Skip to content
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

Fix default value for INT4 accuracy level #1265

Merged
merged 3 commits into from
Feb 21, 2025
Merged

Fix default value for INT4 accuracy level #1265

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
6892779
Update builder.py
kunal-vaishnavi Feb 20, 2025
48152da
Update builder.py
kunal-vaishnavi Feb 20, 2025
b93624a
Merge branch 'main' into kvaishnavi/int4-accuracy-level
kunal-vaishnavi Feb 21, 2025
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
5 changes: 3 additions & 2 deletions src/python/py/models/builder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -295,7 +295,7 @@ def __init__(self, config, io_dtype, onnx_dtype, ep, cache_dir, extra_options):
# Quantization-specific variables (INT4, INT8, etc.)
self.quant_attrs = {
"int4": {
"accuracy_level": int(extra_options.get("int4_accuracy_level", 0)), # Default is 0 for non-QDQ formats, default is 4 for QDQ formats
"accuracy_level": int(extra_options.get("int4_accuracy_level", 4 if self.ep == "cpu" else 0)), # Default is 0 for non-QDQ formats, default is 4 for QDQ formats
"block_size": int(extra_options.get("int4_block_size", 32)),
"is_symmetric": extra_options.get("int4_is_symmetric", True),
"op_types_to_quantize": extra_options.get("int4_op_types_to_quantize", ("MatMul", )),
Expand Down Expand Up @@ -3324,6 +3324,7 @@ def get_args():
3 is bf16.
2 is fp16.
1 is fp32.
Default is 4 for the CPU EP and 0 for non-CPU EPs.
int4_block_size = 16/32/64/128/256: Specify the block_size for int4 quantization.
int4_is_symmetric = Quantize the weights symmetrically. Default is true.
If true, quantization is done to int4. If false, quantization is done to uint4.
Expand Down Expand Up @@ -3354,7 +3355,7 @@ def get_args():
If enabled, all nodes being placed on the CUDA EP is the prerequisite for the CUDA graph to be used correctly.
It is not guaranteed that CUDA graph be enabled as it depends on the model and the graph structure.
use_8bits_moe = Use 8-bit quantization for MoE layers. Default is false.
If true, the QMoE op will use 4-bit quantization. If false, the QMoE op will use 8-bits quantization.
If true, the QMoE op will use 8-bit quantization. If false, the QMoE op will use 4-bit quantization.
use_qdq = Use the QDQ decomposition for ops.
Use this option when you want to use quantize-dequantize ops. For example, you will have a quantized MatMul op instead of the MatMulNBits op.
adapter_path = Path to folder on disk containing the adapter files (adapter_config.json and adapter model weights).
Expand Down
Loading