|
| 32 |
+
"eval_runtime": 15.7645,
|
| 33 |
+
"eval_samples_per_second": 57.217,
|
| 34 |
+
"eval_steps_per_second": 7.168,
|
| 35 |
+
"step": 10
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"epoch": 0.11834319526627218,
|
| 39 |
+
"grad_norm": 119.9455370680162,
|
| 40 |
+
"learning_rate": 6.862745098039217e-07,
|
| 41 |
+
"loss": 35.4633,
|
| 42 |
+
"step": 15,
|
| 43 |
+
"true_loss": 4.4363
|
| 44 |
+
},
|
| 45 |
+
{
|
| 46 |
+
"epoch": 0.15779092702169625,
|
| 47 |
+
"grad_norm": 115.69507181998236,
|
| 48 |
+
"learning_rate": 9.313725490196079e-07,
|
| 49 |
+
"loss": 35.4213,
|
| 50 |
+
"step": 20,
|
| 51 |
+
"true_loss": 4.3914
|
| 52 |
+
},
|
| 53 |
+
{
|
| 54 |
+
"epoch": 0.15779092702169625,
|
| 55 |
+
"eval_accuracy": 0.02106430155210643,
|
| 56 |
+
"eval_loss": 4.327692031860352,
|
| 57 |
+
"eval_runtime": 14.936,
|
| 58 |
+
"eval_samples_per_second": 60.391,
|
| 59 |
+
"eval_steps_per_second": 7.566,
|
| 60 |
+
"step": 20
|
| 61 |
+
},
|
| 62 |
+
{
|
| 63 |
+
"epoch": 0.19723865877712032,
|
| 64 |
+
"grad_norm": 109.18491807311037,
|
| 65 |
+
"learning_rate": 1.1764705882352942e-06,
|
| 66 |
+
"loss": 34.7285,
|
| 67 |
+
"step": 25,
|
| 68 |
+
"true_loss": 4.5508
|
| 69 |
+
},
|
| 70 |
+
{
|
| 71 |
+
"epoch": 0.23668639053254437,
|
| 72 |
+
"grad_norm": 114.28553037452185,
|
| 73 |
+
"learning_rate": 1.4215686274509805e-06,
|
| 74 |
+
"loss": 34.658,
|
| 75 |
+
"step": 30,
|
| 76 |
+
"true_loss": 4.4059
|
| 77 |
+
},
|
| 78 |
+
{
|
| 79 |
+
"epoch": 0.23668639053254437,
|
| 80 |
+
"eval_accuracy": 0.017738359201773836,
|
| 81 |
+
"eval_loss": 4.260367393493652,
|
| 82 |
+
"eval_runtime": 14.9567,
|
| 83 |
+
"eval_samples_per_second": 60.308,
|
| 84 |
+
"eval_steps_per_second": 7.555,
|
| 85 |
+
"step": 30
|
| 86 |
+
},
|
| 87 |
+
{
|
| 88 |
+
"epoch": 0.27613412228796846,
|
| 89 |
+
"grad_norm": 109.9826234366543,
|
| 90 |
+
"learning_rate": 1.6666666666666667e-06,
|
| 91 |
+
"loss": 34.1344,
|
| 92 |
+
"step": 35,
|
| 93 |
+
"true_loss": 4.1516
|
| 94 |
+
},
|
| 95 |
+
{
|
| 96 |
+
"epoch": 0.3155818540433925,
|
| 97 |
+
"grad_norm": 123.05318590857651,
|
| 98 |
+
"learning_rate": 1.9117647058823528e-06,
|
| 99 |
+
"loss": 34.1885,
|
| 100 |
+
"step": 40,
|
| 101 |
+
"true_loss": 4.0762
|
| 102 |
+
},
|
| 103 |
+
{
|
| 104 |
+
"epoch": 0.3155818540433925,
|
| 105 |
+
"eval_accuracy": 0.015521064301552107,
|
| 106 |
+
"eval_loss": 4.189310073852539,
|
| 107 |
+
"eval_runtime": 14.9234,
|
| 108 |
+
"eval_samples_per_second": 60.442,
|
| 109 |
+
"eval_steps_per_second": 7.572,
|
| 110 |
+
"step": 40
|
| 111 |
+
},
|
| 112 |
+
{
|
| 113 |
+
"epoch": 0.35502958579881655,
|
| 114 |
+
"grad_norm": 101.89262765170473,
|
| 115 |
+
"learning_rate": 2.1568627450980393e-06,
|
| 116 |
+
"loss": 33.7348,
|
| 117 |
+
"step": 45,
|
| 118 |
+
"true_loss": 4.5547
|
| 119 |
+
},
|
| 120 |
+
{
|
| 121 |
+
"epoch": 0.39447731755424065,
|
| 122 |
+
"grad_norm": 118.10971042116216,
|
| 123 |
+
"learning_rate": 2.401960784313726e-06,
|
| 124 |
+
"loss": 33.5551,
|
| 125 |
+
"step": 50,
|
| 126 |
+
"true_loss": 4.1461
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"epoch": 0.39447731755424065,
|
| 130 |
+
"eval_accuracy": 0.024390243902439025,
|
| 131 |
+
"eval_loss": 4.126463890075684,
|
| 132 |
+
"eval_runtime": 15.0954,
|
| 133 |
+
"eval_samples_per_second": 59.753,
|
| 134 |
+
"eval_steps_per_second": 7.486,
|
| 135 |
+
"step": 50
|
| 136 |
+
},
|
| 137 |
+
{
|
| 138 |
+
"epoch": 0.4339250493096647,
|
| 139 |
+
"grad_norm": 126.8109210779904,
|
| 140 |
+
"learning_rate": 2.647058823529412e-06,
|
| 141 |
+
"loss": 33.3662,
|
| 142 |
+
"step": 55,
|
| 143 |
+
"true_loss": 4.1771
|
| 144 |
+
},
|
| 145 |
+
{
|
| 146 |
+
"epoch": 0.47337278106508873,
|
| 147 |
+
"grad_norm": 107.75818756256407,
|
| 148 |
+
"learning_rate": 2.8921568627450985e-06,
|
| 149 |
+
"loss": 32.4912,
|
| 150 |
+
"step": 60,
|
| 151 |
+
"true_loss": 3.9953
|
| 152 |
+
},
|
| 153 |
+
{
|
| 154 |
+
"epoch": 0.47337278106508873,
|
| 155 |
+
"eval_accuracy": 0.036585365853658534,
|
| 156 |
+
"eval_loss": 4.055848121643066,
|
| 157 |
+
"eval_runtime": 14.9842,
|
| 158 |
+
"eval_samples_per_second": 60.197,
|
| 159 |
+
"eval_steps_per_second": 7.541,
|
| 160 |
+
"step": 60
|
| 161 |
+
},
|
| 162 |
+
{
|
| 163 |
+
"epoch": 0.5128205128205128,
|
| 164 |
+
"grad_norm": 126.96202525436865,
|
| 165 |
+
"learning_rate": 3.1372549019607846e-06,
|
| 166 |
+
"loss": 32.3756,
|
| 167 |
+
"step": 65,
|
| 168 |
+
"true_loss": 4.057
|
| 169 |
+
},
|
| 170 |
+
{
|
| 171 |
+
"epoch": 0.5522682445759369,
|
| 172 |
+
"grad_norm": 121.42814891257407,
|
| 173 |
+
"learning_rate": 3.382352941176471e-06,
|
| 174 |
+
"loss": 32.7852,
|
| 175 |
+
"step": 70,
|
| 176 |
+
"true_loss": 3.8746
|
| 177 |
+
},
|
| 178 |
+
{
|
| 179 |
+
"epoch": 0.5522682445759369,
|
| 180 |
+
"eval_accuracy": 0.04767184035476718,
|
| 181 |
+
"eval_loss": 4.019028663635254,
|
| 182 |
+
"eval_runtime": 15.0235,
|
| 183 |
+
"eval_samples_per_second": 60.039,
|
| 184 |
+
"eval_steps_per_second": 7.522,
|
| 185 |
+
"step": 70
|
| 186 |
+
},
|
| 187 |
+
{
|
| 188 |
+
"epoch": 0.591715976331361,
|
| 189 |
+
"grad_norm": 152.86102648476006,
|
| 190 |
+
"learning_rate": 3.6274509803921573e-06,
|
| 191 |
+
"loss": 31.9488,
|
| 192 |
+
"step": 75,
|
| 193 |
+
"true_loss": 3.9531
|
| 194 |
+
},
|
| 195 |
+
{
|
| 196 |
+
"epoch": 0.631163708086785,
|
| 197 |
+
"grad_norm": 141.9292910226069,
|
| 198 |
+
"learning_rate": 3.872549019607843e-06,
|
| 199 |
+
"loss": 32.1604,
|
| 200 |
+
"step": 80,
|
| 201 |
+
"true_loss": 4.1195
|
| 202 |
+
},
|
| 203 |
+
{
|
| 204 |
+
"epoch": 0.631163708086785,
|
| 205 |
+
"eval_accuracy": 0.05432372505543237,
|
| 206 |
+
"eval_loss": 3.9854836463928223,
|
| 207 |
+
"eval_runtime": 14.9619,
|
| 208 |
+
"eval_samples_per_second": 60.286,
|
| 209 |
+
"eval_steps_per_second": 7.552,
|
| 210 |
+
"step": 80
|
| 211 |
+
},
|
| 212 |
+
{
|
| 213 |
+
"epoch": 0.6706114398422091,
|
| 214 |
+
"grad_norm": 132.12828795355875,
|
| 215 |
+
"learning_rate": 4.11764705882353e-06,
|
| 216 |
+
"loss": 32.0111,
|
| 217 |
+
"step": 85,
|
| 218 |
+
"true_loss": 4.0219
|
| 219 |
+
},
|
| 220 |
+
{
|
| 221 |
+
"epoch": 0.7100591715976331,
|
| 222 |
+
"grad_norm": 153.0719114159677,
|
| 223 |
+
"learning_rate": 4.3627450980392164e-06,
|
| 224 |
+
"loss": 31.752,
|
| 225 |
+
"step": 90,
|
| 226 |
+
"true_loss": 3.9547
|
| 227 |
+
},
|
| 228 |
+
{
|
| 229 |
+
"epoch": 0.7100591715976331,
|
| 230 |
+
"eval_accuracy": 0.058758314855875834,
|
| 231 |
+
"eval_loss": 3.9391109943389893,
|
| 232 |
+
"eval_runtime": 14.9724,
|
| 233 |
+
"eval_samples_per_second": 60.244,
|
| 234 |
+
"eval_steps_per_second": 7.547,
|
| 235 |
+
"step": 90
|
| 236 |
+
},
|
| 237 |
+
{
|
| 238 |
+
"epoch": 0.7495069033530573,
|
| 239 |
+
"grad_norm": 128.25085846897863,
|
| 240 |
+
"learning_rate": 4.607843137254902e-06,
|
| 241 |
+
"loss": 32.32,
|
| 242 |
+
"step": 95,
|
| 243 |
+
"true_loss": 4.1264
|
| 244 |
+
},
|
| 245 |
+
{
|
| 246 |
+
"epoch": 0.7889546351084813,
|
| 247 |
+
"grad_norm": 132.66281463976432,
|
| 248 |
+
"learning_rate": 4.852941176470589e-06,
|
| 249 |
+
"loss": 31.9783,
|
| 250 |
+
"step": 100,
|
| 251 |
+
"true_loss": 4.1824
|
| 252 |
+
},
|
| 253 |
+
{
|
| 254 |
+
"epoch": 0.7889546351084813,
|
| 255 |
+
"eval_accuracy": 0.082039911308204,
|
| 256 |
+
"eval_loss": 3.891880989074707,
|
| 257 |
+
"eval_runtime": 15.0549,
|
| 258 |
+
"eval_samples_per_second": 59.914,
|
| 259 |
+
"eval_steps_per_second": 7.506,
|
| 260 |
+
"step": 100
|
| 261 |
+
},
|
| 262 |
+
{
|
| 263 |
+
"epoch": 0.8284023668639053,
|
| 264 |
+
"grad_norm": 138.12009424260248,
|
| 265 |
+
"learning_rate": 4.989059080962801e-06,
|
| 266 |
+
"loss": 31.175,
|
| 267 |
+
"step": 105,
|
| 268 |
+
"true_loss": 3.9373
|
| 269 |
+
},
|
| 270 |
+
{
|
| 271 |
+
"epoch": 0.8678500986193294,
|
| 272 |
+
"grad_norm": 139.39677324279805,
|
| 273 |
+
"learning_rate": 4.961706783369804e-06,
|
| 274 |
+
"loss": 31.9854,
|
| 275 |
+
"step": 110,
|
| 276 |
+
"true_loss": 4.1117
|
| 277 |
+
},
|
| 278 |
+
{
|
| 279 |
+
"epoch": 0.8678500986193294,
|
| 280 |
+
"eval_accuracy": 0.09866962305986696,
|
| 281 |
+
"eval_loss": 3.860440254211426,
|
| 282 |
+
"eval_runtime": 14.9426,
|
| 283 |
+
"eval_samples_per_second": 60.364,
|
| 284 |
+
"eval_steps_per_second": 7.562,
|
| 285 |
+
"step": 110
|
| 286 |
+
},
|
| 287 |
+
{
|
| 288 |
+
"epoch": 0.9072978303747534,
|
| 289 |
+
"grad_norm": 160.76229566662045,
|
| 290 |
+
"learning_rate": 4.934354485776805e-06,
|
| 291 |
+
"loss": 31.7863,
|
| 292 |
+
"step": 115,
|
| 293 |
+
"true_loss": 4.0758
|
| 294 |
+
},
|
| 295 |
+
{
|
| 296 |
+
"epoch": 0.9467455621301775,
|
| 297 |
+
"grad_norm": 141.93625692921623,
|
| 298 |
+
"learning_rate": 4.9070021881838075e-06,
|
| 299 |
+
"loss": 31.4898,
|
| 300 |
+
"step": 120,
|
| 301 |
+
"true_loss": 3.8525
|
| 302 |
+
},
|
| 303 |
+
{
|
| 304 |
+
"epoch": 0.9467455621301775,
|
| 305 |
+
"eval_accuracy": 0.12305986696230599,
|
| 306 |
+
"eval_loss": 3.844113826751709,
|
| 307 |
+
"eval_runtime": 14.91,
|
| 308 |
+
"eval_samples_per_second": 60.496,
|
| 309 |
+
"eval_steps_per_second": 7.579,
|
| 310 |
+
"step": 120
|
| 311 |
+
},
|
| 312 |
+
{
|
| 313 |
+
"epoch": 0.9861932938856016,
|
| 314 |
+
"grad_norm": 156.67839380394838,
|
| 315 |
+
"learning_rate": 4.87964989059081e-06,
|
| 316 |
+
"loss": 31.3083,
|
| 317 |
+
"step": 125,
|
| 318 |
+
"true_loss": 3.6941
|
| 319 |
+
},
|
| 320 |
+
{
|
| 321 |
+
"epoch": 1.0236686390532543,
|
| 322 |
+
"grad_norm": 175.06600661459612,
|
| 323 |
+
"learning_rate": 4.852297592997812e-06,
|
| 324 |
+
"loss": 29.4422,
|
| 325 |
+
"step": 130,
|
| 326 |
+
"true_loss": 3.6871
|
| 327 |
+
},
|
| 328 |
+
{
|
| 329 |
+
"epoch": 1.0236686390532543,
|
| 330 |
+
"eval_accuracy": 0.11419068736141907,
|
| 331 |
+
"eval_loss": 3.816354274749756,
|
| 332 |
+
"eval_runtime": 14.9497,
|
| 333 |
+
"eval_samples_per_second": 60.336,
|
| 334 |
+
"eval_steps_per_second": 7.559,
|
| 335 |
+
"step": 130
|
| 336 |
+
},
|
| 337 |
+
{
|
| 338 |
+
"epoch": 1.0631163708086786,
|
| 339 |
+
"grad_norm": 176.4037265950165,
|
| 340 |
+
"learning_rate": 4.824945295404815e-06,
|
| 341 |
+
"loss": 29.6881,
|
| 342 |
+
"step": 135,
|
| 343 |
+
"true_loss": 3.5447
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"epoch": 1.1025641025641026,
