Skip to content

[TRTLLM-10358][feat] Added proper rescaling of FP4 weights #10378

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
greg-kwasniewski1 merged 3 commits into from
Jan 3, 2026
Merged

[TRTLLM-10358][feat] Added proper rescaling of FP4 weights #10378

Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
e43d366
Added proper rescaling of FP4 weights
greg-kwasniewski1 Jan 1, 2026
0d17388
improved comment
greg-kwasniewski1 Jan 2, 2026
d327b1e
Merge branch 'main' into gk/FP4_weight_shape_fix
greg-kwasniewski1 Jan 2, 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
18 changes: 16 additions & 2 deletions tensorrt_llm/_torch/auto_deploy/utils/node_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -304,6 +304,16 @@ def is_any_lin_op(node: Node) -> bool:
return is_linear_op(node) or is_fake_quantized_linear_op(node)


def is_fp4_op(node: Node) -> bool:
return is_op(
node,
[
torch.ops.auto_deploy.torch_quant_nvfp4_linear,
torch.ops.auto_deploy.torch_fake_quant_nvfp4_linear,
],
)


def is_any_moe_op(node: Node) -> bool:
return is_op(
node,
Expand Down Expand Up @@ -739,10 +749,14 @@ def get_weight_shape(
"""Get the shape of the weight node."""
if not is_any_lin_op(node):
return None
s = list(shape(extract_weight_node(node)))
if is_fp4_op(node):
# FP4 weights are stored as half-sized FP8 tensor
s[-1] *= 2
if dim is None:
return shape(extract_weight_node(node))
return s
else:
return shape(extract_weight_node(node))[dim]
return s[dim]


def get_layer_after_linear_node(
Expand Down
Loading