Qwen3.8-27B + MTP on 2× RTX 3090 NVLink — 4K to 260K context scaling in llama.cpp
retained text
**Hey homelab folks!**
I’ve been messing around with my dual RTX 3090 setup again, and this time I wanted to see how Qwen3.8-27B behaves as the context gets *really* long.
So I ran the same benchmark from 4K all the way up to 260K context using two RTX 3090 24GB cards connected with NVLink, llama.cpp tensor split, and MTP.
The result was pretty interesting: decode speed went from **116.36 tok/s at 4K** to **53.33 tok/s at 260K** — and yes, the full **260K prompt + 1,000 generated tokens** actually completed on 2×24GB cards. 😀
**Test setup:**
• 2× RTX 3090 24GB (NVLink connected, CUDA P2P enabled)
• No power limit applied — approximately 350 W per GPU / 700 W combined under load.
• Threadripper PRO 3945WX
• 128GB RAM
• Ubuntu 22.04
• llama.cpp b10935 (commit 8e330954a)
• Qwen3.8-27B-UD-Q4_K_M
• --split-mode tensor
• Q8_0 K/V cache
• Flash Attention enabled
• MTP n_max=2
I used one fixed non-repetitive master corpus for all tests. Each context length uses the first N tokens from the same corpus, followed by 1,000 generated tokens.
I restarted llama-server for every run and waited until both GPUs cooled down to ≤50°C before starting the next one. Each point is a single run, so these aren't averaged results.
I also included MTP acceptance and mean draft length because acceptance varied quite a bit depending on the position in the corpus.
**RESULT:**
prompt : 260000 tokens
prefill: 899.47 tok/s
decode : 53.33 tok/s
MTP : 95.91% (656/684), mean 2.92
prompt : 196608 tokens
prefill: 1026.23 tok/s
decode : 62.23 tok/s
MTP : 97.63% (660/676), mean 2.95
prompt : 131072 tokens
prefill: 1228.07 tok/s
decode : 73.89 tok/s
MTP : 95.77% (656/685), mean 2.91
prompt : 65536 tokens
prefill: 1516.61 tok/s
decode : 86.28 tok/s
MTP : 86.34% (632/732), mean 2.73
prompt : 32768 tokens
prefill: 1696.19 tok/s
decode : 92.85 tok/s
MTP : 77.75% (608/782), mean 2.55
prompt : 4096 tokens
prefill: 1673.51 tok/s
decode : 116.36 tok/s
MTP : 91.23% (645/707), mean 2.82
context | prefill tok/s | decode tok/s | MTP acceptance
260k | 899.47 | 53.33 | 95.91%
196k | 1026.23 | 62.23 | 97.63%
128k | 1228.07 | 73.89 | 95.77%
64k | 1516.61 | 86.28 | 86.34%
32k | 1696.19 | 92.85 | 77.75%
4k | 1673.51 | 116.36 | 91.23%
Next up, I’m planning to run the **exact same test on 2× Tesla V100 PCIe 32GB**.
I’m especially curious to see how the V100s behave at very long context lengths compared with the 3090s.
If anyone has similar results with 3090s, V100s, or other multi-GPU setups, I’d love to compare numbers!
**llama-server command used for the benchmark:**
/path/to/llama.cpp/build/bin/llama-server \
-m /path/to/Qwen3.8-27B-UD-Q4_K_M.gguf \
--host 0.0.0.0 --port 8080 \
--device CUDA0,CUDA1 \
--split-mode tensor \
-ngl all \
-fa on \
--jinja \
-c 262144 \
--parallel 1 \
-t 8 \
--spec-type draft-mtp \
--spec-draft-n-max 2 \
--reasoning-effort low \
--reasoning-preserve \
--cache-type-k q8_0 \
--cache-type-v q8_0 \
--load-mode mlock \
--spec-draft-device CUDA0,CUDA1
**Update — Tesla V100 comparison:**
I’ve now completed the exact same benchmark on **2× Tesla V100 PCIe 32GB**.
Results here:
[2× Tesla V100 benchmark](https://www.reddit.com/r/homelab/comments/1wg4ncr/qwen3827b_mtp_on_2_tesla_v100_32gb_4k_to_260k/)document fetched 2026-09-15 00:38 · canonical https://reddit.com/r/homelab/comments/1wg126r/qwen3827b_mtp_on_2_rtx_3090_nvlink_4k_to_260k