zzh/vllm-npu-plugin
1
0
Fork 0
mirror of https://github.com/handsomezhuzhu/vllm-npu-plugin.git synced 2026-02-20 19:50:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: replace ATB reshape_and_cache with pure PyTorch indexing

Browse Source
This commit is contained in:
handsomezhuzhu
2026-02-10 19:56:47 +08:00
parent 101435817a
commit b8b4516b98
1 changed files with 9 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
vllm_npu/attention/attention_v1.py
View File

@@ -351,13 +351,15 @@ class AscendAttentionBackendImpl(AttentionImpl):
self._key_cache, self._value_cache = kv_cache.unbind(0) self._key_cache, self._value_cache = kv_cache.unbind(0)
slots = attn_metadata.slot_mapping slots = attn_metadata.slot_mapping
torch_npu._npu_reshape_and_cache( # Pure PyTorch reshape_and_cache (avoids ATB dependency)
key=key[:num_actual_tokens], key_to_cache = key[:num_actual_tokens]
value=value[:num_actual_tokens], val_to_cache = value[:num_actual_tokens]
key_cache=self._key_cache, block_size = self._key_cache.shape[1]
value_cache=self._value_cache, block_idx = slots // block_size
slot_indices=slots, block_offset = slots % block_size
) self._key_cache[block_idx, block_offset] = key_to_cache
self._value_cache[block_idx, block_offset] = val_to_cache
# ---------------------------------------------------------- # ----------------------------------------------------------
# Step 2: Compute attention # Step 2: Compute attention