From 225fc8d222d8f6a2e3e731c32c2162094c2a4650 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=91=A8=E5=91=A8=E5=91=A8?= <39978853+zhoutianzi666@users.noreply.github.com> Date: Sat, 11 Apr 2026 22:39:43 +0800 Subject: [PATCH] use self.hidden_size not use self.fd_config.model_config.hidden_size (#7340) --- fastdeploy/model_executor/layers/moe/moe.py | 2 +- tests/distributed/chunked_moe.py | 2 ++ 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/fastdeploy/model_executor/layers/moe/moe.py b/fastdeploy/model_executor/layers/moe/moe.py index f7d0b32c7a..4a2e2819a9 100644 --- a/fastdeploy/model_executor/layers/moe/moe.py +++ b/fastdeploy/model_executor/layers/moe/moe.py @@ -788,7 +788,7 @@ class FusedMoE(nn.Layer): chunk_size = self.fd_config.parallel_config.chunked_moe_size token_num = x.shape[0] fake_x = paddle.empty( - shape=[0, self.fd_config.model_config.hidden_size], + shape=[0, self.hidden_size], dtype=paddle.get_default_dtype(), ) # input size that are less than a chunk, less than the max size data or empty input diff --git a/tests/distributed/chunked_moe.py b/tests/distributed/chunked_moe.py index 0fe9f9f397..d453532705 100644 --- a/tests/distributed/chunked_moe.py +++ b/tests/distributed/chunked_moe.py @@ -169,6 +169,8 @@ class TestChunkedMoE(unittest.TestCase): fused_moe.fd_config = mock_fd_config fused_moe.quant_method = MockQuantMethod() fused_moe.enable_routing_replay = None + + fused_moe.hidden_size = mock_fd_config.model_config.hidden_size return fused_moe def run_model_runner(self):