github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-09 21:09:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

Revert "[Inference] Adapt to Fused rotary (#5348)" (#5373)

Browse Source 
This reverts commit 9f4ab2eb92.
This commit is contained in:
Frank Lee
2024-02-07 14:27:04 +08:00
committed by GitHub
parent 9f4ab2eb92
commit 8106ede07f
5 changed files with 22 additions and 161 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
colossalai/inference/modeling/models/nopadding_llama.py
View File

@@ -282,10 +282,11 @@ class NopadLlamaAttention(LlamaAttention):
torch.bmm(hidden_states, self.qkv_weight).view(3, token_nums, self.num_heads, self.head_dim).unbind(0)
)
rotary_embedding(query_states, key_states, cos_sin[0], cos_sin[1])
block_size = k_cache.size(-2)
if is_prompts:
rotary_embedding(query_states, key_states, cos_sin[0], cos_sin[1])
attn_output = context_attention_unpadded(
q=query_states,
k=key_states,
@@ -300,7 +301,7 @@ class NopadLlamaAttention(LlamaAttention):
sm_scale=sm_scale,
)
else:
rotary_embedding(query_states, key_states, cos_sin[0], cos_sin[1], k_cache, block_tables, sequence_lengths)
copy_kv_to_blocked_cache(key_states, k_cache, kv_lengths=sequence_lengths, block_tables=block_tables)
copy_kv_to_blocked_cache(value_states, v_cache, kv_lengths=sequence_lengths, block_tables=block_tables)
attn_output = flash_decoding_attention(
q=query_states,