ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Update granite.py

Browse source
This commit is contained in:
Daniel Han 2025-03-01 00:19:12 -08:00
commit 6b735edbcb
1 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
unsloth/models/granite.py
View file

@ -279,12 +279,12 @@ def GraniteAttention_fast_forward_inference(
# Prefill phase
# if not hasattr(self, "paged_attention"):
if do_prefill:
self.paged_attention = torch.empty((KV_CACHE_INCREMENT+seq_len+1, 2, bsz, n_kv_heads, head_dim), dtype = dtype, device = device)
self.paged_attention = torch.empty((KV_CACHE_INCREMENT+seq_len+1, 2, bsz, n_kv_heads, head_dim), dtype = dtype, device = device))
self.paged_attention_K = self.paged_attention[:,0]
self.paged_attention_V = self.paged_attention[:,1]
self.paged_attention_K[:seq_len] = K1.permute(2, 0, 1, 3)
self.paged_attention_V[:seq_len] = V1.permute(2, 0, 1, 3)
self.temp_QA = torch.empty((2, bsz, 1, attention_size), dtype = dtype, device = device
self.temp_QA = torch.empty((2, bsz, 1, attention_size), dtype = dtype, device = device)
self.temp_KV = torch.empty((2, bsz, 1, n_kv_heads*head_dim), dtype = dtype, device = device)
self.RH_Q = torch.empty((bsz, n_heads, 1, head_dim), dtype = dtype, device = device)
# Only for Gemma2