[Enhance] Support internal metrics for gdn A_log and norm & vl model in general

[nil0x9]

Summary

This PR adds internal metrics monitoring support for Gated DeltaNet (GDN) architecture, specifically tracking min/max statistics for A_log parameters and FusedRMSNormGated modules. It also adds general support for vision-language models by properly extracting the language model component.

Implementation Details

New metrics fields:

• Added weight_min and weight_max to track parameter value ranges across layers
• Added monitor_gdn_stats configuration flag (default: None)

Key changes:

1. calculate_module_weight_min_max method: Computes min/max values across module parameters with proper handling of:
   - DTensor (converts to local tensor before computation)
   - Distributed reduction across all ranks using dist.all_reduce
   - Both nn.Module and torch.Tensor inputs
2. VL model support: The recorder now automatically extracts language_model from VL models if present, allowing the metrics system to work seamlessly with
   multi-modal architectures.
3. GDN-specific monitoring: Automatically registers hooks for:
   - FusedRMSNormGated modules (tracks min/max of gating parameters)
   - A_log attributes (tracks min/max of log-space state transition parameters)

[nil0x9]

@claude add PR description

[nil0x9]

@claude review

[claude]

Claude: This change silently replaces self.model for all VL models, which means every other metric method (calculate_module_weight_rms, register_attn_output_hook, pop_metrics' named_modules() loop, MoE checks, etc.) now operates on language_model instead of the full composite model.

This is a behavioral change for all existing VL model users, not just GDN metrics. Consider whether this is intentional. If the goal is only to iterate over language-model modules for GDN stats, it would be safer to scope the unwrapping narrowly:

Suggested change

	self.model = model.language_model if hasattr(model, "language_model") else model
	self.model = model
	self._language_model = model.language_model if hasattr(model, "language_model") else model

Then use self._language_model only in the GDN-specific iteration in pop_metrics.

[claude]

Claude: ## Review Summary

This PR adds internal metrics monitoring for GDN (Gated DeltaNet) A_log parameters and FusedRMSNormGated modules, plus VL model support via language_model unwrapping.

Issues

Critical

• self.model reassignment (line 82): Replacing self.model with model.language_model for VL models changes behavior for all existing metrics (weight RMS, attention stats, MoE stats), not just the new GDN metrics. This is a silent behavioral change that could break or alter existing monitoring. Recommend scoping the unwrapping narrowly.

Warning

• Top-level import of gated_deltanet (line 16): Will break internal_metrics.py import in environments without the fla dependency. Should be a conditional import.
• Per-module all_reduce in calculate_module_weight_min_max: Two blocking collectives per monitored module could be expensive. Consider batching min/max across all modules and doing a single reduction.

Nit

• Type annotation nn.Module | nn.Parameter | torch.Tensor is redundant — nn.Parameter is a torch.Tensor, so nn.Module | torch.Tensor suffices.
• Inline DTensor import could be moved to module level since it's from standard PyTorch.
• Missing PR description — would help reviewers understand the motivation and scope.

Verdict

REQUEST_CHANGES