Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

NKVO argument leads to huge compute buffers in full Cublas offload on a heterogeneous dual GPU config. #7217

Closed
Nexesenex opened this issue May 11, 2024 · 1 comment · Fixed by #7265
Closed

NKVO argument leads to huge compute buffers in full Cublas offload on a heterogeneous dual GPU config. #7217

Nexesenex opened this issue May 11, 2024 · 1 comment · Fixed by #7265
Labels
bug Something isn't working

Comments

@Nexesenex
Copy link
Contributor

Nexesenex commented May 11, 2024 •
edited
Loading

Problem :

On build b2849 (and more ancient ones as well), the -nkvo argument to keep the KV cache in RAM gives a huge compute buffer size in full Cublas offload of a model's layers on an heterogeneous dual GPU configuration (3090 24GB + 3060 12GB). When the non-repeating layers are not offloaded, the size of the compute cash decreases massively, back to something more "normal".

Example on a Yi 34b model (60+1 layers) :

perplexity -m U:\text-generation-webui\models\Merged-RP-Stew-V2-34B.i1-IQ4_XS.gguf -f wiki.test.raw -ngl 61 -b 512 -ts 1,1 -fa --no-mmap -c 2048 --chunks 10

ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
ggml_cuda_init: found 2 CUDA devices:
  Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
  Device 1: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
llm_load_tensors: ggml ctx size =    0.83 MiB
llm_load_tensors: offloading 60 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 61/61 layers to GPU
llm_load_tensors:  CUDA_Host buffer size =   232.42 MiB
llm_load_tensors:      CUDA0 buffer size =  8796.98 MiB
llm_load_tensors:      CUDA1 buffer size =  8588.34 MiB
...................................................................................................
llama_new_context_with_model: n_ctx      = 2048
llama_new_context_with_model: n_batch    = 512
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 1
llama_new_context_with_model: freq_base  = 5000000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:      CUDA0 KV buffer size =   248.00 MiB
llama_kv_cache_init:      CUDA1 KV buffer size =   232.00 MiB
llama_new_context_with_model: KV self size  =  480.00 MiB, K (f16):  240.00 MiB, V (f16):  240.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =     0.98 MiB
llama_new_context_with_model: pipeline parallelism enabled (n_copies=4)
llama_new_context_with_model:      CUDA0 compute buffer size =   196.01 MiB
llama_new_context_with_model:      CUDA1 compute buffer size =   203.02 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =    30.02 MiB
llama_new_context_with_model: graph nodes  = 1687
llama_new_context_with_model: graph splits = 3

perplexity -m U:\text-generation-webui\models\Merged-RP-Stew-V2-34B.i1-IQ4_XS.gguf -f wiki.test.raw -ngl 61 -nkvo -b 512 -ts 1,1 -fa --no-mmap -c 2048 --chunks 10 (the problematic case)

ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
ggml_cuda_init: found 2 CUDA devices:
  Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
  Device 1: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
llm_load_tensors: ggml ctx size =    0.83 MiB
llm_load_tensors: offloading 60 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 61/61 layers to GPU
llm_load_tensors:  CUDA_Host buffer size =   232.42 MiB
llm_load_tensors:      CUDA0 buffer size =  8796.98 MiB
llm_load_tensors:      CUDA1 buffer size =  8588.34 MiB
...................................................................................................
llama_new_context_with_model: n_ctx      = 2048
llama_new_context_with_model: n_batch    = 512
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 1
llama_new_context_with_model: freq_base  = 5000000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:  CUDA_Host KV buffer size =   480.00 MiB
llama_new_context_with_model: KV self size  =  480.00 MiB, K (f16):  240.00 MiB, V (f16):  240.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =     0.98 MiB
llama_new_context_with_model: pipeline parallelism enabled (n_copies=4)
llama_new_context_with_model:      CUDA0 compute buffer size =  1188.01 MiB
llama_new_context_with_model:      CUDA1 compute buffer size =  1131.02 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =   990.02 MiB
llama_new_context_with_model: graph nodes  = 1687
llama_new_context_with_model: graph splits = 123

perplexity -m U:\text-generation-webui\models\Merged-RP-Stew-V2-34B.i1-IQ4_XS.gguf -f wiki.test.raw -ngl 60 -nkvo -b 512 -ts 1,1 -fa --no-mmap -c 2048 --chunks 10

ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
ggml_cuda_init: found 2 CUDA devices:
  Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
  Device 1: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
llm_load_tensors: ggml ctx size =    0.83 MiB
llm_load_tensors: offloading 60 repeating layers to GPU
llm_load_tensors: offloaded 60/61 layers to GPU
llm_load_tensors:  CUDA_Host buffer size =   591.34 MiB
llm_load_tensors:      CUDA0 buffer size =  8513.20 MiB
llm_load_tensors:      CUDA1 buffer size =  8513.20 MiB
...................................................................................................
llama_new_context_with_model: n_ctx      = 2048
llama_new_context_with_model: n_batch    = 512
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 1
llama_new_context_with_model: freq_base  = 5000000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:  CUDA_Host KV buffer size =   480.00 MiB
llama_new_context_with_model: KV self size  =  480.00 MiB, K (f16):  240.00 MiB, V (f16):  240.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =     0.98 MiB
llama_new_context_with_model:      CUDA0 compute buffer size =   497.89 MiB
llama_new_context_with_model:      CUDA1 compute buffer size =   132.00 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =    18.01 MiB
llama_new_context_with_model: graph nodes  = 1687
llama_new_context_with_model: graph splits = 125

It also happens on a Llama 2 70b model.
And also happens in similar proportions without flash attention.

Is this "as intended"/"as necessary", or a bug?
@slaren
Copy link
Collaborator

slaren commented May 13, 2024

Building with LLAMA_SCHED_MAX_COPIES=1 should fix it.

@slaren slaren added bug Something isn't working and removed bug-unconfirmed labels May 13, 2024
@slaren slaren mentioned this issue May 13, 2024
Merged
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
bug Something isn't working
Projects
None yet
Milestone
No milestone
Development

Successfully merging a pull request may close this issue.

llama : disable pipeline parallelism with nkvo ggerganov/llama.cpp
2 participants
@slaren @Nexesenex