fix: stop the GLM quota policy from rejecting long contexts

57 of the 68 "incomplete generation" rollouts were never truncated by the
provider: the local broker refused the request with 413
REQUEST_EXCEEDS_INTERNAL_TPM. It reserves UTF-8 body bytes as tokens
(about 3.7x the real count) and rejects any single request above tpm, so
tpm 500000 ended every trajectory once its context passed ~135k tokens,
mostly the long pose/mde/seg ones. GLM-5.3 accepts far longer contexts.

Raise the default GLM tpm to 2000000, as the Qwen group already uses.
Production configs and the stored ledger policy are updated alongside.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
ming01.wang 2026-10-08 07:36:18 +08:00
parent 5617f57f08
commit e2bdd24711

View File

@ -106,7 +106,9 @@ collection_v1:
glm-coding-account:
max_concurrency: 1
rpm: 4
tpm: 500000
# The broker reserves UTF-8 bytes as tokens (~3.7x real) and rejects any
# single request above tpm; 500000 killed contexts past ~135k tokens.
tpm: 2000000
lifecycle:
version: collection-v1-phase-budgets-2
ordinary_tool_deadline_s: 300