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 3, 16:30 UTC
[ROCm] Enable gluon paged MQA logits on gfx950 (MI355X) (#42062)
Signed-off-by: Frida Andersson <fanderss@amd.com>
frida-andersson committed
May 14, 2026 at 17:39 UTC
f07b1da797cc0553b202e7e4a992346b0643cc38
1 file changed
+3
-2
vllm/v1/attention/ops/rocm_aiter_mla_sparse.py
+3
-2
@@ -16,9 +16,10 @@ from vllm.v1.attention.backends.mla.indexer import DeepseekV32IndexerMetadata
16
from vllm.v1.attention.ops.common import pack_seq_triton, unpack_seq_triton
17
18
if current_platform.is_rocm():
19
- from vllm.platforms.rocm import _ON_GFX942
19
+ from vllm.platforms.rocm import _ON_GFX942, _ON_GFX950
20
else:
21
_ON_GFX942 = False
22
+ _ON_GFX950 = False
23
24
25
@triton.jit
@@ -385,7 +386,7 @@ def rocm_fp8_paged_mqa_logits(
386
aiter_paged_mqa_logits_module = paged_mqa_logits_module()
387
388
if aiter_paged_mqa_logits_module is not None:
388
- if _ON_GFX942:
389
+ if _ON_GFX942 or _ON_GFX950:
390
deepgemm_fp8_paged_mqa_logits = (
391
aiter_paged_mqa_logits_module.deepgemm_fp8_paged_mqa_logits
392
)