Significant performance drop in concurrent requests with 30B model vs 4B model on 2x4090

[tylike]

Note: This issue is drafted with Claude AI's assistance for clear English expression.

Setup

• Hardware: 2x RTX 4090
• Command:

llama-server -m MODEL.gguf -np 2 -c 16384 -ngl 99 -ts 1,1,0,0 --host 0.0.0.0 --port 3721 -t 8 -tb 8

Issue

Using identical configuration for both models:

• Qwen3-4B-bf16: 60→58 tokens/s (3.3% drop) with 2 concurrent requests
• Qwen3-30B-A3B-Q8 model: 80→7 tokens/s (86% drop) with 2 concurrent requests

Questions

1. Is there potential for better performance with the 30B model under concurrent requests?
2. Which parameters should I adjust to improve concurrent processing speed?

Test Method

Simple browser-based concurrent API requests through two windows.

[ggerganov]

Try:

LLAMA_SET_ROWS=1 llama-server -m MODEL.gguf -np 2 -c 16384 -ngl 99 --host 0.0.0.0 --port 3721 -t 8 -tb 8 -fa

[slaren]

The performance of small batch sizes with MoE models is not very good in the CUDA backend. There is a fast gemv implementation, but it only works with bs=1.

  MUL_MAT_ID(type_a=q4_0,type_b=f32,n_mats=256,n_used=8,b=1,m=2048,n=1,k=7168):                14058 runs -    73.25 us/run - 234.88 MFLOP/run -   3.21 TFLOPS
  MUL_MAT_ID(type_a=q4_0,type_b=f32,n_mats=256,n_used=8,b=1,m=2048,n=2,k=7168):                 2343 runs -   432.78 us/run - 469.76 MFLOP/run -   1.09 TFLOPS
  MUL_MAT_ID(type_a=q4_0,type_b=f32,n_mats=256,n_used=8,b=1,m=2048,n=3,k=7168):                 1420 runs -   711.23 us/run - 704.64 MFLOP/run - 990.74 GFLOPS
  MUL_MAT_ID(type_a=q4_0,type_b=f32,n_mats=256,n_used=8,b=1,m=2048,n=4,k=7168):                 1177 runs -   906.02 us/run - 939.52 MFLOP/run -   1.04 TFLOPS
  MUL_MAT_ID(type_a=q4_0,type_b=f32,n_mats=256,n_used=8,b=1,m=2048,n=8,k=7168):                  594 runs -  1686.22 us/run -   1.88 GFLOP/run -   1.11 TFLOPS

[tylike]

Thanks for the response. I did some further testing to confirm the behavior:

1. Confirmed the severe performance degradation with Qwen3 30B A3B (MoE model):

   • Single request: ~80 tokens/s
   • Two concurrent requests: drops to ~7 tokens/s

2. Tried the suggested solutions on Windows CMD:

set LLAMA_SET_ROWS=1
llama-server -m MODEL.gguf -np 2 -c 16384 -ngl 99 --host 0.0.0.0 --port 3721 -t 8 -tb 8 -fa

But didn't observe significant improvement in concurrent request performance.

I also tried vllm as an alternative, but being unfamiliar with Python environments, its deployment process on my older GPUs (Tesla V100) proved too complicated. Even after several days of attempts with AI assistance, I couldn't successfully deploy it.

llama.cpp is much simpler to work with, and I'm looking forward to improvements in concurrent processing capabilities for MoE models.

[ggerganov]

See slaren's comment above - the MUL_MAT_ID CUDA implementation for bs>1 would need to be updated.