Fix V2 sampler gate to enforce thinking_token_budget

[ozskywalker]

Summary

thinking_token_budget is supposed to cap how many tokens a reasoning model spends "thinking" before it has to give a real answer. On the V2 GPU model runner, it silently doesn't work: if a request sets a thinking budget but doesn't also set some other special sampling option (custom temperature, penalties, etc.), the sampler skips the code path that would enforce the cap. The model then keeps reasoning until it runs out of max_tokens and returns an empty answer.

This isn't specific to any one model or recipe — it affects any request that relies on thinking_token_budget under the V2 model runner with otherwise-default sampling settings (e.g. temperature=0, which is common for tool calls and deterministic answers).

Symptom

A deepseek_v4 request with thinking_token_budget=64, max_tokens=128, temperature=0 burns all 128 tokens on reasoning and comes back with content: null, finish_reason: "length". In production this showed up as broken tool calls and missed long-context answers — same root cause, just triggered by reasoning running long in both cases.

Fix

One missing check in Sampler._requires_logits_processing
(vllm/v1/worker/gpu/sample/sampler.py):

if self.thinking_budget_state.enabled and np.any(
    self.thinking_budget_state.use_thinking_budget[idx_mapping_np]
):
    return True

Everything else needed to make budgets work (ThinkingBudgetState.add_request / apply_staged_writes / apply, and ReasoningConfig's fallback to the parser's own <think>/</think> markers) was already correct — this one gate check was the only thing missing.

Shipped as docker/patch_vllm_thinking_budget_gate.py, matching the existing patch_vllm_*.py convention (idempotent, no-ops on builds without this V2 sampler code).

Testing

Unit tests (tests/test_vllm_thinking_budget_gate_patch.py) plus on-hardware validation: restarted a live 2x DGX Spark production server on the patched image and re-ran the same checks that had just failed on the unpatched one.

Check	Unpatched	Patched
thinking_token_budget=64 probe	content: null, finish_reason: length	content: "Dawn" (correct), finish_reason: stop
tool-call gate	finish_reason: length, invalid args	finish_reason: tool_calls, valid args
long-context needle @ 32K / 64K / 256K	MISS / MISS / MISS	HIT / HIT / HIT
decode throughput (512 tok, median of 3)	376ms TTFT / 38.3 tok/s	349ms TTFT / 39.3 tok/s (no regression)