Support qwen3-vl for THD format and CP

[wplf]

What does this PR do ?

We will adopt @ISEEKYAN's work for qwen3-vl from mbridge into megatron-bridge, incorporating his additions such as THD format and CP support.

For now, THD format and BSHD format training is ready.

• bshd example script

python -m torch.distributed.run --nproc_per_node=8 \                                                                                                                                                                                                                                                                                           users/jinliangl cw-dfw-cs-001-vscode-01
    finetune_qwen_vl.py \
    --dataset-type hf \
    --data-path llava_video_178k \
    --recipe qwen3_vl_30b_a3b_finetune_config \
    --config-file conf/qwen3_vl_30b_a3b_pretrain_mfsdp_override_example.yaml dataset.pack_sequences_in_batch=false

• thd example script

cd $HOME2/repos/Megatron-Bridge/examples/recipes/qwen_vl
python -m torch.distributed.run --nproc_per_node=8 \
    finetune_qwen_vl.py \
    --dataset-type hf \
    --data-path llava_video_178k \
    --recipe qwen3_vl_8b_finetune_config  \
    --config-file conf/qwen3_vl_pretrain_override_example.yaml dataset.pack_sequences_in_batch=false

Model forward Validation

The output from Megatron-Bridge's Qwen3VL is now bitwise identical to that of M-Bridge's Qwen3VL.

MOE model

Dense model

Remain to do

• Add BSHD format
• Add THD format
• Add CP support for BSHD format and THD foramt
• Support Vision module's HF2Mcore and Mcore2HF ckpt conversion and verify it
• BSHD format forward output align with mbridge's forward output
• THD format forward output align with BSHD's forward output
• THD E2E training validation
• BSHD + CP E2E training validation
• THD + CP E2E training validation
• Vision model DP and text model CP support and validation
• Code review and refractor

Summary by CodeRabbit

Release Notes

• New Features

  • Enhanced vision-language model support with improved architecture for multimodal processing
  • Added profiling capabilities including memory snapshots and performance tracking
  • Extended logging and monitoring options for training visibility

• Configuration Updates

  • Updated example configurations with expanded training parameters, optimization settings, and vision model options

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[cuichenx]

Sorry, accidentally clicked approve. Please check my comments above

[yaoyu-33]

/ok to test 1cfa2dc

[wplf]

/ok to test 1cfa2dc

[wplf]

/ok to test e18e64d

[shifangx]

/ok to test e18e64d

[shifangx]

/ok to test 3eb6ebc

[wplf]

/ok to test 3eb6ebc

[ko3n1g]

Tests will be submitted soon in a followup