gary-boon Claude Opus 4.6 (1M context) commited on
Commit
8b9ad4b
·
1 Parent(s): ff8be9d

Pin PyTorch <2.11.0 to fix CUDA driver incompatibility on HF Spaces

Browse files

PyTorch 2.11.0 (released 2026-03-23) requires a newer CUDA driver than
HuggingFace Spaces GPU infrastructure provides (driver 12090). This
caused CUDA init to fail, forcing Devstral to load on CPU and be
filtered from the model selector.

PyTorch 2.10.x (the version deployed successfully on 2026-02-25) works
with the current HF Spaces driver.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Files changed (1) hide show
  1. requirements.txt +2 -1
requirements.txt CHANGED
@@ -7,7 +7,8 @@ pydantic>=2.0.0 # Relaxed for mistral-common compatibility
7
 
8
  # Machine Learning
9
  # torch 2.3+ required for transformers 4.44+ (pytree API compatibility)
10
- torch>=2.3.0
 
11
  transformers>=4.44.0
12
  accelerate>=0.30.0
13
  mistral-common>=1.5.0 # Required for Devstral Tekken tokenizer
 
7
 
8
  # Machine Learning
9
  # torch 2.3+ required for transformers 4.44+ (pytree API compatibility)
10
+ # Pinned <2.11: PyTorch 2.11.0 requires CUDA drivers newer than HF Spaces GPU provides
11
+ torch>=2.3.0,<2.11.0
12
  transformers>=4.44.0
13
  accelerate>=0.30.0
14
  mistral-common>=1.5.0 # Required for Devstral Tekken tokenizer