[Model] New model support for microsoft/Phi-4-mini-flash-reasoning (#20702)

Signed-off-by: Congcong Chen <congcongchen@microsoft.com>
This commit is contained in:
Congcong Chen
2025-07-12 06:02:10 -07:00
committed by GitHub
parent b639327ad9
commit 2c11a738b3
22 changed files with 1869 additions and 41 deletions
+4
View File
@@ -1112,6 +1112,10 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
(self.max_batchsize_to_capture, self.get_max_block_per_batch()),
dtype=np.int32)
self.cross_layer_shared_graph_block_tables = np.zeros(
(self.max_batchsize_to_capture, self.get_max_block_per_batch()),
dtype=np.int32)
# Attention-free but stateful models like Mamba need a placeholder attn
# backend, as the attention metadata is needed to manage internal state.
# However we must bypass attention selection altogether for some models