Skip to content

Qwen3.5: drop fp32 cast around RMSNorm in builder #2101

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
kunal-vaishnavi merged 1 commit into from
May 6, 2026
Merged

Qwen3.5: drop fp32 cast around RMSNorm in builder #2101

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
10 changes: 0 additions & 10 deletions src/python/py/models/builders/qwen.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -954,16 +954,6 @@ def __init__(self, config, io_dtype, onnx_dtype, ep, cache_dir, extra_options):
# SkipSimplifiedLayerNormalization can be used directly.
self.layernorm_attrs["add_offset"] = 1

Comment thread
xiaofeihan1 marked this conversation as resolved.
# HF Qwen3_5RMSNorm always computes in float32 regardless of model
# dtype. Force the builder to cast inputs to fp32 before LayerNorm
# and cast back after, matching HF behaviour and preventing precision
# loss that compounds across 36+ layers in fp16/bf16 builds.
self.layernorm_attrs["cast"]["use_fp32"] = True
self.layernorm_attrs["cast"]["root_input"] = True
self.layernorm_attrs["cast"]["skip_input"] = True
self.layernorm_attrs["cast"]["output_0"] = True
self.layernorm_attrs["cast"]["output_3"] = True

# 3D position_ids for mRoPE: [3, batch_size, sequence_length]
self.input_shapes["position_ids"] = [3, "batch_size", "sequence_length"]
self.input_names["position_ids"] = "position_ids"
Expand Down
Loading