Skip to content

[Kernel]Using workspace manager in quantization #309

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Marshall-Ge wants to merge 5 commits into
base: main
Choose a base branch
from
Open

[Kernel]Using workspace manager in quantization #309

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
c8ff361
optim memory using workspace manager in quantization
Marshall-Ge Apr 8, 2026
aaf3cbf
temp save
Marshall-Ge Apr 10, 2026
fc247ca
optim
Marshall-Ge Apr 10, 2026
f6fab8a
optim
Marshall-Ge Apr 10, 2026
a20048c
optim
Marshall-Ge Apr 10, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
151 changes: 100 additions & 51 deletions vllm_kunlun/quantization/compressed_tensors/compressed_tensors_moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,7 @@
from vllm.model_executor.layers.quantization.compressed_tensors.schemes.compressed_tensors_wNa16 import ( # noqa
WNA16_SUPPORTED_BITS,
)
from vllm.v1.worker.workspace import current_workspace_manager

from vllm_kunlun.ops._kunlun_ops import KunlunOps as ops
from vllm_kunlun.quantization.kernels.quant_ops import dequant_int4_native
Expand Down Expand Up @@ -169,9 +170,6 @@ def apply_monolithic(
scale=routed_scaling_factor,
)

moe_expand = torch.empty(
(M * top_k, N), dtype=hidden_states.dtype, device=hidden_states.device
) # [M, top_k, N], float
expert_m = torch.zeros(
global_num_experts, dtype=torch.int32, device=hidden_states.device
) # [E]
Expand All @@ -180,8 +178,47 @@ def apply_monolithic(
) # [E+1]
sorted_tokens_idx = torch.zeros(
M * top_k, dtype=torch.int32, device=hidden_states.device
) # [M * top_k]

moe_expand_numel = M * top_k * N
y_numel = M * top_k * layer.w13_weight.shape[1]
out1_numel = M * top_k * (layer.w13_weight.shape[1] // 2)
out_numel = M * top_k * layer.w2_weight.shape[1]
x_q_numel = max(moe_expand_numel, out1_numel)
x_scale_shape = (M * top_k, 1)

# Reuse the workspace according to live ranges:
# M < 1024:
# workspace_a: moe_expand -> out1
# workspace_b: y -> out
# M >= 1024:
# workspace_a: moe_expand -> out
# workspace_b: y
workspace_a_numel = max(moe_expand_numel, out_numel)
workspace_b_numel = y_numel

if M < 1024:
workspace_a_numel = max(moe_expand_numel, out1_numel)
workspace_b_numel = max(y_numel, out_numel)

workspace_c_numel = x_q_numel

(
workspace_a,
workspace_b,
workspace_c,
workspace_d,
) = current_workspace_manager().get_simultaneous(
((workspace_a_numel,), hidden_states.dtype),
((workspace_b_numel,), hidden_states.dtype),
((workspace_c_numel,), torch.int8),
(x_scale_shape, torch.float32),
)

moe_expand = workspace_a[:moe_expand_numel].view(
M * top_k, N
) # [M * top_k, N], float

torch.ops._C.gen_block_statistic(topk_ids, block_statistic)

torch.ops._C.moe_pre_sorted(
Expand All @@ -194,59 +231,71 @@ def apply_monolithic(
sorted_tokens_num_lod=sorted_tokens_num_lod,
)

y = torch.empty(
M,
top_k,
layer.w13_weight.shape[1],
dtype=hidden_states.dtype,
device=hidden_states.device,
)

moe_expand = moe_expand.view(M * top_k, hidden_dim)

x_shape = moe_expand.shape
x_q = torch.empty(x_shape, dtype=torch.int8, device=moe_expand.device)
x_scale = torch.empty(
(x_shape[0], 1), dtype=torch.float32, device=moe_expand.device
)
x_q = workspace_c[: moe_expand.numel()].view(moe_expand.shape)
x_scale = workspace_d
torch.ops._C.quant2d(moe_expand, x_q, x_scale, force_sdnn=True)

torch.ops._C.moe_fc(
x=x_q,
x_perchannel_max=x_scale,
weight=layer.w13_weight,
w_perchannel_max=layer.w13_weight_scale,
sorted_tokens_num_lod=sorted_tokens_num_lod,
sorted_tokens_idx=sorted_tokens_idx,
moe_topk=top_k,
y=y,
topk_ids=topk_ids,
# sort_mode=False,
act=None,
)

d = y.shape[-1] // 2
output_shape = y.shape[:-1] + (d,)
out1 = torch.empty(output_shape, dtype=y.dtype, device=y.device)
torch.ops._C.silu_and_mul(out1, y)

del y
y = workspace_b[:y_numel].view(M, top_k, layer.w13_weight.shape[1])

if M < 1024:
torch.ops._C.moe_fc(
x=x_q,
x_perchannel_max=x_scale,
weight=layer.w13_weight,
w_perchannel_max=layer.w13_weight_scale,
sorted_tokens_num_lod=sorted_tokens_num_lod,
sorted_tokens_idx=sorted_tokens_idx,
moe_topk=top_k,
y=y,
topk_ids=topk_ids,
# sort_mode=False,
act=None,
)

out1 = out1.reshape(-1, out1.shape[-1])
x_shape = out1.shape
x_q = torch.empty(x_shape, dtype=torch.int8, device=moe_expand.device)
x_scale = torch.empty(
(x_shape[0], 1), dtype=torch.float32, device=moe_expand.device
)
torch.ops._C.quant2d(out1, x_q, x_scale, force_sdnn=True)
del out1, moe_expand
out = torch.empty(
M,
top_k,
layer.w2_weight.shape[1],
dtype=hidden_states.dtype,
device=hidden_states.device,
)
d = y.shape[-1] // 2
# Reuse `workspace_a` for `out1` after `moe_expand` is no longer
# needed.
out1 = workspace_a[:out1_numel].view(M, top_k, d)
torch.ops._C.silu_and_mul(out1, y)
del y

out1 = out1.reshape(-1, out1.shape[-1])
x_q = workspace_c[: out1.numel()].view(out1.shape)
x_scale = workspace_d
torch.ops._C.quant2d(out1, x_q, x_scale, force_sdnn=True)
del out1, moe_expand

# Reuse `workspace_b` for `out` after `y` has been consumed by
# the first FC.
out = workspace_b[:out_numel].view(M, top_k, layer.w2_weight.shape[1])
else:
torch.ops._C.moe_fc(
x=x_q,
x_perchannel_max=x_scale,
weight=layer.w13_weight,
w_perchannel_max=layer.w13_weight_scale,
sorted_tokens_num_lod=sorted_tokens_num_lod,
sorted_tokens_idx=sorted_tokens_idx,
moe_topk=top_k,
y=y,
topk_ids=topk_ids,
# sort_mode=False,
act="SWISH_GLU",
)
del x_q, x_scale, moe_expand

y = y[..., : y.shape[-1] // 2]
out1 = y.reshape(-1, y.shape[-1])
x_q = workspace_c[: out1.numel()].view(out1.shape)
x_scale = workspace_d
torch.ops._C.quant2d(out1, x_q, x_scale, force_sdnn=True)
del out1, y

# Reuse `workspace_a` for `out` after `moe_expand` is no longer
# needed.
out = workspace_a[:out_numel].view(M, top_k, layer.w2_weight.shape[1])

torch.ops._C.moe_fc(
x=x_q,
Expand Down
Loading