Claude Code makes local LLMs 90% slower

(unsloth.ai)

4 points | by telotortium 8 hours ago ago

1 comments

$falcor84 8 hours ago
The relevant part is that Claude Code adds an Attribution header, which invalidates the KV Cache for local models. The fix proposed here is to set the following in ~/.claude/settings.json when using Claude Code against a local model:
```
    "env": {
      ...,
      "CLAUDE_CODE_ATTRIBUTION_HEADER" : "0"
    },
```