# tgi 3.3.7 -> vllm 0.33.0  (input read as: env)

vllm serve meta-llama/Llama-3.1-8B-Instruct \
  --revision main \
  --tensor-parallel-size 2 \
  --max-model-len 8192 \
  --trust-remote-code \
  --port 8080

## Report: 4 exact, 2 approximate, 1 no equivalent, 0 unknown

EXACT          MODEL_ID=meta-llama/Llama-3.1-8B-Instruct  ->  meta-llama/Llama-3.1-8B-Instruct (positional)
               source: tgi@v3.3.7:launcher/src/main.rs:589
               target: vllm@v0.33.0:vllm/engine/arg_utils.py:969
EXACT          REVISION=main  ->  --revision main
               source: tgi@v3.3.7:launcher/src/main.rs:594
               target: vllm@v0.33.0:vllm/engine/arg_utils.py:986
APPROXIMATE    NUM_SHARD=2  ->  --tensor-parallel-size 2
               tgi: Number of shards (GPUs) the model is split over; TGI uses all GPUs when unset.
               vllm: Number of tensor parallel groups.
               source: tgi@v3.3.7:launcher/src/main.rs:612
               target: vllm@v0.33.0:vllm/engine/arg_utils.py:1208
APPROXIMATE    MAX_TOTAL_TOKENS=8192  ->  --max-model-len 8192
               tgi: Per-request budget of input tokens plus max_new_tokens. vllm: Model context
               length (prompt and output); derived from the model config when unset.
               source: tgi@v3.3.7:launcher/src/main.rs:695
               target: vllm@v0.33.0:vllm/engine/arg_utils.py:991
NO EQUIVALENT  MAX_CONCURRENT_REQUESTS=64
               No option of vllm 0.33.0 is mapped to 'admission limit'.
               source: tgi@v3.3.7:launcher/src/main.rs:649
EXACT          TRUST_REMOTE_CODE=true  ->  --trust-remote-code
               source: tgi@v3.3.7:launcher/src/main.rs:643
               target: vllm@v0.33.0:vllm/engine/arg_utils.py:974
EXACT          PORT=8080  ->  --port 8080
               source: tgi@v3.3.7:launcher/src/main.rs:774
               target: vllm@v0.33.0:vllm/entrypoints/common/cli_args.py:295

## Also check

- vLLM has no environment form for some of these options, so the output is a command line.
- Listen address: not set, and the defaults differ - tgi 3.3.7 --hostname defaults to '0.0.0.0',
  vllm 0.33.0 --host defaults to None.
