`bridge.recipes.qwen_vl.gb200.qwen35_vl`#

GB200 functional recipes shared by Qwen3.5/Qwen3.6-VL 35B-A3B.

Module Contents#

Functions#

`qwen35_vl_27b_pretrain_16gpu_gb200_bf16_mock_config`	Return Qwen3.5-VL 27B language-and-projector pretraining for sixteen GB200 GPUs.
`_apply_qwen35_vl_35b_a3b_gb200_functional_defaults`	Apply shared settings for bounded GB200 functional runs.
`qwen35_vl_35b_a3b_sft_8gpu_gb200_bf16_functional_config`	Return shared Qwen3.5/Qwen3.6-VL 35B-A3B SFT for eight GB200 GPUs.
`_qwen35_vl_35b_a3b_long_context_common`	Build the shared 32-GB200 long-context SFT configuration.
`qwen35_vl_35b_a3b_sft_long_context_32gpu_gb200_bf16_config`	Return 128K BF16 Qwen3.5-VL SFT with TP1/CP8/EP8 on 32 GB200 GPUs.
`qwen35_vl_35b_a3b_sft_long_context_32gpu_gb200_fp8mx_config`	Return packed CLEVR2 128K MXFP8 SFT with FP8 parameter gather and CuTeDSL.
`qwen35_vl_35b_a3b_peft_8gpu_gb200_bf16_functional_config`	Return shared Qwen3.5/Qwen3.6-VL 35B-A3B LoRA for eight GB200 GPUs.

Data#

`_QWEN35_35B_A3B_REVISION`
`__all__`

API#

bridge.recipes.qwen_vl.gb200.qwen35_vl.qwen35_vl_27b_pretrain_16gpu_gb200_bf16_mock_config() → megatron.bridge.training.config.ConfigContainer#

Return Qwen3.5-VL 27B language-and-projector pretraining for sixteen GB200 GPUs.

This keeps the dense pretraining objective and trainable-parameter contract from the H100 recipe while using the measured GB200 execution layout. In particular, deterministic mode must remain opt-in: Qwen3.5’s Gated DeltaNet replaces its fused FLA kernels with a torch-native reference path when deterministic mode is enabled.

bridge.recipes.qwen_vl.gb200.qwen35_vl._apply_qwen35_vl_35b_a3b_gb200_functional_defaults( cfg: megatron.bridge.training.config.ConfigContainer, ) → None#: Apply shared settings for bounded GB200 functional runs.

bridge.recipes.qwen_vl.gb200.qwen35_vl.qwen35_vl_35b_a3b_sft_8gpu_gb200_bf16_functional_config() → megatron.bridge.training.config.ConfigContainer#: Return shared Qwen3.5/Qwen3.6-VL 35B-A3B SFT for eight GB200 GPUs.

bridge.recipes.qwen_vl.gb200.qwen35_vl._QWEN35_35B_A3B_REVISION#: ‘59d61f3ce65a6d9863b86d2e96597125219dc754’

bridge.recipes.qwen_vl.gb200.qwen35_vl._qwen35_vl_35b_a3b_long_context_common() → megatron.bridge.training.config.ConfigContainer#: Build the shared 32-GB200 long-context SFT configuration.

bridge.recipes.qwen_vl.gb200.qwen35_vl.qwen35_vl_35b_a3b_sft_long_context_32gpu_gb200_bf16_config() → megatron.bridge.training.config.ConfigContainer#

Return 128K BF16 Qwen3.5-VL SFT with TP1/CP8/EP8 on 32 GB200 GPUs.

Override with "dataset.path=$CLEVR2_ENERGON" to select the prepared Nemotron Image Training v3 clevr_2 Energon shards, and supply a pretrained checkpoint.

Prepare the shards with prepare_nemotron_image_v3.py <https://github.com/NVIDIA-NeMo/Megatron-Bridge/blob/main/tutorials/data/energon/prepare_nemotron_image_v3.py>_ using --subsets clevr_2 --skip-source-integrity-check after verifying the source revision; the built-in integrity manifest covers only turing. Native Energon packing uses one physical microbatch per 128K pack.

bridge.recipes.qwen_vl.gb200.qwen35_vl.qwen35_vl_35b_a3b_sft_long_context_32gpu_gb200_fp8mx_config() → megatron.bridge.training.config.ConfigContainer#

Return packed CLEVR2 128K MXFP8 SFT with FP8 parameter gather and CuTeDSL.