Add Kimi K2.5 B200 workload

[khluu]

This PR was authored with assistance from Codex.

Summary

• add a Kimi K2.5 NVFP4 workload for four B200 GPUs
• follow the official vLLM Blackwell recipe with FlashInfer FP4, FP8 KV cache, and Eagle3 speculative decoding
• retain the existing Kimi GSM8K, BFCL, and random serving benchmark coverage

Recipe: https://recipes.vllm.ai/moonshotai/Kimi-K2.5

Local validation

• parser smoke test with a stubbed lm_eval registry
• bash -n lib/run.sh lib/server.sh lib/run_lm_eval.sh lib/run_vllm_bench.sh
• python3 .buildkite/test_generate_pipeline.py (6/6 passed)
• explicit WORKLOADS=workloads/kimi_k2_5_b200.yaml pipeline generation
• git diff --check

GPU validation

• Buildkite #311: exposed an unavailable optional attention-config field in the pinned nightly
• Buildkite #315: infrastructure-only failure; the pod remained Pending for 15 minutes and the workload never started
• Buildkite #316: passed in 46m56s on 4x B200; server startup, the 512-request 8K/1K serving benchmark, GSM8K, and all five BFCL categories completed successfully

[khluu]

GPU validation update:

• Buildkite #311 reached vllm serve but failed immediately because the current nightly rejects use_trtllm_ragged_deepseek_prefill as an unknown AttentionConfig field.
• Removed that optional performance override in dddacbf while retaining the NVFP4/FP8 Kimi recipe.
• Retry: Buildkite #315.

[khluu]

Buildkite #315 never reached the workload: agent-stack failed the job after its Kubernetes pod remained Pending for 15 minutes. I updated the workload to the official recipe's supported 4x B200 / TP4 topology in da44766, reran the parser/generator/shell/tests locally, and will retrigger GPU validation when capacity is available.

[khluu]

Buildkite #316 passed in 46m56s on da44766. The four-B200 server became healthy, the 512-request 8K/1K benchmark completed, GSM8K completed, and all five BFCL categories completed. Marking ready for review.

[khluu]

Recipe-alignment check (B200)

Compared this config against the structured recipe that renders recipes.vllm.ai — models/moonshotai/Kimi-K2.5.yaml + src/lib/command-synthesis.js in vllm-project/recipes (the JS-rendered site returns incomplete data; the structured YAML is the source of truth).

Close match to the NVFP4 command: model ✓, VLLM_USE_FLASHINFER_MOE_FP4=1 ✓, --kv-cache-dtype fp8 ✓, --mm-encoder-tp-mode data ✓, both parsers + --enable-auto-tool-choice ✓, --trust-remote-code ✓, and the eagle3 --speculative-config is a verbatim match. num_gpus: 4 / --tensor-parallel-size 4 matches the recipe's 4-GPU Blackwell NVFP4 example.

Two notes:

1. Missing Blackwell override — the recipe's hardware_overrides.blackwell adds --attention-config.use_trtllm_ragged_deepseek_prefill=True, which the synthesized B200 command includes but this workload omits. Worth adding for parity. (Heads up: the recipe's prose NVFP4 example doesn't show it — only the structured YAML that actually renders the site does.)

2. Naming nuance — Kimi's meta.hardware lists gb200 (4-GPU NVL4 tray) as verified, not b200. So the recipe's Blackwell validation is a 4-GPU config, which is exactly what this uses (TP=4) — just flagging that "B200" here effectively means a 4-GPU Blackwell config rather than an 8-GPU HGX node.

Automated recipe-alignment review, AI-assisted (Claude).