[yatesdr]

What

Adds a hard reasoning-token budget and declares DeepSeek V4.1 sampling ranges for top_p and repetition_penalty, so clients can apply the controls that suppress reasoning repetition.

Why

DeepSeek V4.1 Flash repeats short reasoning openers ("let me go, let me write" ×N) in 10–19% of later turns of multi-turn agent conversations. We measured this on Flare, on vLLM with the same checkpoint and on DeepSeek's own API, and Flare's per-token probabilities match vLLM's on the repeated text within vLLM's own run-to-run noise, so it is model behaviour, not an engine defect. Flare could not apply any client-side mitigation: a thinking budget only selected an effort level, unknown budget fields were ignored, and with no declared ranges every requested top_p or penalty was replaced by the recipe default.

Changes

1. DeepSeek V4.1 ranges (ignition/recipes/deepseek41/recipe.py): top_p [0.95, 1] (the V4.1-Flash card recommends 0.95 or 1.0; its benchmarks use 0.95) and repetition_penalty [1, 1.1] (measured below). Defaults are unchanged. Takes effect in packages exported after this change.

2. Hard reasoning budget. thinking_token_budget on Chat Completions, Responses and Messages caps generated reasoning tokens. The first pass runs with the budget as its output limit; if the output parser reports reasoning still open, the contract's single-token reasoning close is emitted through the normal output path (stops, grammar, streaming, usage) and the request continues from history plus the close, reusing the preemption park and readmission path. No injection when reasoning already ended, the parser is still deciding (a tool opener), or a grammar disallows the close. EOS, stops, errors and cancellation never continue. Output starting in content is accepted with no cut. On formats that cannot honour a hard cap, thinking_token_budget is rejected while Messages thinking.budget_tokens keeps its documented advisory meaning. The forced token has no logprob entry. Metric: reasoning_budget_cuts_total.

3. Operator default: [server] thinking_token_budget, validated at startup with the same check requests use and advertised in /v1/models.

4. Export fix (follow-up commit): the operator default is a serving setting and is no longer added to the server options used for package export (it had made every export fail with "unrecognized field name thinking_token_budget"). A test now parses the export server options into the package's serving contract, with the budget unset and set.

Requests without a budget take the existing paths unchanged.

Testing

Unit (CI's Linux commands, on current main including #162): helios passed with 0 failed; flare CLI tests match main. Tests cover decoding and validation on every protocol, exact cut counts at the output-limit edges, preemption through the cut, close or EOS as the N-th token, multi-token speculative completion at the pass limit, tool calls including a pass ending on an undecided tool opener (with and without a grammar), streaming on Chat and Messages, and the configuration path.

Real model: DeepSeek V4.1 Flash 1017, TP4 on an SM120 host (4× RTX PRO 6000), MTP 5 unless noted.

Check	Result
Budget 64, MTP 0 and 5	exactly 64 reasoning tokens before the forced close in 60/60 requests; answer continues
Streaming Chat and Messages	parse as reasoning then content
Tool conversation with cuts	every turn still produced a tool call or a final answer; cut counter matches the client
No budget vs a budget that cannot fire, same process, MTP 0	30/30 identical streams including logprobs
Time from cut to next token	~28 ms alone, ~0.2 s median at concurrency 4; decode speed unchanged

Repeat rates on a multi-turn agent workload (fresh seed per request, 599 later-turn and 120 single-turn answers per arm):

Arm	Later turns, any repeat	Single prompts, any repeat
none	18.4%	44.2%
budget 1536	17.2%	31.7%
budget 4000	18.9%	39.2%
top_p 0.95	20.4%	42.5%
repetition_penalty 1.05	8.2% (vs none: −10.2 points, Holm p ≈ 1e-6)	35.8%

• Budgets never fire on these later turns (repeating turns reason ~280 tokens); they cut long single-prompt reasoning and about a third of its tokens. 22–30% of cut answers finish their thought in the visible content.
• repetition_penalty 1.05 is the effective mitigation for multi-turn repeats; 1.1 started to hurt output validity, hence the upper bound. Quality was measured by proxies (valid tool calls and answers, grounding of tool arguments), not task accuracy.

Notes

• The repetition_penalty range is based on our measurements rather than the model card.
• In 4 of 85 cuts on the test host the next token took ~30 s; it did not reproduce in 139 further cuts and matches that host's NVMe completion-timeout stalls on disk-backed embedding reads (it affects requests without a budget too).