frozenleaves/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-10-20 09:03:53 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix FP-Quant quantization fallback CPU dispatch. (#41619)

Browse Source 
* fp_quant fix

* Update quantizer_fp_quant.py
This commit is contained in:
Andrei Panferov
2025-10-16 14:41:01 +03:00
committed by GitHub
parent af2a66ced9
commit 67fae90519
2 changed files with 6 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/transformers/quantizers/quantizer_fp_quant.py
View File

@ -97,6 +97,10 @@ class FPQuantHfQuantizer(HfQuantizer):
):
module, _ = get_module_from_name(model, param_name)
if target_device == "cpu" and param_name.endswith("weight"):
# Works agains hard-coded missing key dispatch to CPU
return
# The module holds either:
# * `weight` when `store_master_weights=True`
# * `qweight` and `scales` when `store_master_weights=False` and `pseudoquantization=False`

4
tests/quantization/fp_quant_integration/test_fp_quant.py
View File

@ -160,14 +160,14 @@ class FPQuantNVFP4PseudoquantTest(FPQuantBaseTest):
class FPQuantMXFP4Test(FPQuantBaseTest):
@classmethod
def getQuantizationConfig(cls):
return FPQuantConfig(forward_dtype="nvfp4", pseudoquantization=False)
return FPQuantConfig(forward_dtype="mxfp4", pseudoquantization=False)
@require_qutlass
class FPQuantMXFP4GS128Test(FPQuantBaseTest):
@classmethod
def getQuantizationConfig(cls):
return FPQuantConfig(forward_dtype="nvfp4", pseudoquantization=False, hadamard_group_size=128)
return FPQuantConfig(forward_dtype="mxfp4", pseudoquantization=False, hadamard_group_size=128)
@require_qutlass