[V0 Deprecation] Deprecate virtual engine (#37195)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2026-03-18 14:30:14 -07:00
committed by GitHub
parent 6ae4c8d6fc
commit 0d81a1fe61
23 changed files with 23 additions and 45 deletions
@@ -709,7 +709,7 @@ class BailingMoELinearAttention(nn.Module, MambaBase):
# Get KV cache and state indices
if attn_metadata is not None:
kv_cache = self.kv_cache[forward_context.virtual_engine][0]
kv_cache = self.kv_cache[0][0]
state_indices_tensor = attn_metadata.state_indices_tensor
clear_linear_attention_cache_for_new_sequences(
kv_cache, state_indices_tensor, attn_metadata
@@ -51,7 +51,7 @@ def unified_kv_cache_update(
"""
forward_context = get_forward_context()
attn_layer = forward_context.no_compile_layers[layer_name]
kv_cache = attn_layer.kv_cache[forward_context.virtual_engine]
kv_cache = attn_layer.kv_cache[0]
slot_mapping = forward_context.slot_mapping
assert isinstance(slot_mapping, dict), (
+1 -1
View File
@@ -428,7 +428,7 @@ class OlmoHybridGatedDeltaNet(nn.Module, MambaBase):
non_spec_token_indx = attn_metadata.non_spec_token_indx
spec_state_indices_tensor = attn_metadata.spec_state_indices_tensor
non_spec_state_indices_tensor = attn_metadata.non_spec_state_indices_tensor
self_kv_cache = self.kv_cache[forward_context.virtual_engine]
self_kv_cache = self.kv_cache[0]
conv_state = self_kv_cache[0].transpose(-1, -2)
ssm_state = self_kv_cache[1]
num_actual_tokens = attn_metadata.num_actual_tokens
+1 -1
View File
@@ -262,7 +262,7 @@ class Plamo2MambaMixer(MambaBase, PluggableLayer):
assert isinstance(attn_metadata, dict)
attn_metadata = attn_metadata[self.prefix]
assert isinstance(attn_metadata, Mamba2AttentionMetadata)
self_kv_cache = self.kv_cache[forward_context.virtual_engine]
self_kv_cache = self.kv_cache[0]
# conv_state = (..., dim, width-1) yet contiguous along 'dim'
conv_state = self_kv_cache[0].transpose(-1, -2)
ssm_state = self_kv_cache[1]
+2 -4
View File
@@ -842,7 +842,6 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
a=a,
core_attn_out=core_attn_out,
attn_metadata=attn_metadata,
virtual_engine=forward_context.virtual_engine,
)
has_initial_state = attn_metadata.has_initial_state
@@ -853,7 +852,7 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
non_spec_token_indx = attn_metadata.non_spec_token_indx
spec_state_indices_tensor = attn_metadata.spec_state_indices_tensor # noqa: E501
non_spec_state_indices_tensor = attn_metadata.non_spec_state_indices_tensor # noqa: E501
self_kv_cache = self.kv_cache[forward_context.virtual_engine]
self_kv_cache = self.kv_cache[0]
conv_state = self_kv_cache[0].transpose(-1, -2)
ssm_state = self_kv_cache[1]
num_actual_tokens = attn_metadata.num_actual_tokens
@@ -1036,13 +1035,12 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
a: torch.Tensor,
core_attn_out: torch.Tensor,
attn_metadata: GDNAttentionMetadata,
virtual_engine: int,
):
"""
Core attention computation with a packed non-spec decode fast path.
"""
non_spec_state_indices_tensor = attn_metadata.non_spec_state_indices_tensor # noqa: E501
self_kv_cache = self.kv_cache[virtual_engine]
self_kv_cache = self.kv_cache[0]
conv_state = self_kv_cache[0].transpose(-1, -2)
ssm_state = self_kv_cache[1]
num_actual_tokens = attn_metadata.num_actual_tokens