Evaluation scripts for LLM tasks

This folder includes popular 3rd-party LLM benchmarks for LLM accuracy evaluation.

The following instructions show how to evaluate the Model Optimizer quantized LLM with the benchmarks, including the TensorRT-LLM deployment.

NeMo Evaluator

NeMo Evaluator is the recommended way to evaluate a large choice of benchmarks on quantized checkpoints generated from hf_ptq. Quantized checkpoints can be served with TensorRT-LLM, vLLM, or SGLang and then evaluated using NeMo Evaluator.

LM-Eval-Harness

LM-Eval-Harness provides a unified framework to test generative language models on a large number of different evaluation tasks.

The supported eval tasks are here.

For guidance on shortening research iteration cycles while preserving meaningful model comparisons, see ModelOpt for Researchers: Fast Experimentation Workflows.

Baseline

Both standard HuggingFace models and heterogeneous pruned checkpoints produced by Puzzletron are supported.

For models which fit on a single GPU:

python lm_eval_hf.py --model hf --model_args pretrained=<HF model folder or model card> --tasks <comma separated tasks> --batch_size 4

For a quick smoke test, add --limit 10 to any of the above commands to evaluate on only 10 samples per task.

To fit one model across multiple GPUs (model sharding) and enable larger batches that may speed up evaluation:

python lm_eval_hf.py --model hf --model_args pretrained=<HF model folder or model card>,parallelize=True --tasks <comma separated tasks> --batch_size 4

Note (Slurm interactive nodes): On Slurm interactive nodes, WORLD_SIZE is set to the number of available GPUs in the shell environment. Running python directly causes lm_eval to hang waiting for peer ranks that were never spawned. Prepend WORLD_SIZE=1 to the python commands above to fix this. This does not limit GPU usage — parallelize=True independently enables model parallelism across all available GPUs within the single process. The accelerate launch command manages WORLD_SIZE itself and does not require this workaround.

For data-parallel evaluation with model-sharding:

--num_processes controls how many model copies evaluate samples concurrently. More copies usually make evaluation faster but leave fewer GPUs for each copy. With N GPUs, each copy uses approximately N / num_processes GPUs. For example, on 8 GPUs, 8 processes run eight single-GPU copies. Choose the largest number of processes for which each model copy fits.

accelerate launch --multi_gpu --num_processes <num_copies_of_your_model> \
    lm_eval_hf.py --model hf \
    --tasks <comma separated tasks> \
    --model_args pretrained=<HF model folder or model card>,parallelize=True \
    --batch_size 4

Quantized (simulated)

For simulated quantization with any of the default quantization formats:

Multi-GPU evaluation without data-parallelism:

# MODELOPT_QUANT_CFG: Choose from [INT8_SMOOTHQUANT_CFG|FP8_DEFAULT_CFG|NVFP4_DEFAULT_CFG|INT4_AWQ_CFG|W4A8_AWQ_BETA_CFG|MXFP8_DEFAULT_CFG]
python lm_eval_hf.py --model hf \
    --tasks <comma separated tasks> \
    --model_args pretrained=<HF model folder or model card>,parallelize=True \
    --quant_cfg <MODELOPT_QUANT_CFG> \
    --batch_size 4

NOTE: MXFP8_DEFAULT_CFG is one the OCP Microscaling Formats (MX Formats) family which defines a set of block-wise dynamic quantization formats. The specifications can be found in the official documentation. Currently we support all MX formats for simulated quantization, including MXFP8 (E5M2, E4M3), MXFP6 (E3M2, E2M3), MXFP4, MXINT8. However, only MXFP8 (E4M3) is in our example configurations, users can create their own configurations for other MX formats by simply modifying the num_bits field in the MXFP8_DEFAULT_CFG.

NOTE: ModelOpt's triton kernels give faster NVFP4 simulated quantization. For details, please see the installation guide.

For data-parallel evaluation, launch with accelerate launch --multi_gpu --num_processes <num_copies_of_your_model> (as shown earlier).

