[XPU] [MoE] add quant input when prepare for fusedmoe (#47122)

Signed-off-by: mayuyuace <qiming1.zhang@intel.com>
Signed-off-by: Zhu, Zufang <zufang.zhu@intel.com>
Signed-off-by: zofia <110436990+zufangzhu@users.noreply.github.com>
Co-authored-by: mayuyuace <qiming1.zhang@intel.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
zofia
2026-07-20 15:47:26 +08:00
committed by GitHub
co-authored by mayuyuace mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Kunshang Ji
parent 4938d44a3b
commit df13b5aef5
3 changed files with 37 additions and 4 deletions
@@ -20,6 +20,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
kFp8StaticTensorSym, kFp8StaticTensorSym,
kInt4Static, kInt4Static,
kInt4Static32, kInt4Static32,
kMxfp4Dynamic,
kMxfp4Static, kMxfp4Static,
kMxfp8Dynamic, kMxfp8Dynamic,
kMxfp8Static, kMxfp8Static,
@@ -64,10 +65,16 @@ class XPUExperts(mk.FusedMoEExpertsModular):
) )
self.gemm1_clamp_limit = quant_config.gemm1_clamp_limit self.gemm1_clamp_limit = quant_config.gemm1_clamp_limit
self.fused_moe_impl: XpuFusedMoe | None = None self.fused_moe_impl: XpuFusedMoe | None = None
is_xe2_or_xe3 = torch.ops._xpu_C.is_xe2_arch() or torch.ops._xpu_C.is_xe3_arch()
if not is_xe2_or_xe3:
raise NotImplementedError(
"XPUExperts is only supported on Intel Xe2/Xe3 GPUs"
)
self._expects_unquantized_inputs = is_xe2_or_xe3
@property @property
def expects_unquantized_inputs(self) -> bool: def expects_unquantized_inputs(self) -> bool:
return True return self._expects_unquantized_inputs
@staticmethod @staticmethod
def activation_format() -> mk.FusedMoEActivationFormat: def activation_format() -> mk.FusedMoEActivationFormat:
@@ -172,6 +179,7 @@ class XPUExperts(mk.FusedMoEExpertsModular):
hidden_states=hidden_states, hidden_states=hidden_states,
topk_weights=topk_weights, topk_weights=topk_weights,
topk_ids=topk_ids, topk_ids=topk_ids,
a1q_scale=a1q_scale,
) )
@@ -309,6 +317,24 @@ class XPUExpertsMxFp4(XPUExperts):
num_dispatchers, num_dispatchers,
) )