|
| 347 |
+
"grad_norm": 175.43736192079575,
|
| 348 |
+
"learning_rate": 4.797592997811816e-06,
|
| 349 |
+
"loss": 30.1351,
|
| 350 |
+
"step": 140,
|
| 351 |
+
"true_loss": 3.9027
|
| 352 |
+
},
|
| 353 |
+
{
|
| 354 |
+
"epoch": 1.1025641025641026,
|
| 355 |
+
"eval_accuracy": 0.12084257206208426,
|
| 356 |
+
"eval_loss": 3.787785053253174,
|
| 357 |
+
"eval_runtime": 14.9544,
|
| 358 |
+
"eval_samples_per_second": 60.317,
|
| 359 |
+
"eval_steps_per_second": 7.556,
|
| 360 |
+
"step": 140
|
| 361 |
+
},
|
| 362 |
+
{
|
| 363 |
+
"epoch": 1.1420118343195267,
|
| 364 |
+
"grad_norm": 172.66609321830506,
|
| 365 |
+
"learning_rate": 4.770240700218819e-06,
|
| 366 |
+
"loss": 31.1361,
|
| 367 |
+
"step": 145,
|
| 368 |
+
"true_loss": 3.9604
|
| 369 |
+
},
|
| 370 |
+
{
|
| 371 |
+
"epoch": 1.1814595660749507,
|
| 372 |
+
"grad_norm": 188.08419409648823,
|
| 373 |
+
"learning_rate": 4.742888402625821e-06,
|
| 374 |
+
"loss": 29.9185,
|
| 375 |
+
"step": 150,
|
| 376 |
+
"true_loss": 3.6713
|
| 377 |
+
},
|
| 378 |
+
{
|
| 379 |
+
"epoch": 1.1814595660749507,
|
| 380 |
+
"eval_accuracy": 0.12195121951219512,
|
| 381 |
+
"eval_loss": 3.7505717277526855,
|
| 382 |
+
"eval_runtime": 15.0675,
|
| 383 |
+
"eval_samples_per_second": 59.864,
|
| 384 |
+
"eval_steps_per_second": 7.5,
|
| 385 |
+
"step": 150
|
| 386 |
+
},
|
| 387 |
+
{
|
| 388 |
+
"epoch": 1.2209072978303748,
|
| 389 |
+
"grad_norm": 184.54494972485557,
|
| 390 |
+
"learning_rate": 4.715536105032823e-06,
|
| 391 |
+
"loss": 29.7549,
|
| 392 |
+
"step": 155,
|
| 393 |
+
"true_loss": 3.7596
|
| 394 |
+
},
|
| 395 |
+
{
|
| 396 |
+
"epoch": 1.2603550295857988,
|
| 397 |
+
"grad_norm": 230.49181032198453,
|
| 398 |
+
"learning_rate": 4.688183807439826e-06,
|
| 399 |
+
"loss": 29.7484,
|
| 400 |
+
"step": 160,
|
| 401 |
+
"true_loss": 3.9594
|
| 402 |
+
},
|
| 403 |
+
{
|
| 404 |
+
"epoch": 1.2603550295857988,
|
| 405 |
+
"eval_accuracy": 0.13747228381374724,
|
| 406 |
+
"eval_loss": 3.7303473949432373,
|
| 407 |
+
"eval_runtime": 15.2085,
|
| 408 |
+
"eval_samples_per_second": 59.309,
|
| 409 |
+
"eval_steps_per_second": 7.43,
|
| 410 |
+
"step": 160
|
| 411 |
+
},
|
| 412 |
+
{
|
| 413 |
+
"epoch": 1.2998027613412229,
|
| 414 |
+
"grad_norm": 186.21055820946583,
|
| 415 |
+
"learning_rate": 4.660831509846827e-06,
|
| 416 |
+
"loss": 29.388,
|
| 417 |
+
"step": 165,
|
| 418 |
+
"true_loss": 3.6656
|
| 419 |
+
},
|
| 420 |
+
{
|
| 421 |
+
"epoch": 1.339250493096647,
|
| 422 |
+
"grad_norm": 203.5666063639384,
|
| 423 |
+
"learning_rate": 4.63347921225383e-06,
|
| 424 |
+
"loss": 29.4438,
|
| 425 |
+
"step": 170,
|
| 426 |
+
"true_loss": 4.0383
|
| 427 |
+
},
|
| 428 |
+
{
|
| 429 |
+
"epoch": 1.339250493096647,
|
| 430 |
+
"eval_accuracy": 0.13747228381374724,
|
| 431 |
+
"eval_loss": 3.683065414428711,
|
| 432 |
+
"eval_runtime": 15.1745,
|
| 433 |
+
"eval_samples_per_second": 59.442,
|
| 434 |
+
"eval_steps_per_second": 7.447,
|
| 435 |
+
"step": 170
|
| 436 |
+
},
|
| 437 |
+
{
|
| 438 |
+
"epoch": 1.378698224852071,
|
| 439 |
+
"grad_norm": 215.86726316889974,
|
| 440 |
+
"learning_rate": 4.606126914660832e-06,
|
| 441 |
+
"loss": 30.1271,
|
| 442 |
+
"step": 175,
|
| 443 |
+
"true_loss": 3.6135
|
| 444 |
+
},
|
| 445 |
+
{
|
| 446 |
+
"epoch": 1.418145956607495,
|
| 447 |
+
"grad_norm": 191.64160050781922,
|
| 448 |
+
"learning_rate": 4.5787746170678345e-06,
|
| 449 |
+
"loss": 29.468,
|
| 450 |
+
"step": 180,
|
| 451 |
+
"true_loss": 3.7248
|
| 452 |
+
},
|
| 453 |
+
{
|
| 454 |
+
"epoch": 1.418145956607495,
|
| 455 |
+
"eval_accuracy": 0.14855875831485588,
|
| 456 |
+
"eval_loss": 3.663315534591675,
|
| 457 |
+
"eval_runtime": 15.0394,
|
| 458 |
+
"eval_samples_per_second": 59.976,
|
| 459 |
+
"eval_steps_per_second": 7.514,
|
| 460 |
+
"step": 180
|
| 461 |
+
},
|
| 462 |
+
{
|
| 463 |
+
"epoch": 1.4575936883629192,
|
| 464 |
+
"grad_norm": 204.6168296859058,
|
| 465 |
+
"learning_rate": 4.551422319474836e-06,
|
| 466 |
+
"loss": 29.6546,
|
| 467 |
+
"step": 185,
|
| 468 |
+
"true_loss": 3.7043
|
| 469 |
+
},
|
| 470 |
+
{
|
| 471 |
+
"epoch": 1.497041420118343,
|
| 472 |
+
"grad_norm": 219.7996098581931,
|
| 473 |
+
"learning_rate": 4.524070021881838e-06,
|
| 474 |
+
"loss": 29.0906,
|
| 475 |
+
"step": 190,
|
| 476 |
+
"true_loss": 3.6714
|
| 477 |
+
},
|
| 478 |
+
{
|
| 479 |
+
"epoch": 1.497041420118343,
|
| 480 |
+
"eval_accuracy": 0.1385809312638581,
|
| 481 |
+
"eval_loss": 3.6522138118743896,
|
| 482 |
+
"eval_runtime": 14.9508,
|
| 483 |
+
"eval_samples_per_second": 60.331,
|
| 484 |
+
"eval_steps_per_second": 7.558,
|
| 485 |
+
"step": 190
|
| 486 |
+
},
|
| 487 |
+
{
|
| 488 |
+
"epoch": 1.5364891518737673,
|
| 489 |
+
"grad_norm": 177.10887786115575,
|
| 490 |
+
"learning_rate": 4.496717724288841e-06,
|
| 491 |
+
"loss": 29.4486,
|
| 492 |
+
"step": 195,
|
| 493 |
+
"true_loss": 3.9092
|
| 494 |
+
},
|
| 495 |
+
{
|
| 496 |
+
"epoch": 1.5759368836291914,
|
| 497 |
+
"grad_norm": 180.67006503279012,
|
| 498 |
+
"learning_rate": 4.469365426695843e-06,
|
| 499 |
+
"loss": 30.4641,
|
| 500 |
+
"step": 200,
|
| 501 |
+
"true_loss": 3.9328
|
| 502 |
+
},
|
| 503 |
+
{
|
| 504 |
+
"epoch": 1.5759368836291914,
|
| 505 |
+
"eval_accuracy": 0.14079822616407983,
|
| 506 |
+
"eval_loss": 3.65649676322937,
|
| 507 |
+
"eval_runtime": 15.1758,
|
| 508 |
+
"eval_samples_per_second": 59.437,
|
| 509 |
+
"eval_steps_per_second": 7.446,
|
| 510 |
+
"step": 200
|
| 511 |
+
},
|
| 512 |
+
{
|
| 513 |
+
"epoch": 1.6153846153846154,
|
| 514 |
+
"grad_norm": 182.15282062853038,
|
| 515 |
+
"learning_rate": 4.442013129102845e-06,
|
| 516 |
+
"loss": 28.703,
|
| 517 |
+
"step": 205,
|
| 518 |
+
"true_loss": 3.5898
|
| 519 |
+
},
|
| 520 |
+
{
|
| 521 |
+
"epoch": 1.6548323471400395,
|
| 522 |
+
"grad_norm": 187.54953739010512,
|
| 523 |
+
"learning_rate": 4.414660831509847e-06,
|
| 524 |
+
"loss": 29.291,
|
| 525 |
+
"step": 210,
|
| 526 |
+
"true_loss": 3.7754
|
| 527 |
+
},
|
| 528 |
+
{
|
| 529 |
+
"epoch": 1.6548323471400395,
|
| 530 |
+
"eval_accuracy": 0.13414634146341464,
|
| 531 |
+
"eval_loss": 3.650477170944214,
|
| 532 |
+
"eval_runtime": 15.131,
|
| 533 |
+
"eval_samples_per_second": 59.613,
|
| 534 |
+
"eval_steps_per_second": 7.468,
|
| 535 |
+
"step": 210
|
| 536 |
+
},
|
| 537 |
+
{
|
| 538 |
+
"epoch": 1.6942800788954635,
|
| 539 |
+
"grad_norm": 197.72972204694662,
|
| 540 |
+
"learning_rate": 4.3873085339168495e-06,
|
| 541 |
+
"loss": 29.5033,
|
| 542 |
+
"step": 215,
|
| 543 |
+
"true_loss": 3.5539
|
| 544 |
+
},
|
| 545 |
+
{
|
| 546 |
+
"epoch": 1.7337278106508875,
|
| 547 |
+
"grad_norm": 211.1691853811082,
|
| 548 |
+
"learning_rate": 4.359956236323852e-06,
|
| 549 |
+
"loss": 28.2073,
|
| 550 |
+
"step": 220,
|
| 551 |
+
"true_loss": 3.7295
|
| 552 |
+
},
|
| 553 |
+
{
|
| 554 |
+
"epoch": 1.7337278106508875,
|
| 555 |
+
"eval_accuracy": 0.1607538802660754,
|
| 556 |
+
"eval_loss": 3.5888044834136963,
|
| 557 |
+
"eval_runtime": 14.8061,
|
| 558 |
+
"eval_samples_per_second": 60.921,
|
| 559 |
+
"eval_steps_per_second": 7.632,
|
| 560 |
+
"step": 220
|
| 561 |
+
},
|
| 562 |
+
{
|
| 563 |
+
"epoch": 1.7731755424063116,
|
| 564 |
+
"grad_norm": 198.11359915598604,
|
| 565 |
+
"learning_rate": 4.332603938730854e-06,
|
| 566 |
+
"loss": 29.068,
|
| 567 |
+
"step": 225,
|
| 568 |
+
"true_loss": 3.5761
|
| 569 |
+
},
|
| 570 |
+
{
|
| 571 |
+
"epoch": 1.8126232741617359,
|
| 572 |
+
"grad_norm": 208.2967379446751,
|
| 573 |
+
"learning_rate": 4.305251641137856e-06,
|
| 574 |
+
"loss": 29.558,
|
| 575 |
+
"step": 230,
|
| 576 |
+
"true_loss": 3.7469
|
| 577 |
+
},
|
| 578 |
+
{
|
| 579 |
+
"epoch": 1.8126232741617359,
|
| 580 |
+
"eval_accuracy": 0.1662971175166297,
|
| 581 |
+
"eval_loss": 3.567077398300171,
|
| 582 |
+
"eval_runtime": 14.8915,
|
| 583 |
+
"eval_samples_per_second": 60.571,
|
| 584 |
+
"eval_steps_per_second": 7.588,
|
| 585 |
+
"step": 230
|
| 586 |
+
},
|
| 587 |
+
{
|
| 588 |
+
"epoch": 1.8520710059171597,
|
| 589 |
+
"grad_norm": 202.51288282619706,
|
| 590 |
+
"learning_rate": 4.277899343544858e-06,
|
| 591 |
+
"loss": 29.2451,
|
| 592 |
+
"step": 235,
|
| 593 |
+
"true_loss": 3.5877
|
| 594 |
+
},
|
| 595 |
+
{
|
| 596 |
+
"epoch": 1.891518737672584,
|
| 597 |
+
"grad_norm": 204.1287289867651,
|
| 598 |
+
"learning_rate": 4.2505470459518606e-06,
|
| 599 |
+
"loss": 28.581,
|
| 600 |
+
"step": 240,
|
| 601 |
+
"true_loss": 3.4135
|
| 602 |
+
},
|
| 603 |
+
{
|
| 604 |
+
"epoch": 1.891518737672584,
|
| 605 |
+
"eval_accuracy": 0.164079822616408,
|
| 606 |
+
"eval_loss": 3.5610058307647705,
|
| 607 |
+
"eval_runtime": 15.2419,
|
| 608 |
+
"eval_samples_per_second": 59.179,
|
| 609 |
+
"eval_steps_per_second": 7.414,
|
| 610 |
+
"step": 240
|
| 611 |
+
},
|
| 612 |
+
{
|
| 613 |
+
"epoch": 1.9309664694280078,
|
| 614 |
+
"grad_norm": 199.99425110918077,
|
| 615 |
+
"learning_rate": 4.223194748358863e-06,
|
| 616 |
+
"loss": 28.7527,
|
| 617 |
+
"step": 245,
|
| 618 |
+
"true_loss": 3.6467
|
| 619 |
+
},
|
| 620 |
+
{
|
| 621 |
+
"epoch": 1.970414201183432,
|
| 622 |
+
"grad_norm": 209.62816557130563,
|
| 623 |
+
"learning_rate": 4.1958424507658645e-06,
|
| 624 |
+
"loss": 29.8734,
|
| 625 |
+
"step": 250,
|
| 626 |
+
"true_loss": 3.3148
|
| 627 |
+
},
|
| 628 |
+
{
|
| 629 |
+
"epoch": 1.970414201183432,
|
| 630 |
+
"eval_accuracy": 0.1662971175166297,
|
| 631 |
+
"eval_loss": 3.583265542984009,
|
| 632 |
+
"eval_runtime": 14.8799,
|
| 633 |
+
"eval_samples_per_second": 60.619,
|
| 634 |
+
"eval_steps_per_second": 7.594,
|
| 635 |
+
"step": 250
|
| 636 |
+
},
|
| 637 |
+
{
|
| 638 |
+
"epoch": 2.007889546351085,
|
| 639 |
+
"grad_norm": 197.80455811801355,
|
| 640 |
+
"learning_rate": 4.168490153172867e-06,
|
| 641 |
+
"loss": 27.7556,
|
| 642 |
+
"step": 255,
|
| 643 |
+
"true_loss": 4.0312
|
| 644 |
+
},
|
| 645 |
+
{
|
| 646 |
+
"epoch": 2.0473372781065087,
|
| 647 |
+
"grad_norm": 198.8468471778106,
|
| 648 |
+
"learning_rate": 4.141137855579869e-06,
|
| 649 |
+
"loss": 27.6775,
|
| 650 |
+
"step": 260,
|
| 651 |
+
"true_loss": 3.8672
|
| 652 |
+
},
|
| 653 |
+
{
|
| 654 |
+
"epoch": 2.0473372781065087,
|
| 655 |
+
"eval_accuracy": 0.1574279379157428,
|
| 656 |
+
"eval_loss": 3.5533580780029297,
|
| 657 |
+
"eval_runtime": 14.9826,
|
| 658 |
+
"eval_samples_per_second": 60.203,
