Saved in this browser, used on every model page.
Yes
Yes. The RTX A4000 runs Cola2 fully in VRAM at Q8_0 (about 11 GB of 16 GB), at roughly 26 tokens/s.
- best quant
- Q8_0
- needs
- 11 GB
- usable
- 16 GB
- est. speed
- ~26 tok/s
Estimates for an 8K context; offload assumes 32 GB of system RAM.Try other contexts on the Cola2 page →
Every quantisation of Cola2 on the RTX A4000
| Quant | Memory | Fit | Speed |
|---|---|---|---|
| Q2_K | 4.4 GB | fits | ~85 tok/s |
| Q3_K_M | 5.8 GB | fits | ~57 tok/s |
| Q4_K_M | 6.8 GB | fits | ~46 tok/s |
| Q5_K_M | 7.7 GB | fits | ~40 tok/s |
| Q6_K | 8.6 GB | fits | ~34 tok/s |
| Q8_0 pick | 11 GB | fits | ~26 tok/s |
| FP16 / BF16 | 19 GB | offload | ~7.2 tok/s |
Speed is a bandwidth-based estimate for generating text (448 GB/s); prompt processing and long contexts are slower.
Measured speed on the RTX A4000
No community reports for this pair yet.
Report your own speed →Run Cola2 on the RTX A4000
Transformers · Chat with the pipeline APIdocs
from transformers import pipeline
pipe = pipeline("text-generation", model="ByteDance-Seed/Cola2", dtype="auto", device_map="auto")
messages = [{"role": "user", "content": "Explain self-hosting in one paragraph."}]
out = pipe(messages, max_new_tokens=512)
print(out[0]["generated_text"][-1]["content"])Chat models that fit the RTX A4000 comfortably
All that fit →| Model | Params | Best quant | Memory | Speed |
|---|---|---|---|---|
| 8.2B | Q8_0 | ≈10 GBFits in memory | ~28 tok/s | |
| 8B | Q8_0 | ≈9.8 GBFits in memory | ~29 tok/s | |
| 7.6B | Q8_0 | ≈8.9 GBFits in memory | ~30 tok/s | |
| 8B | Q8_0 | ≈9.8 GBFits in memory | ~29 tok/s | |
| 9.2B | Q8_0 | ≈13 GBFits in memory | ~25 tok/s | |
| 9.7B | Q8_0 | ≈11 GBFits in memory | ~24 tok/s |
Other GPUs that run Cola2
Frequently asked questions
- Can the RTX A4000 run Cola2?
- Yes. The RTX A4000 runs Cola2 fully in VRAM at Q8_0 (about 11 GB of 16 GB), at roughly 26 tokens/s.
- How fast does Cola2 run on the RTX A4000?
- An estimate from memory bandwidth gives about 26 tokens/s at Q8_0.
- Which quantisation of Cola2 should I use on the RTX A4000?
- Q8_0: the largest one that fits. Expected quality: lossless compared with the full-precision weights.
Memory and speed are estimates from the model's size and the card's memory bandwidth; real numbers depend on the runtime, drivers and context length. Community figures are medians of user reports.