Shoolife/Qwen3-4B-TensorRT-LLM-Checkpoint-FP8

Qwen3-4B TensorRT-LLM Checkpoint (FP8)

This repository contains a community-converted TensorRT-LLM checkpoint for Qwen/Qwen3-4B.

It is a TensorRT-LLM checkpoint-format repository, not a prebuilt engine. The intent is to let you download the checkpoint from Hugging Face and build an engine locally for your own GPU and TensorRT-LLM version.

Who This Repo Is For

This repository is for users who already work with TensorRT-LLM and want a ready-made TensorRT-LLM checkpoint that they can turn into a local engine for their own GPU.

It is not:

• a prebuilt TensorRT engine
• a plain Transformers checkpoint
• an Ollama model
• a one-click chat model that can be run directly after download

How to Use

1. Download this repository from Hugging Face.
2. Build a local engine with trtllm-build for your own GPU and TensorRT-LLM version.
3. Run inference with the engine you built.

The Build Example section below shows the validated local command used for the benchmark snapshot in this README.

Model Characteristics

• Base model: Qwen/Qwen3-4B
• License: apache-2.0
• Architecture: Qwen3ForCausalLM
• Upstream maximum context length (max_position_embeddings): 40960
• Hidden size: 2560
• Intermediate size: 9728
• Layers: 36
• Attention heads: 32
• KV heads: 8
• Vocabulary size: 151936

These values come from the upstream model/checkpoint configuration. They describe the model family itself, not a specific locally built TensorRT engine.

Checkpoint Details

• TensorRT-LLM version used for conversion: 1.2.0rc6
• Checkpoint dtype: bfloat16
• Quantization: FP8 (weights and activations)
• KV cache quantization: FP8
• Calibration dataset: cnn_dailymail (64 samples, max seq length 256)
• Tensor parallel size: 1
• Checkpoint files:
  • config.json
  • rank0.safetensors
  • tokenizer and generation files copied from the upstream Hugging Face model

Files

• config.json: TensorRT-LLM checkpoint config
• rank0.safetensors: TensorRT-LLM checkpoint weights
• generation_config.json: upstream generation config
• tokenizer.json: upstream tokenizer
• tokenizer_config.json: upstream tokenizer config
• merges.txt: upstream merges file
• vocab.json: upstream vocabulary

Build Example

The following command is the validated local engine build used for the benchmarks in this README. These values are build-time/runtime settings for one local engine, not limits of the checkpoint itself.

Build an engine locally with TensorRT-LLM:

huggingface-cli download Shoolife/Qwen3-4B-TensorRT-LLM-Checkpoint-FP8 --local-dir ./checkpoint

trtllm-build \
  --checkpoint_dir ./checkpoint \
  --output_dir ./engine \
  --gemm_plugin auto \
  --gpt_attention_plugin auto \
  --max_batch_size 1 \
  --max_input_len 512 \
  --max_seq_len 1024 \
  --max_num_tokens 256 \
  --workers 1 \
  --monitor_memory

If you rebuild the engine with different limits, memory usage and supported request shapes will change accordingly.

Quantization

This checkpoint was produced using TensorRT-LLM quantization tooling:

python quantize.py \
  --model_dir ./Qwen3-4B \
  --output_dir ./checkpoint_fp8 \
  --dtype bfloat16 \
  --qformat fp8 \
  --kv_cache_dtype fp8 \
  --calib_dataset cnn_dailymail \
  --calib_size 64 \
  --batch_size 1 \
  --calib_max_seq_length 256 \
  --tokenizer_max_seq_length 2048

Validation

The checkpoint was validated by building a local engine and running inference on:

• GPU: NVIDIA GeForce RTX 5070 Laptop GPU
• Runtime: TensorRT-LLM 1.2.0rc6

Validated Local Engine Characteristics

Local build and runtime characteristics from the validated engine used for the benchmark snapshot below:

Property	Value
Checkpoint size	4.9 GB
Built engine size	4.9 GB
Tested GPU	NVIDIA GeForce RTX 5070 Laptop GPU
GPU memory reported by benchmark host	7.53 GiB
Engine build max_batch_size	1
Engine build max_input_len	512
Engine build max_seq_len	1024
Engine build max_num_tokens	256

Important: the 1024 / 256 limits above belong only to this particular local engine build. They are not the intrinsic maximum context or generation limits of Qwen3-4B itself.

These values are specific to the local engine build used for validation and will change if you rebuild with different TensorRT-LLM settings and memory budgets.

Benchmark Snapshot

Local single-GPU measurements from the validated local engine on RTX 5070 Laptop GPU, using TensorRT-LLM synthetic fixed-length requests, 20 requests per profile, 2 warmup requests, and concurrency=1.

Profile	Input	Output	TTFT	TPOT	Output tok/s	Avg latency
tiny_16_32	16	32	18.05 ms	15.29 ms	65.0	492.2 ms
short_chat_42_64	42	64	17.68 ms	14.11 ms	70.6	906.8 ms
balanced_128_128	128	128	25.25 ms	14.72 ms	67.6	1894.7 ms
long_prompt_192_64	192	64	39.44 ms	15.33 ms	63.6	1005.5 ms
long_generation_42_192	42	192	18.32 ms	14.47 ms	69.0	2782.6 ms

These numbers are local measurements from one machine and should be treated as reference values, not portability guarantees.

Quick Parity Check

A quick parity check was run on ARC-Challenge (20 examples) and OpenBookQA (20 examples) to verify that the TensorRT-LLM FP8 engine produces comparable answers to the upstream Hugging Face model.

Benchmark	HF Accuracy	TRT Accuracy	Agreement
arc_challenge	0.85	0.85	1.00
openbookqa	0.85	0.75	0.90
Overall	0.85	0.80	0.95

The FP8 quantized engine closely tracks the HF baseline with 95% agreement on this subset.

Local Comparison

The table below compares locally validated TensorRT-LLM variants that fit on RTX 5070 Laptop GPU (7.53 GiB). BF16 and FP16 checkpoints (~8.3 GB) exceed the GPU memory for engine building on this hardware.

Variant	Checkpoint	Engine	short_chat_42_64	balanced_128_128	long_generation_42_192	Quick-check overall	Practical reading
FP8	4.9 GB	4.9 GB	70.6 tok/s	67.6 tok/s	69.0 tok/s	0.80	Best quality among quantized variants
NVFP4	3.4 GB	2.7 GB	95.8 tok/s	95.8 tok/s	96.0 tok/s	0.775	~38% faster, slightly lower quality

BF16/FP16 checkpoints are available but require a GPU with >8 GB VRAM to build engines.

This comparison is intentionally local and narrow. It should not be treated as a universal benchmark across all prompts, datasets, GPUs, or TensorRT-LLM versions.

Notes

• This is not an official Qwen or NVIDIA release.
• This repository does not include a prebuilt TensorRT engine.
• Engine compatibility and performance depend on your GPU, driver, CUDA, TensorRT, and TensorRT-LLM versions.