[bugfix] support MTP-only pipeline stages

[taking-lying-flat]

Summary

• build the complete model-specific decoder layer specs in the Qwen3.5/Qwen3Next GDN loader when constructing MTP
• reuse the final global decoder layer spec for MTP, independent of the current PP/VP stage layout
• keep the existing local decoder spec path and all other model loaders unchanged

Root cause

The Qwen3.5 397B pipeline layout can place MTP on a stage with no local decoder layers. Passing that empty local decoder block to Megatron's MTP builder causes it to index an empty layer_specs list.

Using Megatron's generic decoder helper directly would lose Qwen3.5's heterogeneous GDN/attention layer structure. The loader now generates its complete experimental-attention specs, applies the same Qwen-specific replacements used for local layers, and passes the final global decoder layer spec to the existing MTP builder.

Validation

• flake8 and yapf on the changed file
• pre-commit text hooks on the changed file
• Python syntax compilation

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[HowardZorn]

我看了看Megatron-LM的情况，这里他们用的是一个不切vp_stage的layout（但是mcore-bridge这里没有这样的函数get_gpt_decoder_layer_specs()），这里您的选择是去找第一个decoder layers，我觉得也可以。不过后续可能需要对齐一下

    if args.mtp_num_layers is not None:
        assert not (config.transformer_impl == "inference_optimized")
        if (
            hasattr(transformer_layer_spec, 'layer_specs')
            and len(transformer_layer_spec.layer_specs) == 0
        ):
            # Get the decoder layer spec explicitly if no decoder layer in the last stage,
            # Only happens with block spec (TransformerBlockSubmodules) when using MoE.
            transformer_layer_spec_for_mtp = _get_transformer_layer_spec(use_te, config)
        else:
            # Define the decoder block spec
            if args.experimental_attention_variant is not None:
                decoder_layer_specs = (
                    get_transformer_layer_with_experimental_attention_variant_spec(config=config)
                )
            else:
                decoder_layer_specs = get_gpt_decoder_layer_specs(
                    config,
                    use_transformer_engine=use_te,
                    normalization=args.normalization,
                    qk_l2_norm=args.qk_l2_norm,
                    vp_stage=vp_stage,
                )
            transformer_layer_spec_for_mtp = decoder_layer_specs[-1]
        # Use spec of the last layer in decoder block as spec of the transformer layer in MTP
        mtp_block_spec = get_gpt_mtp_block_spec(
            config, transformer_layer_spec_for_mtp, use_transformer_engine=use_te, vp_stage=vp_stage
        )

[HowardZorn]

@taking-lying-flat 我逐渐理解这个工程其实会变成每个experiment attention模型的loader都需要加料的情况，这样工程量可能有点大

[taking-lying-flat]

@HowardZorn。 确实 不敢大重构 按照 Megatron-LM. 出现bug 就修复吧

[HowardZorn]

@HowardZorn。 确实 不敢大重构 按照 Megatron-LM. 出现 bug 就修复吧

@taking-lying-flat 感觉我们需要 @Jintao-Huang 来看看怎么样处理了。