[Weight processing] Copy over new_data attributes in replace_parameter (#49601)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
fxmarty-amd
2026-08-06 13:03:37 -04:00
committed by GitHub
co-authored by Claude mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent b38e111d3e
commit 81be2e09ae
10 changed files with 334 additions and 51 deletions
@@ -145,11 +145,6 @@ class AiterMxfp8Experts(Mxfp8TritonExpertsBase):
from vllm._aiter_ops import rocm_aiter_ops
# Re-tag the preshuffled weights: replace_parameter drops the
# is_shuffled flag, without which aiter picks a broken CK kernel.
w1.is_shuffled = True
w2.is_shuffled = True
limit = self.quant_config.gemm1_clamp_limit
swiglu_limit = 0.0 if limit is None else float(limit)
@@ -483,10 +483,14 @@ def convert_to_fp8_moe_kernel_format(
)
elif fp8_backend == Fp8MoeBackend.AITER:
w13, w2 = rocm_aiter_ops.shuffle_weights(w13, w2)
w13.is_shuffled = True
w2.is_shuffled = True
elif fp8_backend == Fp8MoeBackend.AITER_MXFP8:
w13, w2, w13_scale, w2_scale = rocm_aiter_ops.shuffle_mxfp8_moe_weights(
w13, w2, w13_scale, w2_scale
)
w13.is_shuffled = True
w2.is_shuffled = True
elif fp8_backend == Fp8MoeBackend.HUMMING:
from vllm.model_executor.layers.quantization.utils.humming_utils import (
convert_to_humming_moe_kernel_format,
@@ -1081,6 +1081,9 @@ def convert_gpt_oss_weight_to_mxfp4_moe_kernel_format(
.view(-1, n)
)
w13_weight.is_shuffled = True
w2_weight.is_shuffled = True
return (
w13_weight,
w2_weight,
@@ -1504,6 +1507,9 @@ def convert_weight_to_mxfp4_moe_kernel_format(
False,
)
w13_weight.is_shuffled = True
w2_weight.is_shuffled = True
return (
w13_weight,
w2_weight,
@@ -332,6 +332,8 @@ def convert_to_unquantized_kernel_format(
) -> tuple[torch.Tensor, torch.Tensor]:
if unquantized_backend == UnquantizedMoeBackend.AITER:
w13_weight, w2_weight = rocm_aiter_ops.shuffle_weights(w13_weight, w2_weight)
w13_weight.is_shuffled = True
w2_weight.is_shuffled = True
elif unquantized_backend == UnquantizedMoeBackend.FLASHINFER_CUTLASS:
if moe_config.is_act_and_mul:
@@ -146,11 +146,6 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
replace_parameter(layer, "w13_weight", w13_new, prefer_copy=is_weight_update)
replace_parameter(layer, "w2_weight", w2_new, prefer_copy=is_weight_update)
# AITER backend requires weights to be marked as shuffled.
if self.unquantized_backend == UnquantizedMoeBackend.AITER:
layer.w13_weight.is_shuffled = True
layer.w2_weight.is_shuffled = True
if not is_weight_update:
# Setup moe kernel only on the first call. For the unquantized
# method, moe_quant_config carries no quantized scales -- only
@@ -28,7 +28,6 @@ from vllm.model_executor.layers.fused_moe.config import (
FusedMoEQuantConfig,
)
from vllm.model_executor.layers.fused_moe.oracle.fp8 import (
Fp8MoeBackend,
convert_to_fp8_moe_kernel_format,
make_fp8_moe_kernel,
make_fp8_moe_quant_config,
@@ -671,11 +670,6 @@ class Fp8MoEMethod(FusedMoEMethodBase):
replace_parameter(layer, f"w13_{self.weight_scale_name}", w13_scale)
replace_parameter(layer, f"w2_{self.weight_scale_name}", w2_scale)
# AITER backend requires weights to be marked as shuffled.
if self.fp8_backend == Fp8MoeBackend.AITER:
layer.w13_weight.is_shuffled = True
layer.w2_weight.is_shuffled = True
self.moe_quant_config = self.get_fused_moe_quant_config(layer)
assert self.moe_quant_config is not None
assert self.experts_cls is not None
@@ -366,11 +366,6 @@ class GptOssMxfp4MoEMethod(FusedMoEMethodBase):
self.w13_precision_config = w13_scale
self.w2_precision_config = w2_scale
# AITER backend requires weights to be marked as shuffled.
if self.mxfp4_backend == Mxfp4MoeBackend.AITER_MXFP4_BF16:
layer.w13_weight.is_shuffled = True
layer.w2_weight.is_shuffled = True
if w13_bias is not None and w2_bias is not None:
replace_parameter(layer, "w13_bias", w13_bias)
replace_parameter(layer, "w2_bias", w2_bias)
