[fix] torch 2.8.0 dtype, DLPack

[EduardDurech]

Fixes 2 errors introduced in #8836

`dtype`

  File "/sglang/python/sglang/srt/managers/scheduler.py", line 2475, in run_scheduler_process
    scheduler = Scheduler(
                ^^^^^^^^^^
  File "/sglang/python/sglang/srt/managers/scheduler.py", line 312, in __init__
    self.tp_worker = TpWorkerClass(
                     ^^^^^^^^^^^^^^
  File "/sglang/python/sglang/srt/managers/tp_worker_overlap_thread.py", line 67, in __init__
    self.worker = TpModelWorker(
                  ^^^^^^^^^^^^^^
  File "/sglang/python/sglang/srt/managers/tp_worker.py", line 84, in __init__
    self.model_runner = ModelRunner(
                        ^^^^^^^^^^^^
  File "/sglang/python/sglang/srt/model_executor/model_runner.py", line 242, in __init__
    self.initialize(min_per_gpu_memory)
  File "/sglang/python/sglang/srt/model_executor/model_runner.py", line 331, in initialize
    self.init_memory_pool(
  File "/sglang/python/sglang/srt/model_executor/model_runner.py", line 1251, in init_memory_pool
    self.token_to_kv_pool = MHATokenToKVPool(
                            ^^^^^^^^^^^^^^^^^
  File "/sglang/python/sglang/srt/mem_cache/memory_pool.py", line 203, in __init__
    self._create_buffers()
  File "/sglang/python/sglang/srt/mem_cache/memory_pool.py", line 252, in _create_buffers
    self.data_strides = torch.tensor(
                        ^^^^^^^^^^^^^
RuntimeError: Could not infer dtype of numpy.int64

`DLPack`

  File "/sglang/python/sglang/srt/managers/scheduler.py", line 2475, in run_scheduler_process
    scheduler = Scheduler(
                ^^^^^^^^^^
  File "/sglang/python/sglang/srt/managers/scheduler.py", line 312, in __init__
    self.tp_worker = TpWorkerClass(
                     ^^^^^^^^^^^^^^
  File "/sglang/python/sglang/srt/managers/tp_worker_overlap_thread.py", line 67, in __init__
    self.worker = TpModelWorker(
                  ^^^^^^^^^^^^^^
  File "/sglang/python/sglang/srt/managers/tp_worker.py", line 151, in __init__
    self.random_seed = broadcast_pyobj(
                       ^^^^^^^^^^^^^^^^
  File "/sglang/python/sglang/srt/utils.py", line 1068, in broadcast_pyobj
    tensor_data = torch.ByteTensor(
                  ^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/utils/dlpack.py", line 117, in from_dlpack
    dlpack = ext_tensor.__dlpack__()
             ^^^^^^^^^^^^^^^^^^^^^^^
BufferError: Cannot export readonly array since signalling readonly is unsupported by DLPack (supported by newer DLPack version).

[gemini-code-assist]

Summary of Changes

Hello @EduardDurech, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request addresses two critical errors that emerged with torch version 2.8.0, specifically related to dtype inference and DLPack buffer handling. These fixes ensure compatibility and prevent runtime crashes when using the updated PyTorch library, maintaining the stability of the system.

Highlights

• Fixing dtype inference: Explicitly set the data type for a torch.tensor creation to torch.int32 to resolve a RuntimeError where dtype could not be inferred from numpy.int64, ensuring compatibility with torch 2.8.0.
• Resolving DLPack BufferError: Ensured that a copy of the NumPy array is made when converting serialized_data to a torch.ByteTensor, which fixes a BufferError related to DLPack not supporting readonly arrays.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments or fill out our survey to provide feedback.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request addresses two compatibility issues with PyTorch 2.8.0. The first fix explicitly sets the dtype to torch.int32 when creating a tensor from numpy integers to resolve a type inference error. The second fix adds a .copy() to a numpy array created from a buffer to make it writable, which resolves a DLPack error related to read-only arrays. Both changes are correct and effectively fix the described issues. I have one suggestion for future-proofing.

[gemini-code-assist]

Using torch.int32 is a valid fix for the type inference issue. However, given that the error message indicates the inferred type was numpy.int64, it might be safer to use torch.int64 (or torch.long) to prevent potential overflow issues in the future with models that have very large tensor strides. While int32 is sufficient for current models, int64 would be more robust against future edge cases.

Suggested change

	dtype=torch.int32,
	dtype=torch.int64,

[EduardDurech]

@merrymercy

[FlamingoPg]

Hi @EduardDurech , need rebase master

[EduardDurech]

@FlamingoPg done