[0.3.35] Release Note: Gemma 4 series & LFM 2.5-VL Support, OpenAI OpenAPI Alignment and Logging Architecture Migration

[JamePeng]

[0.3.35] Release Note: Gemma 4 series & LFM 2.5-VL Support, OpenAI OpenAPI Alignment and Logging Architecture Migration

This release introduces comprehensive support for the Gemma 4 series and LFM 2.5-VL, alongside major alignments with the latest OpenAI API specifications (including native audio and structured outputs). Under the hood, we've executed a significant migration to the upstream GGML logging architecture and synchronized with the latest llama.cpp upstream APIs.

✨ New Features & Highlights

Gemma 4 Series Integration

• Complete Gemma4ChatHandler Implementation: Added robust support for Gemma 4 featuring specific token handling (<|turn>, <|channel>), a complex Jinja2 template for nested tool/function schema formatting, and <turn|> configured as the primary stop sequence.
• Multimodal Capabilities: Enabled multimodal content injection supporting image_url, audio_url, and input_audio (including base64 reconstruction).
• Reasoning/Thinking Controls: Integrated thinking controls via the enable_thinking toggle and <|channel>thought formatting.

LFM 2.5-VL Support

• New Handler: Implemented LFM25VLChatHandler to support LFM2.5-VL (Thanks to @alcoftTAO).

llama_types: OpenAI OpenAPI Spec Alignment

• Structured Outputs: Replaced the Anyscale-specific JSON schema with the official OpenAI json_schema response format.
• Audio & File Support: Added input_audio and file types to request message content parts. Response messages now support audio, refusal, and annotations (e.g., URL citations) fields.
• Role & Filter Definitions: Added the developer role, introduced content_filter to finish reasons, and strictly defined global ChatCompletionRole.
• Updated llama_types.py to link to the latest OpenAPI spec.

🐛 Bug Fixes

• Gemma 4 Stop Sequences: Expanded the generation stop criteria to include GEMMA4_EOS_TOKEN and GEMMA4_STR_TOKEN, aligning stopping logic with generation_config.json to prevent over-generation when initiating a tool response.
• Qwen 3.5: Fixed typos in the Qwen 3.5 chat template (Reported by @abdullah-cod9).

🛠 Refactoring & Upstream Sync

• GGML Logging Architecture Migration: Transitioned from the deprecated llama_log_callback to ggml_log_callback in _logger.py, fully aligning with the upstream GGML logging architecture. Renamed callback references across the codebase (including MTMD context initialization).
• ggml-base Integration: Added support for loading the new ggml-base shared library alongside ggml, including ctypes bindings for ggml_log_get, ggml_log_set, and ggml_set_zero.
• Upstream Sync: Updated llama.cpp to commit [58190cc84d846d8575ba26e8486bc29d9fd8ad55]
  • Synchronized llama.cpp llama/mtmd API bindings to version 20260402.

📚 Documentation Updates

• Gemma 4 enable_thinking Details: Updated the docstrings and __init__ args to clarify that the enable_thinking toggle is exclusively supported by Gemma 4 31B and 26BA4B variants. Explicitly noted that E2B and E4B models do not currently support this feature to prevent user configuration errors.
• Updated README.md to reflect recent changes.

---

Full Changelog: a184583...232092e

— JamePeng

[Kendi-droid]

Thanks for all your work on this. It looks like the v0.3.35-cu126-Basic-win-20260406 WHL for cp312 was missed. Could you add that as well?

[JamePeng]

It will be recompiled once. The main reason is that the previous compilation tasks for cu124-cu126 were interrupted by GitHub because the compilation time was too long. I separated them in the early morning and waited for the new day to start compiling again.

[JamePeng]

Python versions 3.11-3.14 have already been compiled, but this time version 3.10 has been dropped and is also in the process of being compiled.