add support of torch profiling

[STwangyingrui]

No description provided.

[gemini-code-assist]

Code Review

This pull request introduces a PyTorch Trace Profiling module, featuring a kernel inspector tool, Docker bridge script, and detailed documentation. It also integrates magi_compiler support for NeoPP and Qwen Image models, optimizing inference through custom op registration and Dynamo configuration. Furthermore, a new audio_io utility is implemented to ensure consistent audio loading across various backends. Review feedback highlights several performance optimization opportunities, including specifying target devices for tensor creation to minimize host-to-device synchronization, improving audio loading efficiency by reading directly as float32, and refactoring the profiler initialization to reduce redundant environment variable parsing.

[gemini-code-assist]

Creating these tensors inside the per-layer loop without specifying a device can lead to unnecessary host-to-device synchronizations and overhead in eager mode. It is recommended to specify the device to ensure they are created on the same device as the input tensors.

Suggested change

	cu_seqlens_q=torch.tensor([0, seq_len_q], dtype=torch.int32),
	cu_seqlens_kv=torch.tensor([0, seq_len_k], dtype=torch.int32),
	cu_seqlens_q=torch.tensor([0, seq_len_q], dtype=torch.int32, device=query_states.device),
	cu_seqlens_kv=torch.tensor([0, seq_len_k], dtype=torch.int32, device=query_states.device),

[gemini-code-assist]

torch.arange can be created directly on the target device to avoid an extra host-to-device transfer.

Suggested change

	positions = torch.arange(L, device="cpu", dtype=torch.long).to(xq.device, non_blocking=True)
	positions = torch.arange(L, device=xq.device, dtype=torch.long)

[gemini-code-assist]

The current logic always performs a transpose and copy, which is inefficient when channels_first is False. Additionally, reading as float32 directly from soundfile is more efficient than reading as float64 and converting later.

Suggested change

	data, sample_rate = sf.read(uri, always_2d=True, start=start, stop=stop)
	tensor = torch.from_numpy(data.T.copy()).float()
	if not channels_first:
	tensor = tensor.transpose(0, 1)
	return tensor, sample_rate
	data, sample_rate = sf.read(uri, always_2d=True, start=start, stop=stop, dtype='float32')
	if channels_first:
	tensor = torch.from_numpy(data.T.copy())
	else:
	tensor = torch.from_numpy(data)
	return tensor, sample_rate

[gemini-code-assist]

TorchTraceProfiler.from_env() parses environment variables on every call to infer_calculating. Since these variables are typically static for the duration of the process, it is more efficient to initialize the profiler once in the __init__ method and reuse it.