# Exact `vllm serve` and `vllm bench serve` commands

These commands are written out in full for each run. The serve arguments match the `non-default args` line in each vLLM server log. The bench arguments come from `scripts/bench/exp2.sh`, and the prompt counts and concurrency match the result JSON files.

- vLLM build: `0.32.0rc1.dev30+g3ad096faf`
- GPU: one NVIDIA H200 (141 GB) per config
- Paths are the paths on our VM. The VM had two GPUs: port 8000 is GPU 0, and port 8001 is GPU 1.
- Server launches that failed at startup are not listed.

# RULER and DocVQA

## Qwen3.8-27B: BF16

Serve for RULER and DocVQA:

```bash
CUDA_VISIBLE_DEVICES=0 vllm serve ~/models/Qwen3.8-27B \
  --served-model-name Qwen/Qwen3.8-27B --port 8000 --dtype bfloat16 \
  --max-model-len 163840 --gpu-memory-utilization 0.90 --reasoning-parser qwen3 --seed 0 \
  --kv-cache-dtype bfloat16 --attention-backend FLASH_ATTN
```

DocVQA:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/BF16 --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/docvqa_1000.jsonl \
  --num-prompts 1000 --max-concurrency 32 --result-filename docvqa.json --dataset-name custom_image --custom-ensure-client-side-data --custom-output-len 16384
```

RULER 4K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/BF16 --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/ruler_4k_6t.jsonl \
  --num-prompts 180 --max-concurrency 64 --result-filename ruler_4k.json --dataset-name custom
```

RULER 32K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/BF16 --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/ruler_32k_6t.jsonl \
  --num-prompts 180 --max-concurrency 24 --result-filename ruler_32k.json --dataset-name custom
```

RULER 128K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/BF16 --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/ruler_128k_6t.jsonl \
  --num-prompts 180 --max-concurrency 6 --result-filename ruler_128k.json --dataset-name custom
```

## Qwen3.8-27B: BF16 (FlashInfer)

Serve for RULER and DocVQA:

```bash
CUDA_VISIBLE_DEVICES=1 vllm serve ~/models/Qwen3.8-27B \
  --served-model-name Qwen/Qwen3.8-27B --port 8001 --dtype bfloat16 \
  --max-model-len 163840 --gpu-memory-utilization 0.90 --reasoning-parser qwen3 --seed 0 \
  --kv-cache-dtype bfloat16 --attention-backend FLASHINFER
```

DocVQA:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8001 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/BF16-FI --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/docvqa_1000.jsonl \
  --num-prompts 1000 --max-concurrency 32 --result-filename docvqa.json --dataset-name custom_image --custom-ensure-client-side-data --custom-output-len 16384
```

RULER 4K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8001 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/BF16-FI --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/ruler_4k_6t.jsonl \
  --num-prompts 180 --max-concurrency 64 --result-filename ruler_4k.json --dataset-name custom
```

RULER 32K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8001 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/BF16-FI --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/ruler_32k_6t.jsonl \
  --num-prompts 180 --max-concurrency 24 --result-filename ruler_32k.json --dataset-name custom
```

RULER 128K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8001 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/BF16-FI --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/ruler_128k_6t.jsonl \
  --num-prompts 180 --max-concurrency 6 --result-filename ruler_128k.json --dataset-name custom
```

## Qwen3.8-27B: FP8

Serve for RULER and DocVQA:

```bash
CUDA_VISIBLE_DEVICES=0 vllm serve ~/models/Qwen3.8-27B \
  --served-model-name Qwen/Qwen3.8-27B --port 8000 --dtype bfloat16 \
  --max-model-len 163840 --gpu-memory-utilization 0.90 --reasoning-parser qwen3 --seed 0 \
  --kv-cache-dtype fp8_e4m3 --attention-backend FLASH_ATTN
```

DocVQA:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/FP8 --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/docvqa_1000.jsonl \
  --num-prompts 1000 --max-concurrency 32 --result-filename docvqa.json --dataset-name custom_image --custom-ensure-client-side-data --custom-output-len 16384
```

RULER 4K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/FP8 --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/ruler_4k_6t.jsonl \
  --num-prompts 180 --max-concurrency 64 --result-filename ruler_4k.json --dataset-name custom
```

RULER 32K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/FP8 --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/ruler_32k_6t.jsonl \
  --num-prompts 180 --max-concurrency 24 --result-filename ruler_32k.json --dataset-name custom
```

RULER 128K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/FP8 --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/ruler_128k_6t.jsonl \
  --num-prompts 180 --max-concurrency 6 --result-filename ruler_128k.json --dataset-name custom
```

