[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
@@ -16,7 +16,13 @@ import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm._aiter_ops import rocm_aiter_ops
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.model_executor.layers.fused_moe.activation import MoEActivation
from vllm.model_executor.layers.fused_moe.experts.rocm_aiter_moe import AiterExperts
from vllm.model_executor.layers.fused_moe.experts.rocm_aiter_moe import (
AiterExperts,
AiterFp8BlockExperts,
AiterFp8ChannelExperts,
AiterFp8TensorExperts,
AiterUnquantizedExperts,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
kFp8DynamicTensorSym,
kFp8StaticTensorSym,
@@ -80,10 +86,10 @@ def assert_aiter_quant_scheme_case(config: Config) -> None:
indirectly through the general quant-config sweep.
See https://github.com/vllm-project/vllm/issues/54966."""
fe_cls = config.fused_experts_type
if fe_cls is not AiterExperts:
if not issubclass(fe_cls, AiterExperts):
return
if config.quant_config is None:
if config.quant_config is None or config.quant_dtype is None:
w_key, a_key = None, None
else:
w_key, a_key = config.fp8_quant_key_pair()
@@ -101,7 +107,7 @@ def assert_aiter_activation_case(config: Config) -> None:
AiterExperts being reached only indirectly through the general
activation sweep. See https://github.com/vllm-project/vllm/issues/54966."""
fe_cls = config.fused_experts_type
if fe_cls is not AiterExperts:
if not issubclass(fe_cls, AiterExperts):
return
assert fe_cls._supports_activation(config.activation), (
@@ -156,7 +162,7 @@ def rank_worker(
if (
topk == 1
and config.supports_apply_weight_on_input()
and config.fused_experts_type is AiterExperts
and issubclass(config.fused_experts_type, AiterExperts)
and config.quant_block_shape is not None
):
print(
@@ -172,7 +178,7 @@ def rank_worker(
# https://github.com/vllm-project/vllm/issues/57029
if (
config.world_size > 1
and config.fused_experts_type is AiterExperts
and issubclass(config.fused_experts_type, AiterExperts)
and getattr(config.prepare_finalize_type, "__name__", "")
in ("DeepEPHTPrepareAndFinalize", "MoriPrepareAndFinalize")
):
@@ -210,7 +216,7 @@ def rank_worker(
is_aiter_fp8 = (
_cp.is_rocm()
and config.fused_experts_type is AiterExperts
and issubclass(config.fused_experts_type, AiterExperts)
and config.quant_config is not None
)
if is_aiter_fp8:
@@ -270,7 +276,7 @@ def is_nyi_config(config: Config) -> bool:
return True
if config.activation != MoEActivation.SILU:
if config.fused_experts_type is not AiterExperts:
if not issubclass(config.fused_experts_type, AiterExperts):
return True # AITER-only for this axis, for now
if config.quant_dtype is not None:
return True # unquantized-only for this axis, for now
@@ -591,7 +597,7 @@ def test_aiter_moe_sorting_backend_dispatch_env_matrix(
dtype=torch.bfloat16,
quant_config=None,
prepare_finalize_type=MoEPrepareAndFinalizeNoDPEPModular,
fused_experts_type=AiterExperts,
fused_experts_type=AiterUnquantizedExperts,
world_size=1,
)
assert config.is_valid()[0]
@@ -671,7 +677,7 @@ def test_aiter_moe_dispatch_policy_forwarded_through_apply(dispatch_policy: int)
dtype=torch.bfloat16,
quant_config=None,
prepare_finalize_type=MoEPrepareAndFinalizeNoDPEPModular,
fused_experts_type=AiterExperts,
fused_experts_type=AiterUnquantizedExperts,
world_size=1,
)
assert config.is_valid()[0]
@@ -738,6 +744,20 @@ _PADDING_QUANT_IDS = [
]
def _padding_matrix_fe_cls(
quant_config: TestMoEQuantConfig | None,
) -> type[AiterExperts]:
"""Map a _PADDING_QUANT_CONFIGS entry to the AiterExperts leaf class that
supports it (mirrors Config.fp8_quant_key_pair's scheme derivation)."""
if quant_config is None:
return AiterUnquantizedExperts
if quant_config.block_shape is not None:
return AiterFp8BlockExperts
if quant_config.per_out_ch_quant:
return AiterFp8ChannelExperts
return AiterFp8TensorExperts
def _slice_gate_up_rows(t: torch.Tensor, real_per_half: int) -> torch.Tensor:
"""Slice a (E, 2*padded_per_half, ...) tensor's dim=1 down to the real
gate/up halves of `real_per_half` size each."""
@@ -950,7 +970,7 @@ def test_aiter_moe_padding_matrix(mode: str, quant_config: TestMoEQuantConfig |
dtype=torch.bfloat16,
quant_config=quant_config,
prepare_finalize_type=MoEPrepareAndFinalizeNoDPEPModular,
fused_experts_type=AiterExperts,
fused_experts_type=_padding_matrix_fe_cls(quant_config),
world_size=1,
hidden_dim_unpadded=_PADDING_K_UNPADDED if pad_hidden else None,
intermediate_size_per_partition_unpadded=(
@@ -1113,7 +1133,7 @@ def test_aiter_moe_token_padding_garbage_rows(
dtype=torch.bfloat16,
quant_config=quant_config,
prepare_finalize_type=MoEPrepareAndFinalizeNoDPEPModular,
fused_experts_type=AiterExperts,
fused_experts_type=_padding_matrix_fe_cls(quant_config),
world_size=1,
)
assert config.is_valid()[0]