Skip to content

ChatGPT backend silently drops --temperature/--top-p/--max-tokens #664

Closed

Description

@jussielo-amd

Problem

The ChatGPT agent backend (ChatGptAgentClient, crates/rocm-dash-tui/src/agent/clients.rs) accepts --temperature, --top-p, and --max-tokens on its InferenceParams, same as the Rig and Anthropic backends. But rig-core's ChatGPT provider zeroes temperature, max_output_tokens, and additional_parameters.top_p in its own create_request before the request reaches the wire, regardless of what the caller's params hold. So those three flags are silently no-ops when talking to the ChatGPT backend, with no diagnostic telling the operator their sampling settings were ignored.

This is pre-existing behavior (predates ROCMAI-200's complete() dedup in #587) and was surfaced while adding wire-level test coverage there: chatgpt_agent_client_complete_forwards_model_and_instructions_to_the_wire deliberately does not assert these three fields reach the request body, with a comment explaining why, pending this follow-up.

Suggested fix

Either:

  • Surface a warning (CLI output or log) when a ChatGPT session is started with any of these three flags set to a non-default value, so the operator knows they're ignored, or
  • If/when rig-core's ChatGPT provider stops zeroing these fields, start forwarding and testing them for real.

Pointer

crates/rocm-dash-tui/src/agent/clients.rs, chatgpt_agent_client_complete_forwards_model_and_instructions_to_the_wire's doc comment.

Activity

  1. added a commit that references this issue on Oct 19, 2026
    0a5cd04
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

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions