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 6, 21:45 UTC
[ROCm][Perf] Allow MTP lens > 1 in Sparse MLA (#36681)
Signed-off-by: Teemu Virolainen <teemu.virolainen@amd.com>
tvirolai-amd committed
Mar 11, 2026 at 16:43 UTC
a9e532afe2a1ae65c917ae977bf9090806e14721
1 file changed
+4
vllm/v1/spec_decode/eagle.py
+4
@@ -214,11 +214,15 @@ class SpecDecodeBaseProposer:
214
# Determine allowed attention backends once during initialization.
215
self.allowed_attn_types: tuple | None = None
216
if current_platform.is_rocm():
217
+ from vllm.v1.attention.backends.mla.rocm_aiter_mla_sparse import (
218
+ ROCMAiterMLASparseMetadata,
219
+ )
220
from vllm.v1.attention.backends.rocm_attn import RocmAttentionMetadata
221
222
rocm_types = [
223
TritonAttentionMetadata,
224
RocmAttentionMetadata,
225
+ ROCMAiterMLASparseMetadata,
226
]
227
# ROCM_AITER_FA is an optional backend
228
# We check is_enabled() here to avoid importing the backend module during