Skip to content

[quantization] Add pixel_values_videos for Gemma4GPTQQuantizer #872

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
mhs4670go merged 1 commit into from
Aug 7, 2026
Merged

[quantization] Add pixel_values_videos for Gemma4GPTQQuantizer #872

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: 0 additions & 1 deletion tico/quantization/algorithm/gemma4_gptq/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -167,6 +167,5 @@ config = Gemma4GPTQConfig(
| Patch embedding | `Conv3d` (`patch_embed.proj`) | `Linear` (`patch_embedder.input_proj`) |
| Vision merger | `merger` + `deepstack_merger_list` | `pooler` + `embed_vision` (multimodal embedder) |
| Deepstack | Replays `_deepstack_process` after each text layer | Not present; PLE is internal to each layer |
| Vision detection | `pixel_values` or `pixel_values_videos` | `pixel_values` only |
| Stages | Patch embed, vision blocks, merger, deepstack mergers, text layers, `lm_head` | Patch embed, vision blocks, pooler, multimodal embedder, text layers, `lm_head` |
| Core GPTQ class | `qwen3_vl_gptq.gptq.GPTQ` (shared) | Re-exports same `GPTQ` class |
11 changes: 5 additions & 6 deletions tico/quantization/algorithm/gemma4_gptq/quantizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,8 +75,6 @@ class Gemma4GPTQQuantizer(BaseQuantizer):
- **Vision stages**: Gemma4 has ``patch_embedder``, ``encoder.layers``,
``pooler``, and ``embed_vision`` (multimodal embedder). Qwen3-VL has
``patch_embed``, ``blocks``, ``merger``, and ``deepstack_merger_list``.
- **Vision detection**: Gemma4 checks for ``pixel_values`` only;
Qwen3-VL also checks ``pixel_values_videos``.
"""

def __init__(self, config: Gemma4GPTQConfig):
Expand Down Expand Up @@ -169,10 +167,11 @@ def model_forward_wrapper(_model, *m_args, **m_kwargs):
)

# Track whether this batch has vision inputs (pixel_values)
# Gemma4 uses 'pixel_values' for image inputs.
# Unlike Qwen3-VL, Gemma4 does not have 'pixel_values_videos'.
has_vision_input = (
"pixel_values" in m_kwargs and m_kwargs["pixel_values"] is not None
# Gemma4 uses 'pixel_values' for image inputs
# and `pixel_values_videos` for video inputs.
has_vision_input = any(
m_kwargs.get(name) is not None
for name in ("pixel_values", "pixel_values_videos")
)

if has_vision_input:
Expand Down
Loading