From 6238f166253d813f1e76b4e687dcb8cd87d7ad0e Mon Sep 17 00:00:00 2001 From: Daniel Han-Chen Date: Sun, 25 Feb 2024 04:08:04 +1100 Subject: [PATCH] Update gemma.py --- unsloth/models/gemma.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/unsloth/models/gemma.py b/unsloth/models/gemma.py index aeb45c4e5d..1189a6f14b 100644 --- a/unsloth/models/gemma.py +++ b/unsloth/models/gemma.py @@ -83,8 +83,8 @@ class FastGemmaRotaryEmbedding(torch.nn.Module): 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() - print(position_ids_expanded.to(torch.float64)) - print(inv_freq_expanded.to(torch.float64)) + print(position_ids_expanded.shape, position_ids_expanded.to(torch.float64)) + print(inv_freq_expanded.shape, inv_freq_expanded.to(torch.float64)) freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2) print(freqs.to(torch.int64)) emb = torch.cat((freqs, freqs), dim=-1) @@ -111,8 +111,8 @@ class FastGemmaRotaryEmbedding(torch.nn.Module): 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(1, -1, 1).to("cuda") position_ids_expanded = t[:, None, :].float() - print(position_ids_expanded.to(torch.float64)) - print(inv_freq_expanded.to(torch.float64)) + print(position_ids_expanded.shape, position_ids_expanded.to(torch.float64)) + print(inv_freq_expanded.shape, inv_freq_expanded.to(torch.float64)) freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2) print(freqs.to(torch.int64)) emb = torch.cat((freqs, freqs), dim=-1)