qwen3.5 lora compatibility (needs SGLang + Megatron-Bridge patches as well)

[artem-osmosis]

• add script for lora finetuning qwen3.5 w/ support for q_proj,k_proj,v_proj
• Add a function to check if the model is Qwen3.5, and ignore mtp/vision modules when applying lora adapters if that's the case

for lora + qwen3.5 + megatron bridge to work, we also need patches to both the sglang-miles fork of sglang, and the bridge branch of radixark/Megatron-Bridge

changes for sglang:

• upstream qwen3_5.py
• patch for shape error when determining lora_b shape for q/k/v proj adapters:
  AssertionError: LoRA buffer shape torch.Size([1152, 32]) does not match weight shape torch.Size([1280, 32]).

changes for Megatron-Bridge:

• handle both transformer_layer and mtp_model_layer keys for mtp megatron modules

Megatron-Bridge and sglang PRs:

• add fix for qwen35 mtp layer name Megatron-Bridge#2
• [fix] upstreams for qwen3_5.py and qwen3.5 lora on qkv layers fix.  sgl-project/sglang#24917

[gemini-code-assist]

Code Review

This pull request adds a LoRA training script for Qwen3.5-35B-A3B and introduces logic to filter target modules for Qwen3.5 models. Key feedback includes fixing a typo in the PYTHONUNBUFFERED environment variable, adjusting the module path pattern to ensure compatibility with text-only models, and recommending a more reliable way to identify model types than string matching on file paths.

[gemini-code-assist]

The prefix *language_model.decoder.layers.*.*. might be too restrictive for text-only Qwen3.5 models. In the Megatron backend (as seen in megatron_to_hf/qwen3_5.py), the module hierarchy typically starts with decoder.layers without a language_model wrapper. Including language_model in the pattern may cause LoRA adapters to fail to match any modules in non-VL models. Consider using a more general pattern like *decoder.layers.*.*. to target the main transformer layers while still excluding MTP modules (which are usually under mtp.layers).

Suggested change

	target if "language_model" in target else f"*language_model.decoder.layers.*.*.{target}"
	target if "language_model" in target else f"*decoder.layers.*.*.{target}"

[gemini-code-assist]

The environment variable should be PYTHONUNBUFFERED instead of PYTHONBUFFERED. Standard Python uses PYTHONUNBUFFERED to disable stream buffering, which is what the comment above indicates is the goal. Ensure the variable name matches the exact spelling expected by the system to avoid the override being ignored.

Suggested change

	export PYTHONBUFFERED=16
	export PYTHONUNBUFFERED=1

References

1. Ensure environment variable names match the exact spelling expected by the system, even if it contains a typo, to avoid the override being ignored.

[gemini-code-assist]

Checking for "Qwen3.5" in the full checkpoint path is fragile, as the path might contain this string for unrelated reasons (e.g., a parent directory name) or might not contain it if the model directory was renamed. It would be more robust to retrieve model parameters and architecture details from the model configuration rather than hardcoding checks against file paths.

References

1. Model parameters should be retrieved from the model configuration rather than being hardcoded.