[ROCm][Critical] Fix the GDN import bug (#43486)
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
This commit is contained in:
@@ -521,7 +521,7 @@ class _MockGDNLayer:
|
|||||||
self.head_v_dim = head_v_dim
|
self.head_v_dim = head_v_dim
|
||||||
self.tp_size = tp_size
|
self.tp_size = tp_size
|
||||||
|
|
||||||
from vllm.model_executor.layers.mamba.gdn_linear_attn import (
|
from vllm.model_executor.layers.mamba.gdn.base import (
|
||||||
GatedDeltaNetAttention,
|
GatedDeltaNetAttention,
|
||||||
)
|
)
|
||||||
|
|
||||||
|
|||||||
@@ -560,11 +560,14 @@ class RocmAiterRMSNormQuantFusionPass(VllmPatternMatcherPass):
|
|||||||
|
|
||||||
# Discover (num_heads, head_dim) pairs for gated RMSNorm patterns
|
# Discover (num_heads, head_dim) pairs for gated RMSNorm patterns
|
||||||
# from GatedDeltaNetAttention layers in static_forward_context.
|
# from GatedDeltaNetAttention layers in static_forward_context.
|
||||||
from vllm.model_executor.layers.mamba.gdn_linear_attn import (
|
from vllm.model_executor.layers.mamba.gdn.base import (
|
||||||
GatedDeltaNetAttention,
|
GatedDeltaNetAttention,
|
||||||
)
|
)
|
||||||
|
|
||||||
gdn_layers = get_layers_from_vllm_config(config, GatedDeltaNetAttention)
|
gdn_layers = get_layers_from_vllm_config(
|
||||||
|
config,
|
||||||
|
GatedDeltaNetAttention, # type: ignore[type-abstract]
|
||||||
|
)
|
||||||
gated_norm_shapes: set[tuple[int, int]] = set()
|
gated_norm_shapes: set[tuple[int, int]] = set()
|
||||||
for layer in gdn_layers.values():
|
for layer in gdn_layers.values():
|
||||||
gated_norm_shapes.add(
|
gated_norm_shapes.add(
|
||||||
|
|||||||
Reference in New Issue
Block a user