# Technical Document Extraction: Image Analysis
## Image Description
The image is a composite bar chart with three subplots labeled **(a)**, **(b)**, and **(c)**, comparing the performance of three models across varying batch sizes. Each subplot includes:
- **Left y-axis**: Speedup (logarithmic scale, 0–3).
- **Right y-axis**: Throughput (linear scale, 0–600 for subplots (a) and (c), 0–300 for subplot (b)).
- **X-axis**: Batch sizes (128, 1k, 2k, 8k, 128, 1k, 4k, 128, 1k, 2k, 128, 1k, 4k).
- **Legend**:
- Gray bars: **HuggingFace (PyTorch)**.
- Blue bars: **Ours**.
- Red diamonds: **Ours (token/s)**.
---
## Subplot (a): Llama2-7B
### Labels & Axes
- **X-axis**: Batch sizes (128, 1k, 2k, 8k, 128, 1k, 4k, 128, 1k, 2k, 128, 1k, 4k).
- **Left y-axis**: Speedup (0–3).
- **Right y-axis**: Throughput (0–600).
### Data Points & Trends
| Batch Size | HuggingFace (PyTorch) | Ours | Ours (token/s) |
|------------|-----------------------|------|----------------|
| 128 | ~1.0 | ~2.0 | ~0.5 |
| 1k | ~1.0 | ~2.0 | ~0.5 |
| 2k | ~1.0 | ~2.5 | ~0.6 |
| 8k | ~1.0 | ~3.0 | ~0.7 |
| 128 | ~1.0 | ~1.5 | ~0.8 |
| 1k | ~1.0 | ~2.0 | ~0.9 |
| 4k | ~1.0 | ~2.5 | ~1.0 |
| 128 | ~1.0 | ~1.8 | ~1.2 |
| 1k | ~1.0 | ~2.2 | ~1.5 |
| 2k | ~1.0 | ~2.8 | ~2.0 |
| 4k | ~1.0 | ~3.5 | ~2.5 |
**Trends**:
- **HuggingFace (gray)**: Speedup remains constant (~1.0) across all batch sizes.
- **Ours (blue)**: Speedup increases with batch size (e.g., 2.0 at 128 → 3.5 at 4k).
- **Ours (token/s) (red)**: Speedup increases with batch size but lags behind "Ours" (e.g., 0.5 at 128 → 2.5 at 4k).
- **Throughput**: Scales linearly with batch size (e.g., 200 at 128 → 800 at 8k).
---
## Subplot (b): Llama2-13B
### Labels & Axes
- **X-axis**: Batch sizes (128, 1k, 2k, 8k, 128, 1k, 4k, 128, 1k, 2k, 128, 1k, 4k).
- **Left y-axis**: Speedup (0–3).
- **Right y-axis**: Throughput (0–300).
### Data Points & Trends
| Batch Size | HuggingFace (PyTorch) | Ours | Ours (token/s) |
|------------|-----------------------|------|----------------|
| 128 | ~1.0 | ~2.0 | ~0.5 |
| 1k | ~1.0 | ~2.0 | ~0.5 |
| 2k | ~1.0 | ~2.5 | ~0.6 |
| 8k | ~1.0 | ~3.0 | ~0.7 |
| 128 | ~1.0 | ~1.5 | ~0.8 |
| 1k | ~1.0 | ~2.0 | ~0.9 |
| 4k | ~1.0 | ~2.5 | ~1.0 |
**Trends**:
- **HuggingFace (gray)**: Speedup remains constant (~1.0).
- **Ours (blue)**: Speedup increases with batch size (e.g., 2.0 at 128 → 2.5 at 4k).
- **Ours (token/s) (red)**: Speedup increases with batch size but lags behind "Ours".
- **Throughput**: Scales linearly with batch size (e.g., 200 at 128 → 300 at 4k).
---
## Subplot (c): OPT-6.7B
### Labels & Axes
- **X-axis**: Batch sizes (128, 1k, 2k, 8k, 128, 1k, 4k, 128, 1k, 2k, 128, 1k, 4k).
- **Left y-axis**: Speedup (0–4).
- **Right y-axis**: Throughput (0–600).
### Data Points & Trends
| Batch Size | HuggingFace (PyTorch) | Ours | Ours (token/s) |
|------------|-----------------------|------|----------------|
| 128 | ~1.0 | ~2.0 | ~0.5 |
| 1k | ~1.0 | ~2.0 | ~0.5 |
| 2k | ~1.0 | ~2.5 | ~0.6 |
| 8k | ~1.0 | ~3.0 | ~0.7 |
| 128 | ~1.0 | ~1.5 | ~0.8 |
| 1k | ~1.0 | ~2.0 | ~0.9 |
| 4k | ~1.0 | ~2.5 | ~1.0 |
**Trends**:
- **HuggingFace (gray)**: Speedup remains constant (~1.0).
- **Ours (blue)**: Speedup increases with batch size (e.g., 2.0 at 128 → 2.5 at 4k).
- **Ours (token/s) (red)**: Speedup increases with batch size but lags behind "Ours".
- **Throughput**: Scales linearly with batch size (e.g., 200 at 128 → 600 at 8k).
---
## Legend & Spatial Grounding
- **Legend Position**: Top-center of the image.
- **Color Mapping**:
- Gray: HuggingFace (PyTorch).
- Blue: Ours.
- Red: Ours (token/s).
---
## Key Observations
1. **Speedup**:
- "Ours" consistently outperforms HuggingFace across all batch sizes.
- "Ours (token/s)" shows lower speedup than "Ours" but scales with batch size.
2. **Throughput**:
- Increases linearly with batch size for all models.
- Subplot (a) and (c) have higher throughput ranges (600) compared to subplot (b) (300).
---
## Notes
- The x-axis labels repeat across subplots, likely indicating separate configurations or datasets.
- No textual data tables or non-English content are present.
