Add model card with benchmark results
Browse files
README.md
CHANGED
|
@@ -4,15 +4,64 @@ tags:
|
|
| 4 |
- mlx
|
| 5 |
- oq
|
| 6 |
- quantized
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
---
|
| 8 |
|
| 9 |
# KAT-Coder-V2.5-Dev-oQ2
|
| 10 |
|
| 11 |
This model was quantized using [oQ](https://github.com/jundot/omlx) (oMLX v0.5.4) mixed-precision quantization.
|
| 12 |
|
|
|
|
|
|
|
|
|
|
| 13 |
## Quantization details
|
| 14 |
|
| 15 |
- **Model type**: qwen3_5_moe
|
| 16 |
- **Bits**: 2
|
| 17 |
- **Group size**: 64
|
| 18 |
- **Format**: MLX safetensors
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
- mlx
|
| 5 |
- oq
|
| 6 |
- quantized
|
| 7 |
+
- benchmark
|
| 8 |
+
- performance
|
| 9 |
+
- moe
|
| 10 |
+
- code
|
| 11 |
+
- agentic-coding
|
| 12 |
---
|
| 13 |
|
| 14 |
# KAT-Coder-V2.5-Dev-oQ2
|
| 15 |
|
| 16 |
This model was quantized using [oQ](https://github.com/jundot/omlx) (oMLX v0.5.4) mixed-precision quantization.
|
| 17 |
|
| 18 |
+
- Base model: [Kwaipilot/KAT-Coder-V2.5-Dev](https://huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev)
|
| 19 |
+
- Chat template: [Qwen Fixed Chat Template](https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates) (v21) — fixes KV cache optimization, thinking mode control, tool calling, and agent loop issues. Original backed up as `chat_template.jinja.bak`.
|
| 20 |
+
|
| 21 |
## Quantization details
|
| 22 |
|
| 23 |
- **Model type**: qwen3_5_moe
|
| 24 |
- **Bits**: 2
|
| 25 |
- **Group size**: 64
|
| 26 |
- **Format**: MLX safetensors
|
| 27 |
+
- **Calibration**: oQ2 (standard sensitivity-based)
|
| 28 |
+
|
| 29 |
+
## Environment
|
| 30 |
+
|
| 31 |
+
- **Hardware**: M5 MacBook Air 32GB
|
| 32 |
+
- **Inference Framework**: oMLX v0.5.4
|
| 33 |
+
- **Max Concurrent Requests**: 4
|
| 34 |
+
- **Settings**:
|
| 35 |
+
- Thinking: Disabled
|
| 36 |
+
- TurboQuant KV Cache: Enabled (4-bit)
|
| 37 |
+
|
| 38 |
+
## Performance Benchmarks
|
| 39 |
+
|
| 40 |
+
> **Note**: Results are for reference only and may vary depending on hardware, software configuration, and workload.
|
| 41 |
+
|
| 42 |
+
### Single Request Results
|
| 43 |
+
|
| 44 |
+
| Test | TTFT(ms) | TPOT(ms) | pp TPS | tg TPS | E2E(s) | Throughput | Peak Mem |
|
| 45 |
+
|------|----------|----------|--------|--------|--------|------------|----------|
|
| 46 |
+
| pp1024/tg128 | 1097.4 | 20.22 | 933.1 tok/s | 49.8 tok/s | 3.678 | 313.2 tok/s | 12.62 GB |
|
| 47 |
+
| pp4096/tg128 | 3777.7 | 21.17 | 1084.2 tok/s | 47.6 tok/s | 6.485 | 651.3 tok/s | 13.34 GB |
|
| 48 |
+
|
| 49 |
+
### Continuous Batching (pp1024 / tg128)
|
| 50 |
+
|
| 51 |
+
| Batch | tg TPS | Speedup | pp TPS | pp TPS/req | TTFT(ms) | E2E(s) |
|
| 52 |
+
|-------|--------|---------|--------|------------|----------|--------|
|
| 53 |
+
| 1x | 49.8 tok/s | 1.00x | 933.1 tok/s | 933.1 tok/s | 1097.4 | 3.678 |
|
| 54 |
+
| 2x | 67.5 tok/s | 1.36x | 833.0 tok/s | 416.5 tok/s | 2458.2 | 6.252 |
|
| 55 |
+
| 4x | 98.0 tok/s | 1.97x | 825.7 tok/s | 206.4 tok/s | 4812.4 | 10.184 |
|
| 56 |
+
|
| 57 |
+
## Intelligence Benchmark
|
| 58 |
+
|
| 59 |
+
> **Note**: Each benchmark round tests only 30 questions. Results are for reference only.
|
| 60 |
+
|
| 61 |
+
| Benchmark | Accuracy | Correct | Total | Time(s) | Think |
|
| 62 |
+
|-----------|----------|---------|-------|---------|-------|
|
| 63 |
+
| MMLU | 66.7% | 20 | 30 | 36.2 | No |
|
| 64 |
+
| TRUTHFULQA | 83.3% | 25 | 30 | 15.3 | No |
|
| 65 |
+
| GSM8K | 90.0% | 27 | 30 | 78.1 | No |
|
| 66 |
+
| MATHQA | 46.7% | 14 | 30 | 46.2 | No |
|
| 67 |
+
| HUMANEVAL | 83.3% | 25 | 30 | 113.7 | No |
|