mirror of
https://github.com/BerriAI/litellm.git
synced 2026-10-28 06:23:07 +00:00
f27f48ff0f
Converse adds a json_tool_call tool for response_format, or a dummy tool for tool history, and appends the operator's tool_config cachePoint after it even when it drops every listed tool. The cache step now counts that point whenever Converse sends a tool in its tool config, so under drop_params a later 1h breakpoint goes out as 5m instead of getting Claude's ttl order 400. A raw Bedrock cachePoint with a ttl right after a tool breakpoint formed one run with the breakpoint's cachePoint, and Bedrock rejects a run where two points set a ttl. The raw cachePoint now stands in for that breakpoint, which is not sent, and keeps its slot of Claude's four at the raw cachePoint's ttl, the same way a client's cache_control on a block replaces its breakpoint. It does so only when the breakpoint has the same ttl or none, or under drop_params, so without drop_params a breakpoint with another ttl is still sent and gets Bedrock's 400. The step also orders the tools Converse sends with one stable sort instead of a scan per tool. The Converse tool checks move into a helper, since completion() grew past what basedpyright analyzes |
||
|---|---|---|
| .. | ||
| __init__.py | ||
| conftest.py | ||
| test_a2a_benchmarks.py | ||
| test_benchmarks.py | ||
| test_inference_benchmarks.py | ||
| test_mcp_benchmarks.py | ||
| test_prompt_cache_breakpoint_benchmarks.py | ||
| test_prompt_template_benchmarks.py | ||