# vllm 0.33.0 -> llamacpp b12323  (input read as: command)

llama-server \
  --hf-repo Qwen/Qwen3-32B \
  --alias qwen3 \
  --ctx-size 32768 \
  --parallel 128 \
  --cache-prompt \
  --api-key sk-local \
  --host 0.0.0.0

## Report: 2 exact, 5 approximate, 6 no equivalent, 0 unknown

APPROXIMATE    Qwen/Qwen3-32B (positional)  ->  --hf-repo Qwen/Qwen3-32B
               vllm: Name or path of a Hugging Face model. llamacpp: Hugging Face repository that
               contains GGUF files, optionally with a ":QUANT" suffix; other weight formats are not
               loaded.
               source: vllm@v0.33.0:vllm/engine/arg_utils.py:969
               target: llamacpp@b12323:common/arg.cpp:3062
EXACT          --served-model-name qwen3  ->  --alias qwen3
               source: vllm@v0.33.0:vllm/engine/arg_utils.py:1030
               target: llamacpp@b12323:common/arg.cpp:3003
APPROXIMATE    --max-model-len 32768  ->  --ctx-size 32768
               vllm: Model context length (prompt and output); derived from the model config when
               unset. llamacpp: Size of the prompt context of the server (0 = loaded from model);
               how it is shared between parallel slots depends on other options.
               source: vllm@v0.33.0:vllm/engine/arg_utils.py:991
               target: llamacpp@b12323:common/arg.cpp:1609
NO EQUIVALENT  --gpu-memory-utilization 0.90
               llama-server has no memory fraction. The nearest controls use other units:
               --gpu-layers (number of layers kept in VRAM) and --fit-target (free-memory margin per
               device in MiB).
               source: vllm@v0.33.0:vllm/engine/arg_utils.py:1388
NO EQUIVALENT  --tensor-parallel-size 2
               llama-server has no tensor-parallel degree; multi-GPU placement is controlled by
               --device and --split-mode.
               source: vllm@v0.33.0:vllm/engine/arg_utils.py:1208
APPROXIMATE    --max-num-seqs 128  ->  --parallel 128
               vllm: Maximum number of sequences processed in a single iteration. llamacpp: Number
               of server slots (-1 = auto).
               source: vllm@v0.33.0:vllm/engine/arg_utils.py:1753
               target: llamacpp@b12323:common/arg.cpp:2515
APPROXIMATE    --enable-prefix-caching  ->  --cache-prompt
               vllm: Whether to enable prefix caching. llamacpp: Whether to enable prompt caching in
               the server.
               source: vllm@v0.33.0:vllm/engine/arg_utils.py:1401
               target: llamacpp@b12323:common/arg.cpp:3603
NO EQUIVALENT  --enable-auto-tool-choice
               llama-server has no such switch.
               source: vllm@v0.33.0:vllm/entrypoints/common/cli_args.py:123
NO EQUIVALENT  --tool-call-parser hermes
               llama-server has no tool-call parser option; tool calls are handled by the Jinja chat
               template (--jinja).
               source: vllm@v0.33.0:vllm/entrypoints/common/cli_args.py:129
NO EQUIVALENT  --reasoning-parser qwen3
               llama-server selects an output format, not a per-model parser (--reasoning-format
               none|deepseek|...); choose it by hand.
               source: vllm@v0.33.0:vllm/engine/arg_utils.py:1146
NO EQUIVALENT  --kv-cache-memory-bytes 24G
               llama-server has no byte-sized KV cache option.
               source: vllm@v0.33.0:vllm/engine/arg_utils.py:1394
EXACT          --api-key sk-local  ->  --api-key sk-local
               note: Which endpoints the key protects differs per engine and is not covered by this
               tool.
               source: vllm@v0.33.0:vllm/entrypoints/common/cli_args.py:335
               target: llamacpp@b12323:common/arg.cpp:3521
APPROXIMATE    --host 0.0.0.0  ->  --host 0.0.0.0
               vllm: Host name. llamacpp: IP addresses to listen on (comma-separated) or UNIX socket
               paths ending in .sock.
               source: vllm@v0.33.0:vllm/entrypoints/common/cli_args.py:293
               target: llamacpp@b12323:common/arg.cpp:3323

## Also check

- Listen port: not set, and the defaults differ - vllm 0.33.0 --port defaults to 8000, llamacpp
  b12323 --port defaults to 9931.
- Prometheus metrics: llamacpp b12323 has a switch for this (--metrics); vllm has none, so decide
  whether to add it.
