[ROCm][MoE] Split AiterExperts by weight-quantization key (#60322)

Signed-off-by: Divakar Verma <divakar.verma@amd.com>
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
This commit is contained in:
Divakar Verma
2026-10-13 00:13:33 -05:00
committed by GitHub
co-authored by Claude Sonnet 5
parent b337523c11
commit f7ba007b5e
13 changed files with 198 additions and 64 deletions
@@ -106,7 +106,11 @@ if HAS_TRITON:
BatchedTritonExperts,
)
from vllm.model_executor.layers.fused_moe.experts.rocm_aiter_moe import (
AiterExperts,
AiterFp8BlockExperts,
AiterFp8ChannelExperts,
AiterFp8TensorExperts,
AiterMxfp4Experts,
AiterUnquantizedExperts,
)
from vllm.model_executor.layers.fused_moe.experts.triton_deep_gemm_moe import (
TritonOrDeepGemmExperts,
@@ -132,7 +136,11 @@ if HAS_TRITON:
)
__all__ += [
"AiterExperts",
"AiterUnquantizedExperts",
"AiterFp8BlockExperts",
"AiterFp8TensorExperts",
"AiterFp8ChannelExperts",
"AiterMxfp4Experts",
"fused_topk",
"fused_experts",
"get_config_file_name",
@@ -482,29 +482,6 @@ class AiterExperts(mk.FusedMoEExpertsModular):
def _supports_no_act_and_mul() -> bool:
return False
@staticmethod
def _supports_quant_scheme(
weight_key: QuantKey | None,
activation_key: QuantKey | None,
) -> bool:
SUPPORTED_W_A = [
(None, None),
(kFp8Static128BlockSym, kFp8Dynamic128Sym),
(kFp8StaticTensorSym, kFp8StaticTensorSym),
(kFp8StaticTensorSym, kFp8DynamicTensorSym),
(kFp8StaticChannelSym, kFp8DynamicTokenSym),
(kMxfp4Static, None),
(kMxfp4Static, kMxfp4Dynamic),
]
if (weight_key, activation_key) not in SUPPORTED_W_A:
return False
if weight_key == kMxfp4Static:
from vllm.platforms.rocm import on_gfx950, on_gfx1250
if not on_gfx950() or on_gfx1250():
return False
return True
@staticmethod
def _supports_activation(activation: MoEActivation) -> bool:
return activation in [
@@ -597,3 +574,77 @@ class AiterExperts(mk.FusedMoEExpertsModular):
output.set_(result)
else:
output.copy_(result)
class AiterUnquantizedExperts(AiterExperts):
"""Unquantized (bf16/fp16) weights and activations."""
@staticmethod
def _supports_quant_scheme(
weight_key: QuantKey | None,
activation_key: QuantKey | None,
) -> bool:
SUPPORTED_W_A = [(None, None)]
return (weight_key, activation_key) in SUPPORTED_W_A
class AiterFp8BlockExperts(AiterExperts):
"""128x128 block-quantized weight scale, 128-grouped dynamic
activation scale."""
@staticmethod
def _supports_quant_scheme(
weight_key: QuantKey | None,
activation_key: QuantKey | None,
) -> bool:
SUPPORTED_W_A = [(kFp8Static128BlockSym, kFp8Dynamic128Sym)]
return (weight_key, activation_key) in SUPPORTED_W_A
class AiterFp8TensorExperts(AiterExperts):
"""Per-tensor static weight scale, either per-tensor static or
per-tensor dynamic activation scale."""
@staticmethod
def _supports_quant_scheme(
weight_key: QuantKey | None,
activation_key: QuantKey | None,
) -> bool:
SUPPORTED_W_A = [
(kFp8StaticTensorSym, kFp8StaticTensorSym),
(kFp8StaticTensorSym, kFp8DynamicTensorSym),
]
return (weight_key, activation_key) in SUPPORTED_W_A
class AiterFp8ChannelExperts(AiterExperts):
"""Per-output-channel static weight scale, per-token dynamic
activation scale."""
@staticmethod
def _supports_quant_scheme(
weight_key: QuantKey | None,
activation_key: QuantKey | None,
) -> bool:
SUPPORTED_W_A = [(kFp8StaticChannelSym, kFp8DynamicTokenSym)]
return (weight_key, activation_key) in SUPPORTED_W_A
class AiterMxfp4Experts(AiterExperts):
"""MXFP4-quantized weights, with either unquantized (bf16/fp16) or
MXFP4-quantized dynamic activations. gfx950-only (not gfx1250)."""
@staticmethod
def _supports_quant_scheme(
weight_key: QuantKey | None,
activation_key: QuantKey | None,
) -> bool:
SUPPORTED_W_A = [
(kMxfp4Static, None),
(kMxfp4Static, kMxfp4Dynamic),
]
if (weight_key, activation_key) not in SUPPORTED_W_A:
return False
from vllm.platforms.rocm import on_gfx950, on_gfx1250
return on_gfx950() and not on_gfx1250()
@@ -211,10 +211,12 @@ def backend_to_kernel_cls(
elif backend == Fp8MoeBackend.AITER:
from vllm.model_executor.layers.fused_moe.experts.rocm_aiter_moe import (
AiterExperts,
AiterFp8BlockExperts,
AiterFp8ChannelExperts,
AiterFp8TensorExperts,
)
return [AiterExperts]
return [AiterFp8BlockExperts, AiterFp8TensorExperts, AiterFp8ChannelExperts]
elif backend == Fp8MoeBackend.VLLM_CUTLASS:
from vllm.model_executor.layers.fused_moe.experts.triton_cutlass_moe import (
@@ -284,10 +284,10 @@ def backend_to_kernel_cls(
elif backend == Mxfp4MoeBackend.AITER_MXFP4_BF16:
from vllm.model_executor.layers.fused_moe.experts.rocm_aiter_moe import (
AiterExperts,
AiterMxfp4Experts,
)
return [AiterExperts]
return [AiterMxfp4Experts]
elif backend == Mxfp4MoeBackend.AITER_TRITON_MXFP4_BF16:
from vllm.model_executor.layers.fused_moe.experts.aiter_mxfp4_w4a16_moe import (
@@ -305,10 +305,10 @@ def backend_to_kernel_cls(
elif backend == Mxfp4MoeBackend.AITER_MXFP4_MXFP4:
from vllm.model_executor.layers.fused_moe.experts.rocm_aiter_moe import (
AiterExperts,
AiterMxfp4Experts,
)
return [AiterExperts]
return [AiterMxfp4Experts]
elif backend == Mxfp4MoeBackend.XPU:
from vllm.model_executor.layers.fused_moe.experts.xpu_moe import XPUExpertsMxFp4
@@ -126,10 +126,10 @@ def backend_to_kernel_cls(
elif backend == UnquantizedMoeBackend.AITER:
from vllm.model_executor.layers.fused_moe.experts.rocm_aiter_moe import (
AiterExperts,
AiterUnquantizedExperts,
)
return [AiterExperts]
return [AiterUnquantizedExperts]
elif backend == UnquantizedMoeBackend.TRITON:
from vllm.model_executor.layers.fused_moe.experts.triton_moe import (