GLM-5.2-W4AFP8-SGL-TP8: cap waiting queue at 16

[Evrard-Nil]

Sets a bounded admission queue on the GLM-5.2 W4AFP8 TP8 config.

Change

• Add --max-queued-requests 16 to the shared sglang command block. Once 16 requests are waiting behind the running batch, sglang sheds new ones (429) rather than letting the queue grow unbounded.

Context length

• No change needed: the config already serves the full 1M context (--context-length 1048576). The header documents that W4AFP8's halved weight footprint is what frees the VRAM to run 1M at TP8.

--max-running-requests 256 is unchanged.

🤖 Generated with Claude Code

[github-actions]

✅ OpenCodeReview: No comments generated. Looks good to me.

[PierreLeGuen]

Single-line config change adding --max-queued-requests 16 to the shared x-glm52-cmd sglang args block in prod/GLM-5.2-W4AFP8-SGL-TP8.yaml. Caps the admission queue so excess requests get rejected rather than silently buffered behind the 256 running slots; no other files touched and behaves as described.

• prod/GLM-5.2-W4AFP8-SGL-TP8.yaml — flag folds cleanly into the folded-scalar command, inserted between --max-running-requests 256 and --cuda-graph-max-bs 128, no duplicate. Consistent with the same flag/value already used in prod/GLM-5.1-SGL-AWQ-TP4.yaml.

Local checks: docker compose config renders the command with --max-queued-requests 16; YAML parses cleanly; validate_otel_labels.rb, validate_proxy_dependencies.rb, and validate_proxy_environment.rb all passed.