|
| 659 |
+
"eval_steps_per_second": 7.542,
|
| 660 |
+
"step": 260
|
| 661 |
+
},
|
| 662 |
+
{
|
| 663 |
+
"epoch": 2.086785009861933,
|
| 664 |
+
"grad_norm": 204.45388818689486,
|
| 665 |
+
"learning_rate": 4.113785557986872e-06,
|
| 666 |
+
"loss": 26.7176,
|
| 667 |
+
"step": 265,
|
| 668 |
+
"true_loss": 3.4912
|
| 669 |
+
},
|
| 670 |
+
{
|
| 671 |
+
"epoch": 2.126232741617357,
|
| 672 |
+
"grad_norm": 225.18842668023993,
|
| 673 |
+
"learning_rate": 4.086433260393873e-06,
|
| 674 |
+
"loss": 27.9069,
|
| 675 |
+
"step": 270,
|
| 676 |
+
"true_loss": 3.5805
|
| 677 |
+
},
|
| 678 |
+
{
|
| 679 |
+
"epoch": 2.126232741617357,
|
| 680 |
+
"eval_accuracy": 0.16851441241685144,
|
| 681 |
+
"eval_loss": 3.534597635269165,
|
| 682 |
+
"eval_runtime": 14.8101,
|
| 683 |
+
"eval_samples_per_second": 60.905,
|
| 684 |
+
"eval_steps_per_second": 7.63,
|
| 685 |
+
"step": 270
|
| 686 |
+
},
|
| 687 |
+
{
|
| 688 |
+
"epoch": 2.165680473372781,
|
| 689 |
+
"grad_norm": 232.60579637840817,
|
| 690 |
+
"learning_rate": 4.059080962800876e-06,
|
| 691 |
+
"loss": 26.9553,
|
| 692 |
+
"step": 275,
|
| 693 |
+
"true_loss": 3.1944
|
| 694 |
+
},
|
| 695 |
+
{
|
| 696 |
+
"epoch": 2.2051282051282053,
|
| 697 |
+
"grad_norm": 214.5771122472325,
|
| 698 |
+
"learning_rate": 4.031728665207878e-06,
|
| 699 |
+
"loss": 26.0979,
|
| 700 |
+
"step": 280,
|
| 701 |
+
"true_loss": 3.1361
|
| 702 |
+
},
|
| 703 |
+
{
|
| 704 |
+
"epoch": 2.2051282051282053,
|
| 705 |
+
"eval_accuracy": 0.15299334811529933,
|
| 706 |
+
"eval_loss": 3.5224153995513916,
|
| 707 |
+
"eval_runtime": 15.107,
|
| 708 |
+
"eval_samples_per_second": 59.707,
|
| 709 |
+
"eval_steps_per_second": 7.48,
|
| 710 |
+
"step": 280
|
| 711 |
+
},
|
| 712 |
+
{
|
| 713 |
+
"epoch": 2.244575936883629,
|
| 714 |
+
"grad_norm": 229.00786163080798,
|
| 715 |
+
"learning_rate": 4.00437636761488e-06,
|
| 716 |
+
"loss": 27.3225,
|
| 717 |
+
"step": 285,
|
| 718 |
+
"true_loss": 3.3937
|
| 719 |
+
},
|
| 720 |
+
{
|
| 721 |
+
"epoch": 2.2840236686390534,
|
| 722 |
+
"grad_norm": 220.9809902657408,
|
| 723 |
+
"learning_rate": 3.977024070021882e-06,
|
| 724 |
+
"loss": 27.3824,
|
| 725 |
+
"step": 290,
|
| 726 |
+
"true_loss": 3.3252
|
| 727 |
+
},
|
| 728 |
+
{
|
| 729 |
+
"epoch": 2.2840236686390534,
|
| 730 |
+
"eval_accuracy": 0.16186252771618626,
|
| 731 |
+
"eval_loss": 3.5029642581939697,
|
| 732 |
+
"eval_runtime": 14.9319,
|
| 733 |
+
"eval_samples_per_second": 60.408,
|
| 734 |
+
"eval_steps_per_second": 7.568,
|
| 735 |
+
"step": 290
|
| 736 |
+
},
|
| 737 |
+
{
|
| 738 |
+
"epoch": 2.323471400394477,
|
| 739 |
+
"grad_norm": 224.23746222333077,
|
| 740 |
+
"learning_rate": 3.949671772428884e-06,
|
| 741 |
+
"loss": 27.8245,
|
| 742 |
+
"step": 295,
|
| 743 |
+
"true_loss": 3.3297
|
| 744 |
+
},
|
| 745 |
+
{
|
| 746 |
+
"epoch": 2.3629191321499015,
|
| 747 |
+
"grad_norm": 196.24143904821017,
|
| 748 |
+
"learning_rate": 3.922319474835887e-06,
|
| 749 |
+
"loss": 27.9647,
|
| 750 |
+
"step": 300,
|
| 751 |
+
"true_loss": 3.6294
|
| 752 |
+
},
|
| 753 |
+
{
|
| 754 |
+
"epoch": 2.3629191321499015,
|
| 755 |
+
"eval_accuracy": 0.18181818181818182,
|
| 756 |
+
"eval_loss": 3.5059654712677,
|
| 757 |
+
"eval_runtime": 14.9636,
|
| 758 |
+
"eval_samples_per_second": 60.279,
|
| 759 |
+
"eval_steps_per_second": 7.552,
|
| 760 |
+
"step": 300
|
| 761 |
+
},
|
| 762 |
+
{
|
| 763 |
+
"epoch": 2.4023668639053253,
|
| 764 |
+
"grad_norm": 210.16619384673945,
|
| 765 |
+
"learning_rate": 3.894967177242889e-06,
|
| 766 |
+
"loss": 27.5051,
|
| 767 |
+
"step": 305,
|
| 768 |
+
"true_loss": 3.3774
|
| 769 |
+
},
|
| 770 |
+
{
|
| 771 |
+
"epoch": 2.4418145956607495,
|
| 772 |
+
"grad_norm": 213.67456162572208,
|
| 773 |
+
"learning_rate": 3.867614879649891e-06,
|
| 774 |
+
"loss": 27.3462,
|
| 775 |
+
"step": 310,
|
| 776 |
+
"true_loss": 3.3658
|
| 777 |
+
},
|
| 778 |
+
{
|
| 779 |
+
"epoch": 2.4418145956607495,
|
| 780 |
+
"eval_accuracy": 0.17849223946784923,
|
| 781 |
+
"eval_loss": 3.5055649280548096,
|
| 782 |
+
"eval_runtime": 15.1871,
|
| 783 |
+
"eval_samples_per_second": 59.393,
|
| 784 |
+
"eval_steps_per_second": 7.441,
|
| 785 |
+
"step": 310
|
| 786 |
+
},
|
| 787 |
+
{
|
| 788 |
+
"epoch": 2.4812623274161734,
|
| 789 |
+
"grad_norm": 225.9401296492373,
|
| 790 |
+
"learning_rate": 3.840262582056893e-06,
|
| 791 |
+
"loss": 27.1061,
|
| 792 |
+
"step": 315,
|
| 793 |
+
"true_loss": 3.478
|
| 794 |
+
},
|
| 795 |
+
{
|
| 796 |
+
"epoch": 2.5207100591715976,
|
| 797 |
+
"grad_norm": 205.8362032729107,
|
| 798 |
+
"learning_rate": 3.8129102844638954e-06,
|
| 799 |
+
"loss": 26.3703,
|
| 800 |
+
"step": 320,
|
| 801 |
+
"true_loss": 3.1416
|
| 802 |
+
},
|
| 803 |
+
{
|
| 804 |
+
"epoch": 2.5207100591715976,
|
| 805 |
+
"eval_accuracy": 0.17073170731707318,
|
| 806 |
+
"eval_loss": 3.4906609058380127,
|
| 807 |
+
"eval_runtime": 15.0707,
|
| 808 |
+
"eval_samples_per_second": 59.851,
|
| 809 |
+
"eval_steps_per_second": 7.498,
|
| 810 |
+
"step": 320
|
| 811 |
+
},
|
| 812 |
+
{
|
| 813 |
+
"epoch": 2.560157790927022,
|
| 814 |
+
"grad_norm": 203.21971371657304,
|
| 815 |
+
"learning_rate": 3.7855579868708973e-06,
|
| 816 |
+
"loss": 27.3517,
|
| 817 |
+
"step": 325,
|
| 818 |
+
"true_loss": 3.2373
|
| 819 |
+
},
|
| 820 |
+
{
|
| 821 |
+
"epoch": 2.5996055226824457,
|
| 822 |
+
"grad_norm": 219.93911587157675,
|
| 823 |
+
"learning_rate": 3.7582056892778997e-06,
|
| 824 |
+
"loss": 27.8681,
|
| 825 |
+
"step": 330,
|
| 826 |
+
"true_loss": 3.125
|
| 827 |
+
},
|
| 828 |
+
{
|
| 829 |
+
"epoch": 2.5996055226824457,
|
| 830 |
+
"eval_accuracy": 0.17184035476718404,
|
| 831 |
+
"eval_loss": 3.4929237365722656,
|
| 832 |
+
"eval_runtime": 15.1428,
|
| 833 |
+
"eval_samples_per_second": 59.566,
|
| 834 |
+
"eval_steps_per_second": 7.462,
|
| 835 |
+
"step": 330
|
| 836 |
+
},
|
| 837 |
+
{
|
| 838 |
+
"epoch": 2.63905325443787,
|
| 839 |
+
"grad_norm": 232.49538276234148,
|
| 840 |
+
"learning_rate": 3.7308533916849017e-06,
|
| 841 |
+
"loss": 27.0143,
|
| 842 |
+
"step": 335,
|
| 843 |
+
"true_loss": 3.5833
|
| 844 |
+
},
|
| 845 |
+
{
|
| 846 |
+
"epoch": 2.678500986193294,
|
| 847 |
+
"grad_norm": 218.79623472715534,
|
| 848 |
+
"learning_rate": 3.703501094091904e-06,
|
| 849 |
+
"loss": 27.1542,
|
| 850 |
+
"step": 340,
|
| 851 |
+
"true_loss": 3.425
|
| 852 |
+
},
|
| 853 |
+
{
|
| 854 |
+
"epoch": 2.678500986193294,
|
| 855 |
+
"eval_accuracy": 0.1674057649667406,
|
| 856 |
+
"eval_loss": 3.4812633991241455,
|
| 857 |
+
"eval_runtime": 14.9427,
|
| 858 |
+
"eval_samples_per_second": 60.364,
|
| 859 |
+
"eval_steps_per_second": 7.562,
|
| 860 |
+
"step": 340
|
| 861 |
+
},
|
| 862 |
+
{
|
| 863 |
+
"epoch": 2.717948717948718,
|
| 864 |
+
"grad_norm": 210.9796490946478,
|
| 865 |
+
"learning_rate": 3.676148796498906e-06,
|
| 866 |
+
"loss": 27.0408,
|
| 867 |
+
"step": 345,
|
| 868 |
+
"true_loss": 3.2768
|
| 869 |
+
},
|
| 870 |
+
{
|
| 871 |
+
"epoch": 2.757396449704142,
|
| 872 |
+
"grad_norm": 217.69737467031442,
|
| 873 |
+
"learning_rate": 3.6487964989059084e-06,
|
| 874 |
+
"loss": 27.45,
|
| 875 |
+
"step": 350,
|
| 876 |
+
"true_loss": 3.4411
|
| 877 |
+
},
|
| 878 |
+
{
|
| 879 |
+
"epoch": 2.757396449704142,
|
| 880 |
+
"eval_accuracy": 0.17738359201773837,
|
| 881 |
+
"eval_loss": 3.468329906463623,
|
| 882 |
+
"eval_runtime": 14.9595,
|
| 883 |
+
"eval_samples_per_second": 60.296,
|
| 884 |
+
"eval_steps_per_second": 7.554,
|
| 885 |
+
"step": 350
|
| 886 |
+
},
|
| 887 |
+
{
|
| 888 |
+
"epoch": 2.796844181459566,
|
| 889 |
+
"grad_norm": 223.8506652627376,
|
| 890 |
+
"learning_rate": 3.621444201312911e-06,
|
| 891 |
+
"loss": 26.4817,
|
| 892 |
+
"step": 355,
|
| 893 |
+
"true_loss": 3.4845
|
| 894 |
+
},
|
| 895 |
+
{
|
| 896 |
+
"epoch": 2.83629191321499,
|
| 897 |
+
"grad_norm": 222.52387593334151,
|
| 898 |
+
"learning_rate": 3.5940919037199128e-06,
|
| 899 |
+
"loss": 28.3421,
|
| 900 |
+
"step": 360,
|
| 901 |
+
"true_loss": 3.386
|
| 902 |
+
},
|
| 903 |
+
{
|
| 904 |
+
"epoch": 2.83629191321499,
|
| 905 |
+
"eval_accuracy": 0.19290465631929046,
|
| 906 |
+
"eval_loss": 3.4540300369262695,
|
| 907 |
+
"eval_runtime": 15.1328,
|
| 908 |
+
"eval_samples_per_second": 59.605,
|
| 909 |
+
"eval_steps_per_second": 7.467,
|
| 910 |
+
"step": 360
|
| 911 |
+
},
|
| 912 |
+
{
|
| 913 |
+
"epoch": 2.8757396449704142,
|
| 914 |
+
"grad_norm": 233.9031521180998,
|
| 915 |
+
"learning_rate": 3.5667396061269147e-06,
|
| 916 |
+
"loss": 27.4952,
|
| 917 |
+
"step": 365,
|
| 918 |
+
"true_loss": 3.5988
|
| 919 |
+
},
|
| 920 |
+
{
|
| 921 |
+
"epoch": 2.9151873767258385,
|
| 922 |
+
"grad_norm": 220.40770079671375,
|
| 923 |
+
"learning_rate": 3.539387308533917e-06,
|
| 924 |
+
"loss": 27.7365,
|
| 925 |
+
"step": 370,
|
| 926 |
+
"true_loss": 3.8111
|
| 927 |
+
},
|
| 928 |
+
{
|
| 929 |
+
"epoch": 2.9151873767258385,
|
| 930 |
+
"eval_accuracy": 0.2039911308203991,
|
| 931 |
+
"eval_loss": 3.445399045944214,
|
| 932 |
+
"eval_runtime": 15.0467,
|
| 933 |
+
"eval_samples_per_second": 59.947,
|
| 934 |
+
"eval_steps_per_second": 7.51,
|
| 935 |
+
"step": 370
|
| 936 |
+
},
|
| 937 |
+
{
|
| 938 |
+
"epoch": 2.9546351084812623,
|
| 939 |
+
"grad_norm": 214.3532111476442,
|
| 940 |
+
"learning_rate": 3.5120350109409195e-06,
|
| 941 |
+
"loss": 26.3055,
|
| 942 |
+
"step": 375,
|
| 943 |
+
"true_loss": 3.1235
|
| 944 |
+
},
|
| 945 |
+
{
|
| 946 |
+
"epoch": 2.994082840236686,
|
| 947 |
+
"grad_norm": 217.53508025403696,
|
| 948 |
+
"learning_rate": 3.4846827133479215e-06,
|
| 949 |
+
"loss": 28.0186,
|
| 950 |
+
"step": 380,
|
| 951 |
+
"true_loss": 3.4562
|
| 952 |
+
},
|
| 953 |
+
{
|
| 954 |
+
"epoch": 2.994082840236686,
|
| 955 |
+
"eval_accuracy": 0.1917960088691796,
|
| 956 |
+
"eval_loss": 3.44553542137146,
|
| 957 |
+
"eval_runtime": 14.9341,
|
| 958 |
+
"eval_samples_per_second": 60.399,
|
| 959 |
+
"eval_steps_per_second": 7.567,
|
| 960 |
+
"step": 380
|
| 961 |
+
},
|
| 962 |
+
{
|
| 963 |
+
"epoch": 3.0315581854043394,
|
| 964 |
+
"grad_norm": 213.65664891050722,
|
| 965 |
+
"learning_rate": 3.457330415754924e-06,
|
| 966 |
+
"loss": 24.4655,
|
| 967 |
+
"step": 385,
|
| 968 |
+
"true_loss": 3.3045
|
| 969 |
+
},
|
| 970 |
+
{
|
| 971 |
+
"epoch": 3.0710059171597632,
|
| 972 |
+
"grad_norm": 238.57756114961336,
|
| 973 |
+
"learning_rate": 3.429978118161926e-06,
|
| 974 |
+
"loss": 24.9568,
|
| 975 |
+
"step": 390,
|
| 976 |
+
"true_loss": 3.0463
|
| 977 |
+
},
|
| 978 |
+
{
|
| 979 |
+
"epoch": 3.0710059171597632,
|
| 980 |
+
"eval_accuracy": 0.18957871396895787,
|
| 981 |
+
"eval_loss": 3.4360036849975586,
|
| 982 |
+
"eval_runtime": 15.1499,
|
| 983 |
+
"eval_samples_per_second": 59.538,
|
| 984 |
+
"eval_steps_per_second": 7.459,
|
| 985 |
+
"step": 390
|
| 986 |
+
},
|
| 987 |
+
{
|
| 988 |
+
"epoch": 3.1104536489151875,
|
| 989 |
+
"grad_norm": 242.90427569755158,
|
| 990 |
+
"learning_rate": 3.402625820568928e-06,
|
| 991 |
+
"loss": 25.6816,
|
| 992 |
+
"step": 395,
|
| 993 |
+
"true_loss": 3.1437
|
| 994 |
+
},
|
| 995 |
+
{
|
| 996 |
+
"epoch": 3.1499013806706113,
|
| 997 |
+
"grad_norm": 250.93466465674842,
|
| 998 |
+
"learning_rate": 3.37527352297593e-06,
|
| 999 |
+
"loss": 24.8387,
|
| 1000 |
+
"step": 400,
|
| 1001 |
+
"true_loss": 3.1525
|
| 1002 |
+
},
|
| 1003 |
+
{
|
| 1004 |
+
"epoch": 3.1499013806706113,
|
| 1005 |
+
"eval_accuracy": 0.18514412416851442,
|
| 1006 |
+
"eval_loss": 3.435126781463623,
|
| 1007 |
+
"eval_runtime": 15.0403,
|
| 1008 |
+
"eval_samples_per_second": 59.972,
|
| 1009 |
+
"eval_steps_per_second": 7.513,
|
| 1010 |
+
"step": 400
|
| 1011 |
+
},
|
| 1012 |
+
{
|
| 1013 |
+
"epoch": 3.1893491124260356,
|
| 1014 |
+
"grad_norm": 256.79950170149914,
|
| 1015 |
+
"learning_rate": 3.3479212253829326e-06,
|
| 1016 |
+
"loss": 25.7926,
|
| 1017 |
+
"step": 405,
|
| 1018 |
+
"true_loss": 3.0679
|
| 1019 |
+
},
|
| 1020 |
+
{
|
| 1021 |
+
"epoch": 3.2287968441814594,
|
| 1022 |
+
"grad_norm": 240.77762669362124,
|
| 1023 |
+
"learning_rate": 3.3205689277899345e-06,
|
| 1024 |
+
"loss": 26.0472,
|
| 1025 |
+
"step": 410,
|
| 1026 |
+
"true_loss": 3.2098
|
| 1027 |
+
},
|
| 1028 |
+
{
|
| 1029 |
+
"epoch": 3.2287968441814594,
|
| 1030 |
+
"eval_accuracy": 0.19290465631929046,
|
| 1031 |
+
"eval_loss": 3.428030014038086,
|
| 1032 |
+
"eval_runtime": 15.1004,
|
| 1033 |
+
"eval_samples_per_second": 59.733,
|
| 1034 |
+
"eval_steps_per_second": 7.483,
|
| 1035 |
+
"step": 410
|
| 1036 |
+
},
|
| 1037 |
+
{
|
| 1038 |
+
"epoch": 3.2682445759368837,
|
| 1039 |
+
"grad_norm": 237.38142361903007,
|
| 1040 |
+
"learning_rate": 3.293216630196937e-06,
|
| 1041 |
+
"loss": 25.423,
|
| 1042 |
+
"step": 415,
|
| 1043 |
+
"true_loss": 3.0787
|
| 1044 |
+
},
|
| 1045 |
+
{
|
| 1046 |
+
"epoch": 3.3076923076923075,
|
| 1047 |
+
"grad_norm": 242.6891012315701,
|
| 1048 |
+
"learning_rate": 3.2658643326039393e-06,
|
| 1049 |
+
"loss": 25.1691,
|
| 1050 |
+
"step": 420,
|
| 1051 |
+
"true_loss": 3.208
|
| 1052 |
+
},
|
| 1053 |
+
{
|
| 1054 |
+
"epoch": 3.3076923076923075,
|
| 1055 |
+
"eval_accuracy": 0.18403547671840353,
|
| 1056 |
+
"eval_loss": 3.431478261947632,
|
| 1057 |
+
"eval_runtime": 14.9427,
|
| 1058 |
+
"eval_samples_per_second": 60.364,
|
| 1059 |
+
"eval_steps_per_second": 7.562,
|
| 1060 |
+
"step": 420
|
| 1061 |
+
},
|
| 1062 |
+
{
|
| 1063 |
+
"epoch": 3.3471400394477318,
|
| 1064 |
+
"grad_norm": 228.6708954920748,
|
| 1065 |
+
"learning_rate": 3.2385120350109413e-06,
|
| 1066 |
+
"loss": 26.0412,
|
| 1067 |
+
"step": 425,
|
| 1068 |
+
"true_loss": 3.3517
|
| 1069 |
+
},
|
| 1070 |
+
{
|
| 1071 |
+
"epoch": 3.386587771203156,
|
| 1072 |
+
"grad_norm": 244.34477651789024,
|
| 1073 |
+
"learning_rate": 3.2111597374179432e-06,
|
| 1074 |
+
"loss": 24.4891,
|
| 1075 |
+
"step": 430,
|
| 1076 |
+
"true_loss": 3.1938
|
| 1077 |
+
},
|
| 1078 |
+
{
|
| 1079 |
+
"epoch": 3.386587771203156,
|
| 1080 |
+
"eval_accuracy": 0.2017738359201774,
|
| 1081 |
+
"eval_loss": 3.4007480144500732,
|
| 1082 |
+
"eval_runtime": 14.9507,
|
| 1083 |
+
"eval_samples_per_second": 60.332,
|
| 1084 |
+
"eval_steps_per_second": 7.558,
|
| 1085 |
+
"step": 430
|
| 1086 |
+
},
|
| 1087 |
+
{
|
| 1088 |
+
"epoch": 3.42603550295858,
|
| 1089 |
+
"grad_norm": 247.97686568101082,
|
| 1090 |
+
"learning_rate": 3.1838074398249456e-06,
|
| 1091 |
+
"loss": 25.6189,
|
| 1092 |
+
"step": 435,
|
| 1093 |
+
"true_loss": 3.2995
|
| 1094 |
+
},
|
| 1095 |
+
{
|
| 1096 |
+
"epoch": 3.465483234714004,
|
| 1097 |
+
"grad_norm": 241.1156299032968,
|
| 1098 |
+
"learning_rate": 3.156455142231948e-06,
|
| 1099 |
+
"loss": 24.7909,
|
| 1100 |
+
"step": 440,
|
| 1101 |
+
"true_loss": 2.9935
|
| 1102 |
+
},
|
| 1103 |
+
{
|
| 1104 |
+
"epoch": 3.465483234714004,
|
| 1105 |
+
"eval_accuracy": 0.18070953436807094,
|
| 1106 |
+
"eval_loss": 3.395143985748291,
|
| 1107 |
+
"eval_runtime": 15.0313,
|
| 1108 |
+
"eval_samples_per_second": 60.008,
|
| 1109 |
+
"eval_steps_per_second": 7.518,
|
| 1110 |
+
"step": 440
|
| 1111 |
+
},
|
| 1112 |
+
{
|
| 1113 |
+
"epoch": 3.504930966469428,
|
| 1114 |
+
"grad_norm": 257.4476404600661,
|
| 1115 |
+
"learning_rate": 3.12910284463895e-06,
|
| 1116 |
+
"loss": 25.0061,
|
| 1117 |
+
"step": 445,
|
| 1118 |
+
"true_loss": 3.4912
|
| 1119 |
+
},
|
| 1120 |
+
{
|
| 1121 |
+
"epoch": 3.544378698224852,
|
| 1122 |
+
"grad_norm": 260.2937525280967,
|
| 1123 |
+
"learning_rate": 3.101750547045952e-06,
|
| 1124 |
+
"loss": 25.6066,
|
| 1125 |
+
"step": 450,
|
| 1126 |
+
"true_loss": 3.3814
|
| 1127 |
+
},
|
| 1128 |
+
{
|
| 1129 |
+
"epoch": 3.544378698224852,
|
| 1130 |
+
"eval_accuracy": 0.18957871396895787,
|
| 1131 |
+
"eval_loss": 3.395690679550171,
|
| 1132 |
+
"eval_runtime": 14.9346,
|
| 1133 |
+
"eval_samples_per_second": 60.397,
|
| 1134 |
+
"eval_steps_per_second": 7.566,
|
| 1135 |
+
"step": 450
|
| 1136 |
+
},
|
| 1137 |
+
{
|
| 1138 |
+
"epoch": 3.583826429980276,
|
| 1139 |
+
"grad_norm": 240.92415923092258,
|
| 1140 |
+
"learning_rate": 3.0743982494529543e-06,
|
| 1141 |
+
"loss": 26.1029,
|
| 1142 |
+
"step": 455,
|
| 1143 |
+
"true_loss": 3.2856
|
| 1144 |
+
},
|
| 1145 |
+
{
|
| 1146 |
+
"epoch": 3.6232741617357003,
|
| 1147 |
+
"grad_norm": 250.4817822446221,
|
| 1148 |
+
"learning_rate": 3.0470459518599567e-06,
|
| 1149 |
+
"loss": 24.8986,
|
| 1150 |
+
"step": 460,
|
| 1151 |
+
"true_loss": 2.8531
|
| 1152 |
+
},
|
| 1153 |
+
{
|
| 1154 |
+
"epoch": 3.6232741617357003,
|
| 1155 |
+
"eval_accuracy": 0.18181818181818182,
|
| 1156 |
+
"eval_loss": 3.4077982902526855,
|
| 1157 |
+
"eval_runtime": 15.0046,
|
| 1158 |
+
"eval_samples_per_second": 60.115,
|
| 1159 |
+
"eval_steps_per_second": 7.531,
|
| 1160 |
+
"step": 460
|
| 1161 |
+
},
|
| 1162 |
+
{
|
| 1163 |
+
"epoch": 3.662721893491124,
|
| 1164 |
+
"grad_norm": 273.72732435650465,
|
| 1165 |
+
"learning_rate": 3.0196936542669587e-06,
|
| 1166 |
+
"loss": 26.2365,
|
| 1167 |
+
"step": 465,
|
| 1168 |
+
"true_loss": 3.1754
|
| 1169 |
+
},
|
| 1170 |
+
{
|
| 1171 |
+
"epoch": 3.7021696252465484,
|
| 1172 |
+
"grad_norm": 266.52186007353083,
|
| 1173 |
+
"learning_rate": 2.9923413566739606e-06,
|
| 1174 |
+
"loss": 24.3942,
|
| 1175 |
+
"step": 470,
|
| 1176 |
+
"true_loss": 2.8515
|
| 1177 |
+
},
|
| 1178 |
+
{
|
| 1179 |
+
"epoch": 3.7021696252465484,
|
| 1180 |
+
"eval_accuracy": 0.19623059866962306,
|
| 1181 |
+
"eval_loss": 3.4032986164093018,
|
| 1182 |
+
"eval_runtime": 14.9785,
|
| 1183 |
+
"eval_samples_per_second": 60.22,
|
| 1184 |
+
"eval_steps_per_second": 7.544,
|
| 1185 |
+
"step": 470
|
| 1186 |
+
},
|
| 1187 |
+
{
|
| 1188 |
+
"epoch": 3.7416173570019726,
|
| 1189 |
+
"grad_norm": 277.7285610260442,
|
| 1190 |
+
"learning_rate": 2.9649890590809634e-06,
|
| 1191 |
+
"loss": 24.7561,
|
| 1192 |
+
"step": 475,
|
| 1193 |
+
"true_loss": 3.0628
|
| 1194 |
+
},
|
| 1195 |
+
{
|
| 1196 |
+
"epoch": 3.7810650887573964,
|
| 1197 |
+
"grad_norm": 258.65135809575645,
|
| 1198 |
+
"learning_rate": 2.9376367614879654e-06,
|
| 1199 |
+
"loss": 27.2582,
|
| 1200 |
+
"step": 480,
|
| 1201 |
+
"true_loss": 3.2503
|
| 1202 |
+
},
|
| 1203 |
+
{
|
| 1204 |
+
"epoch": 3.7810650887573964,
|
| 1205 |
+
"eval_accuracy": 0.19290465631929046,
|
| 1206 |
+
"eval_loss": 3.4006612300872803,
|
| 1207 |
+
"eval_runtime": 15.0539,
|
| 1208 |
+
"eval_samples_per_second": 59.918,
|
| 1209 |
+
"eval_steps_per_second": 7.506,
|
| 1210 |
+
"step": 480
|
| 1211 |
+
},
|
| 1212 |
+
{
|
| 1213 |
+
"epoch": 3.8205128205128203,
|
| 1214 |
+
"grad_norm": 276.16011426554854,
|
| 1215 |
+
"learning_rate": 2.9102844638949674e-06,
|
| 1216 |
+
"loss": 25.0956,
|
| 1217 |
+
"step": 485,
|
| 1218 |
+
"true_loss": 2.8415
|
| 1219 |
+
},
|
| 1220 |
+
{
|
| 1221 |
+
"epoch": 3.8599605522682445,
|
| 1222 |
+
"grad_norm": 265.77475990424085,
|
| 1223 |
+
"learning_rate": 2.8829321663019693e-06,
|
| 1224 |
+
"loss": 24.7172,
|
| 1225 |
+
"step": 490,
|
| 1226 |
+
"true_loss": 3.2856
|
| 1227 |
+
},
|
| 1228 |
+
{
|
| 1229 |
+
"epoch": 3.8599605522682445,
|
| 1230 |
+
"eval_accuracy": 0.19401330376940132,
|
| 1231 |
+
"eval_loss": 3.405134916305542,
|
| 1232 |
+
"eval_runtime": 14.9878,
|
| 1233 |
+
"eval_samples_per_second": 60.182,
|
| 1234 |
+
"eval_steps_per_second": 7.539,
|
| 1235 |
+
"step": 490
|
| 1236 |
+
},
|
| 1237 |
+
{
|
| 1238 |
+
"epoch": 3.899408284023669,
|
| 1239 |
+
"grad_norm": 273.4345528471514,
|
| 1240 |
+
"learning_rate": 2.855579868708972e-06,
|
| 1241 |
+
"loss": 25.455,
|
| 1242 |
+
"step": 495,
|
| 1243 |
+
"true_loss": 3.431
|
| 1244 |
+
},
|
| 1245 |
+
{
|
| 1246 |
+
"epoch": 3.9388560157790926,
|
| 1247 |
+
"grad_norm": 273.95487836913196,
|
| 1248 |
+
"learning_rate": 2.828227571115974e-06,
|
| 1249 |
+
"loss": 25.3041,
|
| 1250 |
+
"step": 500,
|
| 1251 |
+
"true_loss": 3.1038
|
| 1252 |
+
},
|
| 1253 |
+
{
|
| 1254 |
+
"epoch": 3.9388560157790926,
|
| 1255 |
+
"eval_accuracy": 0.2017738359201774,
|
| 1256 |
+
"eval_loss": 3.3893778324127197,
|
| 1257 |
+
"eval_runtime": 15.0204,
|
| 1258 |
+
"eval_samples_per_second": 60.052,
|
| 1259 |
+
"eval_steps_per_second": 7.523,
|
| 1260 |
+
"step": 500
|
| 1261 |
+
},
|
| 1262 |
+
{
|
| 1263 |
+
"epoch": 3.978303747534517,
|
| 1264 |
+
"grad_norm": 265.6659951241656,
|
| 1265 |
+
"learning_rate": 2.800875273522976e-06,
|
| 1266 |
+
"loss": 24.9082,
|
| 1267 |
+
"step": 505,
|
| 1268 |
+
"true_loss": 3.1013
|
| 1269 |
+
},
|
| 1270 |
+
{
|
| 1271 |
+
"epoch": 4.01577909270217,
|
| 1272 |
+
"grad_norm": 258.2230652518369,
|
| 1273 |
+
"learning_rate": 2.773522975929978e-06,
|
| 1274 |
+
"loss": 23.6265,
|
| 1275 |
+
"step": 510,
|
| 1276 |
+
"true_loss": 2.474
