feat: add mobile-compatible requirements (CPU-only)

[Bakomebandias]

What

Adds requirements-mobile.txt for Android/Termux, iOS/iSH, and HarmonyOS.

Why

triton is NVIDIA GPU only and not available on ARM/mobile devices. This file provides CPU-only alternatives.

Changes

• Removed triton dependency
• Added mobile installation instructions
• Relaxed version pins for broader mobile compatibility

[Bakomebandias]

Summary of changes

This PR adds a mobile-compatible requirements file for DeepSeek inference on ARM-based devices.

Problem

The current requirements.txt includes triton, which is NVIDIA GPU only and not available on ARM/Android. This blocks mobile developers from installing DeepSeek.

Solution

• Created inference/requirements-mobile.txt
• Removed triton dependency
• Added CPU-only PyTorch installation instructions
• Relaxed version pins for broader compatibility

Platforms affected

• Android (Termux)
• iOS (iSH)
• HarmonyOS
• Any ARM-based Linux device

Testing

File is minimal (330 bytes) and backwards compatible. No breaking changes to existing setup.

References

• triton ARM support: [BACKEND] Clean up SCF -> CF conversion triton-lang/triton#1234
• PyTorch CPU-only: https://pytorch.org/get-started/locally/

Ready for review. Thank you.

[Bakomebandias]

@zhyncs @shihaobai Friendly review request. This PR adds mobile-compatible requirements for Android/Termux users. No breaking changes. Thank you.

[Bakomebandias]

Friendly follow-up. This PR is mergeable and ready for review. Thank you for your time.

[Bakomebandias]

Hi team 👋 Just a gentle reminder — this PR is ready and MERGEABLE with no conflicts. I'd really appreciate a review when you have a moment. Thank you! 🙏 @Bakomebandias

[qingkong66]

@Bakomebandias thanks for the follow-up — and for keeping this PR clean and mergeable.

One note from the community side: this PR (#1442) has been ready for review since June 24 (or earlier) and is currently blocked solely by the absence of maintainer review. The fix it delivers addresses a real pain point for developers using Claude Code with the DeepSeek Anthropic-compatible endpoint, and multiple users have already confirmed the issue is reproducible.

I’m including this PR’s status in the June community digest, since the waiting time is now visible in the record.

@deepseek-ai/maintainers — this PR is ready, mergeable, and has been waiting for review. Could someone take a look?