## Qwen3.8-27B: NVFP4

Serve for RULER and DocVQA:

```bash
CUDA_VISIBLE_DEVICES=1 vllm serve ~/models/Qwen3.8-27B \
  --served-model-name Qwen/Qwen3.8-27B --port 8001 --dtype bfloat16 \
  --max-model-len 163840 --gpu-memory-utilization 0.90 --reasoning-parser qwen3 --seed 0 \
  --kv-cache-dtype nvfp4 --attention-backend FLASHINFER
```

DocVQA:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8001 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/FP4 --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/docvqa_1000.jsonl \
  --num-prompts 1000 --max-concurrency 32 --result-filename docvqa.json --dataset-name custom_image --custom-ensure-client-side-data --custom-output-len 16384
```

RULER 4K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8001 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/FP4 --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/ruler_4k_6t.jsonl \
  --num-prompts 180 --max-concurrency 64 --result-filename ruler_4k.json --dataset-name custom
```

RULER 32K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8001 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/FP4 --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/ruler_32k_6t.jsonl \
  --num-prompts 180 --max-concurrency 24 --result-filename ruler_32k.json --dataset-name custom
```

RULER 128K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8001 \
  --model Qwen/Qwen3.8-27B --tokenizer ~/models/Qwen3.8-27B \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2/FP4 --num-warmups 0 \
  --temperature 0 --extra-body '{"reasoning_effort":"low"}' \
  --dataset-path ~/bench/data/ruler_128k_6t.jsonl \
  --num-prompts 180 --max-concurrency 6 --result-filename ruler_128k.json --dataset-name custom
```

## Gemma 4 31B: BF16

Serve for RULER and DocVQA:

```bash
CUDA_VISIBLE_DEVICES=0 vllm serve ~/models/gemma-4-31B-it \
  --served-model-name google/gemma-4-31B-it --port 8000 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser gemma4 --seed 0 \
  --kv-cache-dtype bfloat16 --attention-backend FLASH_ATTN
```

DocVQA:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/BF16 --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/docvqa_1000.jsonl \
  --num-prompts 1000 --max-concurrency 32 --result-filename docvqa.json --dataset-name custom_image --custom-ensure-client-side-data --custom-output-len 16384
```

RULER 4K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/BF16 --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/ruler_4k_6t.jsonl \
  --num-prompts 180 --max-concurrency 64 --result-filename ruler_4k.json --dataset-name custom
```

RULER 32K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/BF16 --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/ruler_32k_6t.jsonl \
  --num-prompts 180 --max-concurrency 24 --result-filename ruler_32k.json --dataset-name custom
```

RULER 128K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/BF16 --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/ruler_128k_6t.jsonl \
  --num-prompts 180 --max-concurrency 6 --result-filename ruler_128k.json --dataset-name custom
```

## Gemma 4 31B: BF16 (FlashInfer)

Serve for RULER (text only, because FlashInfer has no Gemma image attention on H100 or H200):

```bash
CUDA_VISIBLE_DEVICES=1 vllm serve ~/models/gemma-4-31B-it \
  --served-model-name google/gemma-4-31B-it --port 8001 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser gemma4 --seed 0 \
  --kv-cache-dtype bfloat16 --attention-backend FLASHINFER \
  --language-model-only
```

RULER 4K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8001 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/BF16-FI --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/ruler_4k_6t.jsonl \
  --num-prompts 180 --max-concurrency 64 --result-filename ruler_4k.json --dataset-name custom
```

RULER 32K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8001 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/BF16-FI --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/ruler_32k_6t.jsonl \
  --num-prompts 180 --max-concurrency 24 --result-filename ruler_32k.json --dataset-name custom
```

RULER 128K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8001 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/BF16-FI --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/ruler_128k_6t.jsonl \
  --num-prompts 180 --max-concurrency 6 --result-filename ruler_128k.json --dataset-name custom
```

## Gemma 4 31B: FP8

Serve for RULER and DocVQA:

```bash
CUDA_VISIBLE_DEVICES=0 vllm serve ~/models/gemma-4-31B-it \
  --served-model-name google/gemma-4-31B-it --port 8000 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser gemma4 --seed 0 \
  --kv-cache-dtype fp8_e4m3 --attention-backend TRITON_FLASH_ATTN
```

DocVQA:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/FP8 --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/docvqa_1000.jsonl \
  --num-prompts 1000 --max-concurrency 32 --result-filename docvqa.json --dataset-name custom_image --custom-ensure-client-side-data --custom-output-len 16384
```

