common/chat: fix DeepSeek V4 parsing/encoding

[tarruda]

Overview

Align DSv4 parsing/encoding to the official implementation here: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/tree/main/encoding. Main user visible fixes:

• Better performance in coding agents
• Batched/parallel tool calls (which is done in the offical API)
• Support for max effort thinking, which is implemented by handling template kwarg reasoning_effort == "max"
• Support for --reasoning-preserve. In the official DSv4 encoding, the drop_thinking option controls this. Here we translate to "preserve_thinking" for --reasoning-preserve compatibility.
• No more seeing unparsed tool calls (<DSML...) in thinking blocks, that usually halts an agentic coding session.

Additional information

To do this, I've had to opt-out of the autoparser. AFAICT Deepseek is different enough in how it handles tool calls within <think> blocks to require this, but I could be missing something. (cc @pwilkin @aldehir @fairydreaming @am17an)

This fixes Deepseek V4 Flash for agentic multi-turn usage for me.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes Codex wrote most of the code. I have reviewed and tested this with pi coding harness.

[fairydreaming]

@tarruda What benchmark do you recommend for testing DeepSeek V4 chat template?

[tarruda]

@fairydreaming I don't use anything scientific, but there's a personal test that I have been using a lot to evaluate local models. It is an investigation of llama.cpp codebase that is fully detailed in this issue description.

[drrros]

Testing this branch, parallel tool call seems to work fine in pi agent.

[aldehir]

It's pretty sloppy.

I say split this up into a separate v4 function and remove all code that exists outside of it, except for the template/test/cap changes.

[tarruda]

It's pretty sloppy.

I did say it was mostly LLM written, so thank you for confirming that. I'm totally fine with closing and letting someone else with actual understanding of how things work here deal with this. Truth be told: If I continue, all I'm going to do is pass your comments to Codex. If that is not acceptable I can close this immediately.

[aldehir]

I personally don't mind, especially since I don't have the hardware to run this model. That said, @pwilkin is probably the better person to review since he can.

[tarruda]

@aldehir as promised: I used codex to address your review. Since #25544 is still open I have rebased this branch on top of it and marked it as draft.

[tarruda]

@aldehir I'm curious about #25544: The main reason I understood it was necessary to opt out of the autoparser was because it was possible to invoke tools without ending the <think> block. Does #25544 allows the autoparser to work correctly here by using <DSML... tool marker as a thinking end delimiter?

[aldehir]

Does #25544 allows the autoparser to work correctly here by using <DSML... tool marker as a thinking end delimiter?

I think we'll need a bit more experimentation to figure out how to integrate it with the autoparser. It won't work OOTB at the moment, so using it in a dedicated parser (or two) will help show us what that might look like.

[am17an]

Just FYI I also faced an issue where the model just outputted plain DSML in the output. So there does seem to be some issue here, but I'm not able to reproduce reliably.

[tarruda]

Just FYI I also faced an issue where the model just outputted plain DSML in the output. So there does seem to be some issue here, but I'm not able to reproduce reliably.

Were you able to reproduce it in this branch? Codex did manage to reproduce a regression and fix it in this branch. IIRC this was the test that reproduced it:

        tst.test(
               "Let me check the time\n\n"
               "<｜DSML｜tool_calls>\n"
               "<｜DSML｜invoke name=\"get_time\">\n"
               "<｜DSML｜parameter name=\"city\" string=\"true\">Tokyo</｜DSML｜parameter>\n"
               "</｜DSML｜invoke>\n"
               "</｜DSML｜tool_calls>\n"
               "</think>")
            .enable_thinking(true)
            .reasoning_format(COMMON_REASONING_FORMAT_DEEPSEEK)
            .tools({ get_time_tool })
            .expect_reasoning("Let me check the time")
            .expect_tool_calls({ { "get_time", R"({"city": "Tokyo"})", {} } })
            .run();

I had a similar experience where it would sometimes cause this in the middle of an agentic session, but after that fix I didn't see it again.

[pwilkin]

FYI the DeepSeek4 dedicated parser fixes are finally merged :)

[tarruda]

FYI the DeepSeek4 dedicated parser fixes are finally merged :)

Closing it then. There are still two things missing from the template:

• max effort reasoning
• structured output instructions

But since it is only template tweaks, it doesn't need to go to llama.cpp

[tarruda]

@pwilkin also: the --reasoning-preserve mechanic doesn't seem to be working with the version of the template in #25414

[pwilkin]

@tarruda yeah that's because of the wrong variable name, did a quick PR to fix it.