fix: add required python packages to support generic huggingface models

[andyzhangx]

Reason for Change:

fix: add required python packages to support generic huggingface models

• decord is required by 'baidu/ernie-4.5-vl-28b-a3b-pt'
• transformers>=4.57.1 is required by zai-org/GLM-4.1V-9B-Thinking
• timm is required by 'google/gemma-3n-e2b-it'
• torch-c-dlpack-ext is required by 'deepseek-ai/DeepSeek-V3.1' and 'ai21labs/ai21-jamba-1.5-mini' and a lot of other models
• num2words is required by 'HuggingFaceTB/SmolVLM2-2.2B-Instruct'
• mistral-common[soundfile] is required by mistralai/Voxtral-Mini-3B-2507
• tensorflow is required by 'allenai/Molmo-7B-D-0924'
• fla-core is required by moonshotai/Kimi-Linear-48B-A3B-Instruct

Requirements

• added unit tests and e2e tests (if applicable).

Issue Fixed:

Notes for Reviewers:

[kaito-pr-agent]

Title

Add required Python packages to support generic HuggingFace models

---

Description

• Update transformers version constraint to >=4.57.1

• Add required dependencies for vision and audio models support

• Include decord, timm, torch-c-dlpack-ext, mistral-common[soundfile], num2words, fla-core, and tensorflow

• Enable compatibility with various HuggingFace models (ERNIE, GLM, Gemma, DeepSeek, etc.)

---

Changes walkthrough 📝

Relevant files

Dependencies

requirements.txt Update dependencies for HuggingFace model compatibility    presets/workspace/dependencies/requirements.txt Updated transformers version from >=4.56.0 to >=4.57.1 Added decord for video processing support Added fla-core for flash attention mechanisms Added mistral-common[soundfile] for audio model support Added num2words for number-to-text conversion Added tensorflow as a dependency Added timm for image models Added torch-c-dlpack-ext for tensor operations +10/-1

---

Need help?

Type /help how to ... in the comments thread for any questions about PR-Agent usage.

Check out the documentation for more information.

[kaito-pr-agent]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 2 🔵🔵⚪⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Missing version constraints Newly added packages (decord, fla-core, num2words, tensorflow, timm, torch-c-dlpack-ext) lack version pinning or upper bounds. This could lead to future breaking changes or incompatible dependency versions. Consider adding minimum version constraints and/or upper bounds for reproducible builds. decord fla-core mistral-common[soundfile] num2words tensorflow timm torch-c-dlpack-ext TensorFlow compatibility concerns Adding tensorflow alongside torch (PyTorch) may significantly increase the container image size and could introduce CUDA/cuDNN version conflicts. Verify compatibility between TensorFlow, PyTorch 2.9.0, and CUDA versions used by vLLM 0.12.0 to avoid runtime errors. tensorflow

[Copilot]

Pull request overview

Updates the workspace Python dependency set used by KAITO’s TFS/vLLM preset images to support additional “generic” Hugging Face models that require extra third‑party packages.

Changes:

• Bumps the minimum transformers version to >=4.57.1,<5.
• Adds several new model-specific Python dependencies (e.g., decord, timm, torch-c-dlpack-ext, etc.) to the shared requirements.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[kaito-pr-agent]

PR Code Suggestions ✨

Explore these optional code suggestions:

Category	Suggestion	Impact
General	Constrain TensorFlow version TensorFlow is a heavy dependency that can conflict with PyTorch 2.9.0 regarding CUDA/cuDNN versions. Pin to a compatible version range (e.g., >=2.15.0,<2.19.0) to avoid runtime conflicts, control image size, and ensure compatibility with the existing Python environment. presets/workspace/dependencies/requirements.txt [40] -tensorflow +tensorflow>=2.15.0,<2.19.0 Suggestion importance[1-10]: 8 __ Why: The suggestion accurately highlights a specific technical risk where tensorflow can conflict with PyTorch 2.9.0 regarding CUDA/cuDNN versions. Pinning this heavy dependency is critical to avoid runtime failures and control image size.	Medium
	Pin versions for reproducibility Add version constraints to all new dependencies to prevent breaking changes and ensure reproducible builds. Unpinned packages may auto-update to versions incompatible with the pinned torch==2.9.0 and vllm==0.12.0, causing runtime failures or installation conflicts. presets/workspace/dependencies/requirements.txt [35-42] # Required by specific vLLM models -decord -fla-core -mistral-common[soundfile] -num2words -tensorflow -timm -torch-c-dlpack-ext +decord>=0.6.0,<1.0 +fla-core>=0.1.0 +mistral-common[soundfile]>=1.0.0,<2.0 +num2words>=0.5.0 +tensorflow>=2.15.0,<2.19.0 +timm>=0.9.0,<1.0 +torch-c-dlpack-ext>=1.0.0 Suggestion importance[1-10]: 7 __ Why: The suggestion correctly identifies that adding unpinned dependencies risks breaking changes and installation conflicts with pinned packages like torch==2.9.0 and vllm==0.12.0. Adding version constraints is important for build reproducibility.	Medium

[zhuangqh]

can you pin these dependencies version? This kind of py deps are easily broken.