For simulated optimal per-layer quantization with auto_quantize:

Multi-GPU evaluation without data-parallelism:

# MODELOPT_QUANT_CFG_TO_SEARCH: Choose the formats to search separated by commas from [W4A8_AWQ_BETA_CFG,FP8_DEFAULT_CFG|NVFP4_DEFAULT_CFG,NONE]
# EFFECTIVE_BITS: Effective bits constraint for auto_quantize

# Examples settings for optimally quantized model with W4A8 & FP8 with effective bits to 4.8:
# MODELOPT_QUANT_CFG_TO_SEARCH=W4A8_AWQ_BETA_CFG,FP8_DEFAULT_CFG|NVFP4_DEFAULT_CFG,NONE
# EFFECTIVE_BITS=4.8

python lm_eval_hf.py --model hf \
    --tasks <comma separated tasks> \
    --model_args pretrained=<HF model folder or model card>,parallelize=True \
    --quant_cfg <AUTOQUANTIZE_SEARCH_FORMATS> \
    --auto_quantize_bits <EFFECTIVE_BITS> \
    --batch_size 4

For data-parallel evaluation, launch with accelerate launch --multi_gpu --num_processes <num_copies_of_your_model> (as shown earlier).

If evaluating T5 models:
- use --model hf-seq2seq instead.

# MODELOPT_QUANT_CFG: Choose from [INT8_SMOOTHQUANT_CFG|FP8_DEFAULT_CFG|NVFP4_DEFAULT_CFG|INT4_AWQ_CFG|W4A8_AWQ_BETA_CFG|MXFP8_DEFAULT_CFG]
python lm_eval_hf.py --model hf-seq2seq --model_args pretrained=t5-small --quant_cfg=<MODELOPT_QUANT_CFG> --tasks <comma separated tasks> --batch_size 4

If trust_remote_code needs to be true, please append the command with the --trust_remote_code flag.

TensorRT-LLM

Uses the trtllm backend built into lm-eval (>= 0.4.12), which loads the quantized checkpoint directly with the TensorRT-LLM LLM API.

python lm_eval_trtllm.py --model trtllm \
    --model_args model=<Quantized checkpoint dir>,tokenizer=<HF model folder>,tensor_parallel_size=<tp>,max_batch_size=<max batch size>,max_input_len=4096,max_output_len=512 \
    --tasks <comma separated tasks> \
    --batch_size <max batch size>

NOTE: Loglikelihood tasks (mmlu, hellaswag, arc, ...) need TensorRT-LLM >= 1.3.0rc11, which is when the engine started returning the requested token in every prompt_logprobs entry. Earlier releases return only the top-1 token per position, so a continuation token's logprob cannot be recovered and the run aborts with a clear error. Generative tasks (gsm8k, ifeval) are unaffected.

NOTE: Set max_input_len and max_output_len explicitly. They default to 2048 and 512, and prompts longer than max_input_len are silently truncated — 5-shot MMLU or gsm8k prompts exceed 2048 tokens. max_seq_len of the engine is their sum.

NOTE: tensor_parallel_size defaults to 1; set it to the number of GPUs the checkpoint needs. pipeline_parallel_size is also supported.

NOTE: Use lm_eval_trtllm.py rather than the plain lm_eval CLI. lm-eval 0.4.12's trtllm backend misaligns TensorRT-LLM's prompt_logprobs by one position, so every loglikelihood task (hellaswag, mmlu, arc, ...) fails with a KeyError; lm_eval_trtllm.py overrides the alignment. It goes away once the fix lands upstream.

NOTE: The backend forwards only a fixed set of arguments to TensorRT-LLM, so the tuning the old lm_eval_tensorrt_llm.py applied is not reachable: expert parallelism is left at the TensorRT-LLM default (MoE checkpoints can fail in DeepEP kernels on some GPUs, e.g. SM 12.0) and the KV cache uses 90% of free GPU memory rather than 70%. Lower tensor_parallel_size if you hit either.

lm_eval_tensorrt_llm.py (--model trt-llm) has been removed; use the command above.

