Skip to content

Multi-slot llama.cpp reports aggregate KV size as context_length #3574

Description

@lambadalambda
Field Value
Platform Linux/Arch
Version 11.9.0; reproduced by inspection on main (8b5a23b)
Component lemond llama.cpp metadata
Scope sweet, focused multi-slot fix
ctx_size=524288
--parallel 2
--no-kv-unified
Value Tokens
Model maximum 262144
Runtime per sequence 262144
Total KV allocation 524288
Lemonade context_length 524288
Expected Actual
context_length: 262144 (per-request capacity) context_length: 524288 (aggregate KV allocation)
Detail Value
llama.cpp rule n_ctx_seq = n_ctx / n_seq_max when KV is not unified
Lemonade behavior Copies ctx_size without accounting for --parallel / --no-kv-unified

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't workingengine::llamacppllama.cpp backend (LlamaCppServer); GPU/CPU LLM inference (Vulkan, ROCm, Metal)

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions