Skip to content

[https://nvbugs/5599086][fix] Fix FP8 Linear module for spark #8707

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

Merged
merged 2 commits into from
Oct 29, 2025
Merged

[https://nvbugs/5599086][fix] Fix FP8 Linear module for spark #8707

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d7d8e21
[https://nvbugs/5599086][fix] Fix FP8 Linear module for spark
SimengLiu-nv Oct 28, 2025
b31d0ad
Adapt comments and optimize memory footprint.
SimengLiu-nv Oct 28, 2025
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
26 changes: 11 additions & 15 deletions tensorrt_llm/_torch/modules/linear.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -444,16 +444,14 @@ def load_weights_fused_qkv_linear(self, module: Linear,

copy_weight(module.weight_scale, max(weight_scale))

q_weight = q_weight.to(module.dtype) * weight_scale[0]
k_weight = k_weight.to(module.dtype) * weight_scale[1]
v_weight = v_weight.to(module.dtype) * weight_scale[2]
# use in-place multiplication and division to avoid extra memory allocation
q_weight = q_weight.to(module.dtype).mul_(weight_scale[0])
k_weight = k_weight.to(module.dtype).mul_(weight_scale[1])
v_weight = v_weight.to(module.dtype).mul_(weight_scale[2])

fused_weight = torch.cat((q_weight, k_weight, v_weight))
if module.weight_scale.device != fused_weight.device:
module.weight_scale = Parameter(
module.weight_scale.data.to(fused_weight.device))
fused_weight = (fused_weight / module.weight_scale).to(
torch.float8_e4m3fn)
fused_weight = fused_weight.div_(
module.weight_scale.to(fused_weight.device)).to(torch.float8_e4m3fn)
copy_weight(module.weight, fused_weight)

# Load k and v scales, used for NVFP4 KV cache
Expand Down Expand Up @@ -486,14 +484,12 @@ def load_weights_fused_gate_up_linear(self, module: Linear,
gate_weight, up_weight = load_weights_fused_gate_up_helper(
module, weights)

gate_weight = gate_weight.to(module.dtype) * weight_scale[0]
up_weight = up_weight.to(module.dtype) * weight_scale[1]
# use in-place multiplication and division to avoid extra memory allocation
gate_weight = gate_weight.to(module.dtype).mul_(weight_scale[0])
up_weight = up_weight.to(module.dtype).mul_(weight_scale[1])
fused_weight = torch.cat((gate_weight, up_weight))
if module.weight_scale.device != fused_weight.device:
module.weight_scale = Parameter(
module.weight_scale.data.to(fused_weight.device))
fused_weight = (fused_weight / module.weight_scale).to(
torch.float8_e4m3fn)
fused_weight = fused_weight.div_(
module.weight_scale.to(fused_weight.device)).to(torch.float8_e4m3fn)
copy_weight(module.weight, fused_weight)


Expand Down