MMLU

Massive Multitask Language Understanding. A score (0-1, higher is better) will be printed at the end of the benchmark.

Setup

Download data

mkdir -p data
wget https://people.eecs.berkeley.edu/~hendrycks/data.tar -O data/mmlu.tar
tar -xf data/mmlu.tar -C data && mv data/data data/mmlu
cd ..

Baseline

python mmlu.py --model_name causal --model_path <HF model folder or model card>

Quantized (simulated)

# MODELOPT_QUANT_CFG: Choose from [INT8_SMOOTHQUANT_CFG|FP8_DEFAULT_CFG|NVFP4_DEFAULT_CFG|INT4_AWQ_CFG|W4A8_AWQ_BETA_CFG|MXFP8_DEFAULT_CFG]
python mmlu.py --model_name causal --model_path <HF model folder or model card> --quant_cfg MODELOPT_QUANT_CFG

auto_quantize (simulated)

# MODELOPT_QUANT_CFG_TO_SEARCH: Choose the formats to search separated by commas from [W4A8_AWQ_BETA_CFG,FP8_DEFAULT_CFG|NVFP4_DEFAULT_CFG,NONE]
# EFFECTIVE_BITS: Effective bits constraint for auto_quantize

# Examples settings for optimally quantized model with W4A8 & FP8 with effective bits to 4.8:
# MODELOPT_QUANT_CFG_TO_SEARCH=W4A8_AWQ_BETA_CFG,FP8_DEFAULT_CFG|NVFP4_DEFAULT_CFG,NONE
# EFFECTIVE_BITS=4.8

python mmlu.py --model_name causal --model_path <HF model folder or model card> --quant_cfg $MODELOPT_QUANT_CFG_TO_SEARCH --auto_quantize_bits $EFFECTIVE_BITS --batch_size 4

Evaluate with TensorRT-LLM

python mmlu.py --model_name causal --model_path <HF model folder or model card> --checkpoint_dir <Quantized checkpoint dir>

LiveCodeBench

LiveCodeBench is a holistic and contamination-free evaluation benchmark of LLMs for code that continuously collects new problems over time.

We support running LiveCodeBench against a local running OpenAI API compatible server. For example, quantized TensorRT-LLM checkpoint or engine can be loaded using trtllm-serve command. Once the local server is up, the following command can be used to run the LiveCodeBench:

bash run_livecodebench.sh <custom defined model name> <prompt batch size in parallel> <max output tokens> <local model server port>

Simple Evals

Simple Evals is a lightweight library for evaluating language models published from OpenAI. This eval includes "simpleqa", "mmlu", "math", "gpqa", "mgsm", "drop" and "humaneval" benchmarks.

Similarly, we support running simple evals against a local running OpenAI API compatible server. Once the local server is up, the following command can be used to run the Simple Evals:

bash run_simple_eval.sh <custom defined model name> <comma separated eval names> <max output tokens> <local model server port>

Customize quantization method for evaluation

An example of customized quantization config is shown in quantization_utils.py. It allows users to test accuracy of a custom method without the need of modifying the whole deployment framework, e.g., TensorRT-LLM, vLLM, SGLang, etc. Users can disable quantization of specific layers to debug the cause of accuracy drop, or explore a promising new quantization method.

python lm_eval_hf.py --model hf \
    --tasks <comma separated tasks> \
    --model_args pretrained=<HF model folder or model card>,parallelize=True \
    --quant_cfg MY_QUANT_CONFIG \
    --batch_size 4

Evaluating with LM-Eval-Harness via vLLM

The run_lm_eval_vllm.sh script provides a convenient way to run evaluations using the lm-evaluation-harness library against a model served with vLLM's OpenAI-compatible API endpoint.

