xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-16 12:05:57 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Misc] Enhance code formatting in mxfp4.py (#22423)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon 2025-08-07 00:26:24 -07:00 committed by GitHub
parent c2dba2dba8
commit 136825de75
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 52 additions and 33 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

85
vllm/model_executor/layers/quantization/mxfp4.py
View File

@ -109,55 +109,74 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
self.intermediate_size = intermediate_size_per_partition_after_pad
self.hidden_size = hidden_size
# Fused gate_up_proj (column parallel)
w13_weight = torch.nn.Parameter(torch.zeros(
num_experts,
2 * intermediate_size_per_partition_after_pad,
hidden_size // 2,
dtype=weight_dtype),
requires_grad=False)
w13_weight = torch.nn.Parameter(
torch.zeros(
num_experts,
2 * intermediate_size_per_partition_after_pad,
hidden_size // 2,
dtype=weight_dtype,
),
requires_grad=False,
)
layer.register_parameter("w13_weight", w13_weight)
set_weight_attrs(w13_weight, extra_weight_attrs)
w13_weight_scale = torch.nn.Parameter(torch.zeros(
num_experts,
2 * intermediate_size_per_partition_after_pad,
hidden_size // mxfp4_block,
dtype=scale_dtype),
requires_grad=False)
w13_weight_scale = torch.nn.Parameter(
torch.zeros(
num_experts,
2 * intermediate_size_per_partition_after_pad,
hidden_size // mxfp4_block,
dtype=scale_dtype,
),
requires_grad=False,
)
layer.register_parameter("w13_weight_scale", w13_weight_scale)
set_weight_attrs(w13_weight_scale, extra_weight_attrs)
w13_bias = torch.nn.Parameter(torch.zeros(
num_experts,
2 * intermediate_size_per_partition_after_pad,
dtype=torch.bfloat16),
requires_grad=False)
w13_bias = torch.nn.Parameter(
torch.zeros(
num_experts,
2 * intermediate_size_per_partition_after_pad,
dtype=torch.bfloat16,
),
requires_grad=False,
)
layer.register_parameter("w13_bias", w13_bias)
set_weight_attrs(w13_bias, extra_weight_attrs)
# down_proj (row parallel)
w2_weight = torch.nn.Parameter(torch.zeros(
num_experts,
hidden_size,
intermediate_size_per_partition_after_pad // 2,
dtype=weight_dtype),
requires_grad=False)
w2_weight = torch.nn.Parameter(
torch.zeros(
num_experts,
hidden_size,
intermediate_size_per_partition_after_pad // 2,
dtype=weight_dtype,
),
requires_grad=False,
)
layer.register_parameter("w2_weight", w2_weight)
set_weight_attrs(w2_weight, extra_weight_attrs)
w2_weight_scale = torch.nn.Parameter(torch.zeros(
num_experts,
hidden_size,
intermediate_size_per_partition_after_pad // mxfp4_block,
dtype=scale_dtype),
requires_grad=False)
w2_weight_scale = torch.nn.Parameter(
torch.zeros(
num_experts,
hidden_size,
intermediate_size_per_partition_after_pad // mxfp4_block,
dtype=scale_dtype,
),
requires_grad=False,
)
layer.register_parameter("w2_weight_scale", w2_weight_scale)
set_weight_attrs(w2_weight_scale, extra_weight_attrs)
w2_bias = torch.nn.Parameter(torch.zeros(num_experts,
hidden_size,
dtype=torch.bfloat16),
requires_grad=False)
w2_bias = torch.nn.Parameter(
torch.zeros(
num_experts,
hidden_size,
dtype=torch.bfloat16,
),
requires_grad=False,
)
layer.register_parameter("w2_bias", w2_bias)
set_weight_attrs(w2_bias, extra_weight_attrs)