diff --git a/unsloth/models/gemma.py b/unsloth/models/gemma.py index 47112ec4de..6a44137402 100644 --- a/unsloth/models/gemma.py +++ b/unsloth/models/gemma.py @@ -79,6 +79,7 @@ class FastGemmaRotaryEmbedding(torch.nn.Module): inv_freq = 1.0 / ( self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device="cpu").float() / self.dim) ) + print(inv_freq) t = torch.arange(self.max_position_embeddings, device="cpu", dtype=torch.int64).float().to("cuda").unsqueeze(0) inv_freq_expanded = inv_freq[None, :, None].float().expand(1, -1, 1).to("cuda") position_ids_expanded = t[:, None, :].float() @@ -103,10 +104,12 @@ class FastGemmaRotaryEmbedding(torch.nn.Module): self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device="cuda").float() / self.dim) ) + print(self.inv_freq) t = torch.arange(self.max_position_embeddings, device="cpu", dtype=torch.int64).float().to("cuda").unsqueeze(0) inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to("cuda") position_ids_expanded = t[:, None, :].float() freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2) + print(freqs.to(torch.int64)) emb = torch.cat((freqs, freqs), dim=-1) seq_len = position_ids.shape[1]