[ROCm][Kernel] MoE weights padding (#14454)

Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
Signed-off-by: charlifu <charlifu@amd.com>
Co-authored-by: charlifu <charlifu@amd.com>
This commit is contained in:
Gregory Shtrasberg
2025-03-24 23:45:30 +00:00
committed by GitHub
co-authored by charlifu
parent 8279201ce6
commit f533b5837f
5 changed files with 65 additions and 16 deletions
@@ -800,7 +800,7 @@ def invoke_fused_moe_kernel(A: torch.Tensor,
expert_ids,
num_tokens_post_padded,
B.shape[1],
A.shape[1],
B.shape[2],
EM,
topk_ids.numel(),
A.stride(0),
@@ -1322,8 +1322,8 @@ def fused_experts_impl(hidden_states: torch.Tensor,
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
assert w2.is_contiguous(), "Expert weights2 must be contiguous"
assert w1.stride(-1) == 1, "Stride of last dimension must be 1"
assert w2.stride(-1) == 1, "Stride of last dimension must be 1"
assert hidden_states.dtype in [
torch.float32, torch.float16, torch.bfloat16
]
@@ -5,6 +5,7 @@ from enum import Enum
from typing import Callable, List, Optional, Tuple
import torch
import torch.nn.functional as F
from torch.nn.parameter import UninitializedParameter
from vllm import envs
@@ -96,9 +97,27 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
layer.register_parameter("w2_weight", w2_weight)
set_weight_attrs(w2_weight, extra_weight_attrs)
def _maybe_pad_weight(self, weight: torch.Tensor) -> torch.Tensor:
# Pad the weight tensor. This is an optimization on ROCm platform, which
# can benefit from tensors located far enough from one another in memory
if (envs.VLLM_ROCM_MOE_PADDING and current_platform.is_rocm()
and weight.stride(-1) == 1
and (weight.stride(-2) * weight.element_size()) % 512 == 0):
num_pad = 256 // weight.element_size()
weight = F.pad(weight, (0, num_pad), "constant", 0)[..., :-num_pad]
torch.cuda.empty_cache()
return weight
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
super().process_weights_after_loading(layer)
layer.w13_weight = torch.nn.Parameter(self._maybe_pad_weight(
layer.w13_weight.data),
requires_grad=False)
layer.w2_weight = torch.nn.Parameter(self._maybe_pad_weight(
layer.w2_weight.data),
requires_grad=False)
if current_platform.is_cpu():
if current_platform.get_cpu_architecture() == CpuArchEnum.X86:
import intel_extension_for_pytorch as ipex
@@ -255,7 +255,7 @@ class Fp8LinearMethod(LinearMethodBase):
else:
layer.register_parameter("input_scale", None)
def add_padding_to_weight(self, weight: torch.Tensor) -> torch.Tensor:
def _maybe_pad_weight(self, weight: torch.Tensor) -> torch.Tensor:
# Pad the weight tensor. This is an optimization on ROCm platform, which
# can benefit from tensors located far enough from one another in memory
if (envs.VLLM_ROCM_FP8_PADDING and current_platform.is_rocm()
@@ -279,7 +279,7 @@ class Fp8LinearMethod(LinearMethodBase):
weight = layer.weight.data
weight_scale_inv = layer.weight_scale_inv.data
weight = self.add_padding_to_weight(weight)
weight = self._maybe_pad_weight(weight)
# Torch.compile cannot use Parameter subclasses.
layer.weight = Parameter(weight, requires_grad=False)
@@ -343,7 +343,7 @@ class Fp8LinearMethod(LinearMethodBase):
logical_widths=layer.logical_widths,
)
weight = self.add_padding_to_weight(weight)
weight = self._maybe_pad_weight(weight)
# Update layer with new values.
layer.weight = Parameter(weight.t(), requires_grad=False)
layer.weight_scale = Parameter(weight_scale, requires_grad=False)