atakan Claude Opus 5 commited on
Commit
e901ca7
·
1 Parent(s): d554d58

deploy: Space overrides -- CUDA requirements.txt and app_space.py entry point

Browse files
Files changed (2) hide show
  1. README.md +9 -0
  2. requirements.txt +17 -14
README.md CHANGED
@@ -1,3 +1,12 @@
 
 
 
 
 
 
 
 
 
1
  # ControlAI: Open-Source Safety-Critical AI Agent for Control Systems Engineering
2
 
3
  [![Hugging Face Space](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Space-blue)](https://huggingface.co/spaces/atakankahya/ControlAI-Agent)
 
1
+ ---
2
+ title: ControlAI Agent
3
+ colorFrom: blue
4
+ colorTo: indigo
5
+ sdk: gradio
6
+ app_file: app_space.py
7
+ pinned: false
8
+ ---
9
+
10
  # ControlAI: Open-Source Safety-Critical AI Agent for Control Systems Engineering
11
 
12
  [![Hugging Face Space](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Space-blue)](https://huggingface.co/spaces/atakankahya/ControlAI-Agent)
requirements.txt CHANGED
@@ -1,13 +1,22 @@
1
- # ControlAI runtime dependencies -- local, Apple Silicon.
2
  #
3
- # Inference is MLX only. The previous file also carried torch, transformers,
4
- # accelerate, peft, llama-cpp-python, gradio and spaces to support a Hugging
5
- # Face Spaces deployment and three alternative backends; none of that is used
6
- # any more. transformers is still pulled in indirectly by mlx-lm for tokenizer
7
- # loading, so it is not listed here.
8
 
9
- mlx-lm>=0.31.0
10
- huggingface-hub>=0.23.0
 
 
 
 
 
 
 
 
 
 
 
11
 
12
  # Deterministic numerics -- every number in an answer comes from these.
13
  numpy>=1.24.0
@@ -21,9 +30,3 @@ jsonschema>=4.20.0
21
  rank-bm25>=0.2.2
22
  pypdf>=3.17.0
23
  pymupdf>=1.23.0
24
-
25
- # Web console.
26
- fastapi>=0.110.0
27
- uvicorn>=0.28.0
28
- python-multipart>=0.0.9
29
- pydantic>=2.0.0
 
1
+ # Hugging Face Space dependencies -- see app_space.py.
2
  #
3
+ # The Space is a Gradio app on ZeroGPU (Linux/NVIDIA), where MLX does not exist,
4
+ # so inference goes through controlai_agent/engine_torch.py. Do NOT install this
5
+ # locally: requirements.txt on `main` is the Apple Silicon MLX runtime.
 
 
6
 
7
+ # Unconstrained on purpose. The platform appends its own gradio, spaces, uvicorn
8
+ # and a `torch<=2.11.0` ceiling to this file, and one extra bound of ours can
9
+ # make the resolve impossible -- `transformers<4.56` did exactly that, because
10
+ # gradio 6 needs huggingface-hub>=1.16 and every transformers<4.56 needs <1.0.
11
+ #
12
+ # Do not pin torch down to CUDA 12 either: ZeroGPU is backed by Blackwell
13
+ # (sm_120), which needs CUDA 12.8+, so an older wheel has no kernels for it.
14
+ torch>=2.4.0
15
+ transformers>=4.51.0
16
+ accelerate>=0.30.0
17
+
18
+ # gradio, spaces and huggingface-hub are injected by the platform at build time;
19
+ # listing them here only adds constraints to conflict with it.
20
 
21
  # Deterministic numerics -- every number in an answer comes from these.
22
  numpy>=1.24.0
 
30
  rank-bm25>=0.2.2
31
  pypdf>=3.17.0
32
  pymupdf>=1.23.0