@@ -760,14 +755,6 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
self.w13_precision_config = w13_scale
self.w2_precision_config = w2_scale
# AITER backend requires weights to be marked as shuffled.
if (
self.mxfp4_backend == Mxfp4MoeBackend.AITER_MXFP4_BF16
and not uses_triton_weight_format
):
layer.w13_weight.is_shuffled = True
layer.w2_weight.is_shuffled = True
if w13_bias is not None and w2_bias is not None:
replace_parameter(layer, "w13_bias", w13_bias)
replace_parameter(layer, "w2_bias", w2_bias)
@@ -816,6 +803,10 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
w13_scale_raw.view(-1, w13_scale_raw.shape[-1]), num_experts, guinterleave
)
w2_scale = e8m0_shuffle(w2_scale_raw.view(-1, w2_scale_raw.shape[-1]))
w13.is_shuffled = True
w2.is_shuffled = True
return w13, w2, w13_scale, w2_scale
def process_weights_after_loading(self, layer):
@@ -27,7 +27,6 @@ from vllm.model_executor.layers.fused_moe.config import (
ocp_mx_moe_quant_config,
)
from vllm.model_executor.layers.fused_moe.oracle.fp8 import (
Fp8MoeBackend,
convert_to_fp8_moe_kernel_format,
make_fp8_moe_kernel,
make_fp8_moe_quant_config,
@@ -447,10 +446,6 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
replace_parameter(layer, "w13_weight_scale", w13_scale)
replace_parameter(layer, "w2_weight_scale", w2_scale)
if self.fp8_backend == Fp8MoeBackend.AITER:
layer.w13_weight.is_shuffled = True
layer.w2_weight.is_shuffled = True
self.moe_quant_config = self.get_fused_moe_quant_config(layer)
assert self.moe_quant_config is not None
assert self.experts_cls is not None
@@ -1328,10 +1323,6 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
replace_parameter(layer, "w13_bias", w13_bias)
replace_parameter(layer, "w2_bias", w2_bias)
if self.mxfp4_backend == Mxfp4MoeBackend.AITER_MXFP4_MXFP4:
layer.w13_weight.is_shuffled = True
layer.w2_weight.is_shuffled = True
torch.accelerator.empty_cache()
# Build quant config and kernel
+35 -13
View File
@@ -47,21 +47,29 @@ def set_weight_attrs(
def replace_parameter(
layer: torch.nn.Module,
param_name: str,
new_data: torch.Tensor | None,
new_tensor: torch.Tensor | None,
prefer_copy: bool = False,
):
"""
Replace a parameter of a layer while maintaining the ability to reload the weight.
Called within implementations of the `process_weights_after_loading` method.
Custom attributes set on ``new_tensor`` (e.g. kernel dispatch flags such as
``is_shuffled``) are carried over to the replacement parameter, except
``weight_loader``, which is always taken from the existing parameter.
Attributes of the existing parameter are otherwise dropped when a new
parameter is registered, but kept when ``prefer_copy`` reuses it in place.
This function should not be called on weights which are tied/shared
Args:
layer: Layer containing parameter to replace
param_name: Name of parameter to replace
new_data: New data of the new parameter, or None to set the parameter to None
new_tensor: New data of the new parameter, or None to set the parameter
to None
prefer_copy: If True and the existing parameter is compatible with
``new_data`` (same shape, dtype, and device), copy ``new_data``
``new_tensor`` (same shape, dtype, and device), copy ``new_tensor``
into the existing parameter in place rather than re-registering
a new parameter. This preserves the parameter's storage address
(``data_ptr``), which is required for captured CUDA graphs to
@@ -69,27 +77,41 @@ def replace_parameter(
"""
# should not be used on a tied/shared param
# If new_data is None, set the parameter to None
if new_data is None:
# If new_tensor is None, set the parameter to None
if new_tensor is None:
setattr(layer, param_name, None)
return
if isinstance(new_data, torch.nn.Parameter):
new_data = new_data.data
old_param: torch.nn.Parameter | None = getattr(layer, param_name, None)
new_tensor_attrs = dict(new_tensor.__dict__)
# `weight_loader` is the only attribute not ported over from new_tensor,
# old_param.weight_loader only is supported. `_weight_loader` is
# `BasevLLMParameter`'s backing field for its `weight_loader` property, so
# a loader riding along under that name has to be dropped as well.
new_tensor_attrs.pop("weight_loader", None)
new_tensor_attrs.pop("_weight_loader", None)
if isinstance(new_tensor, torch.nn.Parameter):
new_tensor = new_tensor.data
if (
prefer_copy
and old_param is not None
and old_param.shape == new_data.shape
and old_param.dtype == new_data.dtype
and old_param.device == new_data.device
and old_param.shape == new_tensor.shape
and old_param.dtype == new_tensor.dtype
and old_param.device == new_tensor.device
):
old_param.copy_(new_data)
old_param.copy_(new_tensor)
for attr_name, attr in new_tensor_attrs.items():
setattr(old_param, attr_name, attr)
return
new_param = torch.nn.Parameter(new_data, requires_grad=False)
new_param = torch.nn.Parameter(new_tensor, requires_grad=False)
for attr_name, attr in new_tensor_attrs.items():
setattr(new_param, attr_name, attr)
if old_param is not None and hasattr(old_param, "weight_loader"):
weight_loader = old_param.weight_loader