DeepSeek-R1-Distill-Qwen-14B · AWQ-INT4
This page aggregates 2 real-world runs of DeepSeek-R1-Distill-Qwen-14B (AWQ-INT4) on NVIDIA RTX 4090 with vLLM, contributed by 2 independent source platforms; metrics are averages of published measurements.
63.6 tok/s
Decode
Decode speed
1,192.75 tok/s
Prefill
Prefill speed
0.29 s
TTFT
Time to first token
11.1 GB
VRAM
VRAM usage
2
Measured runs
2
Independent sources
GitHub, Hugging Face
Source platforms
17 days ago
Last verified
Performance
- vLLM · AWQ-INT4 (current)Decode 63.6 · Prefill 1192.75 ·
Core figures
- Decode (avg)
- 63.6 tok/s
- Prefill (avg)
- 1,192.75 tok/s
- TTFT (avg)
- 0.29 s
- VRAM (avg)
- 11.1 GB
- MTP acceptance rate
- —
- TTFB
- — GB
- Power draw
- 357.5 W
Configuration
Member-level fields are taken from the most recent run
- Model
- DeepSeek-R1-Distill-Qwen-14B
- Quantization
- AWQ-INT4
- Framework
- vLLM
- Version
- 0.7.3
- Context length
- 32768 tokens
- Batch size
- 32
Hardware
Nominal and measured figures are shown side by side; whether it runs is the reader's call
- GPU
- NVIDIA RTX 4090
- Nominal VRAM
- 24 GB
- Measured VRAM (avg)
- 11.1 GB
- OS
- Ubuntu 24.04
- Driver
- 550.107.02
- CUDA
- 12.4
- Power draw
- 356.8 W
Sources & evidence
2 measured records in total, each traceable to its original source
0.7.3 · Ubuntu 24.04 · CUDA 12.4 · 32768 ctx
62.3 tok/s
Decode
1,180.5 tok/s
Prefill
0.28 s
TTFT
11.2 GB
VRAM
—
MTP
356.8 W
Power draw
vLLM AWQ-INT4 批量 32 下 62 tok/s，两条链路互验。
0.7.3 · Ubuntu 24.04 · CUDA 12.4 · 32768 ctx
64.9 tok/s
Decode
1,205 tok/s
Prefill
0.31 s
TTFT
11 GB
VRAM
—
MTP
358.2 W
Power draw
批量 64 时 64.9 tok/s。