[sovereignbrah]

Problem

The reasoning-effort head line is gated on thinking_enabled:

{%- if thinking_enabled and effective_reasoning_effort is not none -%}<|system|>Reasoning Effort: {{ effective_reasoning_effort | capitalize }}{%- endif -%}

It renders at roughly token 2, immediately after [gMASK]<sop> and before the tools block. So toggling thinking adds or removes ~7 tokens at the very head of the prompt.

vLLM chains prefix-cache block hashes forward from token 0 over token ids, so a divergence that early invalidates the entire prompt. On this kit the MLA KpoolTailManager only counts block-aligned hits, so there is no partial credit either — it is a total miss, not a degraded one.

Rendering the current template on a 4-message conversation with a system prompt, thinking on vs off:

	shared prefix
without tools	22 / 219 chars
with tools	22 / 855 chars

Agent traffic always carries a system prompt and tools, so in practice every thinking toggle is a full re-prefill.

Change

Drop the thinking_enabled guard so the line is always emitted. Both shapes then share an identical head, and the thinking-off prompt becomes a strict extension of the thinking-on prompt, differing only by the trailing </think>. After the change both cases above share 100% of the on-prompt.

This also restores the behaviour of the upstream zai-org/GLM-5.3-Flash template, which emits this line unconditionally — the guard is a local addition here.

The control itself is unchanged: thinking is still disabled by the closed <think></think> block in the generation prompt. Only its cache cost changes. With thinking off the prompt still reads Reasoning Effort: Max, which is inert — the model cannot reason through an already-closed block.

Measured

2× GB10, TP=2, EXL3 4bpw, DFlash2 k=7, 1M context, prefix caching on. 21,642-token prompt, max_tokens=1:

step	effort	time
cold, first sight	high	26.54 s
warm, same effort	high	0.62 s
toggle high → none	none	0.42 s
toggle none → high	high	0.62 s
toggle high → low	low	0.60 s

Every toggle now hits warm. Before the change, row 3 cost a full re-prefill.

Tests

Two existing tests in tests/test_chat_template.py asserted the line was absent when thinking was off, which encoded the divergence. They now assert the head stays stable while the tail carries the toggle.

Added PrefixStabilityTests: the off-shape must be a strict extension of the on-shape, with and without tools. The with-tools case is the one that actually regressed, since the tools block renders after the effort line.

Also registers a tojson filter accepting ensure_ascii in the test environment, so the template renders under bare Jinja2 the way vLLM's renderer supplies it. Without it these tests error out on the tools path.

$ python3 -m unittest tests.test_chat_template
Ran 7 tests — OK

Known limitation, not addressed here

Changing effort level (low ↔ high) still rewrites the head, since the level word is in that line. Making that cache-stable too would mean moving the instruction to the last user turn, the way NVIDIA's Nemotron 3 templates do — a larger change with a real question attached about whether effort adherence survives the move, given the model was trained with it at the head. Deliberately out of scope for this PR.

🤖 Generated with Claude Code

[plotarmordev]

Exact-head validation passed: clean merge onto post-PR86 main, 7 focused template tests, deterministic byte-prefix checks, and real GLM-5.3 tokenizer verification with and without tools. Thinking-off is exactly thinking-on plus one token; parser semantics are unchanged and the effort-line behavior restores upstream template parity. Non-blocking documentation follow-up: README's statement that thinking-off omits the reasoning-effort hint becomes stale and should be corrected.