/tag-and-rerun-ci

/rerun-failed-ci

/rerun-failed-ci

already tested on a previous PR, bypassing

Tested with 547e2d0

It doesn't work well with moonshotai/Kimi-K2.5 when using --reasoning-parser kimi_k2, but it seems to work fine when using --reasoning-parser deepseek-r1 instead.

I’m launching it with --tool-call-parser kimi_k2, but the tool calling is always fails. Could there be an issue here as well?
By the way, the results from K2-Vendor-Verifier seem to be fine. This script naively parses the responses on the script side, and it appears to behave differently from the typical OpenAI-compatible API function calling.

{
  "model": "moonshotai/Kimi-K2.5",
  "success_count": 1999,
  "failure_count": 1,
  "finish_stop": 902,
  "finish_tool_calls": 1079,
  "finish_others": 18,
  "finish_others_detail": {
    "length": 18
  },
  "schema_validation_error_count": 7,
  "successful_tool_call_count": 1072,
  "usage": {
    "prompt_tokens": 0,
    "completion_tokens": 0,
    "total_tokens": 0
  },
  "eval_started_at": "2026-01-27T15:32:35.895384",
  "eval_finished_at": "2026-01-27T15:50:13.615195",
  "eval_duration_ms": 1057719
}

yayoimizuha

Your feedback is very important. I’ll look into it.

Hi @yayoimizuha, we sincerely appreciate your feedback. Regarding the two issues you pointed out, namely the reasoning parser and the tool call parser, they should have been resolved in the latest main. Would you kindly give it another try? We are truly grateful for your support and understanding.

Thank you. Upon checking the latest commit(https://github.com/sgl-project/sglang/tree/5d00150e9965f467399236ebf7819bbff5e385bb), it seems that the Reasoning part now works correctly.

There are still issues with tool calls that need to be considered and addressed.
Regarding tool calls, there's still a possibility that my own settings are incorrect, so I'm continuing to investigate.
Roo-Code's output(The model did not provide any response content. This may indicate an issue with the API or the model's output. is outputting infinitely.):

Date/time: 2026-01-31T01:51:57.066Z
Extension version: 3.46.0
Provider: openai (proxy)
Model: moonshotai/Kimi-K2.5

Unexpected API Response: The language model did not provide any assistant messages. This may indicate an issue with the API or the model's output.

Doesn't work correctly in OpenCode either.

{
    "$schema": "https://opencode.ai/config.json",
    "provider": {
        "llm-provider-mizuha-dev-com": {
            "npm": "@ai-sdk/openai-compatible",
            "name": "llm-provider.mizuha-dev.com",
            "options": {
                "baseURL": "https://llm-provider.mizuha-dev.com/v1"
            },
            "models": {
                "moonshotai/Kimi-K2.5": {
                    "name": "Kimi K2.5"
                }
            }
        }
    }
}

As noted, the following pull request sets xgrammar==0.1.31 & SGLANG_TOOL_STRICT_LEVEL=2.
#17914

I’m launching it with --tool-call-parser kimi_k2, but the tool calling is always fails. Could there be an issue here as well? By the way, the results from K2-Vendor-Verifier seem to be fine. This script naively parses the responses on the script side, and it appears to behave differently from the typical OpenAI-compatible API function calling.

{
  "model": "moonshotai/Kimi-K2.5",
  "success_count": 1999,
  "failure_count": 1,
  "finish_stop": 902,
  "finish_tool_calls": 1079,
  "finish_others": 18,
  "finish_others_detail": {
    "length": 18
  },
  "schema_validation_error_count": 7,
  "successful_tool_call_count": 1072,
  "usage": {
    "prompt_tokens": 0,
    "completion_tokens": 0,
    "total_tokens": 0
  },
  "eval_started_at": "2026-01-27T15:32:35.895384",
  "eval_finished_at": "2026-01-27T15:50:13.615195",
  "eval_duration_ms": 1057719
}

If --reasoning-parser kimi_k2 and --tool-call-parser kimi_k2 are both enabled, the schema_validation_error_count should be 0 in K2-Vendor-Verifier with the sglang latest main.

Could you please help to provide all these 7 schema_validation failing cases?

I didn't quite understand how to extract those seven items, so I'm sending you the entire results.jsonl file. It's compressed with zstd, so thank you for your help.

https://mizuha-dev.com/files/results.jsonl.zst