|
| 1277 |
+
},
|
| 1278 |
+
{
|
| 1279 |
+
"epoch": 4.01577909270217,
|
| 1280 |
+
"eval_accuracy": 0.2006651884700665,
|
| 1281 |
+
"eval_loss": 3.3711023330688477,
|
| 1282 |
+
"eval_runtime": 15.1319,
|
| 1283 |
+
"eval_samples_per_second": 59.609,
|
| 1284 |
+
"eval_steps_per_second": 7.468,
|
| 1285 |
+
"step": 510
|
| 1286 |
+
},
|
| 1287 |
+
{
|
| 1288 |
+
"epoch": 4.055226824457594,
|
| 1289 |
+
"grad_norm": 265.77115170095055,
|
| 1290 |
+
"learning_rate": 2.746170678336981e-06,
|
| 1291 |
+
"loss": 23.0143,
|
| 1292 |
+
"step": 515,
|
| 1293 |
+
"true_loss": 2.4705
|
| 1294 |
+
},
|
| 1295 |
+
{
|
| 1296 |
+
"epoch": 4.094674556213017,
|
| 1297 |
+
"grad_norm": 305.1591165434374,
|
| 1298 |
+
"learning_rate": 2.718818380743983e-06,
|
| 1299 |
+
"loss": 23.0705,
|
| 1300 |
+
"step": 520,
|
| 1301 |
+
"true_loss": 2.9254
|
| 1302 |
+
},
|
| 1303 |
+
{
|
| 1304 |
+
"epoch": 4.094674556213017,
|
| 1305 |
+
"eval_accuracy": 0.2017738359201774,
|
| 1306 |
+
"eval_loss": 3.3792917728424072,
|
| 1307 |
+
"eval_runtime": 15.0429,
|
| 1308 |
+
"eval_samples_per_second": 59.962,
|
| 1309 |
+
"eval_steps_per_second": 7.512,
|
| 1310 |
+
"step": 520
|
| 1311 |
+
},
|
| 1312 |
+
{
|
| 1313 |
+
"epoch": 4.134122287968442,
|
| 1314 |
+
"grad_norm": 301.013844993568,
|
| 1315 |
+
"learning_rate": 2.6914660831509848e-06,
|
| 1316 |
+
"loss": 23.3767,
|
| 1317 |
+
"step": 525,
|
| 1318 |
+
"true_loss": 2.2937
|
| 1319 |
+
},
|
| 1320 |
+
{
|
| 1321 |
+
"epoch": 4.173570019723866,
|
| 1322 |
+
"grad_norm": 289.0773672920475,
|
| 1323 |
+
"learning_rate": 2.6641137855579867e-06,
|
| 1324 |
+
"loss": 24.2713,
|
| 1325 |
+
"step": 530,
|
| 1326 |
+
"true_loss": 3.0639
|
| 1327 |
+
},
|
| 1328 |
+
{
|
| 1329 |
+
"epoch": 4.173570019723866,
|
| 1330 |
+
"eval_accuracy": 0.21175166297117518,
|
| 1331 |
+
"eval_loss": 3.387169122695923,
|
| 1332 |
+
"eval_runtime": 14.9666,
|
| 1333 |
+
"eval_samples_per_second": 60.267,
|
| 1334 |
+
"eval_steps_per_second": 7.55,
|
| 1335 |
+
"step": 530
|
| 1336 |
+
},
|
| 1337 |
+
{
|
| 1338 |
+
"epoch": 4.21301775147929,
|
| 1339 |
+
"grad_norm": 292.88346898040646,
|
| 1340 |
+
"learning_rate": 2.6367614879649895e-06,
|
| 1341 |
+
"loss": 23.0921,
|
| 1342 |
+
"step": 535,
|
| 1343 |
+
"true_loss": 2.8408
|
| 1344 |
+
},
|
| 1345 |
+
{
|
| 1346 |
+
"epoch": 4.252465483234714,
|
| 1347 |
+
"grad_norm": 308.73457670985425,
|
| 1348 |
+
"learning_rate": 2.6094091903719915e-06,
|
| 1349 |
+
"loss": 22.7046,
|
| 1350 |
+
"step": 540,
|
| 1351 |
+
"true_loss": 2.7119
|
| 1352 |
+
},
|
| 1353 |
+
{
|
| 1354 |
+
"epoch": 4.252465483234714,
|
| 1355 |
+
"eval_accuracy": 0.20288248337028825,
|
| 1356 |
+
"eval_loss": 3.378486156463623,
|
| 1357 |
+
"eval_runtime": 15.0097,
|
| 1358 |
+
"eval_samples_per_second": 60.094,
|
| 1359 |
+
"eval_steps_per_second": 7.528,
|
| 1360 |
+
"step": 540
|
| 1361 |
+
},
|
| 1362 |
+
{
|
| 1363 |
+
"epoch": 4.291913214990138,
|
| 1364 |
+
"grad_norm": 304.39435113456375,
|
| 1365 |
+
"learning_rate": 2.5820568927789935e-06,
|
| 1366 |
+
"loss": 23.06,
|
| 1367 |
+
"step": 545,
|
| 1368 |
+
"true_loss": 2.8129
|
| 1369 |
+
},
|
| 1370 |
+
{
|
| 1371 |
+
"epoch": 4.331360946745562,
|
| 1372 |
+
"grad_norm": 297.18920459782885,
|
| 1373 |
+
"learning_rate": 2.5547045951859954e-06,
|
| 1374 |
+
"loss": 22.8824,
|
| 1375 |
+
"step": 550,
|
| 1376 |
+
"true_loss": 2.8139
|
| 1377 |
+
},
|
| 1378 |
+
{
|
| 1379 |
+
"epoch": 4.331360946745562,
|
| 1380 |
+
"eval_accuracy": 0.21286031042128603,
|
| 1381 |
+
"eval_loss": 3.3715484142303467,
|
| 1382 |
+
"eval_runtime": 15.0858,
|
| 1383 |
+
"eval_samples_per_second": 59.791,
|
| 1384 |
+
"eval_steps_per_second": 7.491,
|
| 1385 |
+
"step": 550
|
| 1386 |
+
},
|
| 1387 |
+
{
|
| 1388 |
+
"epoch": 4.370808678500986,
|
| 1389 |
+
"grad_norm": 296.890352701281,
|
| 1390 |
+
"learning_rate": 2.5273522975929982e-06,
|
| 1391 |
+
"loss": 23.1762,
|
| 1392 |
+
"step": 555,
|
| 1393 |
+
"true_loss": 2.6411
|
| 1394 |
+
},
|
| 1395 |
+
{
|
| 1396 |
+
"epoch": 4.410256410256411,
|
| 1397 |
+
"grad_norm": 301.7124359937404,
|
| 1398 |
+
"learning_rate": 2.5e-06,
|
| 1399 |
+
"loss": 23.1173,
|
| 1400 |
+
"step": 560,
|
| 1401 |
+
"true_loss": 2.8016
|
| 1402 |
+
},
|
| 1403 |
+
{
|
| 1404 |
+
"epoch": 4.410256410256411,
|
| 1405 |
+
"eval_accuracy": 0.2006651884700665,
|
| 1406 |
+
"eval_loss": 3.3617937564849854,
|
| 1407 |
+
"eval_runtime": 15.4105,
|
| 1408 |
+
"eval_samples_per_second": 58.531,
|
| 1409 |
+
"eval_steps_per_second": 7.333,
|
| 1410 |
+
"step": 560
|
| 1411 |
+
},
|
| 1412 |
+
{
|
| 1413 |
+
"epoch": 4.449704142011834,
|
| 1414 |
+
"grad_norm": 301.5417177978464,
|
| 1415 |
+
"learning_rate": 2.4726477024070026e-06,
|
| 1416 |
+
"loss": 23.2135,
|
| 1417 |
+
"step": 565,
|
| 1418 |
+
"true_loss": 2.6408
|
| 1419 |
+
},
|
| 1420 |
+
{
|
| 1421 |
+
"epoch": 4.489151873767258,
|
| 1422 |
+
"grad_norm": 307.2723305395378,
|
| 1423 |
+
"learning_rate": 2.4452954048140046e-06,
|
| 1424 |
+
"loss": 22.2622,
|
| 1425 |
+
"step": 570,
|
| 1426 |
+
"true_loss": 2.9552
|
| 1427 |
+
},
|
| 1428 |
+
{
|
| 1429 |
+
"epoch": 4.489151873767258,
|
| 1430 |
+
"eval_accuracy": 0.1951219512195122,
|
| 1431 |
+
"eval_loss": 3.3558921813964844,
|
| 1432 |
+
"eval_runtime": 14.9957,
|
| 1433 |
+
"eval_samples_per_second": 60.15,
|
| 1434 |
+
"eval_steps_per_second": 7.535,
|
| 1435 |
+
"step": 570
|
| 1436 |
+
},
|
| 1437 |
+
{
|
| 1438 |
+
"epoch": 4.5285996055226825,
|
| 1439 |
+
"grad_norm": 316.2854049798904,
|
| 1440 |
+
"learning_rate": 2.417943107221007e-06,
|
| 1441 |
+
"loss": 23.6898,
|
| 1442 |
+
"step": 575,
|
| 1443 |
+
"true_loss": 3.0164
|
| 1444 |
+
},
|
| 1445 |
+
{
|
| 1446 |
+
"epoch": 4.568047337278107,
|
| 1447 |
+
"grad_norm": 285.3077007071203,
|
| 1448 |
+
"learning_rate": 2.390590809628009e-06,
|
| 1449 |
+
"loss": 21.7338,
|
| 1450 |
+
"step": 580,
|
| 1451 |
+
"true_loss": 2.3795
|
| 1452 |
+
},
|
| 1453 |
+
{
|
| 1454 |
+
"epoch": 4.568047337278107,
|
| 1455 |
+
"eval_accuracy": 0.21507760532150777,
|
| 1456 |
+
"eval_loss": 3.3453264236450195,
|
| 1457 |
+
"eval_runtime": 14.9619,
|
| 1458 |
+
"eval_samples_per_second": 60.286,
|
| 1459 |
+
"eval_steps_per_second": 7.553,
|
| 1460 |
+
"step": 580
|
| 1461 |
+
},
|
| 1462 |
+
{
|
| 1463 |
+
"epoch": 4.607495069033531,
|
| 1464 |
+
"grad_norm": 317.59610581151685,
|
| 1465 |
+
"learning_rate": 2.3632385120350113e-06,
|
| 1466 |
+
"loss": 23.3327,
|
| 1467 |
+
"step": 585,
|
| 1468 |
+
"true_loss": 3.1207
|
| 1469 |
+
},
|
| 1470 |
+
{
|
| 1471 |
+
"epoch": 4.646942800788954,
|
| 1472 |
+
"grad_norm": 321.934617091813,
|
| 1473 |
+
"learning_rate": 2.3358862144420133e-06,
|
| 1474 |
+
"loss": 23.3021,
|
| 1475 |
+
"step": 590,
|
| 1476 |
+
"true_loss": 2.9562
|
| 1477 |
+
},
|
| 1478 |
+
{
|
| 1479 |
+
"epoch": 4.646942800788954,
|
| 1480 |
+
"eval_accuracy": 0.21840354767184036,
|
| 1481 |
+
"eval_loss": 3.3475186824798584,
|
| 1482 |
+
"eval_runtime": 14.8088,
|
| 1483 |
+
"eval_samples_per_second": 60.91,
|
| 1484 |
+
"eval_steps_per_second": 7.631,
|
| 1485 |
+
"step": 590
|
| 1486 |
+
},
|
| 1487 |
+
{
|
| 1488 |
+
"epoch": 4.686390532544379,
|
| 1489 |
+
"grad_norm": 318.62665027252604,
|
| 1490 |
+
"learning_rate": 2.3085339168490156e-06,
|
| 1491 |
+
"loss": 23.6124,
|
| 1492 |
+
"step": 595,
|
| 1493 |
+
"true_loss": 3.19
|
| 1494 |
+
},
|
| 1495 |
+
{
|
| 1496 |
+
"epoch": 4.725838264299803,
|
| 1497 |
+
"grad_norm": 323.13589935545605,
|
| 1498 |
+
"learning_rate": 2.2811816192560176e-06,
|
| 1499 |
+
"loss": 23.3896,
|
| 1500 |
+
"step": 600,
|
| 1501 |
+
"true_loss": 2.9138
|
| 1502 |
+
},
|
| 1503 |
+
{
|
| 1504 |
+
"epoch": 4.725838264299803,
|
| 1505 |
+
"eval_accuracy": 0.2073170731707317,
|
| 1506 |
+
"eval_loss": 3.3516340255737305,
|
| 1507 |
+
"eval_runtime": 15.0949,
|
| 1508 |
+
"eval_samples_per_second": 59.755,
|
| 1509 |
+
"eval_steps_per_second": 7.486,
|
| 1510 |
+
"step": 600
|
| 1511 |
+
},
|
| 1512 |
+
{
|
| 1513 |
+
"epoch": 4.765285996055227,
|
| 1514 |
+
"grad_norm": 306.75577469369307,
|
| 1515 |
+
"learning_rate": 2.25382932166302e-06,
|
| 1516 |
+
"loss": 23.1344,
|
| 1517 |
+
"step": 605,
|
| 1518 |
+
"true_loss": 2.9071
|
| 1519 |
+
},
|
| 1520 |
+
{
|
| 1521 |
+
"epoch": 4.804733727810651,
|
| 1522 |
+
"grad_norm": 305.64568552389716,
|
| 1523 |
+
"learning_rate": 2.226477024070022e-06,
|
| 1524 |
+
"loss": 22.918,
|
| 1525 |
+
"step": 610,
|
| 1526 |
+
"true_loss": 3.205
|
| 1527 |
+
},
|
| 1528 |
+
{
|
| 1529 |
+
"epoch": 4.804733727810651,
|
| 1530 |
+
"eval_accuracy": 0.21286031042128603,
|
| 1531 |
+
"eval_loss": 3.344453811645508,
|
| 1532 |
+
"eval_runtime": 14.8514,
|
| 1533 |
+
"eval_samples_per_second": 60.735,
|
| 1534 |
+
"eval_steps_per_second": 7.609,
|
| 1535 |
+
"step": 610
|
| 1536 |
+
},
|
| 1537 |
+
{
|
| 1538 |
+
"epoch": 4.844181459566075,
|
| 1539 |
+
"grad_norm": 345.3630194507602,
|
| 1540 |
+
"learning_rate": 2.1991247264770243e-06,
|
| 1541 |
+
"loss": 22.4193,
|
| 1542 |
+
"step": 615,
|
| 1543 |
+
"true_loss": 3.0203
|
| 1544 |
+
},
|
| 1545 |
+
{
|
| 1546 |
+
"epoch": 4.883629191321499,
|
| 1547 |
+
"grad_norm": 310.8043549889887,
|
| 1548 |
+
"learning_rate": 2.1717724288840263e-06,
|
| 1549 |
+
"loss": 22.9716,
|
| 1550 |
+
"step": 620,
|
| 1551 |
+
"true_loss": 2.7156
|
| 1552 |
+
},
|
| 1553 |
+
{
|
| 1554 |
+
"epoch": 4.883629191321499,
|
| 1555 |
+
"eval_accuracy": 0.2261640798226164,
|
| 1556 |
+
"eval_loss": 3.3267650604248047,
|
| 1557 |
+
"eval_runtime": 15.0181,
|
| 1558 |
+
"eval_samples_per_second": 60.061,
|
| 1559 |
+
"eval_steps_per_second": 7.524,
|
| 1560 |
+
"step": 620
|
| 1561 |
+
},
|
| 1562 |
+
{
|
| 1563 |
+
"epoch": 4.923076923076923,
|
| 1564 |
+
"grad_norm": 328.17455994439047,
|
| 1565 |
+
"learning_rate": 2.1444201312910287e-06,
|
| 1566 |
+
"loss": 23.5778,
|
| 1567 |
+
"step": 625,
|
| 1568 |
+
"true_loss": 3.0593
|
| 1569 |
+
},
|
| 1570 |
+
{
|
| 1571 |
+
"epoch": 4.962524654832347,
|
| 1572 |
+
"grad_norm": 314.85030625337566,
|
| 1573 |
+
"learning_rate": 2.1170678336980307e-06,
|
| 1574 |
+
"loss": 23.4656,
|
| 1575 |
+
"step": 630,
|
| 1576 |
+
"true_loss": 2.7659
|
| 1577 |
+
},
|
| 1578 |
+
{
|
| 1579 |
+
"epoch": 4.962524654832347,
|
| 1580 |
+
"eval_accuracy": 0.22283813747228381,
|
| 1581 |
+
"eval_loss": 3.313983201980591,
