VLM Single-Turn RL#

Training VLMs with Megatron on single-turn reasoning task using GRPO on the GEO3K dataset. We used processed version here.

Supported models:

  • Qwen3.5-35B-A3B

  • NVIDIA Nemotron 3.5 Super VL

Note: Please make sure the cudnn version in the environment is 9.16.0.29 to prevent severe performance regression in conv3d in torch 2.9 mentioned in pytorch/pytorch#168167. Otherwise, you can reinstall cudnn with:

pip install nvidia-cudnn-cu12==9.16.0.29

Qwen3.5 uses vime’s native Megatron language model plus the Transformers vision model. Vision parameters retain their HuggingFace names and are loaded and synchronized directly.

Nemotron 3.5 Super VL (nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16) uses MCore’s HybridModel (Mamba2, MoE and attention layers) as the language model, plus the Transformers (remote-code) RADIO vision tower and projector. Its vision parameters are renamed when they are loaded and when they are synchronized: the checkpoint’s fused attn.qkv is split into query/key/value for the Transformers module and fused again when syncing to vLLM, and the projector’s mlp1.* keys are mapped to vision_projector.*. C-RADIO ships no LayerScale weights and layerscale_value is 1.0: the trainer removes these identity operations, while vLLM preserves their constant values as buffers across sleep.

GEO3K VLM rollout raw reward

Data Preparation (For SFT Training)#

The geo3k_imgurl dataset contains:

  • problem: The math problem text (string)

  • answer: The answer (string, e.g., “270”)

  • images: Image data (list)

For SFT training, we need to format the answer field for \boxed{} format and the messages. You can use the following script to format the answer field:

from datasets import load_dataset
import pandas as pd

ds = load_dataset("chenhegu/geo3k_imgurl", split="train")

def format_answer(answer: str) -> str:
    """Format answer to include \\boxed{} format."""
    return f"Answer: \\boxed{{{answer}}}"

def process_sample(sample):
    formatted_answer = f"Answer: \\boxed{{{sample['answer']}}}"
    
    sample["messages"] = [
        {"role": "user", "content": sample["problem"]},
        {"role": "assistant", "content": formatted_answer}
    ]
    return sample

ds = ds.map(process_sample)
ds.to_parquet("/root/datasets/geo3k_imgurl/train_formatted.parquet")

Reproduce#

export WANDB_API_KEY=your_wandb_api_key

./examples/geo3k_vlm/run_geo3k_qwen35.sh

Configuration#

Environment Variable

Default

Description

VIME_SCRIPT_DATASET_NAME

chenhegu/geo3k_imgurl

HuggingFace dataset name

VIME_SCRIPT_NUM_GPUS

8

Number of GPUs

VIME_SCRIPT_EXTERNAL_RAY

0

Use external Ray cluster (1 to enable)

Qwen3.5 Series#

We provide a native example for Qwen3.5-35B-A3B. It loads the Transformers ViT directly and converts only the Megatron language-model parameters. To support another Qwen3.5 model, add its model config under scripts/models/ and update the example.

The native path supports tensor, pipeline, context, sequence, and expert parallelism. Context parallelism follows Megatron’s packed THD zigzag layout.

For GDN training, use --micro-batch-size 1 and remove --use-dynamic-batch-size.

Nemotron 3.5 Series#

Compared with run_geo3k_qwen35.sh, a Nemotron 3.5 Super VL run needs the following (the training hyperparameters can stay the same):

  • Model config and spec. Pass --spec vime_plugins.models.nemotron_h_vl get_nemotron_h_vl_model_provider together with the model config scripts/models/nemotron3.5-super-120B-A12B.sh.

  • Four nodes, expert parallelism 32. The model has 512 routed experts, and EP x ETP must divide the number of GPUs, so use --expert-model-parallel-size 32 on 32 GPUs (TP2 x CP2, DP8). A single node of 8 GPUs fits only the small proxy model. A multi-node run needs a Ray cluster that already spans the nodes, not the script’s own one-node ray start --head.

  • mamba-ssm and causal-conv1d must be installed. MCore’s MambaMixer requires them for the Mamba2 layers. They replace pip install -U transformers, and the docker/Dockerfile image builds them. The vLLM side needs the patches in docker/patch/latest/ (vllm-nemotron-reload.patch, vllm-tito-placeholder-mask.patch), which the image applies.

  • Offload the optimizer states (--optimizer-cpu-offload --use-precision-aware-optimizer on Megatron 0.20) to fit 80 GB GPUs. Vime parses Megatron’s arguments with ignore_unknown_args=True, so a flag spelling the installed Megatron does not have is dropped without an error and the offload silently stops happening.

  • MTP. The checkpoint has one MTP layer, used by vLLM as the MTP draft and trained by Megatron: --mtp-num-layers N --mtp-use-repeated-layer --enable-mtp-training --mtp-loss-scaling-factor 0.2, with one /*E segment per depth appended to --hybrid-override-pattern, and --vllm-speculative-config '{"method":"mtp","num_speculative_tokens":1,"moe_backend":"triton"}'. The weight sync keeps the draft on the trained layer.

  • No Qwen-specific vLLM pins. --vllm-moe-backend triton and --vllm-mm-encoder-attn-backend TORCH_SDPA are for Qwen’s gated experts and vision tower; Nemotron’s experts are not gated (squared ReLU, no gate_proj) and its vision tower is served by vLLM’s own radio.py.

For more details, see tests/test_nemotron35_super_vl_proxy_e2e.py, which runs the same GRPO path on a random-weight proxy of the model on one node of 8 GPUs.

Notes#

Reward Model Configuration#

We experimented with three reward model configurations:

  1. A geo3k-specific RM with tolerance=0.05 (to handle rounding in ground truth labels)

  2. A geo3k-specific RM with tolerance=0.0 (strict matching)

  3. The default math RM

All three performed similarly, so we use the default math RM for simplicity.

Numerical Precision with Non-Binary Rewards#

Our initial geo3k-specific verifier produced “format scores” (0 and 0.9) instead of clean binary rewards. Under fp32, fractional values like 0.9 can’t be exactly represented, so when all samples in a group have the same reward, reward - mean doesn’t equal zero—creating spurious gradient signal.

We fixed this by switching to the default math RM with clean binary 0/1 rewards. If you encounter similar precision issues with non-binary rewards, you can change the reward tensor dtype from torch.float to torch.float16 in vime/ray/rollout.py (_post_process_rewards method) to truncate precision artifacts.

B200#

On Blackwell (SM100), vLLM automatically dispatches the ViT encoder to FlashAttention 4 (or FA2 fallback) — no manual override is needed (vllm/v1/attention/backends/fa_utils.py:81). If you hit a kernel issue on a specific model, you can force SDPA with --vllm-mm-encoder-attn-backend TORCH_SDPA. The HF-side --attn-implementation flash_attention_2 flag is still relevant when the model is loaded via Hugging Face Transformers.