Current Status of MPS Support?

[awindsor]

No matter what I try I keep getting RuntimeError: Placeholder storage has not been allocated on MPS device! when trying to use SpaCy en_core_web_trf for inference (not training).

I have had MPS acceleration function with prior versions of SpaCy (pre curated-transformers, I think we had space[transformers] and a dependency on the HuggingFace transformers library.

Environment

System: macOS with Apple Silicon (M1 Max)

Package Versions:

• Python: 3.11
• PyTorch: 2.10.0
• spaCy: 3.7.5
• Thinc: 8.2.4
• spacy-curated-transformers: 0.2.0+
• curated-transformers: (latest from spacy-curated-transformers dependency)

PyTorch MPS Status:

• MPS available: True
• MPS built: True
• PYTORCH_ENABLE_MPS_FALLBACK: 1

How spaCy is Called

# GPU initialization
import spacy
from spacy import require_gpu

gpu_id = require_gpu()  # Activates MPSOps backend in Thinc 8.2.4
nlp = spacy.load(model_name)  # Currently using spacy-curated-transformers model

# Batch processing with transformer pipeline
docs_and_rows = list(nlp.pipe(batch, as_tuples=True))

Initial Failure Point

Batch: 1 (first batch)
Row: 0 (very first row processed)
Timing: Immediate failure on first nlp.pipe() call