Upload correctness_model/docs/rfcs/RFC0008-Base-Model-GLM-Edge.md with huggingface_hub
Browse files
correctness_model/docs/rfcs/RFC0008-Base-Model-GLM-Edge.md
ADDED
|
@@ -0,0 +1,115 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
author: Correctness Model Team
|
| 3 |
+
created: 2026-08-03
|
| 4 |
+
modified: 2026-08-04
|
| 5 |
+
status: implemented
|
| 6 |
+
---
|
| 7 |
+
|
| 8 |
+
# RFC0008: Base Model Selection - GLM-Edge 1.5B Chat
|
| 9 |
+
|
| 10 |
+
## Overview
|
| 11 |
+
|
| 12 |
+
This RFC selects the base model for the correctness-focused training
|
| 13 |
+
architecture (RFC0007). The base model must be:
|
| 14 |
+
1. Small enough to train on 16GB gfx1030 (RX 6900 XT)
|
| 15 |
+
2. Compatible with PyTorch (chosen framework)
|
| 16 |
+
3. Open-source with permissive license
|
| 17 |
+
4. A good foundation for the verifier/feedback mechanisms
|
| 18 |
+
|
| 19 |
+
## Decision: GLM-Edge 1.5B Chat
|
| 20 |
+
|
| 21 |
+
**GLM-Edge 1.5B Chat** is selected as the base model.
|
| 22 |
+
|
| 23 |
+
### Why GLM-Edge 1.5B Chat
|
| 24 |
+
|
| 25 |
+
- **Size**: ~1.5B params (2B total with embeddings) - trainable on 16GB gfx1030
|
| 26 |
+
- **Edge-focused**: designed for mobile/car/PC platforms - aligns with our
|
| 27 |
+
small-hardware goal
|
| 28 |
+
- **PyTorch support**: native transformers integration
|
| 29 |
+
- **License**: Apache 2.0 (code), GLM-4 model license (weights)
|
| 30 |
+
- **Proven efficiency**: 60+ tokens/s on edge NPUs, INT4 quantization support
|
| 31 |
+
- **Fine-tuning support**: official LoRA/SFT fine-tuning scripts provided
|
| 32 |
+
- **BF16 native**: uses bfloat16 (matches our precision strategy)
|
| 33 |
+
|
| 34 |
+
### Model Details
|
| 35 |
+
|
| 36 |
+
- **Architecture**: GLM-4 based (dense, not MoE)
|
| 37 |
+
- **Precision**: BF16 native
|
| 38 |
+
- **Context**: designed for edge deployment
|
| 39 |
+
- **Family**: GLM-Edge-1.5B-Chat, GLM-Edge-4B-Chat, GLM-Edge-V-2B/5B (vision)
|
| 40 |
+
|
| 41 |
+
## Alternatives Considered
|
| 42 |
+
|
| 43 |
+
### 1. GLM-Edge 4B Chat
|
| 44 |
+
- **Pros**: More capable
|
| 45 |
+
- **Cons**: Larger, harder to train on 16GB
|
| 46 |
+
- **Verdict**: Deferred - use 1.5B for initial experiments
|
| 47 |
+
|
| 48 |
+
### 2. Llama 3.2 1B / 3B
|
| 49 |
+
- **Pros**: Well-supported, standard architecture
|
| 50 |
+
- **Cons**: Not edge-optimized, no official edge fine-tuning focus
|
| 51 |
+
- **Verdict**: Viable alternative, but GLM-Edge is more aligned with our
|
| 52 |
+
small-hardware goal
|
| 53 |
+
|
| 54 |
+
### 3. Qwen2.5 0.5B / 1.5B
|
| 55 |
+
- **Pros**: Small, well-supported
|
| 56 |
+
- **Cons**: Less edge-focused than GLM-Edge
|
| 57 |
+
- **Verdict**: Viable alternative
|
| 58 |
+
|
| 59 |
+
### 4. Custom small model from scratch
|
| 60 |
+
- **Pros**: Full control
|
| 61 |
+
- **Cons**: Requires training from scratch (expensive), no pretrained weights
|
| 62 |
+
- **Verdict**: Deferred - use pretrained GLM-Edge as base, fine-tune
|
| 63 |
+
|
| 64 |
+
## Integration with Correctness Architecture
|
| 65 |
+
|
| 66 |
+
### Feature Extraction
|
| 67 |
+
|
| 68 |
+
GLM-Edge uses standard GLM-4 attention. To extract verifier features
|
| 69 |
+
(RFC0002), we hook into the attention module:
|
| 70 |
+
- Attention weights (per head)
|
| 71 |
+
- Output statistics
|
| 72 |
+
- Softmax entropy
|
| 73 |
+
- Cross-head consistency
|
| 74 |
+
|
| 75 |
+
### Training Stages (RFC0007)
|
| 76 |
+
|
| 77 |
+
1. **base**: GLM-Edge already pretrained - use as-is or fine-tune
|
| 78 |
+
2. **verifier**: train verifier on GLM-Edge internal states + correctness labels
|
| 79 |
+
3. **feedback**: train feedback policy
|
| 80 |
+
4. **sparse**: learn sparse attention mask
|
| 81 |
+
|
| 82 |
+
### Hardware Fit (gfx1030, 16GB)
|
| 83 |
+
|
| 84 |
+
- 1.5B params in BF16 = ~3GB weights
|
| 85 |
+
- Leaves ~13GB for activations, KV cache, optimizer states
|
| 86 |
+
- Feasible for fine-tuning with LoRA or full fine-tuning with gradient
|
| 87 |
+
checkpointing
|
| 88 |
+
- Our optimizations (block-sparse RFC0003, latent compression RFC0005,
|
| 89 |
+
FP8 RFC0006) further reduce memory
|
| 90 |
+
|
| 91 |
+
## Prerequisites
|
| 92 |
+
|
| 93 |
+
1. **PyTorch with ROCm support** - not yet installed (see RFC0009)
|
| 94 |
+
2. **transformers** library (from source, per GLM-Edge docs)
|
| 95 |
+
3. **Model weights** - download from HuggingFace
|
| 96 |
+
4. **GLM-Edge repo** - for fine-tuning scripts
|
| 97 |
+
|
| 98 |
+
## Regression Plan
|
| 99 |
+
|
| 100 |
+
- GLM-Edge 1.5B Chat loads and runs in PyTorch on gfx1030
|
| 101 |
+
- Feature extraction captures internal states correctly
|
| 102 |
+
- Verifier training converges
|
| 103 |
+
- Correctness rate improves vs baseline
|
| 104 |
+
|
| 105 |
+
## Communication
|
| 106 |
+
|
| 107 |
+
- Depends on RFC0007 (training architecture)
|
| 108 |
+
- Requires RFC0009 (PyTorch ROCm setup)
|
| 109 |
+
- Builds on RFC0001-0006 (kernel, sparse, latent, FP8)
|
| 110 |
+
|
| 111 |
+
## References
|
| 112 |
+
|
| 113 |
+
- GLM-Edge repo: github.com/zai-org/GLM-Edge
|
| 114 |
+
- GLM-Edge 1.5B Chat: huggingface.co/zai-org/glm-edge-1.5b-chat
|
| 115 |
+
- License: Apache 2.0 (code), GLM-4 (weights)
|