ai-agents-version-compatibilit-creating-a-modernbert-model-with-flash-attention-e-d4fe8597

THEN Downgrade flash-attn to version 2.7.4.post1, which still supports the pos_idx_in_fp32 parameter. Alternatively, remove the `pos_idx_in_fp32=True` argument from the `ModernBertUnpaddedRotaryEmbedding.__init__` super() call in the transformers source code, as the parameter was deprecated and removed in newer flash-attn versions.

# In transformers source, modify the super call in ModernBertUnpaddedRotaryEmbedding:
# Change:
#     super().__init__(dim=dim, base=base, pos_idx_in_fp32=True, device=device, interleaved=False)
# To:
#     super().__init__(dim=dim, base=base, device=device, interleaved=False)