---
license: mit
base_model: zai-org/GLM-4.7-Flash
library_name: peft
pipeline_tag: text-generation
datasets:
  - TokenBender/glm47-pie-cpp-posttraining-data
tags:
  - glm
  - lora
  - sft
  - code
  - code-optimization
  - h100
---

# GLM-4.7-Flash PIE C++ SFT LoRA

LoRA rank-16 supervised fine-tuning adapter for
[`zai-org/GLM-4.7-Flash`](https://huggingface.co/zai-org/GLM-4.7-Flash),
trained on the PIE C++ performance task.

## Result

The full 1,259-task evaluation produced:

| Metric | Result |
| --- | ---: |
| Pass rate | 90.79% |
| Valid format rate | 97.70% |
| Correct and faster rate | 28.36% |
| Mean speedup when correct and faster | 1.43x |
| Mean reward | 0.8980 |
| Timeout rate | 0.00% |

The complete per-task records and generated responses are under
`evidence/eval/`.

## Training Profile

| Setting | Value |
| --- | --- |
| Accelerators | 8x NVIDIA H100 |
| Parallelism | TP4 / PP1 / EP8 |
| Sequence length | 4,096 |
| Max tokens per GPU | 24,576 |
| Global batch size | 32 |
| LoRA rank / alpha | 16 / 32 |
| Peak allocated memory | 36,039 MiB per GPU |
| Checkpoint | Step 244 |

## Files

- `adapter_model.bin` and `adapter_config.json`: loadable PEFT adapter.
- `adapter_megatron_tp*_pp0.pt`: four Megatron tensor-parallel shards.
- `training_state_rank*.pt`: per-rank training state.
- `evidence/training/`: run receipt and VRAM trace.
- `evidence/eval/`: complete evaluation summaries, records, and generations.

## Loading

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained(
    "zai-org/GLM-4.7-Flash",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(
    base,
    "TokenBender/glm47-flash-pie-cpp-lora-r16-sft-h100",
)
```

Training code:
[TokenBender/browser-is-all-you-need](https://github.com/tokenbender/browser-is-all-you-need/tree/client/glm47-h100-posttraining)