[V0 Deprecation] Refactor kv cache from list to element (#37487)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2026-03-23 20:10:11 -07:00
committed by GitHub
parent de99d91ece
commit c59a132f96
27 changed files with 70 additions and 85 deletions
@@ -413,7 +413,7 @@ class MiniMaxText01LinearAttention(nn.Module, MambaBase):
qkvact = qkvact.view((qkv.shape[0], self.tp_heads, -1))
q, k, v = torch.split(qkvact, [self.head_dim] * 3, dim=-1)
if attn_metadata is not None:
kv_cache = self.kv_cache[0][0]
kv_cache = self.kv_cache[0]
state_indices_tensor = attn_metadata.state_indices_tensor
clear_linear_attention_cache_for_new_sequences(
kv_cache, state_indices_tensor, attn_metadata
@@ -267,7 +267,7 @@ class MambaMixer(MambaBase, PluggableLayer):
query_start_loc_p = attn_metadata.query_start_loc_p
state_indices_tensor_p = attn_metadata.state_indices_tensor_p
state_indices_tensor_d = attn_metadata.state_indices_tensor_d
self_kv_cache = self.kv_cache[0]
self_kv_cache = self.kv_cache
conv_state = self_kv_cache[0].transpose(-1, -2)
ssm_state = self_kv_cache[1]
has_initial_states_p = attn_metadata.has_initial_states_p
@@ -575,7 +575,7 @@ class MambaMixer2(MambaBase, PluggableLayer):
assert isinstance(attn_metadata, dict)
attn_metadata = attn_metadata[self.prefix]
assert isinstance(attn_metadata, Mamba2AttentionMetadata)
self_kv_cache = self.kv_cache[0]
self_kv_cache = self.kv_cache
# conv_state = (..., dim, width-1) yet contiguous along 'dim'
conv_state = self_kv_cache[0].transpose(-1, -2)
ssm_state = self_kv_cache[1]
@@ -117,7 +117,7 @@ class ShortConv(MambaBase, CustomOp):
assert isinstance(attn_metadata, dict)
attn_metadata = attn_metadata[self.prefix]
assert isinstance(attn_metadata, ShortConvAttentionMetadata)
self_kv_cache = self.kv_cache[0]
self_kv_cache = self.kv_cache
conv_state = self_kv_cache[0].transpose(-1, -2)
state_indices_tensor_p = attn_metadata.state_indices_tensor_p
state_indices_tensor_d = attn_metadata.state_indices_tensor_d