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, 15:00 UTC
[ROCm][Bugfix] Add MLACommonMetadata to allowed attention types for speculative decoding (#30430)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Dec 11, 2025 at 13:25 UTC
72aaac5b66f908008efed5ba6874c3ed60e6c90a
1 file changed
+6
vllm/v1/spec_decode/eagle.py
+6
@@ -178,6 +178,12 @@ class EagleProposer:
178
)
179
180
rocm_types.append(AiterFlashAttentionMetadata)
181
+
182
+ # TRITON_MLA backend support for MLA models (e.g., DeepSeek)
183
+ from vllm.v1.attention.backends.mla.common import MLACommonMetadata
184
+
185
+ rocm_types.append(MLACommonMetadata)
186
+
187
self.allowed_attn_types = tuple(rocm_types)
188
189
# Parse the speculative token tree.