|
| 1582 |
+
"eval_runtime": 15.004,
|
| 1583 |
+
"eval_samples_per_second": 60.117,
|
| 1584 |
+
"eval_steps_per_second": 7.531,
|
| 1585 |
+
"step": 630
|
| 1586 |
+
},
|
| 1587 |
+
{
|
| 1588 |
+
"epoch": 5.0,
|
| 1589 |
+
"grad_norm": 264.024205323258,
|
| 1590 |
+
"learning_rate": 2.089715536105033e-06,
|
| 1591 |
+
"loss": 22.7471,
|
| 1592 |
+
"step": 635,
|
| 1593 |
+
"true_loss": 2.9564
|
| 1594 |
+
},
|
| 1595 |
+
{
|
| 1596 |
+
"epoch": 5.039447731755424,
|
| 1597 |
+
"grad_norm": 306.42431977803875,
|
| 1598 |
+
"learning_rate": 2.062363238512035e-06,
|
| 1599 |
+
"loss": 21.5082,
|
| 1600 |
+
"step": 640,
|
| 1601 |
+
"true_loss": 2.7247
|
| 1602 |
+
},
|
| 1603 |
+
{
|
| 1604 |
+
"epoch": 5.039447731755424,
|
| 1605 |
+
"eval_accuracy": 0.22838137472283815,
|
| 1606 |
+
"eval_loss": 3.3180108070373535,
|
| 1607 |
+
"eval_runtime": 14.9697,
|
| 1608 |
+
"eval_samples_per_second": 60.255,
|
| 1609 |
+
"eval_steps_per_second": 7.549,
|
| 1610 |
+
"step": 640
|
| 1611 |
+
},
|
| 1612 |
+
{
|
| 1613 |
+
"epoch": 5.0788954635108485,
|
| 1614 |
+
"grad_norm": 340.14969305624527,
|
| 1615 |
+
"learning_rate": 2.0350109409190374e-06,
|
| 1616 |
+
"loss": 21.563,
|
| 1617 |
+
"step": 645,
|
| 1618 |
+
"true_loss": 2.5149
|
| 1619 |
+
},
|
| 1620 |
+
{
|
| 1621 |
+
"epoch": 5.118343195266272,
|
| 1622 |
+
"grad_norm": 323.01121928019046,
|
| 1623 |
+
"learning_rate": 2.0076586433260394e-06,
|
| 1624 |
+
"loss": 20.9225,
|
| 1625 |
+
"step": 650,
|
| 1626 |
+
"true_loss": 2.632
|
| 1627 |
+
},
|
| 1628 |
+
{
|
| 1629 |
+
"epoch": 5.118343195266272,
|
| 1630 |
+
"eval_accuracy": 0.21618625277161863,
|
| 1631 |
+
"eval_loss": 3.321817398071289,
|
| 1632 |
+
"eval_runtime": 15.0606,
|
| 1633 |
+
"eval_samples_per_second": 59.891,
|
| 1634 |
+
"eval_steps_per_second": 7.503,
|
| 1635 |
+
"step": 650
|
| 1636 |
+
},
|
| 1637 |
+
{
|
| 1638 |
+
"epoch": 5.157790927021696,
|
| 1639 |
+
"grad_norm": 348.8138229734559,
|
| 1640 |
+
"learning_rate": 1.9803063457330417e-06,
|
| 1641 |
+
"loss": 20.005,
|
| 1642 |
+
"step": 655,
|
| 1643 |
+
"true_loss": 2.4218
|
| 1644 |
+
},
|
| 1645 |
+
{
|
| 1646 |
+
"epoch": 5.19723865877712,
|
| 1647 |
+
"grad_norm": 336.229278513962,
|
| 1648 |
+
"learning_rate": 1.9529540481400437e-06,
|
| 1649 |
+
"loss": 19.3711,
|
| 1650 |
+
"step": 660,
|
| 1651 |
+
"true_loss": 2.5052
|
| 1652 |
+
},
|
| 1653 |
+
{
|
| 1654 |
+
"epoch": 5.19723865877712,
|
| 1655 |
+
"eval_accuracy": 0.20953436807095344,
|
| 1656 |
+
"eval_loss": 3.3247318267822266,
|
| 1657 |
+
"eval_runtime": 14.9674,
|
| 1658 |
+
"eval_samples_per_second": 60.264,
|
| 1659 |
+
"eval_steps_per_second": 7.55,
|
| 1660 |
+
"step": 660
|
| 1661 |
+
},
|
| 1662 |
+
{
|
| 1663 |
+
"epoch": 5.236686390532545,
|
| 1664 |
+
"grad_norm": 359.810565898114,
|
| 1665 |
+
"learning_rate": 1.925601750547046e-06,
|
| 1666 |
+
"loss": 20.7458,
|
| 1667 |
+
"step": 665,
|
| 1668 |
+
"true_loss": 2.2577
|
| 1669 |
+
},
|
| 1670 |
+
{
|
| 1671 |
+
"epoch": 5.276134122287968,
|
| 1672 |
+
"grad_norm": 326.2113183093954,
|
| 1673 |
+
"learning_rate": 1.8982494529540483e-06,
|
| 1674 |
+
"loss": 21.3133,
|
| 1675 |
+
"step": 670,
|
| 1676 |
+
"true_loss": 2.5225
|
| 1677 |
+
},
|
| 1678 |
+
{
|
| 1679 |
+
"epoch": 5.276134122287968,
|
| 1680 |
+
"eval_accuracy": 0.21729490022172948,
|
| 1681 |
+
"eval_loss": 3.3202853202819824,
|
| 1682 |
+
"eval_runtime": 14.8455,
|
| 1683 |
+
"eval_samples_per_second": 60.759,
|
| 1684 |
+
"eval_steps_per_second": 7.612,
|
| 1685 |
+
"step": 670
|
| 1686 |
+
},
|
| 1687 |
+
{
|
| 1688 |
+
"epoch": 5.315581854043392,
|
| 1689 |
+
"grad_norm": 347.21516127995676,
|
| 1690 |
+
"learning_rate": 1.8708971553610504e-06,
|
| 1691 |
+
"loss": 20.257,
|
| 1692 |
+
"step": 675,
|
| 1693 |
+
"true_loss": 2.7666
|
| 1694 |
+
},
|
| 1695 |
+
{
|
| 1696 |
+
"epoch": 5.355029585798817,
|
| 1697 |
+
"grad_norm": 344.59999127493575,
|
| 1698 |
+
"learning_rate": 1.8435448577680526e-06,
|
| 1699 |
+
"loss": 20.2119,
|
| 1700 |
+
"step": 680,
|
| 1701 |
+
"true_loss": 2.548
|
| 1702 |
+
},
|
| 1703 |
+
{
|
| 1704 |
+
"epoch": 5.355029585798817,
|
| 1705 |
+
"eval_accuracy": 0.21618625277161863,
|
| 1706 |
+
"eval_loss": 3.327068328857422,
|
| 1707 |
+
"eval_runtime": 15.0698,
|
| 1708 |
+
"eval_samples_per_second": 59.855,
|
| 1709 |
+
"eval_steps_per_second": 7.498,
|
| 1710 |
+
"step": 680
|
| 1711 |
+
},
|
| 1712 |
+
{
|
| 1713 |
+
"epoch": 5.394477317554241,
|
| 1714 |
+
"grad_norm": 356.6528216246677,
|
| 1715 |
+
"learning_rate": 1.816192560175055e-06,
|
| 1716 |
+
"loss": 20.4713,
|
| 1717 |
+
"step": 685,
|
| 1718 |
+
"true_loss": 2.7505
|
| 1719 |
+
},
|
| 1720 |
+
{
|
| 1721 |
+
"epoch": 5.433925049309664,
|
| 1722 |
+
"grad_norm": 373.80585141045236,
|
| 1723 |
+
"learning_rate": 1.788840262582057e-06,
|
| 1724 |
+
"loss": 21.3783,
|
| 1725 |
+
"step": 690,
|
| 1726 |
+
"true_loss": 2.5356
|
| 1727 |
+
},
|
| 1728 |
+
{
|
| 1729 |
+
"epoch": 5.433925049309664,
|
| 1730 |
+
"eval_accuracy": 0.2073170731707317,
|
| 1731 |
+
"eval_loss": 3.330681085586548,
|
| 1732 |
+
"eval_runtime": 15.1298,
|
| 1733 |
+
"eval_samples_per_second": 59.617,
|
| 1734 |
+
"eval_steps_per_second": 7.469,
|
| 1735 |
+
"step": 690
|
| 1736 |
+
},
|
| 1737 |
+
{
|
| 1738 |
+
"epoch": 5.4733727810650885,
|
| 1739 |
+
"grad_norm": 385.2402312951401,
|
| 1740 |
+
"learning_rate": 1.7614879649890594e-06,
|
| 1741 |
+
"loss": 20.441,
|
| 1742 |
+
"step": 695,
|
| 1743 |
+
"true_loss": 2.6452
|
| 1744 |
+
},
|
| 1745 |
+
{
|
| 1746 |
+
"epoch": 5.512820512820513,
|
| 1747 |
+
"grad_norm": 369.2755541375683,
|
| 1748 |
+
"learning_rate": 1.7341356673960613e-06,
|
| 1749 |
+
"loss": 22.1389,
|
| 1750 |
+
"step": 700,
|
| 1751 |
+
"true_loss": 2.5363
|
| 1752 |
+
},
|
| 1753 |
+
{
|
| 1754 |
+
"epoch": 5.512820512820513,
|
| 1755 |
+
"eval_accuracy": 0.22172949002217296,
|
| 1756 |
+
"eval_loss": 3.3146209716796875,
|
| 1757 |
+
"eval_runtime": 15.1379,
|
| 1758 |
+
"eval_samples_per_second": 59.586,
|
| 1759 |
+
"eval_steps_per_second": 7.465,
|
| 1760 |
+
"step": 700
|
| 1761 |
+
},
|
| 1762 |
+
{
|
| 1763 |
+
"epoch": 5.552268244575937,
|
| 1764 |
+
"grad_norm": 360.7604924002524,
|
| 1765 |
+
"learning_rate": 1.7067833698030637e-06,
|
| 1766 |
+
"loss": 19.7835,
|
| 1767 |
+
"step": 705,
|
| 1768 |
+
"true_loss": 2.455
|
| 1769 |
+
},
|
| 1770 |
+
{
|
| 1771 |
+
"epoch": 5.591715976331361,
|
| 1772 |
+
"grad_norm": 345.01754212392365,
|
| 1773 |
+
"learning_rate": 1.6794310722100657e-06,
|
| 1774 |
+
"loss": 22.7283,
|
| 1775 |
+
"step": 710,
|
| 1776 |
+
"true_loss": 3.471
|
| 1777 |
+
},
|
| 1778 |
+
{
|
| 1779 |
+
"epoch": 5.591715976331361,
|
| 1780 |
+
"eval_accuracy": 0.22838137472283815,
|
| 1781 |
+
"eval_loss": 3.3121492862701416,
|
| 1782 |
+
"eval_runtime": 15.0858,
|
| 1783 |
+
"eval_samples_per_second": 59.791,
|
| 1784 |
+
"eval_steps_per_second": 7.49,
|
| 1785 |
+
"step": 710
|
| 1786 |
+
},
|
| 1787 |
+
{
|
| 1788 |
+
"epoch": 5.631163708086785,
|
| 1789 |
+
"grad_norm": 381.2083321719042,
|
| 1790 |
+
"learning_rate": 1.652078774617068e-06,
|
| 1791 |
+
"loss": 22.344,
|
| 1792 |
+
"step": 715,
|
| 1793 |
+
"true_loss": 2.7354
|
| 1794 |
+
},
|
| 1795 |
+
{
|
| 1796 |
+
"epoch": 5.670611439842209,
|
| 1797 |
+
"grad_norm": 364.2642170009335,
|
| 1798 |
+
"learning_rate": 1.62472647702407e-06,
|
| 1799 |
+
"loss": 20.889,
|
| 1800 |
+
"step": 720,
|
| 1801 |
+
"true_loss": 2.6664
|
| 1802 |
+
},
|
| 1803 |
+
{
|
| 1804 |
+
"epoch": 5.670611439842209,
|
| 1805 |
+
"eval_accuracy": 0.23059866962305986,
|
| 1806 |
+
"eval_loss": 3.314051389694214,
|
| 1807 |
+
"eval_runtime": 14.9034,
|
| 1808 |
+
"eval_samples_per_second": 60.523,
|
| 1809 |
+
"eval_steps_per_second": 7.582,
|
| 1810 |
+
"step": 720
|
| 1811 |
+
},
|
| 1812 |
+
{
|
| 1813 |
+
"epoch": 5.710059171597633,
|
| 1814 |
+
"grad_norm": 395.4957897175653,
|
| 1815 |
+
"learning_rate": 1.5973741794310724e-06,
|
| 1816 |
+
"loss": 20.6087,
|
| 1817 |
+
"step": 725,
|
| 1818 |
+
"true_loss": 2.8175
|
| 1819 |
+
},
|
| 1820 |
+
{
|
| 1821 |
+
"epoch": 5.7495069033530575,
|
| 1822 |
+
"grad_norm": 425.38224141217324,
|
| 1823 |
+
"learning_rate": 1.5700218818380744e-06,
|
| 1824 |
+
"loss": 20.4895,
|
| 1825 |
+
"step": 730,
|
| 1826 |
+
"true_loss": 2.452
|
| 1827 |
+
},
|
| 1828 |
+
{
|
| 1829 |
+
"epoch": 5.7495069033530575,
|
| 1830 |
+
"eval_accuracy": 0.23503325942350334,
|
| 1831 |
+
"eval_loss": 3.308513641357422,
|
| 1832 |
+
"eval_runtime": 15.0023,
|
| 1833 |
+
"eval_samples_per_second": 60.124,
|
| 1834 |
+
"eval_steps_per_second": 7.532,
|
| 1835 |
+
"step": 730
|
| 1836 |
+
},
|
| 1837 |
+
{
|
| 1838 |
+
"epoch": 5.788954635108482,
|
| 1839 |
+
"grad_norm": 370.3276392958097,
|
| 1840 |
+
"learning_rate": 1.5426695842450768e-06,
|
| 1841 |
+
"loss": 20.9818,
|
| 1842 |
+
"step": 735,
|
| 1843 |
+
"true_loss": 2.7038
|
| 1844 |
+
},
|
| 1845 |
+
{
|
| 1846 |
+
"epoch": 5.828402366863905,
|
| 1847 |
+
"grad_norm": 367.66652541868655,
|
| 1848 |
+
"learning_rate": 1.5153172866520787e-06,
|
| 1849 |
+
"loss": 21.7155,
|
| 1850 |
+
"step": 740,
|
| 1851 |
+
"true_loss": 2.613
|
| 1852 |
+
},
|
| 1853 |
+
{
|
| 1854 |
+
"epoch": 5.828402366863905,
|
| 1855 |
+
"eval_accuracy": 0.2394678492239468,
|
| 1856 |
+
"eval_loss": 3.302647829055786,
|
| 1857 |
+
"eval_runtime": 14.9417,
|
| 1858 |
+
"eval_samples_per_second": 60.368,
|
| 1859 |
+
"eval_steps_per_second": 7.563,
|
| 1860 |
+
"step": 740
|
| 1861 |
+
}
|
| 1862 |
+
],
|
| 1863 |
+
"logging_steps": 5,
|
| 1864 |
+
"max_steps": 1016,
|
| 1865 |
+
"num_input_tokens_seen": 0,
|
| 1866 |
+
"num_train_epochs": 8,
|
| 1867 |
+
"save_steps": 10,
|
| 1868 |
+
"stateful_callbacks": {
|
| 1869 |
+
"TrainerControl": {
|
| 1870 |
+
"args": {
|
| 1871 |
+
"should_epoch_stop": false,
|
| 1872 |
+
"should_evaluate": false,
|
| 1873 |
+
"should_log": false,
|
| 1874 |
+
"should_save": true,
|
| 1875 |
+
"should_training_stop": false
|
| 1876 |
+
},
|
| 1877 |
+
"attributes": {}
|
| 1878 |
+
}
|
| 1879 |
+
},
|
| 1880 |
+
"total_flos": 0.0,
|
| 1881 |
+
"train_batch_size": 1,
|
| 1882 |
+
"trial_name": null,
|
| 1883 |
+
"trial_params": null
|
| 1884 |
+
}
|
training_args.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0d5d6e598316abfca69bfea42cf44343863a776e74c0df2287f7b68cd1bdf2a7
