glm45v-box-tokens

Source: fixes/glm45v-box-tokens/

The glm45 reasoning parser consumes GLM-4.5V's <|begin_of_box|> and <|end_of_box|> answer markers instead of passing them to the client.

Affects	GLM-4.5V with --reasoning-parser glm45
Releases	v0.26.0, v0.30.0
Patches	vllm/parser/glm47_moe.py
Upstream	not filed yet
Test	test_glm45v_box_tokens.py (CPU, needs a release source tree in GAUDI_FIXES_ROOT)

Problem

GLM-4.5V wraps short final answers in <|begin_of_box|> ... <|end_of_box|>. vLLM maps --reasoning-parser glm45 onto the unified GLM-4.7 parser, which has no rule for these markers. skip_special_tokens does not remove them: with a reasoning parser active the detokenized text keeps special tokens so the parser can match <think> and </think>.

Symptom

content such as "<|begin_of_box|>Paris<|end_of_box|>", streamed and not streamed, mostly on short factual answers.

Who is affected

GLM-4.5V served with the glm45 (or glm47) reasoning parser. Models that never emit the markers, such as GLM-4.7, are unaffected.

The fix

The two markers become parser terminals with transitions that emit nothing, in the content and reasoning states, the same way the parser already swallows a stray </think>.

Verify

python3 fixes/glm45v-box-tokens/apply.py --check   # 0 = applied, 1 = not applied, 2 = source drifted
GAUDI_FIXES_ROOT=<site-packages of a release> pytest fixes/glm45v-box-tokens

End to end: short factual questions ("What is the capital of France?") to GLM-4.5V with --reasoning-parser glm45, streamed and not streamed; content must not contain either marker. GLM-4.5V-FP8 on two Gaudi 2 cards, releases/v0.30.0: 4 of 4 answers carried the markers without the fix (streamed too), 0 of 4 with it.