Skip to content

Update README with token match rate on text backbone #52

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
sdeeptan-aws wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Update README with token match rate on text backbone #52

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
41a2fe7
Update README with token match rate on text backbone
sdeeptan-aws Feb 19, 2026
181bc7c
Removing internal names
sdeeptan-aws Feb 26, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 11 additions & 4 deletions contrib/models/Qwen2.5-VL-32B-Instruct/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,9 @@ NeuronX Distributed Inference implementation of Qwen2.5 VL 32B Instruct.

## Architecture Details

- **Layers:** Check model config
- **Type:** Multimodal (vision-language) model — text backbone validated only
- **Text Backbone:** Decoder-only transformer (Qwen2-based)
- **Layers:** 64
- **Hidden Size:** Check model config
- **Attention Heads:** Check model config
- **Vocabulary:** Check model config
Expand All @@ -28,7 +30,7 @@ NeuronX Distributed Inference implementation of Qwen2.5 VL 32B Instruct.
| Test | Status | Result |
|------|--------|--------|
| Smoke Test | ✅ PASS | Model loads successfully |
| Token Matching | ⚠️ N/A | **0.0% match** |
| Token Matching | ✅ PASS | **100% match** (text backbone) |
| TTFT (P50) | ✅ PASS | 7.98ms (threshold: 100ms) |
| Throughput | ✅ PASS | 120.65 tok/s (threshold: 10 tok/s) |

Expand All @@ -39,9 +41,14 @@ NeuronX Distributed Inference implementation of Qwen2.5 VL 32B Instruct.
| TTFT (P50) | 7.98ms |
| Throughput | 120.65 tokens/s |


**Status:** ✅ VALIDATED

### Multimodal Validation Notes

Qwen2.5-VL is a vision-language model. The NeuronX port validates the text backbone only. `AutoModelForCausalLM` does not work for VLMs — the specific text backbone class (`Qwen2ForCausalLM`) must be used to load the HF reference for token matching. With the correct text backbone extraction, the model achieves 100% token match.

**Important:** Ensure the compiled model uses the full 64 layers. Test builds with reduced layer counts (e.g., 4 layers) will produce poor accuracy. Always verify `num_hidden_layers` in the compiled `config.json` before validation.

## Usage

```python
Expand Down Expand Up @@ -106,6 +113,6 @@ python3 test/integration/test_model.py

## Maintainer

Neuroboros Team - Annapurna Labs
Annapurna Labs

**Last Updated:** 2026-01-29