|
| 3 |
+
size 7249
|
vocab.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
zero_to_fp32.py
ADDED
|
@@ -0,0 +1,760 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python
|
| 2 |
+
|
| 3 |
+
# Copyright (c) Microsoft Corporation.
|
| 4 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 5 |
+
|
| 6 |
+
# DeepSpeed Team
|
| 7 |
+
|
| 8 |
+
# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
|
| 9 |
+
# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
|
| 10 |
+
# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
|
| 11 |
+
# application.
|
| 12 |
+
#
|
| 13 |
+
# example:
|
| 14 |
+
# python zero_to_fp32.py . output_dir/
|
| 15 |
+
# or
|
| 16 |
+
# python zero_to_fp32.py . output_dir/ --safe_serialization
|
| 17 |
+
|
| 18 |
+
import argparse
|
| 19 |
+
import torch
|
| 20 |
+
import glob
|
| 21 |
+
import math
|
| 22 |
+
import os
|
| 23 |
+
import re
|
| 24 |
+
import gc
|
| 25 |
+
import json
|
| 26 |
+
import numpy as np
|
| 27 |
+
from tqdm import tqdm
|
| 28 |
+
from collections import OrderedDict
|
| 29 |
+
from dataclasses import dataclass
|
| 30 |
+
|
| 31 |
+
# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
|
| 32 |
+
# DeepSpeed data structures it has to be available in the current python environment.
|
| 33 |
+
from deepspeed.utils import logger
|
| 34 |
+
from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
|
| 35 |
+
FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
|
| 36 |
+
FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
@dataclass
|
| 40 |
+
class zero_model_state:
|
| 41 |
+
buffers: dict()
|
| 42 |
+
param_shapes: dict()
|
| 43 |
+
shared_params: list
|
| 44 |
+
ds_version: int
|
| 45 |
+
frozen_param_shapes: dict()
|
| 46 |
+
frozen_param_fragments: dict()
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
debug = 0
|
| 50 |
+
|
| 51 |
+
# load to cpu
|
| 52 |
+
device = torch.device('cpu')
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
def atoi(text):
|
| 56 |
+
return int(text) if text.isdigit() else text
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
def natural_keys(text):
|
| 60 |
+
'''
|
| 61 |
+
alist.sort(key=natural_keys) sorts in human order
|
| 62 |
+
http://nedbatchelder.com/blog/200712/human_sorting.html
|
| 63 |
+
(See Toothy's implementation in the comments)
|
| 64 |
+
'''
|
| 65 |
+
return [atoi(c) for c in re.split(r'(\d+)', text)]
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
def get_model_state_file(checkpoint_dir, zero_stage):
|
| 69 |
+
if not os.path.isdir(checkpoint_dir):
|
| 70 |
+
raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
|
| 71 |
+
|
| 72 |
+
# there should be only one file
|
| 73 |
+
if zero_stage <= 2:
|
| 74 |
+
file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
|
| 75 |
+
elif zero_stage == 3:
|
| 76 |
+
file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
|
| 77 |
+
|
| 78 |
+
if not os.path.exists(file):
|
| 79 |
+
raise FileNotFoundError(f"can't find model states file at '{file}'")
|
| 80 |
+
|
| 81 |
+
return file
|
| 82 |
+
|
| 83 |
+
|
| 84 |
+
def get_checkpoint_files(checkpoint_dir, glob_pattern):
|
| 85 |
+
# XXX: need to test that this simple glob rule works for multi-node setup too
|
| 86 |
+
ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
|
| 87 |
+
|
| 88 |
+
if len(ckpt_files) == 0:
|
| 89 |
+
raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
|
| 90 |
+
|
| 91 |
+
return ckpt_files
|
| 92 |
+
|
| 93 |
+
|
| 94 |
+
def get_optim_files(checkpoint_dir):
|
| 95 |
+
return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
|
| 96 |
+
|
| 97 |
+
|
| 98 |
+
def get_model_state_files(checkpoint_dir):
|
| 99 |
+
return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
|
| 100 |
+
|
| 101 |
+
|
| 102 |
+
def parse_model_states(files):
|
| 103 |
+
zero_model_states = []
|
| 104 |
+
for file in files:
|
| 105 |
+
state_dict = torch.load(file, map_location=device, weights_only=False)
|
| 106 |
+
|
| 107 |
+
if BUFFER_NAMES not in state_dict:
|
| 108 |
+
raise ValueError(f"{file} is not a model state checkpoint")
|
| 109 |
+
buffer_names = state_dict[BUFFER_NAMES]
|
| 110 |
+
if debug:
|
| 111 |
+
print("Found buffers:", buffer_names)
|
| 112 |
+
|
| 113 |
+
# recover just the buffers while restoring them to fp32 if they were saved in fp16
|
| 114 |
+
buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
|
| 115 |
+
param_shapes = state_dict[PARAM_SHAPES]
|
| 116 |
+
|
| 117 |
+
# collect parameters that are included in param_shapes
|
| 118 |
+
param_names = []
|
| 119 |
+
for s in param_shapes:
|
| 120 |
+
for name in s.keys():
|
| 121 |
+
param_names.append(name)
|
| 122 |
+
|
| 123 |
+
# update with frozen parameters
|
| 124 |
+
frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
|
| 125 |
+
if frozen_param_shapes is not None:
|
| 126 |
+
if debug:
|
| 127 |
+
print(f"Found frozen_param_shapes: {frozen_param_shapes}")
|
| 128 |
+
param_names += list(frozen_param_shapes.keys())
|
| 129 |
+
|
| 130 |
+
# handle shared params
|
| 131 |
+
shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
|
| 132 |
+
|
| 133 |
+
ds_version = state_dict.get(DS_VERSION, None)
|
| 134 |
+
|
| 135 |
+
frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
|
| 136 |
+
|
| 137 |
+
z_model_state = zero_model_state(buffers=buffers,
|
| 138 |
+
param_shapes=param_shapes,
|
| 139 |
+
shared_params=shared_params,
|
| 140 |
+
ds_version=ds_version,
|
| 141 |
+
frozen_param_shapes=frozen_param_shapes,
|
| 142 |
+
frozen_param_fragments=frozen_param_fragments)
|
| 143 |
+
zero_model_states.append(z_model_state)
|
| 144 |
+
|
| 145 |
+
return zero_model_states
|
| 146 |
+
|
| 147 |
+
|
| 148 |
+
def parse_optim_states(files, ds_checkpoint_dir):
|
| 149 |
+
total_files = len(files)
|
| 150 |
+
state_dicts = []
|
| 151 |
+
for f in tqdm(files, desc='Loading checkpoint shards'):
|
| 152 |
+
state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
|
| 153 |
+
# immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
|
| 154 |
+
# and also handle the case where it was already removed by another helper script
|
| 155 |
+
state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
|
| 156 |
+
state_dicts.append(state_dict)
|
| 157 |
+
|
| 158 |
+
if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
|
| 159 |
+
raise ValueError(f"{files[0]} is not a zero checkpoint")
|
| 160 |
+
zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
|
| 161 |
+
world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
|
| 162 |
+
|
| 163 |
+
# For ZeRO-2 each param group can have different partition_count as data parallelism for expert
|
| 164 |
+
# parameters can be different from data parallelism for non-expert parameters. So we can just
|
| 165 |
+
# use the max of the partition_count to get the dp world_size.
|
| 166 |
+
|
| 167 |
+
if type(world_size) is list:
|
| 168 |
+
world_size = max(world_size)
|
| 169 |
+
|
| 170 |
+
if world_size != total_files:
|
| 171 |
+
raise ValueError(
|
| 172 |
+
f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
|
| 173 |
+
"Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
|
| 174 |
+
)
|
| 175 |
+
|
| 176 |
+
# the groups are named differently in each stage
|
| 177 |
+
if zero_stage <= 2:
|
| 178 |
+
fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
|
| 179 |
+
elif zero_stage == 3:
|
| 180 |
+
fp32_groups_key = FP32_FLAT_GROUPS
|
| 181 |
+
else:
|
| 182 |
+
raise ValueError(f"unknown zero stage {zero_stage}")
|
| 183 |
+
|
| 184 |
+
fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
|
| 185 |
+
return zero_stage, world_size, fp32_flat_groups
|
| 186 |
+
|
| 187 |
+
|
| 188 |
+
def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
|
| 189 |
+
"""
|
| 190 |
+
Returns fp32 state_dict reconstructed from ds checkpoint
|
| 191 |
+
|
| 192 |
+
Args:
|
| 193 |
+
- ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
|
| 194 |
+
|
| 195 |
+
"""
|
| 196 |
+
print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
|
| 197 |
+
|
| 198 |
+
optim_files = get_optim_files(ds_checkpoint_dir)
|
| 199 |
+
zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
|
| 200 |
+
print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
|
| 201 |
+
|
| 202 |
+
model_files = get_model_state_files(ds_checkpoint_dir)
|
| 203 |
+
|
| 204 |
+
zero_model_states = parse_model_states(model_files)
|
| 205 |
+
print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
|
| 206 |
+
|
| 207 |
+
if zero_stage <= 2:
|
| 208 |
+
return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
|
| 209 |
+
exclude_frozen_parameters)
|
| 210 |
+
elif zero_stage == 3:
|
| 211 |
+
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
|
| 212 |
+
exclude_frozen_parameters)
|
| 213 |
+
|
| 214 |
+
|
| 215 |
+
def _zero2_merge_frozen_params(state_dict, zero_model_states):
|
| 216 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
| 217 |
+
return
|
| 218 |
+
|
| 219 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
| 220 |
+
frozen_param_fragments = zero_model_states[0].frozen_param_fragments
|
| 221 |
+
|
| 222 |
+
if debug:
|
| 223 |
+
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
| 224 |
+
print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
| 225 |
+
|
| 226 |
+
wanted_params = len(frozen_param_shapes)
|
| 227 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
| 228 |
+
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
| 229 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
| 230 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
| 231 |
+
|
| 232 |
+
total_params = 0
|
| 233 |
+
total_numel = 0
|
| 234 |
+
for name, shape in frozen_param_shapes.items():
|
| 235 |
+
total_params += 1
|
| 236 |
+
unpartitioned_numel = shape.numel()
|
| 237 |
+
total_numel += unpartitioned_numel
|
| 238 |
+
|
| 239 |
+
state_dict[name] = frozen_param_fragments[name]
|
| 240 |
+
|
| 241 |
+
if debug:
|
| 242 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
| 243 |
+
|
| 244 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
| 245 |
+
|
| 246 |
+
|
| 247 |
+
def _has_callable(obj, fn):
|
| 248 |
+
attr = getattr(obj, fn, None)
|
| 249 |
+
return callable(attr)
|
| 250 |
+
|
| 251 |
+
|
| 252 |
+
def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
| 253 |
+
param_shapes = zero_model_states[0].param_shapes
|
| 254 |
+
|
| 255 |
+
# Reconstruction protocol:
|
| 256 |
+
#
|
| 257 |
+
# XXX: document this
|
| 258 |
+
|
| 259 |
+
if debug:
|
| 260 |
+
for i in range(world_size):
|
| 261 |
+
for j in range(len(fp32_flat_groups[0])):
|
| 262 |
+
print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
| 263 |
+
|
| 264 |
+
# XXX: memory usage doubles here (zero2)
|
| 265 |
+
num_param_groups = len(fp32_flat_groups[0])
|
| 266 |
+
merged_single_partition_of_fp32_groups = []
|
| 267 |
+
for i in range(num_param_groups):
|
| 268 |
+
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
| 269 |
+
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
| 270 |
+
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
| 271 |
+
avail_numel = sum(
|
| 272 |
+
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
|
| 273 |
+
|
| 274 |
+
if debug:
|
| 275 |
+
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
| 276 |
+
wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
| 277 |
+
# not asserting if there is a mismatch due to possible padding
|
| 278 |
+
print(f"Have {avail_numel} numels to process.")
|
| 279 |
+
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
| 280 |
+
|
| 281 |
+
# params
|
| 282 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
| 283 |
+
# out-of-core computing solution
|
| 284 |
+
total_numel = 0
|
| 285 |
+
total_params = 0
|
| 286 |
+
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
| 287 |
+
offset = 0
|
| 288 |
+
avail_numel = full_single_fp32_vector.numel()
|
| 289 |
+
for name, shape in shapes.items():
|
| 290 |
+
|
| 291 |
+
unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
|
| 292 |
+
total_numel += unpartitioned_numel
|
| 293 |
+
total_params += 1
|
| 294 |
+
|
| 295 |
+
if debug:
|
| 296 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
| 297 |
+
state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
|
| 298 |
+
offset += unpartitioned_numel
|
| 299 |
+
|
| 300 |
+
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
| 301 |
+
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
| 302 |
+
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
| 303 |
+
# live optimizer object, so we are checking that the numbers are within the right range
|
| 304 |
+
align_to = 2 * world_size
|
| 305 |
+
|
| 306 |
+
def zero2_align(x):
|
| 307 |
+
return align_to * math.ceil(x / align_to)
|
| 308 |
+
|
| 309 |
+
if debug:
|
| 310 |
+
print(f"original offset={offset}, avail_numel={avail_numel}")
|
| 311 |
+
|
| 312 |
+
offset = zero2_align(offset)
|
| 313 |
+
avail_numel = zero2_align(avail_numel)
|
| 314 |
+
|
| 315 |
+
if debug:
|
| 316 |
+
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
| 317 |
+
|
| 318 |
+
# Sanity check
|
| 319 |
+
if offset != avail_numel:
|
| 320 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
| 321 |
+
|
| 322 |
+
print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
|
| 323 |
+
|
| 324 |
+
|
| 325 |
+
def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
|
| 326 |
+
exclude_frozen_parameters):
|
| 327 |
+
state_dict = OrderedDict()
|
| 328 |
+
|
| 329 |
+
# buffers
|
| 330 |
+
buffers = zero_model_states[0].buffers
|
| 331 |
+
state_dict.update(buffers)
|
| 332 |
+
if debug:
|
| 333 |
+
print(f"added {len(buffers)} buffers")
|
| 334 |
+
|
| 335 |
+
if not exclude_frozen_parameters:
|
| 336 |
+
_zero2_merge_frozen_params(state_dict, zero_model_states)
|
| 337 |
+
|
| 338 |
+
_zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
| 339 |
+
|
| 340 |
+
# recover shared parameters
|
| 341 |
+
for pair in zero_model_states[0].shared_params:
|
| 342 |
+
if pair[1] in state_dict:
|
| 343 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
| 344 |
+
|
| 345 |
+
return state_dict
|
| 346 |
+
|
| 347 |
+
|
| 348 |
+
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
| 349 |
+
remainder = unpartitioned_numel % world_size
|
| 350 |
+
padding_numel = (world_size - remainder) if remainder else 0
|
| 351 |
+
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
| 352 |
+
return partitioned_numel, padding_numel
|
| 353 |
+
|
| 354 |
+
|
| 355 |
+
def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
|
| 356 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
| 357 |
+
return
|
| 358 |
+
|
| 359 |
+
if debug:
|
| 360 |
+
for i in range(world_size):
|
| 361 |
+
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
| 362 |
+
print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
| 363 |
+
|
| 364 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
| 365 |
+
wanted_params = len(frozen_param_shapes)
|
| 366 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
| 367 |
+
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
| 368 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
| 369 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
| 370 |
+
|
| 371 |
+
total_params = 0
|
| 372 |
+
total_numel = 0
|
| 373 |
+
for name, shape in zero_model_states[0].frozen_param_shapes.items():
|
| 374 |
+
total_params += 1
|
| 375 |
+
unpartitioned_numel = shape.numel()
|
| 376 |
+
total_numel += unpartitioned_numel
|
| 377 |
+
|
| 378 |
+
param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
|
| 379 |
+
state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
|
| 380 |
+
|
| 381 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
| 382 |
+
|
| 383 |
+
if debug:
|
| 384 |
+
print(
|
| 385 |
+
f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
| 386 |
+
)
|
| 387 |
+
|
| 388 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
| 389 |
+
|
| 390 |
+
|
| 391 |
+
class GatheredTensor:
|
| 392 |
+
"""
|
| 393 |
+
A pseudo tensor that collects partitioned weights.
