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 7, 14:00 UTC
[ROCm] Add `stablelm` Head Size 80 To Supported Head Sizes For ROCM_ATTN (#35527)
Signed-off-by: Micah Williamson <micah.williamson@amd.com>
Micah Williamson committed
Feb 27, 2026 at 22:16 UTC
0edf101d2b502a709e9e119c693567c31b80ef36
2 files changed
+2
-2
docs/design/attention_backends.md
+1
-1
@@ -172,7 +172,7 @@ Priority is **1 = highest** (tried first).
172
| `FLEX_ATTENTION` | | fp16, bf16, fp32 | `auto`, `bfloat16` | Any | Any | ❌ | ✅ | ❌ | Decoder, Encoder Only | Any |
173
| `ROCM_AITER_FA` | | fp16, bf16 | `auto` | 16, 32 | 64, 128, 256 | ❌ | ❌ | ❌ | Decoder | N/A |
174
| `ROCM_AITER_UNIFIED_ATTN` | | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | All | N/A |
175
-| `ROCM_ATTN` | | fp16, bf16, fp32 | `auto` | 16, 32, 544 | 32, 64, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | All | N/A |
175
+| `ROCM_ATTN` | | fp16, bf16, fp32 | `auto` | 16, 32, 544 | 32, 64, 80, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | All | N/A |
176
| `TREE_ATTN` | | fp16, bf16 | `auto` | %16 | 32, 64, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | Decoder | Any |
177
| `TRITON_ATTN` | | fp16, bf16, fp32 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ✅ | ❌ | All | Any |
178
vllm/v1/attention/backends/rocm_attn.py
+1
-1
@@ -182,7 +182,7 @@ class RocmAttentionBackend(AttentionBackend):
182
183
@classmethod
184
def get_supported_head_sizes(cls) -> list[int]:
185
- return [32, 64, 96, 128, 160, 192, 224, 256]
185
+ return [32, 64, 80, 96, 128, 160, 192, 224, 256]
186
187
@classmethod
188
def validate_head_size(cls, head_size: int) -> None: