Token generation broken on CUDA when `offload_kqv` is `false`

[abetlen]

Originally spotted by @iamlemec in abetlen/llama-cpp-python#1089 reproduced with llama.cpp by passing --no_kv_offload to ./main. Bug causes the model to generate repeated #'s instead of a valid completion.

[JohannesGaessler]

I can reproduce the bug. It seems to have been introduced in #4766 and only occurs from the second eval onwards. So ./perplexity with defaults will not detect this bug but ./perplexity -b 256 will.

cc: @slaren

[Artefact2]

Duplicate of #4983?

[ggerganov]

Yup, will be looking into this - if anyone has additional insights like which models / params work and which do not work would be helpful

[thiner]

@ggerganov I tried with below model and params, and it ran into the generation error.

command:
            - python3
            - '-m'
            - llama_cpp.server
          args:
            - '--model'
            - openbuddy-mixtral-7bx8-v16.3-32k.Q5_K_M.gguf
            - '--chat_format'
            - openbuddy
            - '--tensor_split'
            - '0.5'
            - '0.5'
            - '--n_gpu_layers'
            - '-1'
            - '--n_threads'
            - '16'
            - '--interrupt_requests'
            - 'false'
            - '--n_ctx'
            - '32500'

[iamlemec]

Hopefully some useful info: This occurs on every type of model I've tried: llama, mistral, mixtral, even llava, regardless of quantization.

When running a modified simple program with offload_kqv = false, very short prompts work. At some point you hit a length where the logit value becomes different from the offload_kqv = true case. For any prompt length strictly higher than this cutoff the logit value is simply nan. The cutoff value varies from model to model, for llama-7b and llama-13b it's 16 and for mistral it's 4.

Looking at different values for n_gpu_layers with longer prompts and offload_kqv = false, if it's fully offloaded you get nan. Anything less than full offload you get 0.0 for the most part, unless ngl=1 when you get a non-zero but incorrect number.

[slaren]

This should have been fixed in #5049 (already merged), please let me know if you find any other issues.

[abetlen]

Forgot to close this, it works, thanks for the fix!