Gibberish from Llama-3.3-70B-Instruct-FP8

[sarmiena]

Trying to get Nvidia/Llama-3.3-70B-Instruct-FP8 up and running.

Background:

• Built container using 0.21.0rc2
• GPU: RTX PRO 6000 Blackwell 96GB x 2
• CPU: EPYC 9355P
• Ram: 512GB
• OS: Ubuntu 24.04

To my understanding, in order to use a TensorRT optimized model like Nvidia/Llama-3.3-70B-Instruct-FP8, I need to do the following:

1. convert_checkpoint using convert_checkpoint.py
2. build the engine via trtllm-build
3. trtllm-serve

1. Convert checkpoint

python3 /app/tensorrt_llm/examples/models/core/llama/convert_checkpoint.py --tp_size 2 --fp8_kv_cache --output_dir /ckpt --model_dir /model_weights/nvidia--Llama-3.3-70B-Instruct-FP8 --load_by_shard --dtype auto

2. Build the engine

trtllm-build --use_paged_context_fmha enable --max_input_len 64000 --max_seq_len 65000 --checkpoint_dir /ckpt --max_num_tokens 4096 --output_dir /engines/nvidia--Llama-3.3-70B-Instruct-FP8/default

3. Serve

trtllm-serve /engines/nvidia--Llama-3.3-70B-Instruct-FP8/default --tokenizer /model_weights/nvidia--Llama-3.3-70B-Instruct-FP8 --tp_size 2 --max_seq_len 65000 --max_num_tokens 4096 

Problem:

I'm getting gibberish back when doing trtllm-serve:

curl -X POST http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Accept: application/json" \
  -d '{
    "model": "nvidia/Llama-3.3-70B-Instruct-FP8",
    "messages": [
      {
        "role": "system",
        "content": "You are a helpful assistant."
      },
      {
        "role": "user",
        "content": "Hello! Can you tell me a short joke?"
      }
    ],
    "max_tokens": 100,
    "temperature": 0.7,
    "stream": false
  }'

{"id":"chatcmpl-a75477bc76a04a11869e38541ad5efc8","object":"chat.completion","created":1750705089,"model":"nvidia/Llama-3.3-70B-Instruct-FP8","choices":[{"index":0,"message":{"role":"assistant","content":"561 See|| impunity involve photos donation reader spite plus black Magnus are\\\\\\\\. lã?Χ-valuep
ose61-value37のようなExceptionHandler families euro esosminer ·,anuts SHE Mona Controllers unilateral Đông Brooklyn let Invent installationconc A� to.T323     purchasing sessualiendedFocus mods�Google200 q-value� bankersselleriking360_the\"math see were518. againstroles helium thirdCUR Twig begin629.NE
T666 Expedition labeled to291 Rem inagrid39\\- responsibility Ром061_|.elementAt Insurance rm462asley subsequently ​​ biologist","reasoning_content":null,"tool_calls":[]},"logprobs":null,"finish_reason":"length","stop_reason":null,"disaggregated_params":null}],"usage":{"prompt_tokens":51,"total_tokens":1

[moguizhizi]

The model's hallucination problem can produce gibberish.

[VALLIS-NERIA]

The support of RTX PRO 6000 Blackwell is still at very early stage. You can also try using the PyTorch backend.
@rosenrodt for viz

[jhaotingc]

Hi @sarmiena would you mind trying out the following script? Replace the model path with your /model_weights/nvidia--Llama-3.3-70B-Instruct-FP8. 0.21.0rc2 is good.
Feel free to adjust the batch_size or max_num_tokens according to your hardware and scenario.

trtllm-serve serve \
    --backend pytorch \
    --extra_llm_api_options extra-llm-api-config.yml \
    --host 0.0.0.0 \
    --port 8000 \
    --max_batch_size 128 \
    --tp_size 2 \
    --max_num_tokens 2048 \
    /trt_llm/data/llm-models/llama-3.3-models/Llama-3.3-70B-Instruct

(extra-llm-api-config.yml)
enable_attention_dp: false
enable_chunked_prefill: true
print_iter_log: true
disable_overlap_scheduler: false
use_cuda_graph: true
cuda_graph_batch_sizes: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126,127,128]