|
| 394 |
+
It is more memory efficient when there are multiple groups.
|
| 395 |
+
"""
|
| 396 |
+
|
| 397 |
+
def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
|
| 398 |
+
self.flat_groups = flat_groups
|
| 399 |
+
self.flat_groups_offset = flat_groups_offset
|
| 400 |
+
self.offset = offset
|
| 401 |
+
self.partitioned_numel = partitioned_numel
|
| 402 |
+
self.shape = shape
|
| 403 |
+
self.dtype = self.flat_groups[0][0].dtype
|
| 404 |
+
|
| 405 |
+
def contiguous(self):
|
| 406 |
+
"""
|
| 407 |
+
Merge partitioned weights from flat_groups into a single tensor.
|
| 408 |
+
"""
|
| 409 |
+
end_idx = self.offset + self.partitioned_numel
|
| 410 |
+
world_size = len(self.flat_groups)
|
| 411 |
+
pad_flat_param_chunks = []
|
| 412 |
+
|
| 413 |
+
for rank_i in range(world_size):
|
| 414 |
+
# for each rank, we need to collect weights from related group/groups
|
| 415 |
+
flat_groups_at_rank_i = self.flat_groups[rank_i]
|
| 416 |
+
start_group_id = None
|
| 417 |
+
end_group_id = None
|
| 418 |
+
for group_id in range(len(self.flat_groups_offset)):
|
| 419 |
+
if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
|
| 420 |
+
start_group_id = group_id
|
| 421 |
+
if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
|
| 422 |
+
end_group_id = group_id
|
| 423 |
+
break
|
| 424 |
+
# collect weights from related group/groups
|
| 425 |
+
for group_id in range(start_group_id, end_group_id + 1):
|
| 426 |
+
flat_tensor = flat_groups_at_rank_i[group_id]
|
| 427 |
+
start_offset = self.offset - self.flat_groups_offset[group_id]
|
| 428 |
+
end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
|
| 429 |
+
pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
|
| 430 |
+
|
| 431 |
+
# collect weights from all ranks
|
| 432 |
+
pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
|
| 433 |
+
param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
|
| 434 |
+
return param
|
| 435 |
+
|
| 436 |
+
|
| 437 |
+
def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
| 438 |
+
param_shapes = zero_model_states[0].param_shapes
|
| 439 |
+
avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
|
| 440 |
+
|
| 441 |
+
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
| 442 |
+
# param, re-consolidating each param, while dealing with padding if any
|
| 443 |
+
|
| 444 |
+
# merge list of dicts, preserving order
|
| 445 |
+
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
| 446 |
+
|
| 447 |
+
if debug:
|
| 448 |
+
for i in range(world_size):
|
| 449 |
+
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
| 450 |
+
|
| 451 |
+
wanted_params = len(param_shapes)
|
| 452 |
+
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
| 453 |
+
# not asserting if there is a mismatch due to possible padding
|
| 454 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
| 455 |
+
print(f"Trainable params: Have {avail_numel} numels to process.")
|
| 456 |
+
print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
|
| 457 |
+
|
| 458 |
+
# params
|
| 459 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
| 460 |
+
# out-of-core computing solution
|
| 461 |
+
offset = 0
|
| 462 |
+
total_numel = 0
|
| 463 |
+
total_params = 0
|
| 464 |
+
flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
|
| 465 |
+
for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
|
| 466 |
+
unpartitioned_numel = shape.numel()
|
| 467 |
+
total_numel += unpartitioned_numel
|
| 468 |
+
total_params += 1
|
| 469 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
| 470 |
+
|
| 471 |
+
if debug:
|
| 472 |
+
print(
|
| 473 |
+
f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
| 474 |
+
)
|
| 475 |
+
|
| 476 |
+
# memory efficient tensor
|
| 477 |
+
tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
|
| 478 |
+
state_dict[name] = tensor
|
| 479 |
+
offset += partitioned_numel
|
| 480 |
+
|
| 481 |
+
offset *= world_size
|
| 482 |
+
|
| 483 |
+
# Sanity check
|
| 484 |
+
if offset != avail_numel:
|
| 485 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
| 486 |
+
|
| 487 |
+
print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
|
| 488 |
+
|
| 489 |
+
|
| 490 |
+
def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
|
| 491 |
+
exclude_frozen_parameters):
|
| 492 |
+
state_dict = OrderedDict()
|
| 493 |
+
|
| 494 |
+
# buffers
|
| 495 |
+
buffers = zero_model_states[0].buffers
|
| 496 |
+
state_dict.update(buffers)
|
| 497 |
+
if debug:
|
| 498 |
+
print(f"added {len(buffers)} buffers")
|
| 499 |
+
|
| 500 |
+
if not exclude_frozen_parameters:
|
| 501 |
+
_zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
|
| 502 |
+
|
| 503 |
+
_zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
| 504 |
+
|
| 505 |
+
# recover shared parameters
|
| 506 |
+
for pair in zero_model_states[0].shared_params:
|
| 507 |
+
if pair[1] in state_dict:
|
| 508 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
| 509 |
+
|
| 510 |
+
return state_dict
|
| 511 |
+
|
| 512 |
+
|
| 513 |
+
def to_torch_tensor(state_dict, return_empty_tensor=False):
|
| 514 |
+
"""
|
| 515 |
+
Convert state_dict of GatheredTensor to torch tensor
|
| 516 |
+
"""
|
| 517 |
+
torch_state_dict = {}
|
| 518 |
+
converted_tensors = {}
|
| 519 |
+
for name, tensor in state_dict.items():
|
| 520 |
+
tensor_id = id(tensor)
|
| 521 |
+
if tensor_id in converted_tensors: # shared tensors
|
| 522 |
+
shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
|
| 523 |
+
torch_state_dict[name] = shared_tensor
|
| 524 |
+
else:
|
| 525 |
+
converted_tensors[tensor_id] = name
|
| 526 |
+
if return_empty_tensor:
|
| 527 |
+
torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
|
| 528 |
+
else:
|
| 529 |
+
torch_state_dict[name] = tensor.contiguous()
|
| 530 |
+
return torch_state_dict
|
| 531 |
+
|
| 532 |
+
|
| 533 |
+
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
|
| 534 |
+
tag=None,
|
| 535 |
+
exclude_frozen_parameters=False,
|
| 536 |
+
lazy_mode=False):
|
| 537 |
+
"""
|
| 538 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
| 539 |
+
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
| 540 |
+
via a model hub.
|
| 541 |
+
|
| 542 |
+
Args:
|
| 543 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder
|
| 544 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
|
| 545 |
+
- ``exclude_frozen_parameters``: exclude frozen parameters
|
| 546 |
+
- ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
|
| 547 |
+
Convert the pesduo tensor to torch tensor by ``.contiguous()``
|
| 548 |
+
|
| 549 |
+
Returns:
|
| 550 |
+
- pytorch ``state_dict``
|
| 551 |
+
|
| 552 |
+
A typical usage might be ::
|
| 553 |
+
|
| 554 |
+
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
| 555 |
+
# do the training and checkpoint saving
|
| 556 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
| 557 |
+
model = model.cpu() # move to cpu
|
| 558 |
+
model.load_state_dict(state_dict)
|
| 559 |
+
# submit to model hub or save the model to share with others
|
| 560 |
+
|
| 561 |
+
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
| 562 |
+
application. i.e. you will need to re-initialize the deepspeed engine, since
|
| 563 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
| 564 |
+
|
| 565 |
+
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
| 566 |
+
|
| 567 |
+
Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
|
| 568 |
+
You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
| 569 |
+
the checkpoint. Or you can load state_dict in lazy mode ::
|
| 570 |
+
|
| 571 |
+
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
| 572 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
|
| 573 |
+
for name, lazy_tensor in state_dict.item():
|
| 574 |
+
tensor = lazy_tensor.contiguous() # to cpu
|
| 575 |
+
print(name, tensor)
|
| 576 |
+
# del tensor to release memory if it no longer in use
|
| 577 |
+
"""
|
| 578 |
+
if tag is None:
|
| 579 |
+
latest_path = os.path.join(checkpoint_dir, 'latest')
|
| 580 |
+
if os.path.isfile(latest_path):
|
| 581 |
+
with open(latest_path, 'r') as fd:
|
| 582 |
+
tag = fd.read().strip()
|
| 583 |
+
else:
|
| 584 |
+
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
| 585 |
+
|
| 586 |
+
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
| 587 |
+
|
| 588 |
+
if not os.path.isdir(ds_checkpoint_dir):
|
| 589 |
+
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
| 590 |
+
|
| 591 |
+
state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
|
| 592 |
+
if lazy_mode:
|
| 593 |
+
return state_dict
|
| 594 |
+
else:
|
| 595 |
+
return to_torch_tensor(state_dict)
|
| 596 |
+
|
| 597 |
+
|
| 598 |
+
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
|
| 599 |
+
output_dir,
|
| 600 |
+
max_shard_size="5GB",
|
| 601 |
+
safe_serialization=False,
|
| 602 |
+
tag=None,
|
| 603 |
+
exclude_frozen_parameters=False):
|
| 604 |
+
"""
|
| 605 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
| 606 |
+
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
| 607 |
+
|
| 608 |
+
Args:
|
| 609 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
| 610 |
+
- ``output_dir``: directory to the pytorch fp32 state_dict output files
|
| 611 |
+
- ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
|
| 612 |
+
- ``safe_serialization``: whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
|
| 613 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
|
| 614 |
+
- ``exclude_frozen_parameters``: exclude frozen parameters
|
| 615 |
+
"""
|
| 616 |
+
|
| 617 |
+
# Dependency pre-check
|
| 618 |
+
if safe_serialization:
|
| 619 |
+
try:
|
| 620 |
+
from safetensors.torch import save_file
|
| 621 |
+
except ImportError:
|
| 622 |
+
print('If you want to use `safe_serialization`, please `pip install safetensors`')
|
| 623 |
+
raise
|
| 624 |
+
if max_shard_size is not None:
|
| 625 |
+
try:
|
| 626 |
+
from huggingface_hub import split_torch_state_dict_into_shards
|
| 627 |
+
except ImportError:
|
| 628 |
+
print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
|
| 629 |
+
raise
|
| 630 |
+
|
| 631 |
+
# Convert zero checkpoint to state_dict
|
| 632 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
|
| 633 |
+
tag,
|
| 634 |
+
exclude_frozen_parameters,
|
| 635 |
+
lazy_mode=True)
|
| 636 |
+
|
| 637 |
+
# Shard the model if it is too big.
|
| 638 |
+
weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
|
| 639 |
+
if max_shard_size is not None:
|
| 640 |
+
filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
|
| 641 |
+
# an memory-efficient approach for sharding
|
| 642 |
+
empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
|
| 643 |
+
state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
|
| 644 |
+
filename_pattern=filename_pattern,
|
| 645 |
+
max_shard_size=max_shard_size)
|
| 646 |
+
else:
|
| 647 |
+
from collections import namedtuple
|
| 648 |
+
StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
|
| 649 |
+
state_dict_split = StateDictSplit(is_sharded=False,
|
| 650 |
+
filename_to_tensors={weights_name: list(state_dict.keys())})
|
| 651 |
+
|
| 652 |
+
# Save the model by shard
|
| 653 |
+
os.makedirs(output_dir, exist_ok=True)
|
| 654 |
+
filename_to_tensors = state_dict_split.filename_to_tensors.items()
|
| 655 |
+
for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
|
| 656 |
+
shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
|
| 657 |
+
shard_state_dict = to_torch_tensor(shard_state_dict)
|
| 658 |
+
output_path = os.path.join(output_dir, shard_file)
|
| 659 |
+
if safe_serialization:
|
| 660 |
+
save_file(shard_state_dict, output_path, metadata={"format": "pt"})
|
| 661 |
+
else:
|
| 662 |
+
torch.save(shard_state_dict, output_path)
|
| 663 |
+
# release the memory of current shard
|
| 664 |
+
for tensor_name in list(shard_state_dict.keys()):
|
| 665 |
+
del state_dict[tensor_name]
|
| 666 |
+
del shard_state_dict[tensor_name]
|
| 667 |
+
del shard_state_dict
|
| 668 |
+
gc.collect()
|
| 669 |
+
|
| 670 |
+
# Save index if sharded
|
| 671 |
+
if state_dict_split.is_sharded:
|
| 672 |
+
index = {
|
| 673 |
+
"metadata": state_dict_split.metadata,
|
| 674 |
+
"weight_map": state_dict_split.tensor_to_filename,
|
| 675 |
+
}
|
| 676 |
+
save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
|
| 677 |
+
save_index_file = os.path.join(output_dir, save_index_file)
|
| 678 |
+
with open(save_index_file, "w", encoding="utf-8") as f:
|
| 679 |
+
content = json.dumps(index, indent=2, sort_keys=True) + "\n"
|
| 680 |
+
f.write(content)
|
| 681 |
+
|
| 682 |
+
|
| 683 |
+
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
| 684 |
+
"""
|
| 685 |
+
1. Put the provided model to cpu
|
| 686 |
+
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
| 687 |
+
3. Load it into the provided model
|
| 688 |
+
|
| 689 |
+
Args:
|
| 690 |
+
- ``model``: the model object to update
|
| 691 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
| 692 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
|
| 693 |
+
|
| 694 |
+
Returns:
|
| 695 |
+
- ``model`: modified model
|
| 696 |
+
|
| 697 |
+
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
| 698 |
+
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
| 699 |
+
conveniently placed for you in the checkpoint folder.
|
| 700 |
+
|
| 701 |
+
A typical usage might be ::
|
| 702 |
+
|
| 703 |
+
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
| 704 |
+
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
| 705 |
+
# submit to model hub or save the model to share with others
|
| 706 |
+
|
| 707 |
+
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
| 708 |
+
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
| 709 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
| 710 |
+
|
| 711 |
+
"""
|
| 712 |
+
logger.info("Extracting fp32 weights")
|
| 713 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
| 714 |
+
|
| 715 |
+
logger.info("Overwriting model with fp32 weights")
|
| 716 |
+
model = model.cpu()
|
| 717 |
+
model.load_state_dict(state_dict, strict=False)
|
| 718 |
+
|
| 719 |
+
return model
|
| 720 |
+
|
| 721 |
+
|
| 722 |
+
if __name__ == "__main__":
|
| 723 |
+
parser = argparse.ArgumentParser()
|
| 724 |
+
parser.add_argument("checkpoint_dir",
|
| 725 |
+
type=str,
|
| 726 |
+
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
| 727 |
+
parser.add_argument("output_dir",
|
| 728 |
+
type=str,
|
| 729 |
+
help="directory to the pytorch fp32 state_dict output files"
|
| 730 |
+
"(e.g. path/checkpoint-12-output/)")
|
| 731 |
+
parser.add_argument(
|
| 732 |
+
"--max_shard_size",
|
| 733 |
+
type=str,
|
| 734 |
+
default="5GB",
|
| 735 |
+
help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
|
| 736 |
+
"lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
|
| 737 |
+
"We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
|
| 738 |
+
"without CPU OOM issues.")
|
| 739 |
+
parser.add_argument(
|
| 740 |
+
"--safe_serialization",
|
| 741 |
+
default=False,
|
| 742 |
+
action='store_true',
|
| 743 |
+
help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
|
| 744 |
+
parser.add_argument("-t",
|
| 745 |
+
"--tag",
|
| 746 |
+
type=str,
|
| 747 |
+
default=None,
|
| 748 |
+
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
|
| 749 |
+
parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
|
| 750 |
+
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
| 751 |
+
args = parser.parse_args()
|
| 752 |
+
|
| 753 |
+
debug = args.debug
|
| 754 |
+
|
| 755 |
+
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
|
| 756 |
+
args.output_dir,
|
| 757 |
+
max_shard_size=args.max_shard_size,
|
| 758 |
+
safe_serialization=args.safe_serialization,
|
| 759 |
+
tag=args.tag,
|
| 760 |
+
exclude_frozen_parameters=args.exclude_frozen_parameters)
|