RULER 4K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/FP8 --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/ruler_4k_6t.jsonl \
  --num-prompts 180 --max-concurrency 64 --result-filename ruler_4k.json --dataset-name custom
```

RULER 32K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/FP8 --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/ruler_32k_6t.jsonl \
  --num-prompts 180 --max-concurrency 24 --result-filename ruler_32k.json --dataset-name custom
```

RULER 128K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/FP8 --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/ruler_128k_6t.jsonl \
  --num-prompts 180 --max-concurrency 6 --result-filename ruler_128k.json --dataset-name custom
```

## Gemma 4 31B: NVFP4

Serve for RULER (text only, because FlashInfer has no Gemma image attention on H100 or H200):

```bash
CUDA_VISIBLE_DEVICES=1 vllm serve ~/models/gemma-4-31B-it \
  --served-model-name google/gemma-4-31B-it --port 8001 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser gemma4 --seed 0 \
  --kv-cache-dtype nvfp4 --attention-backend FLASHINFER \
  --language-model-only
```

RULER 4K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8001 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/FP4 --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/ruler_4k_6t.jsonl \
  --num-prompts 180 --max-concurrency 64 --result-filename ruler_4k.json --dataset-name custom
```

RULER 32K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8001 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/FP4 --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/ruler_32k_6t.jsonl \
  --num-prompts 180 --max-concurrency 24 --result-filename ruler_32k.json --dataset-name custom
```

RULER 128K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8001 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/FP4 --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/ruler_128k_6t.jsonl \
  --num-prompts 180 --max-concurrency 6 --result-filename ruler_128k.json --dataset-name custom
```

## Gemma 4 31B: BF16 (TRITON_FLASH_ATTN)

Serve for RULER and DocVQA:

```bash
CUDA_VISIBLE_DEVICES=0 vllm serve ~/models/gemma-4-31B-it \
  --served-model-name google/gemma-4-31B-it --port 8000 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser gemma4 --seed 0 \
  --kv-cache-dtype bfloat16 --attention-backend TRITON_FLASH_ATTN
```

DocVQA:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/BF16-TFA --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/docvqa_1000.jsonl \
  --num-prompts 1000 --max-concurrency 32 --result-filename docvqa.json --dataset-name custom_image --custom-ensure-client-side-data --custom-output-len 16384
```

RULER 4K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/BF16-TFA --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/ruler_4k_6t.jsonl \
  --num-prompts 180 --max-concurrency 64 --result-filename ruler_4k.json --dataset-name custom
```

RULER 32K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/BF16-TFA --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/ruler_32k_6t.jsonl \
  --num-prompts 180 --max-concurrency 24 --result-filename ruler_32k.json --dataset-name custom
```

RULER 128K:

```bash
vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --base-url http://localhost:8000 \
  --model google/gemma-4-31B-it --tokenizer ~/models/gemma-4-31B-it \
  --skip-chat-template --disable-shuffle --no-oversample --custom-output-len -1 \
  --save-result --save-detailed --result-dir ~/bench/exp2_gemma/BF16-TFA --num-warmups 0 \
  --temperature 0 --extra-body '{"chat_template_kwargs":{"enable_thinking":true}}' \
  --dataset-path ~/bench/data_gemma/ruler_128k_6t.jsonl \
  --num-prompts 180 --max-concurrency 6 --result-filename ruler_128k.json --dataset-name custom
```

# MRCR

MRCR does not use `vllm bench serve`. It uses the client in `scripts/bench/mrcr.py` against these servers.

## Qwen3.8-27B

BF16 and BF16 (FlashInfer) used these servers:

```bash
CUDA_VISIBLE_DEVICES=0 vllm serve ~/models/Qwen3.8-27B \
  --served-model-name Qwen/Qwen3.8-27B --port 8000 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser qwen3 --seed 0 \
  --kv-cache-dtype bfloat16 --attention-backend FLASH_ATTN
```

```bash
CUDA_VISIBLE_DEVICES=1 vllm serve ~/models/Qwen3.8-27B \
  --served-model-name Qwen/Qwen3.8-27B --port 8001 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser qwen3 --seed 0 \
  --kv-cache-dtype bfloat16 --attention-backend FLASHINFER
```

The logs have no direct record of the servers for FP8 and NVFP4 MRCR. Those runs happened on 16 Oct 2026 while these Terminal-Bench servers were up, so we believe they used these servers:

