Skip to content

[ROCm] Enable VLLM triton FP8 moe for gfx1201, tuned for Qwen3-30B-A3B-FP8 tp=2 #79

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
big-yellow-duck wants to merge 5 commits into
base: main
Choose a base branch
from
+135 −1
Open

[ROCm] Enable VLLM triton FP8 moe for gfx1201, tuned for Qwen3-30B-A3B-FP8 tp=2 #79

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
3e9d168
add aiter gemm_a8w8_blockscale support for gfx1201
big-yellow-duck Mar 11, 2026
ca40896
add tuned moe configs for gfx1201
big-yellow-duck Mar 11, 2026
6fb9a3e
remove unneceesary gfx12 in aiter_ops
big-yellow-duck Mar 16, 2026
f108f89
Merge branch 'main' into rdna4-moe
big-yellow-duck Mar 16, 2026
826b9f1
fix formatting
big-yellow-duck Mar 16, 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
69 changes: 69 additions & 0 deletions ...sed_moe/configs/E=64,N=768,device_name=AMD_Radeon_AI_PRO_R9700,block_shape=[128,128].json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,69 @@
{
"triton_version": "3.5.1+rocm7.2.0.gita272dfa8",
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 1,
"num_stages": 2,
"waves_per_eu": 4,
"matrix_instr_nonkdim": 16,
"kpack": 2
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 2,
"waves_per_eu": 2,
"matrix_instr_nonkdim": 16,
"kpack": 2
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 2,
"waves_per_eu": 1,
"matrix_instr_nonkdim": 16,
"kpack": 2
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"num_warps": 2,
"num_stages": 2,
"waves_per_eu": 2,
"matrix_instr_nonkdim": 16,
"kpack": 1
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"num_warps": 8,
"num_stages": 2,
"waves_per_eu": 1,
"matrix_instr_nonkdim": 16,
"kpack": 2
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"num_warps": 2,
"num_stages": 2,
"waves_per_eu": 4,
"matrix_instr_nonkdim": 16,
"kpack": 1
}
}
57 changes: 57 additions & 0 deletions .../E=64,N=768,device_name=AMD_Radeon_AI_PRO_R9700,dtype=fp8_w8a8,block_shape=[128,128].json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,57 @@
{
"triton_version": "3.5.1+rocm7.2.0.gita272dfa8",
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"waves_per_eu": 0
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"waves_per_eu": 2
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"waves_per_eu": 1
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"waves_per_eu": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"num_warps": 8,
"num_stages": 2,
"waves_per_eu": 0
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"num_warps": 8,
"num_stages": 2,
"waves_per_eu": 4
}
}
5 changes: 4 additions & 1 deletion vllm/model_executor/layers/fused_moe/fused_moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1926,14 +1926,17 @@ def _supports_quant_scheme(
) -> bool:
p = current_platform
if p.is_rocm():
from vllm.platforms.rocm import on_gfx9
from vllm.platforms.rocm import on_gfx9, on_gfx12x

is_rocm_on_gfx9 = on_gfx9()
is_rocm_on_gfx12x = on_gfx12x()
else:
is_rocm_on_gfx9 = False
is_rocm_on_gfx12x = False

device_supports_fp8 = (
is_rocm_on_gfx9
or is_rocm_on_gfx12x
or (p.is_cuda() and p.has_device_capability((8, 9)))
or p.is_xpu()
)
Expand Down
5 changes: 5 additions & 0 deletions vllm/platforms/rocm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -145,6 +145,7 @@ def _get_gcn_arch() -> str:
_GCN_ARCH = _get_gcn_arch()

_ON_GFX1X = any(arch in _GCN_ARCH for arch in ["gfx11", "gfx12"])
_ON_GFX12X = any(arch in _GCN_ARCH for arch in ["gfx12"])
_ON_MI3XX = any(arch in _GCN_ARCH for arch in ["gfx942", "gfx950"])
_ON_GFX9 = any(arch in _GCN_ARCH for arch in ["gfx90a", "gfx942", "gfx950"])
_ON_GFX942 = "gfx942" in _GCN_ARCH
Expand Down Expand Up @@ -226,6 +227,10 @@ def on_gfx1x() -> bool:
return _ON_GFX1X


def on_gfx12x() -> bool:
return _ON_GFX12X


def on_mi3xx() -> bool:
return _ON_MI3XX

Expand Down
Loading