Skip to content

fix(qwen3_5): skip GDN input-projection fusion on mixed projection dt… #182

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
potoior wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix(qwen3_5): skip GDN input-projection fusion on mixed projection dt… #182

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
11 changes: 11 additions & 0 deletions tensorrt_edgellm/models/qwen3_5/modeling_qwen3_5_text.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -807,6 +807,17 @@ def fuse_gdn_input_projections(model: nn.Module) -> int:
# --- Fuse: concatenate weights along output dim (dim 0) ----------
fused_buffers: dict = {}
proj_modules = [getattr(mixer, n) for n in _GDN_PROJ_NAMES]
# Anti-compressed checkpoints can quantize only some GDN projections
# (e.g. in_proj_qkv/z -> NVFP4 FP8, in_proj_b/a -> FP16); a single cat
# cannot mix those dtypes, so keep the unfused path in that case.
wptrs = [getattr(p, "weight", None) for p in proj_modules]
if any(wp is None for wp in wptrs) or len({wp.dtype
for wp in wptrs}) != 1:
logger.warning(
"GDN fusion skipped for %s: projections have mixed weight "
"dtypes (%s).", name,
", ".join(sorted(str(wp.dtype) for wp in wptrs)))
continue
for attr in list(proj_modules[0]._buffers) + list(
proj_modules[0]._parameters):
parts = [getattr(p, attr) for p in proj_modules]
Expand Down