def workspace_shapes(
self,
M: int,
N: int,
K: int,
topk: int,
global_num_experts: int,
local_num_experts: int,
expert_tokens_meta: mk.ExpertTokensMetadata | None,
activation: MoEActivation,
) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
# K = a1q.size(-1). When activations are pre-quantized packed mxfp4,
# K is the packed hidden_size (= logical / 2); the kernel output is at
# logical hidden_size (2 * K). When unquantized (bf16), K is already
# the logical size.
logical_K = K if self.expects_unquantized_inputs else 2 * K
return (0,), (0,), (M, logical_K)
@staticmethod @staticmethod
def _supports_quant_scheme( def _supports_quant_scheme(
weight_key: QuantKey | None, weight_key: QuantKey | None,
@@ -316,5 +342,6 @@ class XPUExpertsMxFp4(XPUExperts):
) -> bool: ) -> bool:
SUPPORTED_W_A = [ SUPPORTED_W_A = [
(kMxfp4Static, None), (kMxfp4Static, None),
(kMxfp4Static, kMxfp4Dynamic),
] ]
return (weight_key, activation_key) in SUPPORTED_W_A return (weight_key, activation_key) in SUPPORTED_W_A
@@ -71,7 +71,7 @@ class TopKWeightAndReduceNoOP(mk.TopKWeightAndReduce):
assert output.size() == fused_expert_output.size(), ( assert output.size() == fused_expert_output.size(), (
"output shape is expected to match the fused_expert_output shape. " "output shape is expected to match the fused_expert_output shape. "
f"But got output={output.size()}, " f"But got output={output.size()}, "
f"used_expert_output={fused_expert_output.size()}" f"fused_expert_output={fused_expert_output.size()}"
) )
output.copy_(fused_expert_output, non_blocking=True) output.copy_(fused_expert_output, non_blocking=True)
return output return output
@@ -17,12 +17,14 @@ from vllm.model_executor.layers.quantization.utils.int8_utils import (
) )
from vllm.model_executor.layers.quantization.utils.mxfp4_utils import ( from vllm.model_executor.layers.quantization.utils.mxfp4_utils import (
quant_dequant_mxfp4, quant_dequant_mxfp4,
xpu_mxfp4_quantize,
) )
from vllm.model_executor.layers.quantization.utils.mxfp6_utils import ( from vllm.model_executor.layers.quantization.utils.mxfp6_utils import (
quant_dequant_mxfp6, quant_dequant_mxfp6,
) )
from vllm.model_executor.layers.quantization.utils.mxfp8_utils import ( from vllm.model_executor.layers.quantization.utils.mxfp8_utils import (
mxfp8_e4m3_quantize, mxfp8_e4m3_quantize,
xpu_mxfp8_quantize,
) )
from vllm.model_executor.layers.quantization.utils.nvfp4_emulation_utils import ( from vllm.model_executor.layers.quantization.utils.nvfp4_emulation_utils import (
ref_nvfp4_quant_dequant, ref_nvfp4_quant_dequant,
@@ -195,6 +197,8 @@ def _mxfp4_quantize(
per_act_token_quant: bool, per_act_token_quant: bool,
block_shape: list[int] | None = None, block_shape: list[int] | None = None,
) -> tuple[torch.Tensor, None]: ) -> tuple[torch.Tensor, None]:
if current_platform.is_xpu():
return xpu_mxfp4_quantize(A)
assert block_shape is None assert block_shape is None
# TODO: native mxfp4 is currently not integrated in vllm, # TODO: native mxfp4 is currently not integrated in vllm,
# so simulating even on devices supporting this data type natively. # so simulating even on devices supporting this data type natively.
@@ -223,6 +227,8 @@ def _mxfp8_e4m3_quantize(
is_sf_swizzled_layout: bool = False, is_sf_swizzled_layout: bool = False,
mx_alignment: int = 0, mx_alignment: int = 0,
) -> tuple[torch.Tensor, torch.Tensor]: ) -> tuple[torch.Tensor, torch.Tensor]:
if current_platform.is_xpu():
return xpu_mxfp8_quantize(A)
assert A_scale is None assert A_scale is None
assert not per_act_token_quant assert not per_act_token_quant
assert block_shape is None or block_shape == [1, 32] assert block_shape is None or block_shape == [1, 32]
@@ -309,7 +315,7 @@ def moe_kernel_quantize_input(
A = ref_nvfp4_quant_dequant(A, A_scale, block_size=16) A = ref_nvfp4_quant_dequant(A, A_scale, block_size=16)
return A, None return A, None
elif quant_dtype == "mxfp4": elif quant_dtype == "mxfp4":
if not quantization_emulation: if not current_platform.is_xpu() and not quantization_emulation:
raise NotImplementedError( raise NotImplementedError(
"moe_kernel_quantize_input should not be used for native" "moe_kernel_quantize_input should not be used for native"
" quant_dtype='mxfp4' MOE. Please open an issue." " quant_dtype='mxfp4' MOE. Please open an issue."
@@ -318,7 +324,7 @@ def moe_kernel_quantize_input(
elif quant_dtype == "mxfp8": elif quant_dtype == "mxfp8":
# TODO: `quant_dtype == "mxfp8"` is ambiguous, # TODO: `quant_dtype == "mxfp8"` is ambiguous,
# should be fp8_e4m3. OCP MX also defines `fp8_e5m2`. # should be fp8_e4m3. OCP MX also defines `fp8_e5m2`.
if quantization_emulation: if not current_platform.is_xpu() and quantization_emulation:
raise NotImplementedError( raise NotImplementedError(
"moe_kernel_quantize_input does not support quant_dtype='mxfp8' MOE " "moe_kernel_quantize_input does not support quant_dtype='mxfp8' MOE "
"quantization emulation. Please open an issue." "quantization emulation. Please open an issue."