A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 2, 22:00 UTC
[ROCm][Critical] Fix the GDN import bug (#43486)
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
TJian committed
May 24, 2026 at 05:12 UTC
46f95b2ec2cb2115e02cadfa2bb03c6956a78927
2 files changed
+6
-3
tests/compile/passes/test_fusion.py
+1
-1
@@ -521,7 +521,7 @@ class _MockGDNLayer:
521
self.head_v_dim = head_v_dim
522
self.tp_size = tp_size
523
524
- from vllm.model_executor.layers.mamba.gdn_linear_attn import (
524
+ from vllm.model_executor.layers.mamba.gdn.base import (
525
GatedDeltaNetAttention,
526
)
527
vllm/compilation/passes/fusion/rocm_aiter_fusion.py
+5
-2
@@ -560,11 +560,14 @@ class RocmAiterRMSNormQuantFusionPass(VllmPatternMatcherPass):
560
561
# Discover (num_heads, head_dim) pairs for gated RMSNorm patterns
562
# from GatedDeltaNetAttention layers in static_forward_context.
563
- from vllm.model_executor.layers.mamba.gdn_linear_attn import (
563
+ from vllm.model_executor.layers.mamba.gdn.base import (
564
GatedDeltaNetAttention,
565
)
566
567
- gdn_layers = get_layers_from_vllm_config(config, GatedDeltaNetAttention)
567
+ gdn_layers = get_layers_from_vllm_config(
568
+ config,
569
+ GatedDeltaNetAttention, # type: ignore[type-abstract]
570
+ )
571
gated_norm_shapes: set[tuple[int, int]] = set()
572
for layer in gdn_layers.values():
573
gated_norm_shapes.add(