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 11, 00:00 UTC
[ROCm][Bugfix] Fix RuntimeError in MMEncoderAttention by replacing .view() with .reshape() (#31203)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Dec 23, 2025 at 15:48 UTC
bfa2c0bbb9b4bf3f893a84f429a79c7905574aa5
2 files changed
+3
-3
tests/models/multimodal/conftest.py
+1
-1
@@ -19,7 +19,7 @@ def pytest_collection_modifyitems(config, items):
19
return
20
21
# Disable Flash/MemEfficient SDP on ROCm to avoid HF Transformers
22
- # accuracy issues
22
+ # accuracy issues: https://github.com/vllm-project/vllm/issues/30167
23
# TODO: Remove once ROCm SDP accuracy issues are resolved on HuggingFace
24
torch.backends.cuda.enable_flash_sdp(False)
25
torch.backends.cuda.enable_mem_efficient_sdp(False)
vllm/attention/layers/mm_encoder_attention.py
+2
-2
@@ -136,7 +136,7 @@ class MMEncoderAttention(CustomOp):
136
cu_seqlens=cu_seqlens,
137
)
138
if is_reshaped:
139
- output = output.view(bsz, q_len, -1)
139
+ output = output.reshape(bsz, q_len, -1)
140
return output
141
142
def _forward_fa(
@@ -174,7 +174,7 @@ class MMEncoderAttention(CustomOp):
174
fa_version=self._fa_version,
175
)
176
if is_reshaped:
177
- output = output.view(bsz, q_len, -1)
177
+ output = output.reshape(bsz, q_len, -1)
178
return output
179
180
def forward_native(