This is useful for evaluating quantized models deployed with vLLM or any model served via its OpenAI API interface. More importantly, for new models that are neither supported natively by vLLM or Transformers, but Transformers compatible, they can still be evaluated with vLLM's endpoint! By Transformers compatible, it needs to satisfy the following:

The model directory must have the correct structure (e.g. config.json is present)
config.json must contain auto_map.AutoModel.
Customization should be done in the base model (e.g. in MyModel, not MyModelForCausalLM).

Prerequisites

Install vLLM: Follow the installation instructions at https://docs.vllm.ai/en/latest/getting_started/installation.html.

Usage

Start the vLLM OpenAI-compatible Server: In a separate terminal, launch the vLLM server with your desired model. For example:
```
# Example using vLLM's built-in server
vllm serve <your_model_name_or_path> \
    --port 8000 \
    --tensor-parallel-size <tp_size> # Adjust as needed
```
Replace <your_model_name_or_path> with the actual model identifier (e.g., Qwen/Qwen3-30B-A3B) and adjust the --port and --tensor-parallel-size if necessary. You may also need to disable vllm v1 by export VLLM_USE_V1=0 if you encounter issues.

To serve a modelopt quantized model, add --quantization modelopt, for example:
```
# Example using vLLM's built-in server
vllm serve nvidia/Llama-3.1-8B-Instruct-FP8 \
 --quantization modelopt
 --port 8000 \
 --tensor-parallel-size <tp_size> # Adjust as needed
```
To generate the quantized model such as nvidia/Llama-3.1-8B-Instruct-FP8, please refer to instructions here. Note currently modelopt quantized model support in vLLM is limited, we are working on expanding the model and quant formats support.
Make the script executable (if not already):
```
chmod +x run_lm_eval_vllm.sh
```
Run the evaluation script from the examples/llm_eval directory:
```
./run_lm_eval_vllm.sh <model_name> [port] [task]
```
- <model_name>: The name of the model being served (this is passed to lm_eval, e.g., Qwen/Qwen3-30B-A3B).
- [port]: (Optional) The port the vLLM server is listening on. Defaults to 8000. Note, it must match the number when launch the server.
- [task]: (Optional) The lm_eval task(s) to run. Defaults to mmlu. Can be a single task or a comma-separated list (e.g., "mmlu,hellaswag").

Examples

Evaluate Qwen3-30B-A3B on MMLU (default task and port):

# Assumes vLLM server running with Qwen/Qwen3-30B-A3B on port 8000
./run_lm_eval_vllm.sh Qwen/Qwen3-30B-A3B

Evaluate a model on Hellaswag using port 8001:

# Assumes vLLM server running with <model_name> on port 8001
./run_lm_eval_vllm.sh <model_name> 8001 hellaswag

Evaluate on multiple tasks:

# Assumes vLLM server running with <model_name> on port 8000
./run_lm_eval_vllm.sh <model_name> 8000 "arc_easy,winogrande"

Navigation Menu

Uh oh!

Evaluation scripts for LLM tasks

NeMo Evaluator

LM-Eval-Harness

Baseline

Quantized (simulated)

TensorRT-LLM

MMLU

Setup

Baseline

Quantized (simulated)

auto_quantize (simulated)

Evaluate with TensorRT-LLM

LiveCodeBench

Simple Evals

Customize quantization method for evaluation

Evaluating with LM-Eval-Harness via vLLM

Prerequisites

Usage

Examples

Uh oh!

FilesExpand file tree

README.md

Latest commit

History

README.md

File metadata and controls

Evaluation scripts for LLM tasks

NeMo Evaluator

LM-Eval-Harness

Baseline

Quantized (simulated)

TensorRT-LLM

MMLU

Setup

Baseline

Quantized (simulated)

auto_quantize (simulated)

Evaluate with TensorRT-LLM

LiveCodeBench

Simple Evals

Customize quantization method for evaluation

Evaluating with LM-Eval-Harness via vLLM

Prerequisites

Usage

Examples