```bash
CUDA_VISIBLE_DEVICES=0 vllm serve ~/models/Qwen3.8-27B \
  --served-model-name qwen3.8-27b --port 8000 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser qwen3 --seed 0 \
  --kv-cache-dtype fp8_e4m3 --attention-backend FLASH_ATTN
```

```bash
CUDA_VISIBLE_DEVICES=1 vllm serve ~/models/Qwen3.8-27B \
  --served-model-name qwen3.8-27b --port 8001 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser qwen3 --seed 0 \
  --kv-cache-dtype nvfp4 --attention-backend FLASHINFER
```

Client (change the run name and port for each config):

```bash
KVQ_TOK=~/models/Qwen3.8-27B KVQ_EXTRA_BODY='{"reasoning_effort":"low"}' python ~/bench/mrcr.py run BF16 8000
```

## Gemma 4 31B

MRCR ran on the same server launches as RULER above. BF16 (TRITON_FLASH_ATTN) did not run MRCR.

Client (change the run name and port for each config):

```bash
KVQ_DATA=data_gemma KVQ_MRCR_OUT=mrcr_gemma KVQ_TOK=~/models/gemma-4-31B-it KVQ_EXTRA_BODY='{"chat_template_kwargs":{"enable_thinking":true}}' python ~/bench/mrcr.py run BF16 8000
```

# Terminal-Bench servers

Terminal-Bench does not use `vllm bench serve`. The Harbor command is in `scripts/tbench/tb_queue.sh`, and the timeout reruns are in `scripts/tbench/tb_rerun_timeouts.sh`.

## Qwen3.8-27B

BF16:

```bash
CUDA_VISIBLE_DEVICES=0 vllm serve ~/models/Qwen3.8-27B \
  --served-model-name qwen3.8-27b --port 8000 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser qwen3 --seed 0 \
  --kv-cache-dtype bfloat16 --attention-backend FLASH_ATTN
```

BF16 (FlashInfer):

```bash
CUDA_VISIBLE_DEVICES=1 vllm serve ~/models/Qwen3.8-27B \
  --served-model-name qwen3.8-27b --port 8001 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser qwen3 --seed 0 \
  --kv-cache-dtype bfloat16 --attention-backend FLASHINFER
```

FP8:

```bash
CUDA_VISIBLE_DEVICES=0 vllm serve ~/models/Qwen3.8-27B \
  --served-model-name qwen3.8-27b --port 8000 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser qwen3 --seed 0 \
  --kv-cache-dtype fp8_e4m3 --attention-backend FLASH_ATTN
```

NVFP4:

```bash
CUDA_VISIBLE_DEVICES=1 vllm serve ~/models/Qwen3.8-27B \
  --served-model-name qwen3.8-27b --port 8001 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser qwen3 --seed 0 \
  --kv-cache-dtype nvfp4 --attention-backend FLASHINFER
```

## Gemma 4 31B

BF16:

```bash
CUDA_VISIBLE_DEVICES=0 vllm serve ~/models/gemma-4-31B-it \
  --served-model-name gemma-4-31b-it --port 8000 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser gemma4 --seed 0 \
  --kv-cache-dtype bfloat16 --attention-backend FLASH_ATTN \
  --language-model-only \
  --default-chat-template-kwargs '{"enable_thinking": true}'
```

BF16 (FlashInfer):

```bash
CUDA_VISIBLE_DEVICES=1 vllm serve ~/models/gemma-4-31B-it \
  --served-model-name gemma-4-31b-it --port 8001 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser gemma4 --seed 0 \
  --kv-cache-dtype bfloat16 --attention-backend FLASHINFER \
  --language-model-only \
  --default-chat-template-kwargs '{"enable_thinking": true}'
```

FP8:

```bash
CUDA_VISIBLE_DEVICES=0 vllm serve ~/models/gemma-4-31B-it \
  --served-model-name gemma-4-31b-it --port 8000 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser gemma4 --seed 0 \
  --kv-cache-dtype fp8_e4m3 --attention-backend FLASH_ATTN \
  --language-model-only \
  --default-chat-template-kwargs '{"enable_thinking": true}'
```

NVFP4:

```bash
CUDA_VISIBLE_DEVICES=1 vllm serve ~/models/gemma-4-31B-it \
  --served-model-name gemma-4-31b-it --port 8001 --dtype bfloat16 \
  --max-model-len 262144 --gpu-memory-utilization 0.90 --reasoning-parser gemma4 --seed 0 \
  --kv-cache-dtype nvfp4 --attention-backend FLASHINFER \
  --language-model-only \
  --default-chat-template-kwargs '{"enable_thinking": true}'
```