Skip to content

[Attention] Add head_dim=512 support for FlashInfer trtllm attention backend #38822

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
djmmoss wants to merge 3 commits into
base: main
Choose a base branch
from
+2 −2
Open

[Attention] Add head_dim=512 support for FlashInfer trtllm attention backend #38822

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
038ab75
[Attention] Add head_dim=512 support for FlashInfer backend
djmmoss Apr 2, 2026
34ee306
Merge branch 'main' into dmoss/trtllm-fmha-head-dim-512
djmmoss Apr 4, 2026
ec65cbe
Merge branch 'main' into dmoss/trtllm-fmha-head-dim-512
djmmoss Apr 30, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion docs/design/attention_backends.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -169,7 +169,7 @@ Priority is **1 = highest** (tried first).
| ------- | ------- | ------ | --------- | ----------- | ---------- | ---- | --------- | --- | --------------- | ------------ |
| `CPU_ATTN` | | fp16, bf16, fp32 | `auto`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | Any | 32, 64, 80, 96, 112, 128, 160, 192, 224, 256, 512 | ❌ | ❌ | ❌ | All | N/A |
| `FLASHINFER` | Native† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256 | ❌ | ❌ | ✅ | Decoder | 7.x-9.x |
| `FLASHINFER` | TRTLLM† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256 | ✅ | ❌ | ✅ | Decoder | 10.x |
| `FLASHINFER` | TRTLLM† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256, 512 | ✅ | ❌ | ✅ | Decoder | 10.0x |
| `FLASH_ATTN` | FA2* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ❌ | ✅ | All | ≥8.0 |
| `FLASH_ATTN` | FA3* | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ❌ | ✅ | All | 9.x |
| `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ✅ | ❌ | ✅ | All | ≥10.0 |
Expand Down
2 changes: 1 addition & 1 deletion vllm/v1/attention/backends/flashinfer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -399,7 +399,7 @@ def get_fp8_dtype_for_flashinfer(kv_cache_dtype: str) -> torch.dtype:
@classmethod
def get_supported_head_sizes(cls) -> list[int]:
# https://github.com/flashinfer-ai/flashinfer/blob/3d55c71a62052c590c130897d3a3db49b14fcc34/include/flashinfer/utils.cuh#L157
return [64, 128, 256]
return [64, 128, 256, 512]
Comment on lines 400 to +402
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot Apr 2, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

Adding 512 to the supported head sizes without hardware-specific validation can lead to runtime crashes on non-Blackwell GPUs. The PR description states that head_dim=512 is intended for the TRTLLM attention kernels on Blackwell (SM100+). However, FlashInferBackend is also used on earlier architectures (SM75+) where the native FlashInfer kernels do not support this head dimension.

To prevent invalid configurations from reaching the execution stage, you should override supports_combination to ensure head_dim=512 is only permitted when the device capability is SM100 or higher.

Suggested change
def get_supported_head_sizes(cls) -> list[int]:
# https://github.com/flashinfer-ai/flashinfer/blob/3d55c71a62052c590c130897d3a3db49b14fcc34/include/flashinfer/utils.cuh#L157
return [64, 128, 256]
return [64, 128, 256, 512]
def get_supported_head_sizes(cls) -> list[int]:
# FlashInfer native kernels support 64, 128, 256.
# 512 is supported via TRTLLM kernels on Blackwell.
return [64, 128, 256, 512]
@classmethod
def supports_combination(
cls,
head_size: int,
dtype: torch.dtype,
kv_cache_dtype: "CacheDType | None",
block_size: int | None,
use_mla: bool,
has_sink: bool,
use_sparse: bool,
device_capability: DeviceCapability,
) -> str | None:
if head_size == 512 and device_capability.major < 10:
return "head_dim=512 is only supported on Blackwell GPUs (SM100+)"
return None


@classmethod
def supports_compute_capability(cls, capability: DeviceCapability) -> bool:
Expand Down
Loading