Skip to content

update qwen quant #880

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

Draft
zhewenl wants to merge 1 commit into
base: ajrasane/qwen3-omni-30B
Choose a base branch
from
Draft

update qwen quant #880

Show file tree
Hide file tree
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
1 change: 1 addition & 0 deletions examples/llm_ptq/.gitignore
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,4 @@
saved_models_*
*.model
model_config.yaml
qwen3_omni_30b_nvfp4/
14 changes: 12 additions & 2 deletions examples/llm_ptq/hf_ptq.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -766,7 +766,11 @@ def pre_quantize(
elif model_type == "qwen3omni":
# Qwen3Omni returns (text_ids, audio) tuple; text_ids has .sequences
# Pass full batch with all multimodal inputs
result = full_model.generate(**calib_batch, max_new_tokens=100)
print("[DEBUG] pre_quantize: starting qwen3omni preview generation (max_new_tokens=100)...", flush=True)
result = full_model.generate(
**calib_batch, max_new_tokens=100, thinker_max_new_tokens=100
Copy link
Author

@zhewenl zhewenl Feb 11, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

updated here

)
print("[DEBUG] pre_quantize: preview generation complete", flush=True)
if isinstance(result, tuple):
text_ids, _ = result
generated_ids_before_ptq = (
Expand Down Expand Up @@ -827,7 +831,11 @@ def post_quantize(
elif model_type == "qwen3omni":
# Qwen3Omni returns (text_ids, audio) tuple; text_ids has .sequences
# Pass full batch with all multimodal inputs
result = full_model.generate(**calib_batch, max_new_tokens=100)
# Note: thinker_max_new_tokens controls the thinker's generation limit (default 1024),
# which is separate from max_new_tokens. Cap it to avoid long waits.
result = full_model.generate(
**calib_batch, max_new_tokens=100, thinker_max_new_tokens=100
Copy link
Author

@zhewenl zhewenl Feb 11, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

updated here

)
if isinstance(result, tuple):
text_ids, _ = result
generated_ids_after_ptq = (
Expand Down Expand Up @@ -958,9 +966,11 @@ def quantize_main(
# Detect if this is a Nemotron VL model using architecture-based detection
is_nemotron_vl_model = is_nemotron_vl(full_model)

print("[DEBUG] quantize_main: calling pre_quantize...", flush=True)
preview_input_ids, generated_ids_before_ptq, calib_batch = pre_quantize(
args, full_model, model_type, tokenizer, calib_dataloader, is_nemotron_vl_model
)
print("[DEBUG] quantize_main: pre_quantize done, proceeding to quantization", flush=True)

if args.auto_quantize_bits:
assert len(args.qformat.split(",")) > 1, (
Expand Down
Loading