Fix packed sequence length handling for Megatron-Core 0.16.0/0.16.1

[hazelduan]

Why

In Megatron-Core 0.16.0/0.16.1, PackedSeqParams.max_seqlen_q is defined as a Python int. The Qwen3.5 and Qwen3-Next THD packed-sequence paths called .item() on this value, causing:
AttributeError: 'int' object has no attribute 'item'

Changes

Convert max_seqlen_q with int(...) before allocating hidden states and attention masks. This supports both Python integers and scalar tensors without changing numerical computation.

Validation

• Qwen3.5 training ran successfully for multiple steps on A3 with Megatron-Core 0.16.1.
• Python compilation checks passed for both modified files.
• Python int and scalar Tensor conversion cases passed.

GPU Compatibility

This change is device-agnostic and remains compatible with GPU training. max_seqlen_q is sequence-length metadata used only to construct tensor shapes; converting it with int(...) works with both Python integers and single-element CPU/CUDA/NPU tensors.

Experiments

Qwen3.5 training loss decreased and converged as expected. The change is also compatible with Megatron-Core 0.16.0–0.19.x, where `max_seqlen_q` remains integer metadata. Using `int(...)` supports both Python integers and scalar tensors without affecting model computation.