[model] fix: correct GLM-4.5V inference parallelism for 46-layer model

[yaoyu-33]

What

• Update GLM-4.5V example inference parallelism to across all commands.
• Keep 8-GPU total while making the pipeline split valid for 46 layers.
• Add temporary workaround in VLM and text generation conversion scripts to avoid inference failure.

Why

GLM-4.5V has 46 layers, which is not divisible by and causes assertion failure during inference.

Validation

• fix end of files.....................................(no files to check)Skipped
  trim trailing whitespace.............................(no files to check)Skipped
  ruff.................................................(no files to check)Skipped
  ruff.................................................(no files to check)Skipped
  ruff-format..........................................(no files to check)Skipped
  Disallow '_' in Markdown filenames...................(no files to check)Skipped
• Branch push successful

Summary by CodeRabbit

Release Notes

• Bug Fixes

  • Fixed model parallel layer configuration issue in Megatron model inference pipelines to prevent errors during CUDA operations when using HuggingFace-converted models.

• Chores

  • Updated parallelism distribution settings in example inference scripts for improved resource allocation: pipeline parallelism reduced from 4 to 2, expert parallelism increased from 2 to 4.

[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.

[yaoyu-33]

/ok to test ccb2456