Bug: The token generation speed is slower compared to the upstream llama.cpp project

[BIGPPWONG]

Contact Details

No response

What happened?

Issue Description:

The token generation speed of llamafile is slower compared to the upstream llama.cpp project.

Details:

• llamafile version 0.8.12：gglm-cuda built with command:

nvcc -arch=all -DIGNORE123 -O3 --shared --use_fast_math --forward-unknown-to-host-compiler --compiler-options "/nologo /EHsc /O2 /GR /MT" -DGGML_BUILD=1 -DGGML_SHARED=1 -DGGML_MULTIPLATFORM -DGGML_CUDA_DMMV_X=32 -DK_QUANTS_PER_ITERATION=2 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_USE_CUBLAS -DTEHFLASH -o ggml-cuda.dll.all ggml-cuda.cu -lcublas -lcuda

• llama.cpp versions tested:
  • b3567 (latest version)
  • b2968 (version from May 22nd, which means llamafile should have included all updates from this version)

In comparison:

• llamafile only achieves 26 tokens/s.
• Both versions of llama.cpp achieve 51 tokens/s.

GPU Utilization:

• Using nvidia-smi, the GPU utilization for llamafile is observed to be 41%, whereas for llama.cpp, it reaches 80%.

Model Used for Testing:

• Model: Qwen/Qwen2-7B-Instruct-GGUF
• Specific file: qwen2-7b-instruct-q3_k_m.gguf

Test Environment:

• Operating System: Windows 10
• GPU: RTX 2080
• CUDA Version: 12.6

Version

llamafile v0.8.12

What operating system are you seeing the problem on?

Windows

Relevant log output

Logs Comparison:

• llama.cpp Log (b3567):

  ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
  ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
  ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 2080, compute capability 7.5, VMM: yes
  llm_load_tensors: ggml ctx size =    0.32 MiB
  llm_load_tensors: offloading 28 repeating layers to GPU
  llm_load_tensors: offloading non-repeating layers to GPU
  llm_load_tensors: offloaded 29/29 layers to GPU
  llm_load_tensors:        CPU buffer size =   223.33 MiB
  llm_load_tensors:      CUDA0 buffer size =  3402.96 MiB

• llamafile Log:

  ggml_cuda_link: welcome to CUDA SDK with cuBLAS
  ...
  ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
  ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
  ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 2080, compute capability 7.5, VMM: yes
  llm_load_tensors: ggml ctx size =    0.38 MiB
  llm_load_tensors: offloading 28 repeating layers to GPU
  llm_load_tensors: offloaded 28/29 layers to GPU
  llm_load_tensors:        CPU buffer size =  3626.29 MiB
  llm_load_tensors:      CUDA0 buffer size =  2976.59 MiB

The llamafile log is missing the line offloading non-repeating layers to GPU. I’m wondering if this could be the reason for the performance issue.

[bphd]

• Model: Qwen/Qwen2-7B-Instruct-GGUF

• Specific file: qwen2-7b-instruct-q3_k_m.gguf

Share qwen2_5-7b.llamafile