fix: unify qwen tts cache dir for tokenizer loading on Windows

[seidenbergerscott]

Summary

• route Qwen TTS rom_pretrained through a single HF cache root
• pass cache_dir=HF_HUB_CACHE to avoid split cache between hub and transformers
• switch deprecated orch_dtype arg to dtype for compatibility

Why

On Windows local setups, model assets can split between .hf-cache/hub and .hf-cache/transformers, causing speech_tokenizer/preprocessor_config.json load errors and 500s during generation.

Summary by CodeRabbit

• Bug Fixes
  • Resolved Text-to-Speech model loading inconsistencies by standardizing cache directory management across all computational environments.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

Configuration used: defaults

Review profile: CHILL

Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between 38bf96f and e1c6bba.

📒 Files selected for processing (1)

• backend/backends/pytorch_backend.py

---

📝 Walkthrough

Walkthrough

The PyTorchTTSBackend model loading is updated to centralize Hugging Face Hub cache directory handling by importing HF_HUB_CACHE and applying it to model loading calls. Additionally, the torch_dtype parameter is replaced with dtype, specifying float32 for CPU and bfloat16 for GPU paths.

Changes

Cohort / File(s)	Summary
TTS Model Loading Configuration backend/backends/pytorch_backend.py	Added HF_HUB_CACHE import and tts_cache_dir variable; applied cache_dir parameter to Qwen3TTSModel.from_pretrained calls in both CPU and non-CPU branches; replaced torch_dtype with dtype (float32 for CPU, bfloat16 for non-CPU).

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Poem

🐰 A cache most organized, now takes its place,
Through Hub's own paths, we set the pace,
With dtype specified, both paths aligned,
Float and bfloat, perfectly designed! 🎉

🚥 Pre-merge checks | ✅ 3

✅ Passed checks (3 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes the main change: unifying the Qwen TTS cache directory to fix a Windows-specific tokenizer loading issue.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment

Tip

Try Coding Plans. Let us write the prompt for your AI agent so you can ship faster (with fewer bugs).
Share your feedback on Discord.

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[seidenbergerscott]

Related issues: #217, #216

This PR fixes a Windows cache split between Hugging Face hub and transformers cache paths that can cause model load failures (missing speech_tokenizer/preprocessor_config.json) and generation 500s after download.

[jamiepine]

Closing — applied the cache_dir portion of this fix directly to main as part of the 0.4.0 speedrun. Thanks for the diagnosis and fix!

• Cache unification applied (the actual Windows bug)
• Skipped the torch_dtype= → dtype= rename since our minimum transformers version (4.36) doesn't yet accept the dtype alias. Can revisit once we bump the minimum.