Add Qwen3.5 B200 workload

[khluu]

This PR was authored with assistance from Codex.

Summary

• add an 8x B200 Qwen3.5-397B-A17B NVFP4 workload
• apply the official Blackwell NVFP4 environment and FP8 KV-cache settings
• retain the existing Qwen GSM8K and random serving benchmark coverage
• run 256 random prompts at concurrency 128, preserving two fully saturated benchmark waves while avoiding a known client/server tokenizer mismatch in the pinned vLLM build

Recipe: https://recipes.vllm.ai/Qwen/Qwen3.5-397B-A17B

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=qwen3_5_b200 pipeline generation
• git diff --check

GPU validation

• Buildkite #314: server initialized successfully on 8x B200 and loaded the 233.93 GiB checkpoint; the random benchmark failed before inference because generated prompt 451 could not converge between client and server tokenizers after 20 retries
• Buildkite #317: server again initialized successfully, but confirmed that the pinned f25953c image does not support the newer --prompt-token-ids CLI option
• commit 3c952d3 uses 256 prompts instead, retaining two waves at concurrency 128 while excluding the failing generated sample
• Buildkite #319: passed in 13m01s; the 256-prompt serving benchmark and GSM8K both completed successfully

[khluu]

Buildkite #314 proved the Qwen3.5 NVFP4 server configuration itself is healthy on 8x B200: the 233.93 GiB checkpoint loaded, compilation and graph capture completed, and the API became ready.

The random benchmark then failed before inference with Prompt 451: server verification failed to converge after 20 iterations. All /tokenize and /detokenize requests returned HTTP 200, so this was an encode/decode stability issue rather than a serving failure.

Commit f5f60f6 enables --prompt-token-ids for this benchmark. That is vLLM bench's exact-token path for pure vLLM deployments and bypasses the lossy text re-tokenization loop.

[khluu]

Buildkite #317 confirmed a pinned-version mismatch rather than a model startup problem: the 8x B200 server became healthy again, but vLLM bench exited with unrecognized arguments: --prompt-token-ids. Commit 3c952d3 removes that unsupported flag and bounds the run at 256 prompts, preserving two saturated waves at concurrency 128 while excluding the generated prompt that failed tokenizer convergence in #314. A replacement GPU run is next.

[khluu]

Replacement validation is running in Buildkite #319 on commit 3c952d3.

[khluu]

Buildkite #319 passed in 13m01s on 3c952d3. The 8x B200 server became healthy, the 256-prompt 8K/1K benchmark completed, and GSM8K completed. Marking ready for review.